56 results about "Temporal modeling" patented technology

This invention belongs to the field of campus management and discloses a method for monitoring campus personnel behavior trajectories based on multi-source data fusion, including the following steps: acquiring multi-source data within the campus; preprocessing the multi-source data using a data weighted fusion algorithm; performing differentiated scene optimization processing on the collected facial images to output clear facial image frames and effective feature data; constructing real-time behavior trajectories of campus personnel based on a temporal modeling model using clear facial image frames, effective feature data, and preprocessed multi-source data; analyzing the real-time behavior trajectories through an anomaly detection mechanism to determine the trajectory anomaly status and classify the anomaly level; executing multi-channel alarm push according to the anomaly level; visually restoring the real-time behavior trajectories and historical behavior trajectories based on a campus spatial model; and absorbing new data and anomaly judgment results through an incremental learning algorithm to update the parameters of the temporal modeling model, anomaly detection mechanism, and differentiated scene optimization processing related models.

This invention discloses a method and system for predicting turbine disk temperature fields, an electronic device, and a storage medium. The method utilizes an improved Mamba-2 model with bidirectional coupling of an improved Mamba module and a spatiotemporal attention mechanism to predict turbine disk temperature fields. The spatiotemporal attention mechanism can dynamically adjust feature weights according to the temporal evolution of the temperature field, avoiding being limited to static feature weighting. The improved Mamba module performs feature filtering based on the weighted output of the spatiotemporal attention mechanism, which can selectively filter redundant temporal information and compensate for the deficiency of the Mamba-2 model in capturing features sensitive to core parameters. This forms a closed-loop collaborative mechanism of temporal modeling, feature enhancement, and gating optimization, achieving deep integration of the two. This improves both the temporal modeling accuracy of turbine disk temperature fields and the ability to capture core features, thereby improving the prediction accuracy of transient temperature fields of turbine disks.

The invention discloses a tumble risk assessment method and system based on individualized information, and relates to the technical field of medical assessment and artificial intelligence, and the method comprises the following steps: a data acquisition step, a support period detection step, an individualized feature fusion step, a gait period segmentation step, a feature extraction step, and a time sequence coding step. A risk assessment step; firstly, multi-channel plantar pressure time sequence data of a subject and individualized information of the subject are obtained, and then through five-layer progressive processing, the problems of supporting period self-adaptive detection, gait cycle precise segmentation, multi-dimensional feature extraction, time sequence dynamic modeling and self-adaptive model training are solved respectively. And finally, accurate individual fall risk assessment is realized. Thresholds are automatically adapted for subjects with different physiological features, the gait deterioration trend caused by fatigue is captured through periodic-level double-flow time sequence modeling, and self-adaptive personalized feature learning is achieved through cascaded personalized information layer-by-layer fusion.

This invention discloses a time series anomaly detection method and system based on multi-scale spatiotemporal modeling, comprising: extracting multivariate time series from business data; generating multiple sub-time series at different scales using one-dimensional convolution; independently encoding each scale sub-time series using a scale-independent spatiotemporal encoder to obtain spatiotemporal enhanced features of the multi-scale time series; employing a cross-scale hybrid expert mechanism to achieve information exchange between scales, obtaining a multi-scale sequence representation after scale interaction; integrating the interaction representations of each scale and performing decoding and reconstruction; generating anomaly scores by calculating the reconstruction error between the original input and the reconstructed sequence, and obtaining anomaly detection results. This method decomposes complex time series into multiple scales, with each scale collaboratively modeling spatiotemporal dependencies, which helps to more accurately discover and verify anomalous signals, improving the accuracy and reliability of detection.

This application discloses a model training method based on visual reconstruction supervision, comprising: constructing a video anomaly detection model to be trained, the model including a global-local temporal modeling module, a feature reconstruction module, and one or more visual classification branch modules; obtaining an anomaly video training set, and performing weakly supervised training on the video anomaly detection model to be trained using the anomaly video training set to obtain a trained target detection model; wherein, the global-local temporal modeling module is used to obtain target visual features corresponding to each target input video in the anomaly video training set, and the feature reconstruction module and the visual classification branch module are used to perform visual reconstruction supervision based on each target visual feature during the weakly supervised training process. Implementing this application embodiment can fully utilize the original visual signals of the video training set during the model training process, which is beneficial to improving the weakly supervised model's ability to understand fine-grained visual features and its temporal modeling ability.

This invention relates to an asynchronous time series prediction method based on cross-temporal attention and adaptive fusion, belonging to the field of time series prediction technology. Addressing the problem of inconsistent sampling frequencies in multivariate asynchronous time series data, this method proposes a data processing architecture combining a cross-temporal attention mechanism and adaptive fusion, including the following steps: normalizing the original asynchronous time series data and generating data segments using a sliding window; calculating the feature change amplitude of adjacent time steps using a cross-temporal attention mechanism, generating attention weights and weighting the features to highlight key time steps; fusing multivariate features into a unified-dimensional feature matrix using an adaptive fusion matrix; inputting the feature matrix into a lightweight liquid neural network model for training and prediction, which uses the Euler discretization method to update the hidden states, balancing continuous-time modeling and computational efficiency; and finally outputting the inverse-normalized prediction result. This invention has been validated on multiple real-world datasets, demonstrating that it significantly improves computational efficiency while maintaining high prediction accuracy, making it suitable for real-time prediction scenarios.

This invention relates to a method for predicting urban surface temperature by combining LSTM and geographic information, belonging to the field of spatiotemporal modeling technology for urban surface temperature. It includes the following steps: S1: Acquire temporal LST and geospatial feature data to construct a spatiotemporal sequence input; S2: Based on geospatial features, generate the initial hidden state and cell state of a Long Short-Term Memory (LSTM) network through linear mapping; S3: Introduce spatial weights based on spatial distance to spatially weight and adjust the hidden state of the LSTM, modeling spatial heterogeneity; S4: Employ a spatial cross-attention mechanism to model the correlation between geographic location information and the hidden state, fuse spatial information, and output the prediction result. By sequentially introducing the above mechanisms, this invention achieves collaborative modeling of the temporal dependence and spatial non-stationarity of surface temperature without interfering with the LSTM temporal modeling process, significantly improving prediction accuracy and stability, especially in complex urban environments and cross-regional applications.

The present application relates to the technical field of training rehabilitation evaluation, and particularly relates to an intelligent rehabilitation training evaluation method based on a multi-scale spatio-temporal graph convolution network, which comprises: inputting a human posture sequence to be evaluated into a trained rehabilitation evaluation model to output a rehabilitation evaluation prediction result; the processing steps of the rehabilitation evaluation model comprise: extracting joint modal feature data and skeletal modal feature data of each frame of data in the human posture sequence; capturing global dependency and short-term time sequence association in the joint modal feature sequence and the skeletal modal feature sequence respectively through a plurality of multi-scale spatio-temporal modeling modules connected in cascade to obtain joint modal features and skeletal modal features; inputting the joint modal features and the skeletal modal features after fusion into two fully connected layers for processing to obtain a first prediction result and a second prediction result and to sum the first prediction result and the second prediction result. The present application can improve the accuracy and comprehensiveness of rehabilitation training evaluation and significantly reduce training and deployment costs.

This invention provides a method for optimizing minimum EEOI speed of ships based on a PI-BT network, considering ocean current uncertainties, belonging to the field of ship energy efficiency optimization and intelligent navigation technology. Based on measured ocean current data from shipping routes, this invention mines the characteristics and probability distribution of ocean current uncertainties through statistical testing and wavelet decomposition techniques; derives the mapping relationship between EEOI and main engine speed and ocean current velocity, establishing a minimum EEOI speed optimization model; identifies core sensitive parameters through sensitivity analysis; constructs a physically guided Bayesian Transformer (PI-BT) network, designing a dual-channel input embedding layer, a Bayesian Transformer encoder, an EEOI physical information constraint layer, and a multi-objective optimization output module; and constructs a multi-component total loss function to complete network training, achieving robust optimization of minimum EEOI speed of ships under ocean current uncertainties. This invention integrates temporal modeling, uncertainty quantification, and physical constraint capabilities, significantly improving the accuracy, robustness, and computational efficiency of the speed optimization scheme.

This invention relates to the field of computer vision technology, providing a method for training a behavior recognition model, a behavior recognition method, and an apparatus. The method and apparatus utilize a frozen pre-trained image-text model (a second video feature extractor and a text encoder) as a general knowledge base to guide a finely tuned student model (a first video feature extractor) with temporal modeling capabilities. A multi-head residual projection network is used for feature-level knowledge transfer and representation alignment. The behavior recognition model trained by this method can accurately understand dynamic behaviors in videos and possesses strong zero-shot reasoning ability (i.e., generalization ability), effectively recognizing behavior categories not seen during the training phase. Furthermore, the entire model framework is clear, training is stable, and the final model exhibits excellent recognition performance for both known and unknown behavior categories under open-vocabulary settings.

This invention discloses a method and system for evaluating the quality of immersive videos based on multimodal perception, relating to the field of video evaluation. The method includes: acquiring a multi-viewpoint texture and depth format video and extracting keyframes; extracting texture and depth features using ConvNeXt, processing texture features through a frequency-space texture enhancement module, and combining depth features with a cross-modal collaborative representation module to obtain structural texture coupling features; extracting semantic distortion features using a semantic distortion perception module; concatenating the coupling features and semantic distortion features and inputting them into a temporal modeling module to capture temporal features, then generating global perception features through a viewpoint fusion module; and finally outputting the final video quality score through a quality regression module. This invention achieves a more accurate and robust evaluation of immersive video quality by fusing multi-viewpoint texture and depth information and sequentially performing frequency-space texture enhancement, cross-modal collaborative representation, semantic distortion perception, temporal modeling, and viewpoint fusion.

This invention belongs to the field of computer vision technology and provides a video anomaly detection method and system based on semantic modulation and knowledge distillation. The method involves acquiring a video stream to be detected, extracting image frames from the video stream, generating text descriptions corresponding to the image content for each frame, extracting visual features and textual features from each frame, using the visual features as input, and employing a pre-trained student network to perform temporal modeling of the visual features, generating implicit conditional features, generating modulation parameters and feature modulation amounts, enhancing and fusing the visual features, and then passing the fused features through a classifier to obtain frame-level anomaly scores. During pre-training, the student network undergoes multi-objective distillation training with the teacher network, thus improving training stability and cross-scene generalization ability.

This invention relates to the field of dynamic response calculation for engineering structures, specifically disclosing a method for constructing a structural explosion response calculation model based on graph neural networks. The method includes: encoding and decoding the node and edge features of the graph structure; constructing a spatiotemporal graph processor consisting of multiple stacked spatiotemporal graph tiles, with spatial modeling of a physics-enhanced graph attention network and temporal modeling of gated recurrent units coupled within each tile; integrating physical information into the model in the form of soft or hard constraints; establishing a physics-driven message passing mechanism, including designing a physics-enhanced message function, a wave velocity-aware attention mechanism, and a physics-driven message aggregation and node update method; training the model using a multi-step Rollout training strategy, and validating the model through data accuracy evaluation and physical consistency checks. This invention integrates data-driven and physical mechanisms, solving the problems of time-consuming traditional numerical simulations, weak generalization of purely data-driven models, and insufficient physical consistency.

A system for predictive monitoring of patient deterioration by temporal modeling of electronic health records, the system comprising: a housing defining an enclosed volume; a plurality of input interface units arranged on the housing and configured to receive electronic health data including physiological signals, laboratory measurements, and clinical observations from external medical devices; an analog-to-digital converter unit electrically connected to the input interface units and configured to convert received signals into digital data streams; a signal conditioning unit operationally connected to the analog-to-digital converter unit and comprising filtering, normalization, and scaling circuits configured to standardize heterogeneous data into a uniform data representation;a temporal buffer unit consisting of a plurality of storage registers arranged in a sequential configuration, wherein the temporal buffer unit is configured to store time-indexed data samples over a predefined observation interval to generate a temporally ordered data structure; a synchronization unit configured to generate clock signals and align incoming data streams into synchronized timeframes before storage in the temporal buffer unit; a matrix transformation processor electrically coupled to the temporal buffer unit, wherein the matrix transformation processor comprises parallel arithmetic circuits, including multipliers, adders, and accumulators, arranged to perform weighted transformations on the temporally ordered data structure to generate temporal feature representations;a storage unit with non-volatile memory elements configured to store weighting parameters corresponding to the temporal characteristics of physiological data; a sequence evaluation unit operationally connected to the matrix transformation processor, comprising comparator circuits, gradient calculation circuits, and threshold evaluation registers configured to determine temporal deviations, rate-of-change characteristics, and cumulative variations indicative of a deterioration in the patient's condition; a decision output unit configured to generate a deterioration prediction signal based on the outputs of the sequence evaluation unit, the decision output unit further comprising a display unit, an audible alarm unit, and a communication unit for transmitting alarm signals;and an energy management unit configured to supply regulated electrical power to all units within the enclosure, wherein the system is configured to continuously process temporally ordered electronic health data to generate predictive indicators of deterioration in patient condition in real time.

This invention discloses a method for multimodal dynamic temporal error detection and error text output during the inhalation process of a nasal aerosol delivery device, wherein the delivery device is a spacer. The method utilizes consumer electronic devices to collect video and audio data of the patient's inhalation process. After preprocessing and time alignment, audio and video features are extracted and multimodal fusion is performed. Furthermore, a dynamic position code is constructed based on the step duration to adapt to the different operating rhythms of different patients. A Transformer with sequence constraints is used to perform temporal modeling on the fused dynamic code sequence, outputting anomaly scores to complete the inhalation error determination. When an error is detected, an error text description is output based on an error type prototype library. This method requires no additional dedicated sensors, can cover both home and medical institution scenarios, and can jointly model the sequence of actions, temporal rhythm, and the correlation between voice prompts and action execution, improving the intelligence, convenience, and interpretability of inhalation error monitoring, and providing a basis for subsequent intervention.

This invention proposes a customer attitude recognition and operational intervention method based on AI semantic analysis, belonging to the field of data processing technology. The method includes: acquiring real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generating interaction feature vectors and business context data sets respectively; performing a feature-level affine transformation on the current interaction feature vector based on the business context vector of the previous moment to generate a modulated feature vector; using a gated recurrent unit to perform temporal modeling on the modulated feature vector to output the customer attitude vector of the current moment; driving an attention mechanism based on the difference between the attitude vectors of the current moment and the previous moment to dynamically calculate the attention weight of each data item in the business context data set, and generating an updated context vector; feeding the updated context vector back to the feature modulation step, and simultaneously comparing the current attitude vector with a preset threshold vector to trigger corresponding operational intervention instructions.

The purpose of this disclosure is to provide a computer system based on a quantitative pain model, comprising: a preprocessing module, a local temporal feature extraction module, a cross-domain global frequency domain feature extraction module, a cross-domain interactive fusion module, a feature fusion module, and a classification output module. Through a dual-parallel processing architecture of local temporal and cross-domain global pathways, multi-scale convolution and sliding window temporal modeling are used to accurately capture transient and evolutionary temporal features related to pain. Then, through wavelet transform and attention mechanisms, frequency domain oscillation energy closely related to the physiological mechanisms of pain is explicitly extracted. Finally, a cross-domain interactive encoder is used for deep fusion, achieving complementarity and enhancement of temporal and frequency domain features. This disclosure effectively controls the computational scale while maintaining high-precision classification, making it easier to deploy lightweight and perform real-time inference on embedded devices or mobile medical terminals, demonstrating significant practical value.

This invention provides a radar signal modulation recognition method based on temporal modeling and cross-variable fusion, comprising: preprocessing the radar IQ signal to be identified into blocks to obtain block tensors; constructing a hybrid neural network model, which includes a block transform temporal encoder, a modern temporal convolutional network cross-variable fusion module, and a classification head. The block transform temporal encoder is used to capture long-range temporal dependencies, and the modern temporal convolutional network cross-variable fusion module is used to perform cross-variable fusion on the extracted temporal feature tensors; inputting the block tensors into the trained hybrid neural network model for processing using the block transform temporal encoder, the modern temporal convolutional network cross-variable fusion module, and the classification head, and outputting radar signal modulation recognition results. This improves the model's adaptability and generalization ability; compensates for the shortcomings of local segmentation methods in global modeling capabilities; and achieves joint feature extraction of multi-channel radar signals.

The present application relates to the field of financial risk control and anti-fraud technology, and provides a financial fraud detection method and device based on multi-scale space-time feature learning, comprising the following steps: S1, acquiring financial transaction data and performing basic preprocessing; S2, modeling structural features of input features through an encoder; S3, introducing multi-scale time modeling in a latent space; S4, constructing a training-stable latent representation learning model; S5, generating a risk prediction result through a decoder and a classifier; and S6, risk assessment output. The present application has enhanced structure relationship description capability, more sufficient multi-scale time feature expression, improved training stability and generalization capability, and strong engineering applicability.

The application discloses a physical guidance-based spatio-temporal hypergraph condition denoising diffusion probability model construction method, and relates to the field of ecological environment modeling and artificial intelligence prediction. The application innovatively combines the physical heat diffusion process and the spatio-temporal modeling of multi-modal environmental factors by introducing a temperature driving module based on heat conduction and a spatio-temporal perception fusion module. In addition, a lightweight convolution block attention module and a hierarchical hypergraph attention mechanism are embedded in the diffusion denoising network to realize high-order correlation modeling and dynamic fusion of multi-scale spatial features. The method can effectively capture the spatio-temporal evolution characteristics and ecological dependence of species distribution while maintaining reasonable computational complexity, thereby significantly improving the accuracy of distribution prediction. The experimental results on different types of data sets prove the superiority and robustness of the proposed PSTH-CDPM.

This application relates to the field of computer vision technology and discloses a method and system for detecting pain expressions. The method includes: acquiring a facial video stream, segmenting muscle regions and extracting texture features after facial key point localization and inter-frame alignment; calculating motion energy based on these features, extracting enhanced micro-expression temporal segments and multi-scale spatiotemporal features, constructing a muscle dynamics model and completing state estimation; generating an adaptive candidate spatiotemporal window set through spatiotemporal attention weighted fusion; performing temporal modeling for dynamic feature fusion, and selecting the optimal window by fusing micro-expression features; and outputting the final pain level after double consistency verification and loop closure optimization. This application overcomes the limitations of static images, accurately captures facial dynamics and temporal changes, strengthens feature correlation, and significantly improves detection accuracy.

This invention discloses a collaborative sensing asynchronous feature alignment method and system, relating to the field of collaborative sensing technology. The method includes: acquiring a sequence of historical feature maps of collaborative vehicles and calculating the time delay information of each frame of the feature map relative to the current moment; mapping the time delay information to a time embedding vector and performing frame-by-frame correction of the sequence through time-aware feature modulation; constructing a continuous-time state-space model, dynamically adjusting the state decay coefficient based on the physical time difference to achieve temporal feature aggregation under non-uniform delay, and obtaining the hidden state features at the current moment; using a hybrid spatial gating module to extract motion features and static background features in parallel, and generating spatially corrected features through gating fusion; predicting the offset field based on a multi-scale strategy for feature resampling, and outputting the final aligned features through an artifact repair module. Through continuous-time modeling, motion-static separation correction, and multi-scale alignment, the accuracy and robustness of asynchronous collaborative sensing are effectively improved.

This invention relates to the field of engineering structural safety monitoring and spatiotemporal data modeling technology, specifically to a spatiotemporal modeling method for infrastructure based on adaptive temporal and spatial relationships. First, time-series data from each monitoring point are independently modeled, extracting temporal state feature vectors. Then, a spatial relationship graph is dynamically constructed and updated based on the similarity between these vectors. Next, information propagation and aggregation are performed on the dynamic graph to generate spatially enhanced state feature vectors that integrate spatiotemporal context. This invention reduces complexity through decoupled modeling and improves sensitivity and response speed to anomalies and phased risk changes through data-driven dynamic spatial relationships, thereby effectively improving the accuracy and timeliness of infrastructure status assessment and risk warning.

This invention presents a lightweight physical layer authentication method based on BiLSTM-Linformer. It addresses the decline in model adaptability caused by the complex time-varying nature of CSI data in dynamic environments by proposing a data augmentation method based on temporal interpolation and contrastive learning, a dimensionality reduction method combining PCA and shallow autoencoders, and a lightweight authentication scheme centered on BiLSTM and multi-head self-attention mechanisms. The scheme first uses temporal interpolation to augment CSI data, enhancing temporal continuity, and combines contrastive learning to optimize the model's discriminative ability, improving adaptability to mobile environments. Subsequently, PCA and shallow autoencoders are used for feature dimensionality reduction, reducing computational burden while preserving key CSI patterns and improving authentication efficiency. Finally, BiLSTM is introduced for temporal modeling, and a multi-head self-attention mechanism is combined to optimize temporal feature weight allocation, enabling the model to focus on key changes in CSI data, further improving authentication accuracy and robustness. Experimental results show that the proposed authentication scheme achieves an authentication stability score of 0.95 in dynamic environments, an improvement of approximately 27% compared to existing methods.

This invention discloses an intelligent detection system for residual coal unloading in skips based on image recognition, comprising the following modules: a data acquisition module, using an intrinsically safe mining camera positioned at the skip unloading location to acquire unloading video sequences; a calibration and segmentation module, used to perform region calibration and frame-by-frame image segmentation to form a residual feature sequence; a temporal modeling module, including a neural oscillator network, used for rhythm modeling and phase discrimination, outputting temporal discrimination results; a fusion and judgment module, used to combine spatial parameters and temporal results to generate a fused residual score; an alarm recording module, used to trigger an alarm and generate snapshot and video recordings when residual anomalies occur; and a linkage control module, used to transmit the judgment status and output linkage signals to the hoisting system safety loop. This invention achieves accurate detection and safe control of residual coal unloading in skips through camera recognition and neural oscillator network modeling.

This disclosure provides a method and apparatus for training and locating a crime prediction model. The training method is based on multi-source raw data, mapping this data to corresponding grid cells within a target area to construct multi-dimensional features that integrate historical crime data, dynamic environment data, and static spatial data. This multi-dimensional data is then input into the crime prediction model for training, resulting in a well-trained model. This disclosure not only uses historical crime data but also integrates dynamic environment data and static spatial data, mapping the data to corresponding grid cells and time slices from both temporal and spatial dimensions to achieve spatiotemporal alignment. Furthermore, it trains both the spatial and temporal modeling modules end-to-end, enabling the model to automatically learn the optimal representation of spatiotemporal features. In other words, it effectively models spatiotemporal interactions, comprehensively considering the spatial patterns of crime changes in addition to the temporal dimension, thereby improving the model's prediction accuracy.

This invention proposes a method for detecting EEG signals of depression based on a multi-scale SELSTMNet model. The method includes the following steps: 1. Preprocessing the acquired multi-channel EEG signals by using bandpass filtering and notch filtering to remove noise, independent component analysis to remove artifacts, and segmenting the preprocessed signals according to fixed time windows; 2. Constructing a multi-scale SELSTMNet model, combining a multi-scale feature extraction module with a bidirectional long short-term memory network to model the temporal features of the EEG signals; 3. Dividing the preprocessed EEG data into training, validation, and test sets, and training the model to optimize detection performance; 4. Inputting test samples into the trained model and outputting the corresponding depression classification results. Through multi-scale feature extraction and adaptive temporal modeling, this invention can effectively improve the classification accuracy of depression EEG signals and has strong robustness and practicality.

This invention provides a method and system for multi-frame infrared sea surface target detection based on deep temporal modeling. Belonging to the field of image processing and target detection, it is suitable for continuous detection and tracking of weak infrared targets under complex sea conditions. The method relies on the collaborative processing logic of single-frame detection preprocessing, temporal state prediction, cross-frame association matching, and target trajectory decision and management modules. It sequentially executes single-frame detection and redundancy filtering, temporal state prediction, prediction-guided cross-frame association matching, and target trajectory decision and management steps. Through deep temporal modeling, it mines the temporal features of target movement and existence, guides cross-frame association with predicted position, and constructs a unified judgment mechanism for joint confidence verification. This achieves bidirectional optimization of missing target completion and false detection suppression, while simultaneously implementing full lifecycle management of target trajectories. This invention effectively solves the problem of both missed detections and false detections in existing algorithms, improving the continuity and accuracy of weak infrared sea surface target detection under complex sea conditions.

This invention discloses an AI-based early warning method for estuarine wetland degradation risk, belonging to the field of ecological remote sensing monitoring technology. Addressing the shortcomings of existing wetland degradation prediction methods, such as reliance on multi-source data, lack of bidirectional temporal modeling and key-step focusing capabilities, and absence of critical warning mechanisms, this invention uses only MODIS NDVI time series data as the sole data source. It simultaneously captures the forward and backward temporal dependencies of the NDVI sequence through a bidirectional long short-term memory network (BiLSTM) and introduces an attention mechanism to adaptively focus on the key time steps that contribute most to the degradation trend. After model training, multi-step predictions are made for wetland vegetation cover dynamics over the next 6–24 months. Based on this, the degradation rate and acceleration of the predicted sequence are calculated and compared with historical thresholds. When the predicted NDVI consistently falls below the critical degradation threshold and the degradation rate accelerates significantly, a graded early warning signal is automatically issued, particularly identifying critical degradation states about to cross the irreversible inflection point. Using Chongming East Beach in the Yangtze River Estuary as an example, this method achieves a prediction error (MAE) as low as 0.028 when using only the NDVI single indicator and can issue effective critical warnings 6–12 months in advance. This invention has the advantages of simple data acquisition, high prediction accuracy, interpretable attention weights, and strong transferability, providing a new technical means for the protection and management of estuarine wetlands.

This disclosure relates to a method, apparatus, device, and storage medium for audio-visual fatigue prediction based on temporal modeling. The method includes: extracting facial temporal maps and audio temporal features from acquired audio-visual data to be tested according to a preset frequency; determining a target facial feature extraction module, a target audio feature extraction module, and a target fatigue prediction module corresponding to the facial temporal maps; inputting the facial temporal maps into a pre-trained target facial feature extraction module, outputting facial feature change information in the temporal domain; inputting the audio temporal features into a pre-trained target audio feature extraction module, outputting audio feature change information in the temporal domain; inputting the facial feature change information and the audio feature change information in the temporal domain into a pre-trained target fatigue prediction module, outputting the fatigue level corresponding to the audio-visual data to be tested, effectively reflecting the fatigue level of the target face, capturing individual difference information, shortening inference time, and improving robustness.