446 results about "Spatial spectrum" patented technology

The invention discloses an intelligent steel pipe surface defect recognition system based on deep learning, and particularly relates to the technical field of pipe surface defect analysis. An annular polarization light source array and a high-frame-rate CMOS sensor are adopted to synchronously collect visible light and near-infrared multi-polarization images; a surface normal is calculated based on Stokes parameters, mirror surface suppression and diffuse reflection enhancement are realized, a defect candidate area is generated by fusing multi-scale Laplacian pyramid residual error and Renyi entropy segmentation threshold positioning, multi-physical quantity registration is completed through white light interference and infrared thermal imaging, a six-channel feature cube is constructed, and a three-dimensional image is obtained. According to the method, space, spectrum and thermal characteristics are jointly extracted in the multi-head attention convolutional neural network, the confidence coefficient is evaluated in combination with Jensen-Shannon divergence, and the polarization angle and the focal length are dynamically adjusted according to the confidence coefficient, so that closed-loop parameter self-optimization is realized, and the micro-scale pitting corrosion and millimeter-scale crack detection precision is remarkably improved.

The invention discloses a self-supervised hyperspectral image classification method suitable for a low-annotation sample scene, and relates to the technical field of hyperspectral remote sensing image processing, comprising a self-supervised category sensing network oriented to the low-annotation scene; in the pre-training stage, a grouping spectrum enhancement module, a spectrum self-attention module and mask reconstruction are adopted, and the model is guided to focus on category-sensitive space-spectrum features under the label-free condition by minimizing the difference between a reconstructed image and an original shielded area; in the fine tuning stage, pre-trained network parameters are used as initialization parameters, and feature expression is further refined through classification loss. Therefore, by adopting the self-supervised hyperspectral image classification method suitable for the low-label sample scene, the lossless transmission of difficult sample features is realized, the distinguishing feature expression of mixed pixels is enhanced, and the classification balance of few sample categories is improved.

The invention discloses a high-frequency shallow water bottom reverberation signal modeling method, and belongs to the technical field of underwater acoustic engineering and signal processing. The method comprises the following steps: setting a modeling hypothesis; a cylindrical array geometric structure is constructed, scattering units are divided, and channel parameters and scattering functions are obtained in combination with a Bellhop tool and a GABIM model; generating a reverberation signal time domain expression, deducing an array receiving signal spatial domain expression and calculating a spatial spectrum; the model can also adjust underwater environment parameters to realize simulation under different conditions, and performs time domain and space domain theoretical analysis on reverberation signals based on a cylindrical array. Multi-dimensional underwater environment parameters are fused, and the modeling precision is improved; a receiving and transmitting combined cylindrical array is adapted, and spatial gain is embodied; experiments prove that the average error is within 3dB, the practicability is outstanding, support can be provided for design and optimization of the sonar system, and the method is suitable for underwater reverberation signal simulation and characteristic analysis of the sonar system in the high-frequency shallow water environment.

The invention discloses an intermediate infrared polarization spectrum imaging method and device. The method comprises the following steps: carrying out joint coding on an intermediate infrared light beam carrying to-be-measured target space, spectrum and polarization information by adopting an intermediate infrared polarization spectrum modulator consisting of metasurface units which are arranged in an array and have non-correlation response characteristics; projecting the coding light field to a two-dimensional detector array to generate a two-dimensional coding image under single exposure; and decoding the two-dimensional coding image through a reconstruction unit so as to reconstruct and obtain a space-spectrum-polarization four-dimensional data cube of the target to be measured. The problem that polarization information cannot be rapidly obtained in the prior art is solved, target full information can be obtained under single exposure, and the system has the advantages of being high in imaging speed, simple in structure, easy to miniaturize and the like.

The invention relates to the technical field of earthquake prediction, in particular to a three-dimensional earthquake motion field rapid prediction method based on deep learning and physical constraint, and the method comprises the following steps: S1, data collection: obtaining original data related to an earthquake event in real time; s2, data preprocessing and multi-modal data fusion: preprocessing the original data, and fusing the preprocessed original data into multi-modal data; s3, spatial interpolation: realizing seismic oscillation feature interpolation prediction of the epicentral region, and outputting spatial spectrum data; s4, time sequence prediction and physical information optimization: generating a seismic oscillation time domain prediction sequence, and correcting errors by using an SVM model; s5, real-time reasoning and emergency response: realizing real-time minute-level response under the GPU architecture, and outputting three-dimensional seismic motion field data; according to the method, high-precision minute-level prediction of the three-dimensional seismic motion field is realized, and the efficiency and reliability of post-earthquake emergency response are remarkably improved.

According to the method, firstly, SAR data are mapped to an optical image domain by using a conditional diffusion model, pseudo-optical images with consistent spatial spectrums are generated, and fusion distortion caused by difference of SAR imaging mechanisms in a traditional method is overcome; secondly, a refined cloud region detection mechanism of a Fmask cloud mask is introduced, a cloud pollution region and a cloudless region are dynamically distinguished in combination with an adversarial training strategy, and the problem that the cloudless region is mistakenly changed in the reconstruction process of an existing method is effectively solved; besides, a multi-source integrated sample data set is constructed in stages, and a training normal form of multi-index joint optimization of PSNR, SSIM and the like is adopted, so that texture detail recovery and spectrum fidelity of a thick cloud coverage area are realized in a complex scene. In downstream application tasks such as land utilization classification and disaster dynamic monitoring, the visual quality and the quantitative index of the cloud removal result have good effects, and reliable technical support is provided for high-precision reconstruction of remote sensing information of a multi-cloud area.

The invention relates to the technical field of spectrum detection, and discloses a quality control method for traditional Chinese medicine capsules, which comprises the following steps: continuously scanning by using a micro light spot probe during capsule movement to obtain a spatial spectrum response sequence; performing adjacent micro-area differential operation on the sequence, and filtering a shell background and retaining a particle scattering signal by utilizing a microstructure difference between a capsule shell continuous film and a powder content discrete accumulation; according to the method, a traditional path of a shell standard model is established, through a spatial frequency domain decoupling mechanism, reference errors caused by batch drifting of shell physical attributes are avoided, and the detection sensitivity of trace component fluctuation and foreign matter mixing is improved.

The invention belongs to the technical field of spectrum separation identification and spatial information analysis, and discloses a flame environment-oriented multi-depth spatial spectrum light ray measurement system, which comprises a light ray collection module used for converging dynamic flame radiation light rays; the spectrum separation module is used for spatially separating a plurality of different wavelengths of the flame radiation light converged by the light collection module to form multi-wavelength radiation light with spatially separated wavelengths; the space light splitting module is used for uniformly splitting the multi-wavelength radiation light subjected to wavelength space separation into a plurality of light beams and transmitting the light beams to the multi-spectral imaging unit; the multispectral imaging unit comprises a plurality of imaging modules, and each imaging module is used for receiving radiation information of one light beam and performing imaging; and the light analysis module is used for calculating and obtaining multi-wavelength original radiation information of different depth positions of the flame. According to the invention, multi-wavelength radiation information of different depths in a flame three-dimensional space can be obtained, and a basis is provided for establishing a flame environment temperature field model.

The invention discloses an underwater sound environment sensing method based on multi-array-element sparse channel estimation. An underwater sound receiving end receives signals transmitted through an underwater sound multipath channel through a multi-array-element array; performing Hilbert transform on the receiving signal of each array element to obtain an analysis signal, and calculating a cross-correlation function of the analysis signal and the transmitting signal; based on the cross-correlation function, sparse channel parameters, including path amplitude and time delay, of each array element are estimated by adopting an orthogonal matching pursuit algorithm combined with a constant false alarm detection dynamic threshold value; the method comprises the following steps: constructing an array response vector by using sparse channel parameters of a multi-array element array, and searching and estimating angles of arrival, including a direction angle and a pitch angle, of a multipath signal through a spatial spectrum peak value; and based on the arrival angles, amplitudes and time delays of the direct path and the reflection path, inverting an underwater environment structure through a ray acoustic theory, including a reflection point distance and a reflection surface normal vector, so as to realize three-dimensional perception of the underwater reflector. According to the invention, multipath resolution can be improved, and false alarm and missing detection can be effectively reduced.

The invention provides a hyperspectral image reconstruction method based on space-spectral characteristic fusion. The method comprises the following implementation steps: acquiring a training sample set and a test sample set; constructing a reconstruction network model based on space-spectral characteristic fusion; performing iterative training on the reconstructed network model; and obtaining a reconstruction result of the hyperspectral image. According to the invention, the spatial-spectral characteristic guiding network in the encoder carries out spatial-spectral characteristic extraction on input characteristics, the spatial-spectral characteristic guiding network in the bottleneck layer carries out spatial-spectral characteristic fusion, the spatial-spectral characteristic guiding network in the decoder carries out spatial-spectral characteristic detail recovery, spatial structure information and spectral distribution information in the input characteristics are fully extracted, and the spatial-spectral characteristic guiding network in the decoder carries out spatial-spectral characteristic fusion. The spatial features and the spectral features cooperatively participate in the reconstruction process, and cooperative recovery of the spatial information and the spectral information of the hyperspectral image is realized, so that the reconstruction precision of the hyperspectral image is effectively improved.

The invention discloses a coarse cereal aflatoxin detection method combining hyperspectral imaging and deep learning, and belongs to the field of agricultural product quality safety detection. The method comprises the following steps: firstly, acquiring hyperspectral imaging data of coarse cereal grains, extracting space and spectral information, forming basic data, and performing preprocessing, space-time registration and feature alignment; building a hyperspectral space-spectrum double-branch feature fusion convolutional neural network deep learning model, and importing the processed data to complete iterative training; and finally, performing same-standard data acquisition and processing on to-be-detected coarse cereals, importing the trained model, and realizing aflatoxin detection and pollution area positioning through pixel-level analysis. According to the method, deep fusion of space and spectral features is realized, the detection accuracy and generalization ability are effectively improved, the detection result is stable, reliable and traceable, and the fine detection requirement of aflatoxin in coarse cereals is met.

The invention discloses a farmland boundary automatic identification method for an agricultural unmanned aerial vehicle, and the method comprises the steps: comprehensively collecting multi-mode remote sensing data of a visible light image, a near-infrared image and a laser point cloud, combining airborne real-time positioning and attitude determination information, and carrying out the deep fusion through spatial spectrum features and elevation features, thereby obtaining a multi-modal remote sensing image; accurate recognition and vectorization output of farmland boundaries are achieved, and the problems that in the prior art, recognition precision is insufficient, the anti-interference capacity is weak, and universality of complex land parcels is poor are effectively solved.

The invention relates to the field of array signal processing, and particularly discloses a DOA joint estimation method based on a quaternion polarization sensitive array. Electromagnetic wave dual polarization components are captured through the orthogonal dipole and the loop antenna, and a quaternion observation matrix is constructed; calculating a quaternion covariance matrix by adopting a sliding window mechanism, and separating a signal / noise subspace by adopting quaternion singular value decomposition; and constructing a spatial spectrum function in combination with a quaternion steering vector, and realizing joint estimation of an azimuth angle and a pitch angle through two-dimensional search. According to the method, the unified characterization capability of quaternions on polarization-airspace information is fully utilized, and the DOA estimation precision and the anti-interference performance of the multi-polarization signal are remarkably improved.

The invention provides a remote sensing image classification method based on a spatial spectrum attention network, and relates to the technical field of remote sensing image processing, and the method comprises the steps: obtaining a hyperspectral remote sensing satellite image, carrying out the data preprocessing, and constructing a hyperspectral remote sensing image classification data set; constructing a classification model DSAFNet based on a double-branch spectrum-space attention fusion network, and taking the classification model DSAFNet as an initial hyperspectral classification model; training the initial hyperspectral classification model through the hyperspectral remote sensing image classification data set to obtain a trained classification network model; obtaining a to-be-classified hyperspectral remote sensing image; and inputting a to-be-classified hyperspectral remote sensing image into the trained classification network model to obtain an image classification result. According to the technical scheme, generation of the hyperspectral remote sensing image classification model with high recognition accuracy and efficiency is realized.

The invention discloses a FDA-MIMO radar multi-target distance angle joint super-resolution method and system, and mainly solves the problem that the existing phased array only depends on the angle to identify multiple targets, resulting in low resolution ability, and the implementation scheme of the method comprises the steps of obtaining FDA-MIMO radar echo data, and performing separation aliasing on the FDA-MIMO radar echo data to obtain a data matrix; performing conversion stacking on the data matrix to obtain NM * L-dimensional data, and calculating a covariance matrix of the NM * L-dimensional data; performing reverse array decoherence processing on the covariance matrix to obtain a new source signal covariance matrix; constructing a two-dimensional spatial spectrum function of a target distance and a target angle for the new source signal covariance matrix through a MUSIC algorithm; and performing spectrum peak search on the two-dimensional spatial spectrum function, and detecting a local maximum value of the two-dimensional spatial spectrum function to obtain a distance and angle joint estimation value of a plurality of targets. The method can improve the angle resolution capability of the target when the target angle interval is lower than the traditional resolution, reduces the influence of the amplitude-phase error on the resolution capability when the target distance interval is large enough, and can be used for target recognition.

The invention discloses a vector hydrophone array orientation estimation method for performing tensor decomposition by using a propagation operator, relates to the technical field of vector hydrophone array orientation estimation, and discloses a vector hydrophone array orientation estimation method for performing tensor decomposition by using a propagation operator. The method comprises the following steps: firstly, constructing a three-dimensional array manifold tensor composed of an array direction matrix and a vector hydrophone output matrix; respectively expanding received signal tensors according to three modes, solving a propagation operator based on a column block covariance matrix, constructing a normalized signal subspace, establishing a spatial spectrum function with a noise subspace, and obtaining a pitch angle and an azimuth angle of a sound source through spectrum peak search; according to the method, high-order singular value decomposition is avoided, the operand is greatly reduced, meanwhile, high resolution and low sidelobe direction finding performance are kept, and the method is suitable for a real-time underwater acoustic direction finding system of a ship-borne platform, a buoy platform and an unmanned platform.

A half-spectrum search DOA estimation method based on characteristic value gradient jump belongs to the field of array signal processing, and comprises the following steps: modeling a received signal to obtain an array output vector; calculating a received signal covariance matrix; introducing a complex conjugate covariance matrix and a scanning source to construct a new covariance matrix, and constructing a spatial spectrum function; constructing a characteristic value gradient discrimination function, and adaptively estimating a signal source number by using the characteristic value gradient discrimination function; scanning angles are traversed by using a half-spectrum search method, and a signal direction of arrival is verified and determined by combining an MVDR algorithm according to a sharp spectrum peak formed by a spatial spectrum function. According to the method, source number priori knowledge is not needed, the source number can be identified autonomously, the problem that estimation is inaccurate under complex conditions in traditional half-spectrum search is solved, and the method has good robustness under the scene of unknown source number. According to the method, the limitation that few traditional half-spectrum search angle estimation is inaccurate is overcome, and a new thought is provided for half-spectrum search DOA estimation under the condition that the number of information sources is unknown.

The invention provides a space-spectrum Mama hyperspectral image classification method based on a pyramid structure. The problem that an existing method is insufficient in performance in classification of complex backgrounds and fine-grained ground objects is mainly solved. Comprising the following steps: 1) acquiring a hyperspectral image, and constructing a training set and a test set; 2) designing a double-branch model based on multi-scale space-spectrum adaptive fusion, and extracting multi-scale space and spectrum feature information in parallel; 3) respectively designing a spatial feature extraction module and a spectral feature extraction module, capturing target information in a spatial domain and a spectral domain, and optimizing fusion of spatial spectral features by using a multi-scale adaptive weighting mechanism; 4) constructing a space spectrum interactive fusion module for deep interactive fusion of features extracted by space and spectrum branches; and 5) training the model until convergence, and obtaining a final classification result by using the model. The method can effectively improve the processing capability of the complex background in the image, enhance the ground feature classification precision, and significantly improve the hyperspectral image classification performance.

The invention belongs to the technical field of human body posture reconstruction, and particularly provides an MIMO millimeter wave radar and infrared camera fused human body posture reconstruction method which is used for solving the problems that an existing human body posture reconstruction method is low in resolution, poor in stability, poor in reconstruction effect and the like. According to the method, a two-dimensional spatial spectrogram of a human body is extracted through a radar signal preprocessing algorithm, and two human body posture feature representations are obtained in combination with synchronously collected infrared images; then a fusion human body posture reconstruction model based on the MIMO millimeter wave radar and the infrared camera is constructed, effective extraction and fusion of two kinds of human body posture feature data are achieved through the fusion human body posture reconstruction model, high-precision and stable human body posture reconstruction is completed, and finally stable perception of the human body posture is achieved.

The invention belongs to the field of hyperspectral image classification, and discloses a hyperspectral image classification method and system based on multi-scale spatial-spectral joint representation and dynamic context modeling, and the method comprises the steps: carrying out the feature dimension reduction processing of a hyperspectral image through principal component analysis; spatial spectrum collaborative information of hyperspectral data is deeply mined through a multi-scale spatial spectrum joint characterization module, and adaptive fusion and enhancement of spatial spectrum characteristics under different scales are realized; a dynamic context modeling strategy is introduced, and the perception ability of the model to context information is optimized by establishing a long-range dependency relationship between features; advanced feature integration and nonlinear transformation are carried out through a multi-layer perceptron, and precise classification of hyperspectral image ground objects is completed. According to the method, the performance superior to that of a current mainstream method is obtained on three public data sets, and the effectiveness and generalization ability of the method are verified.

The invention discloses a visible light infrared multi-source image target fusion detection method and system, and relates to the technical field of image target detection, and the method comprises the steps: carrying out the radiation correction and spatial registration of a multi-source image set, forming an aligned multi-image sequence, carrying out the coordinate calculation and position correlation through employing a PROSAC optimization algorithm, obtaining the position information of a preliminary screening target, and carrying out the recognition of the preliminary screening target; and inputting the position information of the primarily screened target into a dual-channel decoupling model, executing multi-scale context sensing and space-spectrum joint analysis by a visible light channel layer, performing thermal contour extraction and thermal radiation intensity calibration by an infrared channel layer, outputting a target decoupling feature vector, performing manifold mapping on the target decoupling feature vector, and obtaining a multi-source evidence body. According to the invention, through the dual-channel decoupling model and the D-S evidence fusion engine, the discrimination capability of target feature expression is significantly improved, and the accuracy of target detection is enhanced.

The invention discloses a hyperspectral unsupervised anomaly detection method based on asymmetric consensus learning. The method comprises the following steps of: 1, preprocessing an input hyperspectral image, and constructing a region-level training sample; 2, constructing a hyperspectral anomaly detection model based on asymmetric consensus learning; 3, defining a loss function Loss of the hyperspectral anomaly detection model; 4, designing an asymmetric anomaly detection normal form, and training the hyperspectral anomaly detection model; 5, calculating a reconstruction error graph P of a reconstruction error between the original image and the reconstructed image; step 6, carrying out binarization segmentation on the reconstruction error, outputting a pixel-level anomaly detection result, and obtaining an anomaly detection result R; according to the method, the spatial spectrum cooperation characteristic and the region homogeneity characteristic of the hyperspectral image are fully utilized, so that the hyperspectral anomaly detection precision is remarkably improved.

A digital-physical twin system and method for environmental process modeling and forecasting are disclosed. The system includes a digital twin server configured to receive environmental data from a real-world environment, including hyperspectral and spectroscopic imaging, simulate environmental process transitions using a predictive model based on the environmental data, and generate parameters for a physical experiment designed to validate or refine the predictive model. A physical twin device, comprising a scaled and instrumented representation of the real-world environment, is configured to execute the physical experiment under controlled conditions. Experimental data is returned to the digital twin to iteratively refine the predictive model in a closed-loop learning cycle using self-supervised and reinforcement learning. The system supports spatially-spectrally selective experimentation, including fluorescence spectroscopy, to enhance environmental sensing. This architecture enables scalable, sample-efficient modeling of processes such as post-wildfire hydrology, vegetation regrowth, and soil change.

The invention discloses a hyperspectral reconstruction method based on deep learning, and the method comprises the steps: inputting an RGB image into an encoder through constructing a hyperspectral reconstruction network model, extracting coding features, and carrying out the reconstruction of a hyperspectral image; and a high-frequency compensation module is adopted to process coding features extracted by a non-local two-dimensional attention block in the encoder so as to recover high-frequency details lost in the down-sampling process, and features after high-frequency compensation are obtained. And a reconstructed hyperspectral image is obtained through a decoder. According to the invention, through the spatial spectrum attention and high-frequency compensation module, the mutual relation of multi-dimensional features is fully explored, the loss of high-frequency information in the network model calculation process is made up, and the problems of high cost for obtaining hyperspectral images at present and insufficient reconstruction effect of the existing hyperspectral reconstruction method are solved. High-efficiency and accurate hyperspectral reconstruction is realized, and a low-cost approach is provided for obtaining a hyperspectral image.

This invention discloses a method and system for hyperspectral image denoising based on a spatial-spectral joint self-attention mechanism, belonging to the field of image processing technology. First, based on the characteristics of hyperspectral images, a spatial-spectral joint self-attention mechanism network is constructed. The noisy hyperspectral image is used as input to the network to extract spatial-spectral features. Then, a global spectral self-attention mechanism is used to extract the band correlations of the hyperspectral image. Finally, the extracted spatial-spectral features are reconstructed using a multiple perceptron and residual connections to reconstruct a clean, noise-free hyperspectral image. The system includes a feature extraction subsystem, a non-local spatial self-attention subsystem, a global spectral self-attention subsystem, and an image reconstruction subsystem. This invention can effectively restore noisy hyperspectral images to obtain noise-free hyperspectral images. Compared to convolutional networks, it can better model long-range dependency information and has better adaptability to target hyperspectral images.

The invention discloses a tire surface defect detection method based on multi-scale image sharpening, and the method comprises the steps: collecting a visual detection image of a tire surface, and carrying out the preprocessing of the visual detection image to generate a standardized visual detection image; carrying out multi-scale decomposition and sharpening processing, and carrying out fusion reconstruction to obtain a sharpened enhanced image; texture interference suppression and abrupt change region detail enhancement are carried out to obtain an interference suppression image; executing a TR-MUSIC algorithm to generate a global abnormal spatial spectrogram; constructing an improved U-KAN network model, and generating a defect segmentation mask; extracting defect area characteristic parameters, and outputting a tire surface defect detection result. According to the invention, through combination of multi-scale image sharpening enhancement, the TR-MUSIC algorithm and the improved U-KAN network model, high-precision automatic detection of weak and small defects on the tire surface under a complex texture background is realized.

The invention relates to a high-resolution wavefront measurement method and device for a complex light field, electronic equipment and a storage medium, and the method comprises the steps: obtaining an observation image representing the to-be-measured wavefront intensity distribution based on to-be-measured light waves passing through a scattering sheet; obtaining initial complex amplitude space distribution corresponding to the wavefront to be measured based on the observation image; and inputting the initial complex amplitude space distribution and the observation image into a target function, taking the complex amplitude space distribution under the condition that the numerical value of the target function is minimum as the complex amplitude space distribution corresponding to the wavefront to be measured, and performing wavefront inversion by minimizing the target function to obtain the wavefront to be measured. The objective function comprises a data fidelity term, a spatial smoothness constraint term and a spatial spectrum bandwidth constraint term of the wavefront to be measured. And for a complex light field with strong aberration, turbulent flow or high-order singular points, the high-resolution wavefront is quickly reconstructed. Object information recording and imaging can be supported under the defocus condition, and the method is further suitable for application requirements of rapid and high-throughput detection and the like.

The invention discloses a system and method for generating, regulating and controlling a partially coherent space-time light field. The system comprises a pulse shaping assembly composed of a plurality of optical elements and an incident light source used for providing a completely coherent ultrafast pulse light field. An optical element in the pulse shaping assembly comprises a first blazed grating, a first cylindrical lens, a two-dimensional dynamic holographic modulator, a second cylindrical lens and a second blazed grating which are sequentially arranged in the light field propagation direction. The focal lengths of the first cylindrical lens and the second cylindrical lens are the same, and the distance between the adjacent optical elements is equal to the focal lengths of the first cylindrical lens and the second cylindrical lens; the two-dimensional dynamic holographic modulator is used for loading a specific complex amplitude modulation function to reconstruct the spatial-spectral characteristics of the light field. According to the invention, the two-dimensional space-spectrum function of the pulse is dynamically modulated, the generation of the partially coherent space-time wave packet is realized, and the amplitude and structure distribution of the space-time coherence can be designed and regulated as required.

The invention discloses a hyperspectral image classification method based on light spectrum hybrid adaptive band selection, which belongs to the technical field of remote sensing image processing, and comprises the following steps: based on a data cube sample set and label vectors, carrying out hierarchical random sampling in proportion, and dividing a training set and a test set; inputting the training set into a hyperspectral image classification model based on light spectrum hybrid adaptive band selection, and performing forward propagation to record an optimal weight; inputting a test set into the model, performing forward propagation by using the optimal weight, outputting a pixel-level category probability through a convolution integral category head, and determining a prediction label; the forward propagation comprises the steps of establishing learnable weight vectors for all spectral bands at the first layer of a data loader, performing band-by-band weighting on each cube sample, extracting spatial spectral features by using a light spectrum mixed structure combined with a local convolution global converter, performing cross attention band screening on high-level spatial spectral tensor, and performing data processing on the high-level spatial spectral tensor. And dynamically calculating the importance weight of the spectral band, and updating the learnable band weight vector by using the importance weight of the spectral band.

The invention provides a far-field sound source localization method applied to a transformer substation, and relates to the field of sound source localization, and the method comprises the steps: obtaining a target frequency sound in the transformer substation, and converting the target frequency sound into a spherical harmonic domain multi-channel observation vector; inputting the spherical harmonic domain multichannel observation vector into a preset network model to obtain a spherical harmonic domain mask matrix, and obtaining an in-band smooth covariance based on the spherical harmonic domain mask matrix; the preset network model is used for filtering the spherical harmonic domain multi-channel observation vector; according to the in-band smooth covariance, a spatial spectrum matrix is obtained, and the spatial spectrum matrix comprises spatial spectrum values in different candidate sound source directions; performing multi-sound source distinguishing on the spatial spectrum matrix to obtain a multi-source direction set, and realizing far-field sound source positioning of the transformer substation; the multi-source direction set comprises polar angle and azimuth angle coordinates of each target sound source direction and is used for reflecting the sound source direction in the transformer substation. The method solves the problems that the noise interference of the transformer substation is large and multiple sound sources are difficult to distinguish, and realizes the accurate positioning of the sound sources in the transformer substation.