60 results about "Regularization algorithm" patented technology

The invention discloses a multi-modal data driven power distribution network equipment state evaluation and optimization method, and relates to the technical field of power distribution network equipment, multi-modal data is collected, the data is from multiple equipment and operating environments in a power distribution network, and sensor data comprises real-time monitoring data of the equipment; and performing time sequence correction and synchronization processing on the acquired multi-modal data, and ensuring time alignment between different data sources through an advanced time sequence synchronization algorithm. According to the method, the dynamic time warping algorithm and deep learning are combined, the problem of multi-modal data synchronization is solved, the state of the power distribution network equipment is evaluated in real time, faults are recognized in advance, and early warning is provided. The reinforcement learning algorithm optimizes load distribution and self-healing scheduling, and improves the operation efficiency and fault recovery capability of the power distribution network. The system also generates intelligent early warning and maintenance decisions through the combination of deep learning and reinforcement learning, predicts faults, provides maintenance suggestions, optimizes operation and maintenance, prolongs the service life of equipment, and improves the overall stability of the power distribution network.

The invention discloses a sign language recognition system based on a dynamic time warping algorithm, and the system comprises a data collection module which is used for collecting gesture data in real time through a camera and a depth sensor; the feature extraction and time sequence conversion module is used for converting the gesture data into a time sequence; the dynamic time warping matching module is used for matching the input gesture with the gesture in the template library by utilizing a dynamic time warping algorithm; the matching result selection and template classification module is used for selecting the most matched template according to the similarity and determining the gesture category; the embedded platform and hardware acceleration module is used for executing a sign language recognition algorithm on the embedded platform and improving the recognition speed and real-time performance through hardware acceleration; and the character and voice output and real-time feedback optimization module is used for converting the recognition result into characters and voice output. Through hardware acceleration and data compression optimization, the sign language recognition precision, real-time performance and adaptability are improved, and the sign language recognition method is suitable for an embedded system with limited resources.

The invention discloses a spatial projection regularization method for sound source localization and sound field reconstruction and a related device, and belongs to the technical field of sound source localization. Specifically, according to the method, an equivalent source method model is utilized, and an external radiation sound field of a sound source to be measured is simulated through a series of virtual sources. Then, according to the first sound pressure transfer matrix from the virtual source arrangement surface to the holographic measuring point plane and the sound pressure of the holographic measuring point, a space projection regularization algorithm is used for solving the inverse problem, and therefore the equivalent source intensity is obtained; and then, reconstructing an external radiation sound field of the sound source to be detected by using the second sound pressure transfer matrix and the equivalent source intensity. And finally, by taking the maximum sound pressure amplitude as a standard, screening field points in an external radiation sound field so as to determine the position of the sound source to be detected. According to the space projection regularization algorithm, truncation processing of the vector space is carried out with the projection size of the measurement information in the vector space as the criterion, real information is reserved, and noise interference is restrained to the maximum extent.

The invention discloses a power equipment fault identification method based on multi-physics field coupling and related equipment, relates to the technical field of fault identification, and solves the problem of high false alarm rate of power equipment fault identification. According to the method, the multi-physics field coupling finite element model of the power equipment is constructed, the field distribution characteristic matrix is solved to accurately reflect the interaction of electromagnetic loss-thermal stress-deformation in the power equipment, and then calculation is performed by combining real-time physics field parameter data and the field distribution characteristic matrix through a regularization algorithm. And the fault contribution degree weight vector is obtained to represent the contribution degree of each area of the power equipment to the equipment abnormity, so that the composite fault identification capability can be improved, the false alarm rate is reduced, and the fault identification precision is improved.

A computer implemented method of reconstructing magnetic resonance images (MRI) in Cartesian coordinates uses acquired magnetic resonance data and implements a Fourier transform to place the MRI data in k-space. The method allows for under-sampling the k-space and achieving an accurate output image by selecting an image model to map the sampled data and iteratively converge the model to an output that matches a region of interest subject to the MRI. The image model may be an alternating direction method of multipliers (ADMM) or an ADMM with non-convex low rank regularization algorithm. A de-noising algorithm may be at least one of a plug and play block matching and 3D filtering (PnP-BM3D), a plug and play weighted nuclear norm minimization (WNNM), or a plug and play denoising convolutional neural networks (PnP-DnCNN) algorithm. An iterative optimization of the variables of the model yields an output image.

The invention belongs to the technical field of boiler heating surface high-temperature corrosion monitoring, and relates to a boiler heating surface temperature monitoring method and system based on dynamic heat flux density monitoring. The method comprises the steps of obtaining temperature field data of a plurality of monitoring areas of a boiler heating surface; constructing an improved Tikhonov regularization algorithm based on the temperature field data, and carrying out heat flow field reverse solution by adopting the improved Tikhonov regularization algorithm to obtain reconstructed heat flow field distribution and a heat conduction nuclear matrix after regularization optimization; and generating a temperature evolution trend by adopting a GRU prediction model based on an attention mechanism based on the reconstructed heat flow field distribution and the regularized and optimized heat conduction nuclear matrix. The method can accurately predict the temperature evolution trend through obtaining the comprehensive temperature field data, facilitates the real-time monitoring of the temperature of the heating surface of the boiler, and improves the safety and reliability of the operation of the boiler. Operation personnel can adjust operation parameters in time according to a prediction result, and the operation state of the boiler is optimized.

The invention relates to the technical field of model construction, in particular to a heart model construction method based on multiple physical fields and geometric multiple scales, which comprises the following steps of: performing segmentation processing and three-dimensional model reconstruction on a heart image to generate an initial heart geometric model; performing anisotropic attribute assignment on the initial heart geometric model based on a regularization algorithm to obtain a myocardial tissue structure; embedding a cell electrophysiological model into the computational grid of the myocardial tissue structure through parameter mapping to output a heart model with multi-scale geometric characteristics; an electrophysiological field equation, a mechanical field equation and a heart fluid dynamic field equation are constructed respectively, and coupling among multiple physical fields is achieved; determining whether the feature weight of the myocardial tissue structure texture needs to be increased or not; determining whether the transition function width of cell electrophysiological model parameter mapping needs to be increased or not; and determining a pressure feedback relaxation factor in the coupling iteration. The stability of heart model construction is improved.

The invention relates to the technical field of high-temperature pressure pipeline state monitoring, and discloses a quantitative detection method for oxide skin accumulation in a pipeline, electromagnetic response of the pipeline is excited through a multi-band composite detection signal, and multi-physical field characteristic data is collected in combination with a quantum sensor array. And an improved Bayesian regularization algorithm is adopted to realize oxide skin thickness inversion, and a crack propagation model is constructed to carry out residual life prediction. A dynamic reference calibration technology is innovatively introduced, and environmental interference is suppressed through a temperature-frequency drift compensation model and space-time coding. The system comprises a multi-channel data fusion module, a three-dimensional visualization module and a self-diagnosis unit, and realizes closed-loop processing from signal acquisition to state evaluation. And reconstructing thickness distribution by adopting radial basis function interpolation, and generating a maintenance decision in combination with the risk level matrix. According to the method, the problems of insufficient precision and weak anti-interference capability of a traditional detection method are solved, and the reliability and the intelligent level of high-temperature pipeline state monitoring are remarkably improved.

The embodiment of the invention discloses a thermal station secondary return water temperature prediction method and device based on XGBoost and a neural network. The method comprises the following steps: acquiring original feature data through a temperature sensor; modeling the original feature data set based on an XGBoost algorithm to obtain the importance weight of each feature, and screening out first feature data according to the sequence of the importance weights; processing the first feature data; constructing an XGBoost-GWO-ANN secondary return water temperature prediction model, using the first feature data as model training data, using a Bayesian regularization algorithm to train an ANN artificial neural network model, and using a GWO grey wolf optimization algorithm to find an optimal solution for an initial weight and a threshold value of the ANN; when the performance indexes of the training set and the test set meet set requirements, carrying out reverse normalization processing on an output result of the secondary return water temperature prediction model to obtain a secondary return water temperature prediction result; the prediction performance of the model is improved, and the prediction precision and efficiency of the model are improved.

The invention provides a multi-source data fusion industrial energy-saving management system based on big data, and the system comprises a data collection and edge processing module which is used for collecting on-site multi-source heterogeneous data in parallel and executing filtering and time calibration; the industrial big data storage center is used for storing data by adopting a cold and hot data layered architecture; the spatio-temporal data deep fusion engine is configured to be used for constructing a virtual production takt axis, mapping a continuous energy consumption waveform to the axis by using a nonlinear time warping algorithm, eliminating time lag interference and generating a standard energy consumption fingerprint in combination with a thermal response model; the dynamic energy-saving decision control module is used for calculating a transient energy consumption balance index covering recovery compensation and life loss based on the fingerprint vector, and generating a control instruction when an energy-saving gain threshold value is met; and the terminal execution feedback module is used for instruction issuing and feedback. According to the method, the problem of spatial-temporal dislocation of discrete work orders and continuous energy consumption data is solved, and refined dynamic energy-saving control of equipment is realized.

The present application relates to the technical field of model construction, and particularly relates to a heart model construction method based on multi-physical field and geometric multi-scale, comprising: carrying out segmentation processing and three-dimensional model reconstruction on a heart image to generate an initial heart geometric model, carrying out anisotropic attribute assignment on the initial heart geometric model based on a regularization algorithm to obtain a myocardial tissue structure; embedding a cell electrophysiology model into a calculation grid of the myocardial tissue structure through parameter mapping to output a heart model with multi-scale geometric characteristics; respectively constructing an electrophysiology field equation, a mechanics field equation and a heart fluid dynamics field equation to realize coupling between the multi-physical fields; determining whether the feature weight of the myocardial tissue structure texture needs to be increased; determining whether the transition function width of the cell electrophysiology model parameter mapping needs to be increased; and determining a pressure feedback relaxation factor in the coupling iteration. The present application improves the stability of heart model construction.

The invention belongs to the technical field of seismic exploration, and relates to a strong interference seismic data seismic source function inversion method based on transverse local regularization, and the method comprises the steps: applying transverse local regularization constraint in a seismic source function waveform inversion process to guarantee the transverse continuity and precision of an inversion result; the nonlinearity of inversion and the degree of dependence on a speed model are reduced by using near-channel short-time data, and the overall calculation efficiency of the algorithm is ensured; seismic source function inversion is carried out under a full-waveform inversion framework, and an inversion result theoretically has the advantages of high precision and high resolution; the advantages of a cross-correlation objective function and a transverse local regularization algorithm are combined, a high-quality seismic source function inversion result can be obtained under the condition of strong interference, the transverse continuity of the inversion result is enhanced by utilizing mutual constraint between adjacent seismic source functions, and the problem of dependency of an initial seismic source is overcome to a large extent. According to the method, a seismic source function inversion result with high precision and good transverse consistency can be stably obtained under the condition that strong interference exists in seismic data.

The application discloses a kind of based on neural network's semi-batch reaction thermal behavior identification method, first define thermal behavior label, solve dimensionless mathematical model;Again, extract core numerical features from traditional criterion, and match with NI, TR and QFS thermal behavior to constitute data set;Again, data set is preprocessed, and is divided;Again, the recognition accuracy and generalization ability of different machine learning algorithms to reaction thermal behavior are analyzed and compared, and double-layer BP neural network is selected;Then determine the structure of double-layer BP neural network;Again, the weight and threshold of double-layer BP neural network are optimized using genetic algorithm and bayesian regularization algorithm;Again, confusion matrix is drawn to evaluate the classification performance of double-layer BP neural network;Finally, dimensionless mathematical model and optimized double-layer BP neural network are deployed, simulate and thermal behavior prediction for semi-batch reaction, effectively solve the problem that reaction system thermal behavior identification angle is not comprehensive, with stronger generalization ability and universality.

The invention relates to a dangerous chemical accident consequence prediction method and system. The prediction method comprises the following steps: constructing time sequence characteristic information of dangerous chemicals; optimizing different weight updating modes used in the training process of the long and short term memory network model through a regularization algorithm, and obtaining a fusion model trained by adopting different weight updating modes; and according to the prediction performance of each fusion model on the accident consequence of the dangerous chemicals, inputting the time sequence characteristic information into the fusion model meeting a preset condition, and generating the accident prediction consequence of the dangerous chemicals. According to the method, the consequences of accidents such as leakage, deflagration and fire disasters of the dangerous chemical storage tank can be quickly predicted more efficiently and accurately, and reliable technical support is provided for emergency site handling personnel.

The invention relates to the technical field of highway bridge safety monitoring, and discloses a bridge dynamic weighing algorithm based on a regularization algorithm. According to the method, a penalty term is added into an error function to obtain an axle load calculation formula with a regularization matrix, and the regularization matrix is determined by an axle load covariance matrix and a load response covariance matrix. The method specifically comprises the steps of obtaining influence line distribution through a calibration test; giving an initial regularization matrix to obtain an initial axle weight; calculating a load response covariance matrix based on the influence line distribution, the initial axle load and the measurement noise, and iterating a stable axle load covariance matrix in combination with a preset initial axle load covariance matrix; generating a regularization matrix according to the load response covariance matrix and the updated axle load covariance matrix, and calculating a new axle load; and repeating the iteration process until the axle load update difference value is smaller than a preset value, and finally outputting an axle load result. According to the invention, the axle load identification precision of the bridge dynamic weighing system is effectively improved.

The invention discloses an AI-based digital collection background music recommendation method, and relates to the technical field of AI. The method comprises the following steps: acquiring a key frame average image, a dominant tone vector and a style label of a digital collection, and candidate music samples and audio signals of a music library; extracting dynamic features of the digital collections to generate rhythm representation vectors; processing the music audio to generate a rhythm structure vector, and calculating the rhythm similarity between the rhythm structure vector and the rhythm structure vector through a dynamic time warping algorithm; generating a digital collection visual style vector and a music style vector, and calculating a style consistency score by using a cosine similarity algorithm; an auxiliary music feature vector is generated through the digital collection visual style vector, and a guide score is calculated in combination with the bottom layer audio features; and calculating a comprehensive score index according to the rhythm similarity, the style consistency score and the guide score. According to the invention, background music recommendation is carried out on the digital collections through the comprehensive scoring index.

The invention discloses an electrical impedance tomography method based on parameter optimization and boundary clustering, which combines penalty functions of two regularization algorithms Tikhonov and TV for application, proposes that a particle swarm algorithm is used for regularization parameter optimization of the combined penalty functions, and takes an image quality index (AL) as a fitness value of the particle swarm algorithm, so as to determine an optimal regularization parameter, and finally to obtain the electrical impedance tomography based on the parameter optimization and boundary clustering. The conductivity is obtained through a Newton iteration method, in order to further remove the artifacts, a boundary object B (artifacts) of the target is obtained through a Niblack algorithm, deep retrieval is conducted on the boundary object B of the target through a boundary clustering algorithm, all background pixels A are obtained, and the final conductivity distribution can be obtained. Simulation and actual measurement results show that the image reconstructed by the paper method can more accurately reflect the position information of the target object in the electric field, artifacts are effectively inhibited, and the reconstruction effect is improved.

The invention discloses a curative effect prediction method based on multi-organ metastasis genome data, and belongs to the technical field of medical models, and the method specifically comprises the following steps: collecting clinical pathological characteristics, multi-organ metastasis genome data and a treatment scheme of a breast cancer patient, and recording a metastasis part and a load state; dimensionality reduction is conducted on high-dimensional genome data through a regularization algorithm, feature importance is evaluated in combination with a nonlinear model, and clinical, treatment and genome features related to treatment response are screened out; inputting the screened features into a machine learning and deep learning framework, randomly dividing a training set and a test set in a layered manner, optimizing hyper-parameters through cross validation, and constructing a classic machine learning set model and a deep learning model based on an attention mechanism; disturbing test queue treatment scheme data, evaluating the consistency of model recommendation and an actual scheme, and verifying the prediction capability and clinical practicability of the model; according to the method, multi-dimensional data are integrated, and the curative effect prediction accuracy of the metastatic breast cancer is improved.

The invention is suitable for the technical field of smart cities, and provides a road inspection disease deduplication method, which comprises the following steps: collecting road disease data from different mobile inspection devices; performing space-time fusion processing on the road disease data through a preset space-time fusion model to obtain road disease fusion features; calculating an optimal transmission distance between the road disease fusion features through a preset entropy regularization algorithm, and generating a cost matrix; and based on the cost matrix, determining target road disease data, and performing de-duplication processing on the target road disease data. The method solves the problems that an existing disease duplicate removal method depends on static visual features of diseases, the essential similarity of the diseases under geometric deformation cannot be effectively captured, and the same disease is easily and repeatedly detected.

The application provides a power distribution room dynamic temperature control method and system based on heat tracing, a fine thermodynamic model of the power distribution room is constructed, and inverse problem solving is performed on multi-point temperature monitoring data based on a total variation regularization algorithm, so that accurate tracing and visual positioning of hidden and non-uniformly distributed heat sources are realized; on this basis, the system dynamically solves an optimal control strategy according to the heat source distribution obtained through real-time inversion, in combination with a preset cooling efficiency matrix, so as to drive each cooling device to perform directional, quantitative and collaborative precise cooling. The method and system effectively overcome the inherent defects of the traditional feedback regulation mode based on average temperature, such as difficulty in identifying local hot spots and mismatch of cooling resource allocation, achieve the technical effects of significantly improving temperature control accuracy and equipment operation reliability while greatly reducing the energy consumption of the cooling system, and provide an innovative solution for intelligent and efficient temperature management of closed electrical spaces such as power distribution rooms.

A multi-exponential feature extraction ECT image reconstruction method first reconstructs the image grayscale matrix and sensitivity matrix through zero vector expansion to construct a sparse observation equation. The grayscale vector obtained by the Tikhonov regularization algorithm is then modeled as a finite innovation rate signal. This grayscale vector is filtered using an exponential regeneration sampling kernel and uniformly sampled to extract feature information. After obtaining measurement values from the sampled samples, the sparse observation matrix is combined with the FRI observation matrix, and the sparse observation vector is combined with the FRI observation vector. The row vectors of the integrated observation matrix and integrated observation vector are randomly reconstructed to obtain the integrated observation equation. The capacitance tomography image reconstruction problem is converted into an L0 norm minimization problem. The L0 norm minimization problem is solved to obtain an estimate of the sparse vector, and the grayscale vector estimate is calculated. Finally, the original image is reconstructed. The present invention achieves good image quality and high reconstruction accuracy.

A method includes obtaining k-space data acquired by an MRI scanner from a single channel body coil utilizing a multi-shot EP-DWI pulse sequence and sampling the k-space data for a plurality of shots so that for each shot both a central k-space is fully sampled to form a central calibration region and an outer k-space is partially sampled by a factor equal to a number of shots. The method includes reconstructing an initial fully sampled k-space estimate for each shot utilizing both partial Fourier constant sampling and projection on convex sets reconstruction, wherein the plurality of shots is treated as a plurality of channels for filling in missing k-space for a respective shot. The method includes utilizing a low-rank regularization algorithm in an iterative manner to generate a reconstructed image for each shot, wherein the initial fully sampled k-space estimate for each shot is utilized as an initial guess.

The application belongs to the technical field of industrial manufacturing, and provides a kind of martensitic stainless steel finishing thickness control optimization method, including data collection and arrangement, the data collected and arranged are cleaned and preprocessed, variables are screened using L1 regularization algorithm, a regression model is established and an optimization scheme is proposed by using the regression model combined with the process conditions in production;The application screens out the key variables affecting the thickness change through L1 regularization, and optimizes the parameter setting in the finishing process by using the regression model, which can significantly reduce the phenomenon of thin head and thick tail of the strip steel, thereby improving the thickness consistency and precision of the final product, avoiding the product scrap rate caused by non-standard thickness, thereby improving the production efficiency and reducing the production cost.

The application discloses a coseismic dislocation amount calculation method and system based on monocular monitoring video images, belongs to the technical field of earthquake emergency monitoring and space geodetic survey, and sequentially completes video acquisition preprocessing, feature point partition, camera vibration correction and relative displacement calculation, sub-pixel displacement optimization, pixel-physical displacement conversion, coseismic dislocation parameter solving and sliding track reconstruction and result verification output. Image sequences are extracted from monocular monitoring video streams, tracking and correction windows are selected after distortion correction and scale calibration, the influence of camera shaking is eliminated, sub-pixel level displacement measurement is realized by combining a regularization algorithm, a mapping model is established by recalibrating objects to complete displacement conversion and component decomposition, sliding parameters are solved, and a track curve is drawn. The coseismic dislocation amount measurement is low in cost, high in precision and real-time, can capture complex fault sliding dynamics, and supplements a traditional measurement method, thereby providing key data support for earthquake emergency.

This application discloses an inversion lithography method and system based on adaptive threshold regularization, which relates to the fields of inversion lithography and optical proximity correction in computational lithography. The method includes obtaining an initial mask image; optimizing the initial mask image using a CTM algorithm to obtain a grayscale value mask containing SRAFs; performing regional division to obtain a grayscale value mask after regional division; and using an adaptive threshold regularization algorithm to dynamically adjust the thresholds of each region in the grayscale value mask after regional division to obtain a binary mask containing SRAFs. The dynamic threshold adjustment optimization is to adjust the thresholds corresponding to each region in the grayscale value mask after regional division by adaptively adjusting parameter a to determine the optimal threshold for each region. Parameter a is a parameter in the regularization function in the adaptive threshold regularization algorithm. This application can obtain a binary mask that meets industrial manufacturing needs and improves the accuracy of mask imaging.

The invention discloses an object measurement-based video building detection method, which comprises the following steps of: firstly extracting an image from a video at a fixed frame rate and preprocessing the image, and then extracting a potential object region from the image by using a binarization norm gradient object measurement algorithm to obtain region data; secondly, a deep active learning algorithm is used, and scene feature learning is carried out based on regional data; and then selecting and classifying features obtained by feature learning by using a manifold regularization algorithm to obtain a video building detection result, and defining an optimization objective function. And finally, training and testing a deep active learning algorithm, a manifold regularization algorithm and a classifier based on the optimized objective function. According to the method, the important region with high object property can be efficiently and automatically extracted from the image, so that the interference of redundant information is avoided, and the semantic precision of building detection is improved.

The invention discloses a numerical control machine tool machining process energy sequence anomaly detection method, and belongs to the technical field of numerical control machine tool energy anomaly detection. The method specifically comprises: acquiring a standard energy sequence of a target workpiece in a numerical control machine tool machining process; searching and processing a climbing set and a downhill set in the energy sequence; dividing the energy sequence into k types of standard energy sequences based on a Gaussian hybrid clustering algorithm; obtaining a to-be-detected energy sequence, searching a climbing set and a downhill set in the energy sequence, and processing the climbing set and the downhill set; dividing the energy sequence into k types of to-be-detected energy sequences based on a Gaussian hybrid clustering algorithm; aligning the k types of to-be-detected energy sequences with the corresponding k types of standard energy sequences based on a dynamic time warping algorithm; calculating a power abnormal factor and a length abnormal factor of each type of energy sequence; and judging whether each type of energy sequence is abnormal or not. The application of the invention can reduce energy loss in industrial production and improve energy utilization efficiency.

The invention provides a microbial species identification method and device and a computer readable storage medium, and belongs to the technical field of species identification. The method comprises the following steps: collecting a plurality of single-cell Raman spectrum data of a plurality of microorganisms; analyzing and processing the plurality of single-cell Raman spectrum data, and determining the area of each Raman shift under a curve in a preset range; and inputting the curve area of each Raman shift in the predetermined range into a trained regularization model, and determining the species type of each microbial single cell. According to the method, species identification of microorganisms can be realized in a real-time and lossless manner at the single cell level, the method is suitable for identifying cells of target species in a complex microbiome sample, and industrial transformation can be carried out in cooperation with a single cell sorting platform. In addition, according to the method, Raman spectrums of different microbial species are analyzed by utilizing a regularization algorithm, and synchronous integration of microbial characteristic peak identification and accurate classification is realized.

The invention discloses an EIT image reconstruction quality evaluation-based dike termite nest identification method, and relates to the technical field of dike termite nest identification, and the method comprises the following steps: multi-channel EIT data collection, ant nest multi-physical field positive problem solving, image reconstruction based on a Tikhonov regularization algorithm, image reconstruction quality function calculation and image reconstruction quality evaluation. By designing a multi-channel EIT data acquisition scheme, full-coverage acquisition of electrical signals of the whole dike is realized, weak conductivity abnormal signals caused by ant caves are effectively reserved, comprehensive and pure original data support is provided for subsequent positive problem solving and image reconstruction, and the method has the advantages of being high in accuracy and high in reliability. The method ensures that signals of deep or tiny ant caves are not covered by noise, and solves the problems of limited data acquisition range, large noise interference and difficulty in capturing weak electrical characteristics of the ant caves in the dike in the existing detection technology.

The invention discloses a regularization algorithm and neural network cascaded spectrum reconstruction method and system. The method comprises the following steps: firstly, carrying out spectrum reconstruction on obtained light current data by adopting a regularization algorithm to obtain a pre-reconstructed spectrum; then the pre-reconstructed spectrum is equally divided into a plurality of sub-band spectrums according to a set interval range, and mutual interference between sub-bands is avoided; and inputting each sub-band spectrum into an adaptive neural network, and synchronously performing spectrum reconstruction on each sub-band spectrum through each neural network to obtain a sub-band reconstruction spectrum. And finally, splicing all the sub-band reconstructed spectrums into a final reconstructed spectrum according to a set fusion strategy. Therefore, the pre-reconstruction spectrum with a certain reference is used as the input of the neural network reconstruction spectrum to carry out refined spectrum reconstruction, so that the dependence of the neural network on a data set is greatly reduced, the spectrum reconstruction problem is simplified, the parameter scale of the neural network is reduced, the calculation cost and the storage space occupation are reduced, and the reconstruction rate is also improved.