118 results about "Sigmoid function" patented technology

The invention discloses a forest fire risk prediction method based on multi-source data, and belongs to the technical field of forest fire prevention. Aiming at the problem of low prediction precision caused by one-sided information of a single data source and insufficient multi-source data fusion in the prior art, the method is realized by the following steps: acquiring micrometeorological data including temperature and humidity, wind speed and air pressure, and image data including an infrared image and a visible light image; the data of the mountain fire-prone area comprises historical fire frequency, vegetation type and topographic information; uTC + 8 time synchronization and WGS84 coordinate system space calibration are carried out on the data, and missing values and abnormal values are processed; carrying out feature layer fusion by adopting an attention mechanism, and extracting core features such as a temperature and humidity coupling index and vegetation dryness; spatial correlation features are captured through CNN, a time sequence trend is captured through LSTM, a mountain fire occurrence probability is output by using a Sigmoid function after decision-making layer fusion, and a result is calibrated in combination with sub-region features. Through multi-source data deep fusion and spatial-temporal feature collaborative analysis, the accuracy and timeliness of forest fire risk prediction are improved, a new data source can be expanded and accessed, and the method is suitable for a complex forest fire prevention scene.

The invention discloses a photovoltaic power station defect detection method based on thermal imaging data, and relates to the technical field of artificial intelligence, and the method comprises the steps: S1, collecting the thermal imaging data of a photovoltaic module under different working conditions, and carrying out the preprocessing of the thermal imaging data, and obtaining a thermal imaging data set; s2, based on the thermal imaging data set, dynamically dividing a temperature interval and calculating an adaptive bandwidth, and performing weight distribution and color band center weighted fusion by using a Sigmoid function to generate an adaptive pseudo-color image capable of enhancing local contrast; s3, constructing a double-branch defect detection model based on the original temperature data and the pseudo-color image; and S4, performing defect detection on the real-time thermal imaging data based on the double-branch defect detection model to obtain a defect detection result of the current photovoltaic power station. The method can be effectively applied to large-scale inspection operation and maintenance of the photovoltaic power station.

The invention discloses an AI generated image detection method and device and a storage medium, and the method comprises the steps: obtaining a to-be-detected image, and inputting the to-be-detected image to an image encoder of a counterfeit detection model; extracting a depth visual feature vector of the to-be-detected image through the image encoder; performing linear operation on the depth visual feature vector and a preset weight parameter matrix, and superposing a preset offset parameter in the linear operation process to obtain an intermediate scalar value; the middle scalar value serves as input of a Sigmoid function, and a scalar confidence coefficient score is calculated; and judging whether the to-be-detected image is an AI generation image or not according to a comparison result of the scalar confidence score and a dynamic judgment threshold. According to the method, depth features are extracted through cross-modal alignment training, linear mapping efficient reasoning is carried out, the generalization ability of an unknown generation model is remarkably improved, dynamic threshold value self-adaptive adjustment is combined, lightweight deployment is achieved while high precision is guaranteed, and the actual application requirement is met.

The invention belongs to the technical field related to deep learning, and discloses a training method and application of a three-dimensional flow field prediction system of an underwater vehicle, and the training method comprises the steps: calculating the mass center of a neighbor point set for each surface grid point of a vehicle model, constructing a covariance matrix between the neighbor point set and the centroid of the neighbor point set, and performing eigenvalue decomposition to obtain a normal vector of the point; calculating a normal vector included angle between the surface grid point and each point in the neighbor point set, taking the obtained maximum included angle as the geometric feature measurement of the point, and converting the geometric feature measurement into a weight through a Sigmoid function, thereby obtaining the sampling probability of each point; randomly extracting surface grid points from the original surface grid of the aircraft based on the obtained sampling probability to obtain a point cloud of a corresponding model; and training a point cloud neural network by using the sampled point cloud data to obtain the three-dimensional flow field prediction system of the underwater vehicle. Based on the method, the prediction precision and the flow field detail recovery capability can be improved while the calculation efficiency is maintained.

The invention relates to the technical field of network acceleration, and discloses an edge-end large language model reasoning acceleration method and accelerator, and the method comprises the following steps: reconstructing a calculation process of a decoding stage, and carrying out the deep fusion of a multi-head attention mechanism and the calculation operation of a feedforward network; the weight and key value data are stored in HBM, and the coefficient and the accumulated attention score are stored in DDR; for linear matrix calculation, a unified matrix calculation unit is used for executing multi-precision matrix operation; for nonlinear function calculation, a mathematical transformation and linear fitting method is adopted, a Softmax function is converted into operation with 2 as the bottom through a bottom conversion formula, and truncation and third-order linear fitting are conducted on a Sigmoid function; and constructing a key value screening algorithm based on the accumulated attention score, dynamically adjusting a key value storage position, maintaining a recent key value cache region and an important key value cache region in a limited cache space, and realizing key value efficient cache in long text reasoning.

The invention relates to the technical field of image data processing, in particular to an MBR water production turbidity detection method and system based on image processing, and the method comprises the steps: obtaining an original image of an MBR water production pipeline, and extracting a water body region by using a mask matrix; calculating the local contrast of the pixel points relative to the neighborhood, constructing a suppression weight based on the noise standard deviation of the image sensor, and weighting the local contrast to obtain scattering response intensity; obtaining the texture disorder degree based on the ratio of the geometric mean value to the arithmetic mean value of the feature values of the structure tensor; the particle confidence is obtained based on the scattering response intensity and the texture disorder degree, the weighted particle confidence is obtained through weighting of a Sigmoid function, and the average value of the particle confidence is calculated to serve as the comprehensive turbidity so as to evaluate the water production state. The scattering response intensity and the texture disorder degree are fused, a weight suppression and soft threshold mechanism is introduced, pipe wall scratches and thermal noise interference are reduced, and the detection accuracy is improved.

The invention relates to the technical field of laser measurement and photoelectric detection, and discloses a laser boresight equipment high-precision calibration method and system based on region segmentation mapping, and the method comprises the steps: firstly collecting a standard grid frosted glass target plate image, extracting grid intersection point pixel coordinates through employing an SURF algorithm, and building a data set corresponding to physical coordinates; dividing the field of view into a plurality of sub-regions, and respectively resolving a perspective transformation matrix mapped from a pixel coordinate system to a physical coordinate system; secondly, collecting a light spot image, performing initial positioning by using a Canny operator and a least square method, fitting a sampling sequence through a Sigmoid function, and extracting edge sub-pixel coordinates in combination with a central difference method; and finally, calling a corresponding transformation matrix according to the sub-region where the light spot is located to obtain physical coordinates, and calculating the light spot jerk value, the maximum diameter and the included angle between the optical axis and the mechanical axis. According to the invention, image distortion is compensated through view field partition mapping, and the calibration precision and detection efficiency of the laser boresight equipment are improved in combination with a sub-pixel positioning technology.

The invention discloses a damage identification method and device for an in-service steel wire rope type horizontal lifeline, and belongs to the technical field of high-altitude operation safety facilities. The method comprises the following steps: synchronously acquiring a magnetic flux leakage signal and a surface image of an in-service steel wire rope through magnetic flux leakage detection equipment and a high-definition camera which are carried on a steel wire rope inspection robot; preprocessing the acquired magnetic flux leakage signal, wherein the preprocessing comprises singular value elimination and trend term removal processing; carrying out de-noising processing on the pre-processed signal by adopting an improved wavelet threshold de-noising algorithm fused with a Sigmoid function; extracting a characteristic value for representing the damage of the steel wire rope, and performing normalization processing to form a characteristic vector; and inputting into a BP neural network identification model optimized by a genetic algorithm for identification, and outputting an assessment result of the damage type and positioning of the steel wire rope. The method can realize automatic and quantitative detection and accurate identification of internal and external damages of the steel wire rope.

The invention discloses a gas molecule smell prediction method and system based on an isotropic graph neural network, and relates to the technical field of tobacco production equipment.The method comprises the steps that an SMILES expression of gas molecules is converted into a molecular graph structure containing atomic features and edge features and three-dimensional space coordinates through DGL-LifeSci and RDKit toolkits; then inputting the data into an EGNN isotropic graph neural network, aggregating neighborhood information through a message passing mechanism, and synchronously updating node features and coordinate features; a plurality of EGNN layers are stacked for deep feature extraction, and high-dimensional node representation and final coordinate values are obtained; performing global pooling through an MLPReadout module to obtain a global molecular representation vector of a fixed dimension; and finally, mapping the vector to a 138-dimensional space through a full connection layer, processing through a Sigmoid function to obtain an odor probability vector, and finally judging and outputting a specific odor category through a threshold value.

The invention provides an equipment quality and supplier intelligent evaluation method based on a space-time coupling Bayesian model and a two-ticket graph database, and belongs to the technical field of intelligent databases. The two-ticket graph database is firstly constructed, and a station environment, equipment full life cycle maintenance and supplier data are integrated; obtaining a supplier initial score by using an entropy weight method and a Sigmoid function; constructing a space-time coupling Bayesian model by means of ST-LSTM and VAE, and clustering equipment health states; introducing an environment coefficient and a KL divergence gradient to correct a score, and separating environment responsibilities from supplier responsibilities; calculating an equipment target score by combining an aging model and cosine similarity; and aggregating scores through a graph attention network, considering quality fluctuation, and generating a supplier global score. According to the invention, dynamic, precise and traceable evaluation is realized, a basis is provided for power grid supply chain management and the like, and the method is suitable for supplier evaluation in electric power and other industrial fields.

The application provides an improved multi-agent deep reinforcement learning unloading method, relates to the technical field of Internet of Vehicles and edge computing, and solves the technical problems of computing overhead, energy consumption and time delay in the task unloading process.The technical scheme comprises the following steps: S1: task unloading problem modeling;S2: improved unloading method design and implementation.The task queue smoothing algorithm designed by the application can accurately reflect the task load of the server and the mobile terminal, and assist the deep learning algorithm to make more accurate and stable unloading decisions.The application designs a smoothing penalty mechanism of a Sigmoid function in a reward function, and provides continuous and accurate gradient guidance for agent learning.The application corrects the action output by the deep reinforcement learning algorithm in real time in the early training stage through the closed-loop feedback mechanism of the PI controller, and improves the time delay performance of unloading.

The invention discloses a biological information security authentication method based on deep fusion of fingerprint features and digital passwords. The method comprises the steps that a user customizes the digital passwords; extracting fingerprint features by using a transfer learning model based on a ResNet-50 architecture to obtain five high-dimensional digital feature vectors; a Sigmoid function is used to carry out normalization on all the feature vectors; obtaining a random sequence for encryption; performing scrambling operation on all the gray level images by using three-dimensional Class-Z transformation; diffusion operation based on a semi-tensor product is carried out on all gray level images; storing the ciphertext images in a database according to a sequence input by a user; when the user attempts to verify, the system executes a dynamic step. According to the security authentication framework based on deep fusion of the biological characteristics and the digital passwords, the permanent identity risk caused by inherent non-revocation of the biological characteristics is fundamentally solved, and a multi-protection mechanism with high efficiency and high security for original fingerprint characteristics is constructed with the assistance of a dynamic updating / immediate template generation mechanism.

The present application belongs to the field of marine environment monitoring and signal processing technology, and relates to a real-time processing method for data denoising and time lag correction of a marine dissolved gas sensor, which comprises: identifying a steady-state interval of historical observation data of the sensor, fitting a steady-state jump with a Sigmoid function to generate a virtual true value of gas concentration; constructing a neural network model with observation data as input and a preliminary correction value of gas concentration as output, training the model with historical observation data as input and the virtual true value of gas concentration as a training target; using a sliding window mechanism to intercept observation data segments of the current and historical time of the sensor, inputting the data segments into the model to output a preliminary correction value of gas concentration at the current time; and using a logic state machine for post-processing to obtain a final correction value of gas concentration at the current time. The present application can provide the sensor with real-time compensation functions that take into account high dynamic response and zero steady-state drift, thereby realizing in-situ acquisition and real-time return of high-precision observation data.

The invention discloses an artificial neural network microwave device wide-range modeling method based on multi-dimensional space division and dynamic adaptive sampling, and relates to the field of microwave device modeling. The invention provides a boundary smooth splicing method based on a Sigmoid function. And overlapping unit samples are shared by the adjacent sub-regions, so that rich data support is provided for the boundary region. And on the basis, a self-adaptive correction function is constructed by utilizing a Sigmoid function: according to the position of the sub-model on each dimension, single-side Sigmoid correction is adopted in a boundary partition, and double-side correction is adopted in a middle partition. Weighted fusion is only carried out in the boundary overlapping region in the correction, so that the outputs of adjacent sub-regions are in smooth transition, and the internal precision of each sub-model is not influenced. And finally, all the corrected sub-models are superposed to form a globally continuous and seamlessly spliced ANN model, so that the problem of boundary discontinuity in multi-dimensional and wide-range modeling is effectively solved.

The invention discloses a lightweight non-repudiation model fingerprint efficient traceability method, and belongs to the related technical fields of cryptography application, digital content traceability and the like. The method comprises the following specific steps: inputting an original image data set into a computer system; training and learning a lightweight encoder and a decoder, wherein the encoder maps a binary fingerprint and an image to generate a residual error; the residual error is constrained through a Sigmoid function to generate a fingerprint-containing image, and the model is optimized through dual loss; generating metadata of a PIT parameter binding salt value and a commitment value, and embedding the fingerprint into the image; during detection, a decoder is loaded to extract a decoded fingerprint, binaryzation is carried out through a threshold function, and dual verification is realized through fingerprint polynomial comparison and commitment value verification. According to the method, a fingerprint verification mode is converted into polynomial verification through a polynomial equivalent detection technology, the communication complexity is reduced to sub-linear overhead, lightweight efficient verification is realized, and the method is suitable for efficient traceability and authenticity verification of deep forged images.

The application discloses a kind of fusion well logging data and multi-scale fracture identification method of wavelet transform, belong to oil and gas exploration and development geophysical well logging technical field.This method includes: obtaining the natural gamma, deep and shallow resistivity and acoustic travel time well logging curve of target formation;Curve is adaptively wavelet transformed, and multi-scale feature is extracted;Based on high-frequency component, the fracture response characteristic parameter set containing deep and shallow resistivity difference index, crack coefficient, relative acoustic travel time and natural gamma change rate is constructed;The parameter set is weighted and fused into preliminary fracture intensity parameter;The parameter is nonlinearly enhanced with Sigmoid function and moving average smoothing processing;Finally, the gradient of the smoothed parameter is calculated, and the fracture area is automatically divided by threshold determination and output quantitative evaluation result.The application effectively strengthens the high-frequency abnormal response of fracture, significantly improves the accuracy, stability and anti-interference ability of fracture identification using only conventional logging data.

The invention belongs to the technical field of marine environment monitoring and signal processing, and relates to a marine dissolved gas sensor data denoising and time-delay correction real-time processing method, which comprises the following steps: identifying a steady-state interval of historical observation data of a sensor, and fitting steady-state jump by using a Sigmoid function to generate a gas concentration virtual true value; a neural network model with observation data as input and the gas concentration preliminary correction value as output is constructed, and the model is trained with historical observation data as input and the gas concentration virtual true value as a training target; intercepting observation data fragments of the sensor at current and historical moments by adopting a sliding window mechanism, and inputting the observation data fragments into the model to output a gas concentration preliminary correction value at the current moment; and performing post-processing by using the logic state machine to obtain a final correction value of the gas concentration at the current moment. According to the invention, a real-time compensation function considering high dynamic response and zero steady-state drift can be provided for the sensor, so that in-situ acquisition and real-time return of high-precision observation data are realized.

The invention provides a Sigmoid modulation phase correlation surface wave full waveform inversion method and system. The method comprises the following steps: acquiring a surface wave seismic record and constructing an initial S wave velocity model; performing forward modeling to generate a synthetic record; calculating instantaneous phase correlation of observation and synthetic records to construct an amplitude-independent target function; modulating the target function by using a Sigmoid function, and mapping the value of the target function to a predetermined interval; and calculating a gradient based on the modulated target function and iteratively updating the speed model until a termination condition is met, and outputting a final model. According to the method, the smooth saturation and bounded output characteristics of the Sigmoid function are utilized, the target function response is converted into a self-adaptive gradient scaling mechanism, a self-stabilization optimizer is introduced for the inversion process, and the convergence domain is widened by applying smooth constraint, so that the dependence on the initial model precision is reduced, and the first success rate and the result reliability of inversion under complex conditions are improved.

The invention relates to the technical field of energy storage system abnormity early warning and risk grading, in particular to an energy storage abnormity early warning and risk grading system based on a large language model, which captures an abnormity early warning signal in real time through SCADA-> Kafka flow, extracts a data window, obtains a feature vector through feature quantification, inputs a GBDT model to obtain a risk score in combination with a Sigmoid function, and performs early warning and risk grading according to the risk score. According to the energy storage abnormity early warning and risk grading system based on the large language model, in 60-day operation of 15 stations, the grading accuracy rate is 93.4%, the three-level false alarm rate is 2.7%, the three-level risk is divided according to a threshold value and a hysteresis threshold value, hierarchical interpretation of technicians, operation and maintenance managers and clients is generated based on LLM, operation steps are retrieved and output through SOP, the threshold value and the model are collected, fed back and updated, and the energy storage abnormity early warning and risk grading system based on the large language model has the advantages that the grading accuracy rate is 93.4%, and the three-level false alarm rate is 2.7%. The technical staff score is 4.8 / 5, the customer satisfaction degree is 4.7 / 5, the average end-to-end time delay is 1.1 s, the risk quantification accuracy and the communication efficiency are improved, and stable operation of the energy storage system is guaranteed.

The application discloses a difficult tracheal intubation auxiliary evaluation system and method suitable for a visual laryngoscope based on multi-modal data construction, and belongs to the technical field of data processing systems. The auxiliary evaluation system is provided with a data main line; a calculation unit and a result output unit are sequentially connected on the data main line; the calculation unit stores a computer program; the computer program runs a calculation method of a difficult tracheal intubation probability value when executed by a processor; the value calculation method is obtained through formula I: formula I: wherein, x s structured clinical data of the evaluated person; f s structured clinical feature coding function; W is a weight matrix; b is a bias term; I g,v body position image data of the evaluated person; f θ image feature extraction function; α g,v image attention weight of different visual angles in the same body position group when g takes the same value and v takes different values; β g inter-group attention weight of the body position group g in the body position image data of the evaluated person; and lambda is a modal fusion weight parameter; σ Sigmoid function.

The invention discloses a self-adaptive graph optimization method and system based on MEMS inertial navigation assisted PPP-RTK, and the method comprises the steps: collecting multi-system GNSS original observation data, PPP-RTK precise correction data and MEMS inertial navigation data in real time, and completing time synchronization of the data through a PPS mechanism; performing ionosphere, troposphere and orbit error correction on the GNSS observation value by using the precise correction data, and performing downsampling and pre-integration processing on inertial navigation data; gNSS signal quality and inertial navigation noise level are evaluated in real time, a Sigmoid function is used to calculate a comprehensive interference degree and map the comprehensive interference degree into an adaptive observation weight, and a dynamic information matrix is constructed; constructing a factor graph model containing an inertial navigation pre-integration factor, an adaptive GNSS observation factor and a prior factor; and solving optimal state estimation by adopting an incremental graph optimization algorithm based on a sliding window and marginalization technology. According to the method, by dynamically adjusting the observation weight and introducing inertial navigation strong constraint, the robustness of multi-source fusion positioning is remarkably improved, and continuous high-precision positioning in a complex scene is realized.

The application discloses an AI generated image detection method and device and a storage medium, comprising: acquiring a to-be-detected image, inputting the to-be-detected image into an image encoder of a forgery detection model; extracting a deep visual feature vector of the to-be-detected image through the image encoder; performing linear operation on the deep visual feature vector and a preset weight parameter matrix, and superimposing a preset bias parameter in the linear operation process to obtain an intermediate scalar value; taking the intermediate scalar value as the input of a Sigmoid function to calculate a scalar confidence score; and determining whether the to-be-detected image is an AI generated image according to the comparison result of the scalar confidence score and a dynamic judgment threshold. The application extracts deep features through cross-modal alignment training, performs linear mapping and efficient reasoning, significantly improves the generalization ability of unknown generated models, and realizes lightweight deployment while ensuring high precision by combining dynamic threshold adaptive adjustment, thereby meeting the actual application requirements.

The invention discloses a multi-legged robot radiation source autonomous positioning method based on AI dynamic path optimization, and belongs to the technical field of nuclear radiation monitoring and intelligent robots, and the method comprises the steps: collecting space radiation intensity data, eliminating pseudo sources, calculating a radiation gradient field, and constructing a radiation field in real time; an improved social network search algorithm is adopted to solve a target function of an optimal path, the target function dynamically fuses a path length and a radiation gradient integral, and a path weight coefficient alpha and a radiation weight coefficient beta are adjusted in real time through a Sigmoid function according to environmental complexity; the multi-legged robot conducts gait self-adaptive control according to the optimal path, and the movement speed is adjusted according to the real-time radiation intensity. According to the method, dynamic path optimization, multi-mode perception and pseudo source elimination are fused, the problems of path disjunction, data distortion, pseudo source interference, insufficient dynamic adaptability and the like in the prior art are solved, and an autonomous and credible radiation source positioning solution is provided for nuclear leakage emergency.

The application discloses a method for manufacturing a colorimetric sensing unit and array by using a multi-target Bayesian optimization algorithm. Different formulations of colorimetric sensing units are initially prepared on a hole plate; the hole plate is dried, and the dried hole plate is introduced into a gas test to obtain the values of response time, reversibility, responsivity and sensitivity of each formulation; the values of response time, reversibility, responsivity and sensitivity of each formulation are processed by a Sigmoid function to obtain score values, which are substituted into a multi-target Bayesian optimization algorithm to generate a next round of formulation; colorimetric sensing units are prepared on the hole plate according to the formulation; the steps are repeatedly cycled, and the number of cycles reaches a stop condition to obtain a global optimal formulation; and colorimetric sensing units are prepared according to the global optimal formulation. The method solves the problem that multiple indexes of a sensor are simultaneously optimized and some indexes are easily neglected, and can quickly optimize a CO2 colorimetric sensing array which simultaneously has wide range, high sensitivity, fast response and high reversibility.

The invention discloses a gas foil bearing wear high-fidelity modeling method based on a physical neural network, and relates to the technical field of rotor dynamics and fault diagnosis. The method comprises the following steps: firstly, constructing physical models of gas foil bearing wear, wherein the physical models comprise a gas foil bearing structure and lubrication mechanism model and a gas bearing rotor system nonlinear dynamic model; and then building a hybrid modeling framework fusing a physical mechanism and data driving, training a model mismatch item by adopting a neural network, representing an unknown parameter by adopting a Sigmoid function, introducing a displacement compensation factor, and finishing model parameter updating by combining forward calculation and reverse calculation. According to the method, high-fidelity modeling of wear evolution and key parameter collaborative identification can be realized, the model adaptability and the data matching precision under complex working conditions are improved, and technical support is provided for a gas bearing digital twin system and predictive maintenance.

The invention discloses a three-level T-type inverter grid-connected control method based on a sliding-mode observer, and relates to the technical field of grid-connected control. According to the three-level T-type inverter grid-connected control method based on the sliding-mode observer, a preset time sliding-mode control scheme based on the sliding-mode observer is provided, a continuous Sigmoid function is introduced to replace a nonlinear sign function to construct a sliding-mode observation reaching law, signals can be effectively and smoothly controlled, and the control precision of the three-level T-type inverter grid-connected control method based on the sliding-mode observer is improved. The inherent high-frequency buffeting phenomenon of traditional sliding mode control due to a sign function is remarkably suppressed, and meanwhile, the convergence speed and tracking precision of the system are improved. On the basis of observing power grid voltage serving as external disturbance through a sliding-mode observer, feedforward compensation is performed on the external disturbance, and the feedforward compensation is applied to a sliding-mode control reaching law of preset time so as to suppress buffeting and accelerate the convergence process of output current. And the timeliness and the stability of grid-connected control of the three-level T-type inverter based on the sliding-mode observer are improved.

The application discloses a parameter identification method for improving an industrial robot GMS friction model, and comprises the following steps: step 1, introducing a Sigmoid function to design a GMS friction model satisfying a switch point differential; step 2, constructing an improved Stribeck effect model of the GMS friction model; step 3, designing an excitation trajectory for Stribeck effect identification and identifying parameters based on a SOS algorithm; and step 4, designing a GMS dynamic parameter identification method based on SOS state point determination. The application satisfies the switch point differential, avoids model oscillation, reduces resonance of a servo system for robot joint control, more accurately describes the Stribeck effect, improves the precision of a classical GMS model, avoids empirical determination of state points, improves model fitting precision, and improves the identification precision of the GMS friction model as a whole.

The invention discloses a construction site dynamic self-adaptive target detection method, device and equipment and a storage medium, and the method comprises the steps: combining channel attention and space deformable convolution through a dynamic self-adaptive attention module, generating a channel weight through a Sigmoid function, and introducing a learnable offset to dynamically adjust a convolution kernel receptive field, thereby achieving the target detection of a construction site. Obtaining key feature representation for inhibiting construction site environment noise interference; applying a deformable feature pyramid network according to the key feature representation, dynamically adjusting the size and position of a receptive field according to the target scale and shape in a feature fusion stage, and obtaining a scale-adaptive feature fusion representation suitable for the construction site multi-scale target; according to the feature fusion representation, calculating by applying an adaptive overlap loss function to obtain a topological relation weight between the targets, dynamically adjusting a regression gradient of the overlapped targets according to the topological relation weight between the targets, and obtaining a target detection result; noise suppression and target feature enhancement can be realized, and the detection precision and the recognition accuracy are improved.

This invention relates to the field of optical communication technology. To address the technical problems of slow convergence speed and susceptibility to local optima in existing gray wolf algorithms, a coherent free-space optical communication system based on an improved gray wolf algorithm is proposed. By introducing inertial weights and a nonlinear decay model to construct a velocity-guided position update equation, and combining it with an improved DLH strategy based on a dynamic elite reinforcement mechanism, the algorithm effectively balances its global exploration and local exploitation capabilities, solving the problem of traditional algorithms easily getting trapped in local optima when dealing with high-order Zernike aberrations. At the same time, a population size reduction NPSR strategy based on a nonlinear sigmoid function is designed to rapidly reduce the population size in the later stages of iteration through a nonlinear approach, significantly reducing the running time of the SLAO system.

This invention provides a method and system for large-scale model preference optimization based on reward margin constraints. It constructs a third-order margin partitioning criterion and dynamically estimates the quantile threshold of the reward margin using exponential moving averages, adaptively dividing preference pairs into uncertainty, buffer, and confidence regions. Subsequently, a differentiated mapping strategy is applied to different regions. Finally, the differentiated mapping is integrated into a truncated sigmoid function, and a TruncPO loss function is constructed to optimize model parameters for human preference alignment tasks in large language models. This invention can improve the original win rate of different models by 4% in benchmark tests such as AlpacaEval2 and Arena-Hard, while reducing KL divergence fluctuation by more than one-third, effectively balancing the order consistency and probabilistic calibration of preference optimization. It is suitable for deployment in large language model application systems requiring accurate preference alignment, such as intelligent dialogue, automatic problem solving, and intelligent education.