61 results about "Deep level" patented technology

The invention discloses a chip layout method and system based on reinforcement learning and back-end optimization. The method comprises the following steps: constructing a layout optimization model; generating a corresponding position mask, a line mask and a view mask; fusing the position mask and the line mask to generate a local feature, and fusing the line mask and the view mask to generate a global feature; performing feature fusion on the local features and the global features to generate comprehensive features; according to the embedded vectors of the comprehensive features and the global features, generating the probability of various layout optimization actions which should be adopted, and calculating the excitation corresponding to the state after each round of action is updated and the value of the excitation; performing parameter updating according to newly stored experience; training the layout optimization model through the original data set; and eliminating the layout overlapping part to obtain a final layout map. According to the invention, a network can capture a complex mode in chip layout in a deeper level, the layout efficiency of the chip is improved, a more effective layout scheme can be explored, and the exploration efficiency is improved.

The invention discloses a deep learning-based power utilization abnormal behavior detection method and system. The method comprises the following steps of: obtaining electrical quantity modal, environment modal and log modal data of a target system, and performing data cleaning, adaptive segmentation standardization, missing value filling and timestamp adding preprocessing; respectively inputting each piece of preprocessed modal data into a Transform model based on time decay attention and time perception position coding to carry out deep feature extraction, and obtaining features of corresponding modals; the electrical quantity modal features are enhanced and spliced with the environment modal features and the log modal features to obtain fusion features, the fusion features are utilized to train an anomaly detection model based on deep learning, and the anomaly detection model is updated in real time based on a dynamic threshold strategy to judge the anomaly standard; and using the trained anomaly detection model to carry out electricity consumption anomaly detection. According to the invention, the safety and stability of the power system can be significantly improved, and the intelligent development of the power industry is promoted.

The invention belongs to the technical field of semiconductor defect detection, and relates to a deep learning-based deep energy level defect identification method, system and device, and a medium. According to the method, transient curve data of an external capacitor is modulated to obtain a time sequence feature vector containing time, temperature and bias voltage information, then a multi-defect-induced weighted capacitor feature vector is obtained through calculation and modulation of a neural network model, and the time sequence feature vector and the multi-defect-induced weighted capacitor feature vector are trained after being spliced. Until the neural network model converges, the trained neural network model predicts a defect parameter vector, defect weight distribution and transient capacitance response at the (i + 1) th moment at the ith moment; according to the method, DLTS test data is modeled and analyzed in combination with a deep learning algorithm and a semiconductor deep energy level defect kinetic model, precise recognition and quantification of a complex defect system are achieved, the problem that defect recognition precision is insufficient in an existing DLTS test method is solved, and meanwhile the method is low in requirement for test equipment and high in applicability.

A cross-domain mechanical fault diagnosis method based on multi-channel feature fusion of a CBAM and use thereof includes conducting preliminary feature extraction in a grey-scale graph formed by original signals with convolutional neural network, obtaining high-level features, and compressing the high-level features with a full-connection layer module; conducting deep-level multi-sensor feature extraction with an improved convolutional block attention module (CBAM); conducting fusion for multi-sensor features extracted with an improved convolutional block attention module and obtaining multi-sensor fusion features; and inputting the multi-sensor fusion features into a tag assignor for fault diagnosis results. In the present invention, the latest multi-channel domain adaptation fault diagnosis method is used to realize efficiently intelligent fault diagnosis tasks of bearings in different working states.

The invention discloses a gesture recognition method and device, equipment, a storage medium and a program product, and relates to the technical field of data processing, and the disclosed gesture recognition method comprises the steps: obtaining a gesture thermodynamic diagram collected by a millimeter wave radar; performing dimension reduction processing and space-time compression processing on the gesture thermodynamic diagram in sequence through a space-time feature extraction module of the gesture recognition model to obtain deep space-time features of the gesture; and performing category prediction on the deep spatial-temporal features of the gestures through a classification module of the gesture recognition model to obtain a gesture recognition result. The data size of the gesture deep-level spatial-temporal features after dimension reduction and compression is small, gesture recognition is performed based on the gesture deep-level spatial-temporal features, the requirements for hardware computing power and storage resources are low, a classification module of a gesture recognition model can quickly perform category prediction to obtain a gesture recognition result, and the gesture recognition efficiency is improved. The technical problem of how to process the distance-Doppler thermodynamic diagram and improve the gesture recognition efficiency is solved, and the gesture recognition efficiency is improved.

The invention belongs to the technical field of antenna design, and provides a geological radar antipodal Vivaldi antenna for lining deep detection, which comprises a dielectric substrate, a metal radiation patch and a parasitic patch, the two metal radiation patches are etched on the upper surface and the lower surface of the dielectric substrate; the two metal radiation patches have the same structure and are respectively composed of a radiation part and a feed part; the feed part is etched with a microstrip line, and the microstrip line is connected with the radiation part in a gradual change index curve mode. The inner edge of the radiation part is an index gradual change slot line; the side edges of the two radiation parts are symmetrically engraved with concave semi-elliptical grooves; a plurality of parallel comb-shaped open grooves which are vertical to the side edge of the dielectric substrate are formed in the semi-elliptical groove; the tail ends of the comb-shaped open grooves are attached to the contour lines of the semi-elliptical grooves at different lengths. The method can improve the accuracy and reliability of tunnel lining deep layers.

The invention discloses a multi-sliding-block die structure for a box shell die casting and a die casting method thereof.The die structure comprises four sliding blocks which are located on the four side faces of the die structure correspondingly and used for achieving four-side touch-through forming; the interior of at least one sliding block is cut to form an inclined small sliding block, and the small sliding block and the sliding block form a nested core-pulling structure which is used for forming a reverse groove or hole rib structure on the side face; the feeding hole is formed in the upper die and is used for injecting molten metal from the upper part; and an inductor is configured and used for controlling the small sliding block to exit before the sliding block in the demolding process, so that graded demolding is realized. The problem that a traditional single sliding block cannot process a groove / hole forming an included angle with the main demolding direction or forming a reverse direction with the main demolding direction is solved, and demolding with deeper features is completed while lateral core pulling is achieved through the nested small sliding block; and the small sliding block is tightly matched with the sliding block, so that the size precision and the surface quality of formed fine structures such as reverse grooves and thin ribs are ensured.

The invention relates to the technical field of metal defect detection, in particular to a defect detection method, device and equipment for metal disc parts and a medium. The method comprises the steps that ultrasonic phased array signals of a target metal disc part are obtained and preprocessed, preprocessed ultrasonic phased array signals are obtained, and the ultrasonic phased array signals comprise multi-channel sound wave signals collected by different detection faces of the metal disc part at different angles; and utilizing a pre-constructed CNN-BiLSTM-based feature extraction model to obtain multi-dimensional features fusing a time domain and a space domain, and utilizing a classifier to determine a defect detection result. The problem that different types of defects can be accurately identified and distinguished in a duplicated defect combination scene is solved, the purpose of improving the detection accuracy of deep or hidden defects is achieved, accurate fault positioning and repairing basis can be provided for operation and maintenance detection personnel, hidden defects of disc parts in a mechanical device are effectively identified, and the detection efficiency is improved. And major safety accidents are avoided.

The invention belongs to the technical field of large-scale power grid safety guarantee, and particularly relates to a secondary virtual loop intelligent checking method and system based on a substation configuration file. According to the method, an electrical information fusion graph fusing primary wiring and secondary functions is constructed by analyzing a substation configuration file; secondly, performing reverse tracing to form an independent virtual loop causal chain aiming at an execution outlet node; calculating a full-link time sequence deviation of each virtual loop causal chain; and finally, performing weighted summation on respective time sequence deviations of a plurality of causal chains which are gathered at the same outlet and have mutually exclusive starting end function logics to obtain comprehensive risk weighted time delay, and judging whether virtual loop configuration conflicts exist or not according to a comparison result of an index and a preset threshold value. According to the method, accumulated time delay risks can be found, and hidden deep logic configuration conflicts among different functions can be identified and quantified, so that the substation configuration file checking depth and precision are remarkably improved, and the safety of large-scale power grid operation is ensured.

A transformer-based low-light image enhancement method includes the following steps: constructing and training a network structure comprising a local enhancement module branch and a cascaded branch of a sliding window-based self-attention mechanism module; wherein the local enhancement module comprises symmetrical downsampling and upsampling layers, each of which is composed of a downsampling layer connected by a residual transformer module, and the upsampling and downsampling layers are connected by a residual layer, forming an hourglass residual enhancement module based on the transformer model, thereby obtaining both low-dimensional information and deeper feature information during local correction; wherein the WSAB module performs gradual global enhancement through cascading downsampling and then cascading WSAB modules, ultimately obtaining a trainable gamma parameter and a global joint correction matrix; and using this network structure to enhance low-light images. The network designed by this method can improve the ability to extract deep features, improve the ability to extract global feature parameters, and improve the low-light image enhancement effect.

The invention relates to the technical field of deep learning, and particularly provides a joint classification method for hyperspectral images and laser radar data. The method comprises the following steps: firstly, extracting multi-scale features of multi-modal data through a decoupling three-branch encoder, and then carrying out attention guidance cross-modal fusion on three initial features of the same scale through an autonomously designed feature fusion module to obtain three multi-modal fusion features of different scales; inputting the deepest fusion feature into a weight sharing decoder, and obtaining pseudo data through a reconstruction task; and the real data and the pseudo data are identified by the structure consistency perception discriminator, and the discrimination loss is returned, so that the encoder parameters are updated, and iteration is performed to the optimal. And finally, inputting the three-scale fusion features obtained based on the optimal parameters into a multi-level feature fusion classifier to complete classification, thereby effectively improving the classification precision, and realizing the adaptive fusion and enhancement of the multi-modal features.

The present invention claims protection for a face forgery detection method (CDCL) based on cross-domain consistency learning, which aims to explore the common forgery features of images generated by different forgery methods and achieve highly generalized face forgery detection, belonging to the field of computer vision technology. The method comprises the following steps: Step 1. The present invention proposes a cross-domain learning module, which helps to extract cross-domain consistency features of images generated by different forgery methods. Step 2. The present invention designs a diversion center differential attention, which can generate key and value pairs by aggregating the pixel-level intensity and gradient information of the query, and inject heterogeneous receptive field sizes into the label to capture deep-level fine-grained features. Step 3. The present invention uses data generated by a set of specific forgery techniques (source domain) and different forgery techniques (target domain) to induce the model to learn cross-domain consistency features. Step 4. In order to reduce the intra-class distance while increasing the inter-class distance, the present invention designs a cross-domain consistency center loss, and obtains the final network model through iterative adversarial training and parameter updating.

The invention relates to the technical field of semiconductor photoelectricity, in particular to a blue light LED growing method for improving carrier recombination efficiency, which comprises the following steps of: S1, respectively extending a Buffer layer, a 3D layer, a GR layer and an N-GaN layer on sapphire according to a conventional blue light LED manufacturing process; and S2, a layer of N-Al0. 2Ga0. 8N is additionally arranged behind the N-GaN layer, and SiH4, TMAl and TEGa are introduced at the same time. And S3, continuing to be consistent with the conventional blue light LED manufacturing process, and extending Q1, Q2 and Q3 quantum well layers. And S4, according to the Last Barrier layer, a conventional Al < 0.3 > Ga < 0.7 > N + AlN layer is changed into an Al < 0.7 > Ga < 0.3 > N-Al < 0.65 > Ga < 0.35 > N gradient layer, and the time is shortened by half. S5, a hole storage acceleration layer is formed on the rear portion of the Last Barrier layer in an epitaxial mode, P-In0. 2Ga0. 8N is introduced, a shallow energy level layer is formed, the time is consistent with that of S4, and S6, an Al0. 45Ga0. 55N-Al0. 5Ga0. 5N gradient layer is formed on the rear portion of the Last Barrier layer in an epitaxial mode again, and the time is half of that of S5. The quantum well structure is redesigned, and the n-AlGaN and the hole storage acceleration layer are added to realize the deceleration of electrons in the quantum well and the acceleration of holes, so that the main light-emitting region is migrated to a deep level, the carrier recombination efficiency is improved, and the light-emitting efficiency is further improved.

A hibernation depth management method and a solid state disk (SSD), the SSD periodically determines a set hibernation depth for itself according to a hibernation depth that an application function in the SSD can enter; wherein the set hibernation depth is a deepest level of hibernation depth that each application function in the SSD can enter. Embodiments of the present disclosure enable the SSD to adaptively adjust the hibernation depth, thereby achieving autonomous optimization of power consumption.

The present application relates to the technical field of metal defect detection, and specifically to a defect detection method, device, equipment and medium for metal disc parts. The method includes: obtaining an ultrasonic phased array signal of a target metal disc part and preprocessing it to obtain a preprocessed ultrasonic phased array signal, wherein the ultrasonic phased array signal includes a multi-channel sound wave signal collected at different angles from different detection surfaces of the metal disc part; using a pre-built CNN‑BiLSTM-based feature extraction model to obtain a multi-dimensional feature that integrates the time domain and the spatial domain, and using a classifier to determine the defect detection result. It solves the problem of being able to accurately identify and distinguish different types of defects in the scenario of processing duplicate defect combinations, achieves the purpose of improving the detection accuracy of deep or hidden defects, can provide accurate fault location and repair basis for operation and maintenance inspection personnel, effectively identify hidden defects of disc parts in mechanical devices, and avoid the occurrence of major safety accidents.

This application relates to the technical field of shield machine cutter head evaluation. It discloses an intelligent evaluation system and method for the state of a shield machine cutter head. By collecting multi-source data from torque, current, and vibration sensors, and using the sliding window feature extraction technology, rich statistical features of the operating state are obtained. Further, temporal context encoding is performed on the time queue of the statistical features of the cutter head operating state to obtain the temporal context encoding features of the cutter head operating state, so as to capture the deep-level information of the cutter head operating state. Finally, the encoded temporal context encoding features of the cutter head operating state are input into the trained GBRT model to achieve accurate prediction of the tool wear degree. This method not only improves the accuracy and comprehensiveness of state monitoring, but also enhances the early warning ability for potential faults, providing guarantee for the safe and stable operation of the shield machine.

The present application relates to a method, device, equipment and medium for verifying the commissioning status of power equipment, and relates to the technical fields of power engineering auditing and artificial intelligence. It includes: parsing the submitted review materials using a multimodal document parsing model; inputting the dispatching operation power time series into a variational autoencoder model for sequence reconstruction calculation to obtain the reconstruction error and generate a time series anomaly feature component; generating a trajectory space anomaly feature component based on the spatio-temporal data of project personnel clock-in and the benchmark coordinates of the equipment ledger, and at the same time performing error level analysis on the on-site commissioning image data to generate an image anti-counterfeiting anomaly feature component; submitting the above multi-source heterogeneous data to a multimodal large model for cross-comparison and logical reasoning, breaking the limitation of only relying on the submitted review materials by a single party, identifying forgery behaviors from the source. With the equipment identifier as the only primary key, in the face of false commissioning scenarios with strong concealment and across multiple links, a multi-dimensional cross-verification closed loop is formed to identify deep造假行为 is misspelled. It should be "deep造假行为 is misspelled. It should be "deep forgery behaviors" here. So the corrected translation is as follows: The present application relates to a method, device, equipment and medium for verifying the commissioning status of power equipment, and relates to the technical fields of power engineering auditing and artificial intelligence. It includes: parsing the submitted review materials using a multimodal document parsing model; inputting the dispatching operation power time series into a variational autoencoder model for sequence reconstruction calculation to obtain the reconstruction error and generate a time series anomaly feature component; generating a trajectory space anomaly feature component based on the spatio-temporal data of project personnel clock-in and the benchmark coordinates of the equipment ledger, and at the same time performing error level analysis on the on-site commissioning image data to generate an image anti-counterfeiting anomaly feature component; submitting the above multi-source heterogeneous data to a multimodal large model for cross-comparison and logical reasoning, breaking the limitation of only relying on the submitted review materials by a single party, identifying forgery behaviors from the source. With the equipment identifier as the only primary key, in the face of false commissioning scenarios with strong concealment and across multiple links, a multi-dimensional cross-verification closed loop is formed to identify deep forgery behaviors.

The application discloses a method for isolating GaN-based devices, wherein a Fe film is evaporated on a region requiring device isolation on a surface of a GaN HEMT structure, a pulse laser is used to heat the Fe film covering region, laser pulse energy is reasonably controlled, Fe is accelerated to diffuse into the sample and the sample is recrystallized, so that Fe is incorporated into the sample lattice. Fe impurities are deep level impurities in GaN-based materials, and doping Fe in the region requiring isolation can form a high resistance state, so that the purpose of device isolation is achieved. The method can reduce the defect density of the isolation region by recrystallizing the sample through high-energy laser pulses and introducing Fe impurities in the recrystallization process.

The application discloses an ultrasonic signal classification and identification method, medium and equipment, and belongs to the technical field of signal processing and deep learning; the method comprises the following steps: collecting ultrasonic signals of excitation channels and receiving channels under different working conditions in structural ultrasonic detection, reconstructing channel signals into time sequence samples; inputting the time sequence samples into a double-branch deep learning model, and outputting working condition category probability distribution results; the double-branch deep learning model obtains the working condition category probability distribution results through a time domain feature extraction module, a space domain feature extraction module, a feature fusion module and a classification and prediction module; the application fully considers and utilizes the excitation and receiving signal features in ultrasonic detection, adopts a time and space domain feature fusion mode to fully capture multi-scale feature information and perform deep-level fusion, solves the problem that a single network is difficult to represent ultrasonic signal features from a multi-scale level, has strong engineering practicability, and can be widely applied in the fields of structural nondestructive detection, structural health monitoring and the like.

The application provides a lightweight steel surface defect detection method and device and processing equipment, which are based on a YOLOv11 model, deep adaptability optimization is performed, the steel surface defect detection task has better detection precision, the parameter quantity and the calculation quantity can be effectively reduced, the generalization ability is good, when the industrial terminal is deployed, the high performance and the lightweight can be considered, and thus the industrial application prospect is good.

The invention provides a high-electron-mobility silicon MOS (Metal Oxide Semiconductor) quantum device and a preparation method thereof. The method comprises the following steps: growing a dielectric layer on a silicon substrate; etching the dielectric layer and the silicon substrate to form a mesa structure; removing the dielectric layer by wet etching; cleaning the sample from which the dielectric layer is removed; growing a dielectric layer on the cleaned sample; repeatedly executing the step of removing the dielectric layer by wet etching until growing the dielectric layer on the cleaned sample for preset times; and preparing a metal electrode on the sample which is repeated for preset times. According to the preparation method, after RIE etching, silicon dioxide circulation oxidation is adopted for multiple times, an oxide layer is removed in combination with a buffer oxide etching solution, surface roughness scattering of the MOS structure can be greatly reduced, ionized intrinsic deep energy level defects introduced in the dielectric layer due to RIE table top etching can be removed by corroding original silicon dioxide, and the carrier mobility is effectively improved.

The embodiment of the invention discloses a block chain virtual machine fuzz test method, which guides a generative large model to generate diverse and targeted smart contract codes as virtual machine test input by designing a cue word iterative update mechanism, improves the fuzz test efficiency, and improves the fuzz test accuracy. And the possibility of finding vulnerabilities hidden in the deep region of the block chain virtual machine is increased. The block chain virtual machine fuzz testing device disclosed by the embodiment of the specification also has the above beneficial effects.

The application relates to the field of semiconductor technology and discloses a gallium oxide single crystal substrate defect map construction method and equipment, which comprises the following steps: preparing an ohmic contact on the back surface of a gallium oxide single crystal substrate to form a capacitor structure for deep level transient spectrum measurement; setting a high-density two-dimensional measurement grid point on the surface to be measured; controlling an automatic translation platform to make the gallium oxide single crystal substrate move in sequence, automatically collecting deep level transient spectrum signals on each two-dimensional measurement grid point through a programmable fast DLTS measurement system to obtain DLTS spectra of each two-dimensional measurement grid point; automatically analyzing the DLTS spectra of each two-dimensional measurement grid point, identifying deep level defects of different categories, and calculating the concentration of each category of deep level defects in each two-dimensional measurement grid point; and based on the identified defect categories and the concentration data of each two-dimensional measurement grid point, constructing a full-field defect map for representing the spatial distribution information of deep level defects on the gallium oxide single crystal substrate.

The application discloses a kind of hypersonic gas-solid interface atomic collision event statistics method.The method comprises: simulating the high-speed flight of object in gas environment in NAMD software, output the position and speed information of all atoms in system at different time;Using VMD software, all gas atoms that can interact with the front surface in a certain period after system steady state are counted, and the coordinates, speed and potential energy of these gas atoms are output;Define the time when the gas atom turns in the gas-solid atomic repulsive force layer as the collision point, the incident point of the gas atom into the attractive force layer is obtained by pre-searching from the collision point, and the reflection point when the gas atom leaves is obtained by post-searching, which is recorded as a collision event, and the collision event information of different atoms and surface is obtained by circulation.The application proposes a kind of gas-solid interface atomic collision event statistics method under hypersonic strong shock wave condition, which can study the energy transfer mechanism of gas-solid interface from a deeper level and reveal the nature of aerodynamic heating.

The invention discloses a Transform-based electroencephalogram recognition method, which belongs to the technical field of radar signal processing and comprises the following steps: acquiring an electroencephalogram signal; the electroencephalogram signals serve as input signals, a brain region Transform module is adopted to extract first features, and the first features at least comprise global spatial features of the electroencephalogram signals in all brain regions and on the whole brain region and global spatial dependency among all electrodes; the first feature is used as an input signal, a time Transform module is adopted to extract a second feature, and the second feature at least comprises a global time sequence feature of the electroencephalogram signal; the second features serve as input, a space-time multi-scale convolution module is adopted to extract third features, namely final features, and the third features at least comprise deep-level space-time features of the electroencephalogram signals, so that the problems that space-time information extraction is incomplete and recognition performance is low in the feature extraction process are solved.

This application relates to an acoustic lens, an ultrasonic probe, and an ultrasonic imaging device. The acoustic lens has a front surface, and the curvature of the front surface includes a central region and side regions located on both sides of the central region. The focal length of the central region is greater than that of the side regions. This acoustic lens adjusts the focal length in sections to achieve multi-dimensional, multi-point focusing. Ultrasonic waves passing through the central region are focused at a deeper level of the target, while those passing through the side regions are focused at a shallower level. Furthermore, the beam narrowing is significant, enabling the acquisition of clear images from shallow to deep within complex structures, thus improving the quality of ultrasonic images.

The invention discloses a preparation method of a high-quality surface pure rhenium sample, which comprises the following steps of: cutting and sampling a pure rhenium original material, grinding, polishing and cleaning by ultrasonic vibration, loading the sample on a sample clamping device, then putting the sample into a mixed solution, refrigerating and carrying out electrolytic polishing treatment to obtain the high-quality surface pure rhenium sample. And a pure rhenium sample with a high-quality surface can be obtained. According to the method, the defects of an existing high-quality surface pure rhenium sample preparation method are overcome, single or multiple samples can be prepared at the same time through the method, the quality effect is excellent, the rhenium crystal information analysis rate reaches 99% or above, the urgent requirement of the academic circle for the high-quality surface pure rhenium sample is met, and the method is suitable for popularization and application. And a good technical support and a material preparation basis are provided for macroscopic performance analysis of pure rhenium and deep research work of microscopic deformation mechanism. The method has the advantages of being simple, convenient, efficient, high in safety, low in cost, excellent in technical effect and the like.

The invention discloses a circuit repairing and detecting method. The circuit repairing and detecting method comprises the steps of providing a circuit; the circuit comprises an upper layer area and a target area. Exposing the target area; performing circuit repairing on the to-be-repaired area of the target area; isolating the target area to expose at least two areas to be detected; forming a to-be-detected point on the to-be-detected area; metal impurities in the target area are removed; and using the nano probe machine table to be in contact with the to-be-tested point to carry out an electrical property test. According to the invention, through combining the characteristics of the focused ion beam machine and the nano probe machine, the electrical test of the deep circuit repair micro area is realized, and the combination not only improves the test precision, but also significantly shortens the test time; the to-be-detected points are stacked in the to-be-detected area, so that compared with the prior art needing to depend on the existing test points or large-area lead-out, the short-circuit risk is greatly reduced, the electrical test of the independent target in the circuit repair area can be realized, and the deeper to-be-detected area can be detected.

The application provides a defect detection method and system for an EUV photomask body, which can scan at least one to-be-detected position point of the EUV photomask body (including an EUV mask blank or an EUV photomask plate with a corresponding pattern) by using extreme ultraviolet laser with different incident intensities without damaging the EUV photomask body, obtain corresponding reflectivity, and obtain defect information of the corresponding to-be-detected position point by analyzing the reflectivity, including a distribution range in a transverse direction and deep information.

A temperature noise correction method for CMOS space cameras based on attention mechanisms and LSTM is presented. This method relates to the field of optical remote sensing technology, specifically to the field of temperature noise correction for CMOS space cameras. The core of the method is the use of a multi-level Long Short-Term Memory (LSTM) network with an attention mechanism to explore how temperature changes affect the noise performance of CMOS space cameras under two different operating conditions: dark and bright fields. This deep learning model possesses powerful autonomous learning capabilities, capable of mining and understanding complex noise patterns hidden within massive amounts of data, thereby improving the accuracy of noise identification and calibration. Its key advantage lies in its end-to-end learning approach, automatically learning and establishing a deep, nonlinear relationship between temperature and noise directly from the raw data, without requiring manual intervention for feature extraction or designing complex correction algorithms. This significantly enhances the model's generalization ability and adaptability to unseen data scenarios.