161 results about "Madaline" patented technology

The invention discloses a ship arrival and departure prediction method based on a multi-modal neural network and an adaptive LSTM, and the method comprises the following steps: S1, collecting data, including ship position data, environment data and historical arrival and departure data; s2, data preprocessing: cleaning and standardizing the data; s3, feature extraction optimization is carried out on various types of data; s4, performing multi-modal data fusion, including data splicing fusion, data weighted fusion and data deep fusion; s5, training and constructing a self-adaptive LSTM model, wherein the self-adaptive LSTM model predicts the port entering and leaving behaviors and time of the ship; s6, evaluating and optimizing a self-adaptive LSTM (Long Short Term Memory) model; s7, performing real-time prediction and scheduling; and S8, dynamically predicting and providing scheduling suggestions. Based on the ship position data, the environment data and the historical port entering and leaving data, the multi-mode neural network is used for data fusion, the port entering and leaving prediction model based on the self-adaptive LSTM is constructed, and the prediction accuracy and real-time performance are improved.

The invention discloses a multi-agent non-cooperative game driven high-concurrency task reasoning method and system, and the method comprises the steps: decomposing a heterogeneous task into standardized task units, and carrying out the quantitative modeling of the calculation resource demands of the task; constructing a task auction mechanism of the multi-agent non-cooperative game model, performing task allocation by adopting an anti-strategic auction rule and a fragmentation asynchronous protocol, and constraining resource declaration behaviors of the agents through a credit pledge mechanism; the real-time resource utilization rate of the system is monitored, the task priority is dynamically adjusted in combination with the task preemption historical state, a resource soft preemption strategy is triggered, and a compensation queue is established; a self-adaptive model splitting strategy is adopted to distribute reasoning tasks to end side equipment and edge computing nodes, cooperative execution is achieved through data compression and transmission, and task rescheduling is conducted according to feedback of an execution result; according to the method, the task execution efficiency, the resource utilization rate and the system stability can be remarkably improved in a high-concurrency scene.

The invention discloses a distribution network cable fault early warning method and system based on multi-terminal synchronous traveling wave detection, and relates to the technical field of cable fault detection. The method comprises the following steps: deploying at least three synchronous detection ends at each branch end point of a cable, and synchronously triggering a sampling rate; establishing a master-slave synchronous network, and performing frequency spectrum correction on the acquired data; constructing a real-time correction model based on the data of the distributed temperature sensor; extracting early fault features, and constructing a three-dimensional feature tensor; carrying out preliminary diagnosis on edge nodes and carrying out feature classification by adopting a transfer learning optimization network; and in combination with the cable load rate and the environment temperature and humidity, an early warning threshold self-adaptive model based on fuzzy PID (Proportion Integration Differentiation) is established. Through multi-terminal cooperative perception and intelligent analysis, the scheme converts a fault processing mode from passive first-aid repair to active defense, improves the manual inspection efficiency of the distribution network cable, and finds early fault hidden dangers in advance.

The invention discloses an intelligent auditing method and system based on a multi-modal large model, and relates to the field of artificial intelligence, and the method comprises the following steps: a field adaptive model is constructed by injecting financial and accounting field corpora into an LLaMA-3 architecture, accounting criteria, policies and regulations can be automatically analyzed, executable auditing rules can be generated, the system adopts a multi-channel structure, and the auditing efficiency is improved. According to the method, key entity extraction, logic relation identification and ambiguity resolution are synchronously completed, the accuracy and robustness of rules are ensured, multi-source heterogeneous data are integrated by constructing a finance and accounting knowledge graph, generated SPARQL query is executed to realize automatic auditing testing, block chain anchoring rule hash and a three-dimensional evidence matrix are innovatively introduced, and the auditing efficiency is improved. And the traceable and non-tampering decision traceability of the whole process is realized. According to the method, the auditing efficiency, the coverage and the automation level are remarkably improved, the auditing risk is effectively avoided, and a highly credible solution is provided for intelligent auditing of man-machine cooperation.

The invention belongs to the technical field of new-generation information, and particularly relates to a power plant metal supervision entity relationship extraction method based on dual coding, which comprises the following steps: multi-modal data preprocessing: collecting and cleaning data, and carrying out data labeling; constructing a dual coding joint learning model: designing a network layer architecture, and training the dual coding joint learning model; constructing and querying a dynamic knowledge graph: extracting a model to generate a triple, and performing time sequence evolution analysis, causal reasoning interface and dynamic updating; and incremental knowledge updating and dynamic model optimization: an incremental learning and feedback module forms a bidirectional closed loop between the knowledge graph and the relationship extraction model, and continuous evolution of the system is ensured through dynamic knowledge updating and adaptive model optimization. According to the method, through bidirectional feature modeling of dual paths, collaborative representation optimization between entities and relationships is realized while context information is captured, so that the requirements of complicated data types and diversified semantic associations in an engineering scene are met.

The invention belongs to the technical field of computer artificial intelligence, and relates to a citrus intelligent planting management-oriented large model field quantification and adaptive model deployment method, which comprises the following steps of: firstly, constructing a large model calibration data set and carrying out preprocessing, and inputting a Transform model to execute forward reasoning; then carrying out left multiplication rotation on a weight matrix of a to-be-quantized layer of the model and right multiplication rotation on an activation matrix, executing GPTQ quantization, calculating a dynamic adaptive smoothing factor of each channel of the activation matrix, executing normalization and symmetric quantization according to hidden dimension grouping, and generating a quantized Transform model; then obtaining citrus industry text data to construct a fine tuning instruction set, and performing supervised training on the quantitative model to generate a full-quantitative and full-precision model; and finally, by training a task complexity classifier, selecting a full-quantization or full-precision model as a target deployment model according to a task complexity level, so that efficient quantization and intelligent task adaptation of the model are realized.

The invention discloses an adaptive TimeXer model photovoltaic power prediction method and system based on weather typing, and belongs to the technical field of photovoltaic power generation, and the method comprises the steps: dividing a historical photovoltaic power sequence into time patches, carrying out PatchEmbedded projection on the time patches to generate time stamps, and generating learnable global stamps; the method comprises the following steps: carrying out standardization processing on meteorological data, carrying out clustering to generate a weather classification label, converting the weather classification label into a discrete variable label, and carrying out Variate Embedded projection on the standardized continuous meteorological data to generate a continuous meteorological factor; performing endogenous self-attention calculation on learnable global marks, continuous meteorological factors and discrete variable marks to obtain attention weights among the time marks, aggregating the continuous meteorological factors and the discrete variable marks by using the attention weights among the time marks through cross attention processing, and inputting multi-scale feature marks into a feedforward network to obtain a multi-scale feature mark; and adjusting the frequency domain of the predicted value by using a Fourier mixing layer to obtain a final prediction result.

The invention discloses a juicy peach yield prediction method based on comprehensive data analysis, and particularly relates to the technical field of agricultural intelligent perception. The method comprises the following steps: collecting multi-source data of an orchard, and constructing a data set containing meteorological, physiological, remote sensing and soil information; fruit tree physiological response parameters are extracted, and a bimodal diagram structure fusing the spatial adjacency relation and the physiological state similarity is established in combination with the dynamic climate anomaly index; inputting the graph structure into a graph neural network model, extracting spatial-temporal characteristics, dynamically adjusting an edge weight and a propagation layer number, and constructing an adaptive model; performing region division and weighted summarization according to a model output result, and finally obtaining a predicted value of the total yield of the orchard; the method improves the prediction accuracy under the conditions of complex climate and unstable data, and is suitable for refined orchard management.

The invention belongs to the technical field of network security, and provides a network threat real-time detection and defense method and system based on artificial intelligence. The method comprises the steps of multi-modal data acquisition and preprocessing, dynamic graph feature engineering and knowledge graph collaborative fusion, dual-adaptive model training and optimization, streaming real-time detection and anomaly scoring, DRL-driven hierarchical defense response and automatic disposal, and feedback-driven model adaptive updating and block chain auditing. According to the method, a mixed model of OS-ELM + dual-adaptive ridge regression + federated learning is designed, the training speed is higher than that of CNN, and over-fitting / under-fitting is avoided by dynamically adjusting a regularization coefficient; the federal learning realizes data local training and parameter uploading, and solves the problem of privacy disclosure; knowledge distillation enables the model volume to be reduced, edge equipment deployment is adapted while the accuracy is maintained, and the generalization ability is obviously superior to that of a traditional static model.

The invention provides an interpretable domain adaptive fault diagnosis method based on source domain attribution alignment. The method comprises the following steps: constructing an adversarial domain adaptive network comprising a cross-domain feature extractor, a domain discriminator and a fault classifier; calculating envelope spectrum characteristics of the input monitoring signal based on signal processing knowledge, and calculating integral gradient envelope spectrum characteristics of a source domain sample by means of a decision result in fault diagnosis model training; designing attribution alignment loss, and aligning physical envelope spectrum characteristics of a source domain sample and attribution envelope spectrum characteristics generated by model decision through maximum mean value difference; and synchronously optimizing domain discrimination loss, fault classification loss and source domain attribution alignment loss. The method fully focuses on the consistency of the model decision and the physical characteristics in the source domain training stage, introduces the physical rule constraint for the domain self-adaption process, enables the domain self-adaption model to learn cross-domain data characteristics meeting the physical rule and decision logic at the same time, and can improve the fault diagnosis precision and interpretability of the model.

The invention discloses a dynamic simulation prediction model construction method based on data driving. The dynamic simulation prediction model construction method comprises the following steps: S1, collecting multi-source heterogeneous multi-modal original data and preprocessing to obtain a standardized data set; s2, extracting a long-term dependent feature sequence by adopting a Mamba depth state space model; s3, establishing a parameter diffusion space by using the self-adaptive multi-mode diffusion model, and obtaining diffusion characteristic parameters; s4, performing reverse denoising convergence processing on the diffusion characteristic parameters to generate real-time adaptive model parameters; s5, performing dynamic weighted fusion of the feature sequence based on real-time adaptive model parameters; s6, inputting the fusion sequence into a dynamic simulation prediction model framework for training, and obtaining trained model parameters; and S7, solidifying model parameters to complete model construction. According to the invention, the dynamic prediction precision and generalization performance are improved.

The invention is suitable for the technical field of automatic driving vehicle control, and provides an automatic driving track rapid tracking control method based on online neural network learning, which comprises the following steps: firstly, designing an adaptive model prediction controller, then making a data set and training a fuzzy radial basis function neural network; then, online learning of the fuzzy radial basis function neural network is carried out; and finally, accelerated optimization of the training process of the fuzzy radial basis function neural network is carried out. The automatic driving track rapid tracking controller based on online neural network learning provided by the invention has the remarkable advantages that the neural network model is trained by using control data of MPC (model predictive control), and a nonlinear relation is fitted to approximate the MPC, so that the calculation efficiency is effectively improved, and the rapidity of the controller is improved; the controller can collect real-time data, update a data set and learn new neural network weight parameters online in the process of tracking a reference trajectory, and shows better adaptability and precision compared with a traditional model prediction controller.

The invention discloses a power system distribution robust scheduling optimization method based on optimal configuration, and relates to the field of power system robust scheduling, and the method comprises the steps: carrying out the coupling modeling of a power transmission network extension plan, a DPFC and an energy storage device, and solving a model to generate an optimal configuration scheme; constructing an uncertainty set capable of being dynamically updated, and obtaining a day-ahead scheduling scheme by adopting a dual-stage distribution robust optimization method in combination with the optimal configuration scheme; on the basis of a day-ahead scheduling scheme, a hybrid robust and adaptive model prediction control framework is utilized, and a real-time control parameter of the DPFC and a rapid power adjustment instruction of energy storage are generated through rolling optimization; according to the invention, by combining optimal configuration, dual-stage distribution robust optimization and hybrid robust adaptive model prediction control, collaborative optimization of long-term planning and short-term scheduling is realized, and through real-time model parameter updating and dynamic uncertainty processing, real-time scheduling is realized. The problems that a scheduling scheme is insufficient in robustness, and economical efficiency and real-time adaptability are difficult to consider are solved.

The invention relates to the cross technical field of artificial intelligence and geological disaster monitoring and early warning, in particular to a landslide recognition method and system based on staged feature adaptive transfer learning, and the method comprises the steps: constructing an initial landslide recognition model F1 with an encoder-decoder; a universal visual pre-training weight is migrated, and encoder parameters of the model F1 are finely adjusted by using a remote sensing image data set DS of a source domain, so that a landslide identification optimization model F2 adaptive to the field is obtained; inputting the remote sensing image data set DS of the source domain and the unmanned aerial vehicle image data set DT of the target domain into the model F2, and processing based on a covariance alignment mechanism to obtain a domain adaptive model F3; inputting a small amount of labeled target domain small sample data D 'T into the model F3, and finely adjusting specified parameters in the model F3 through a supervised learning mechanism to obtain a target domain adaptive model F4; and based on the target domain adaptation model F4, executing a landslide identification task on an input to-be-detected unmanned aerial vehicle image.

The invention discloses a federal remote sensing large model training method and system based on low-rank self-adaption. The method comprises the following steps: performing personalized heterogeneous data knowledge learning by using respective remote sensing data, a pre-trained large model and a local low-rank self-adaption model; using the trained local low-rank self-adaptive model parameters to initialize global shared low-rank self-adaptive model parameters, performing collaborative federal training based on the pre-trained large model and the global shared low-rank self-adaptive model, and learning global remote sensing task domain knowledge; after alignment of local heterogeneous data knowledge and global remote sensing task knowledge is realized in a different-rank alignment fusion mode, a federated remote sensing large model composed of parameters of a pre-trained large model and an aligned local low-rank adaptive model is output, so that the parameter quantity and communication overhead uploaded by federated training are greatly reduced, efficient training of the model is ensured, and the reliability of the model is improved. And self-adaptive alignment of local heterogeneous data and global remote sensing tasks is also realized, and the universality and robustness of the model are ensured.

The invention relates to the field of motor fault diagnosis, and particularly discloses a permanent magnet synchronous motor fault diagnosis method and system, which dynamically update motor physical model parameters through an online identification algorithm so as to reflect the characteristic change of a motor in real time. Thirdly, predicting the theoretical voltage of the healthy motor under the working condition by using the self-adaptive model and the real-time operation data, and calculating a residual sequence subjected to working condition normalization between the theoretical predicted voltage and the actual voltage; the residual signal is essentially free from the influence of working condition change, and only the abnormal characteristics caused by the fault are highlighted. And finally, inputting the high-robustness residual error sequence into a convolutional recurrent neural network, deeply fusing space-time fault features, and realizing accurate judgment on a motor fault type and confidence thereof, thereby effectively improving variable working condition adaptability and weak fault detection capability of diagnosis.

The invention relates to the technical field of household equipment intrusion detection, in particular to a mimicry intrusion detection method and system for household equipment, and the method comprises the steps: S0, starting credible verification and initialization; the method comprises the following steps: S1, carrying out multi-modal feature fusion processing; s2, disturbance type joint detection; s3, graded elastic response is carried out; s4, updating the self-adaptive model; s5, identifying a semantic exception instruction; and S6, neural morphology calculation is accelerated. According to the method, damage attack path dependence is detected through dynamic disturbance, multi-modal features are fused through a modal alignment mechanism, low-power-consumption acceleration is achieved through neuromorphic hardware, semantic anomaly instruction recognition and brain-like feedback offline optimization are combined, and the core problem that high-precision detection and high robustness cannot be considered at the same time in the prior art is solved; and particularly, real-time attacks and semantic spoofing attacks of unknown vulnerabilities are effectively resisted.

The invention provides a non-invasive brain-computer interface signal identification method and system combined with deep learning, and relates to the technical field of artificial intelligence and brain-computer interfaces. Firstly, noninvasive neurophysiological signals are collected, and time sequence fragments are obtained through impedance detection, quality scoring and time alignment; denoising the time sequence fragment to generate a time-frequency-channel three-dimensional tensor and manifold domain representation; performing self-supervised pre-training and applying mutual information alignment constraint to construct a double-branch basic model, and outputting joint representation; performing joint optimization on multi-term loss based on the tagged source domain data to obtain a double-branch recognition model; in the deployment period, manifold alignment and optimal transmission calibration are carried out on target manifold domain representation, calibration representation and time-frequency tensor are input into a model, only part of parameters are updated, and prediction entropy is minimized to obtain an adaptive model; and further outputting a final category and confidence coefficient based on dynamic weighted fusion of two paths of judgment results and consistency measurement. According to the method, the cross-session adaptability and the recognition precision are improved, and the interpretability and the real-time security are enhanced.

According to the self-adaption-based automatic control method and system for the numerical control typesetting machine, the dynamic feature tensor of the raw material is obtained, the defect that single data is prone to being interfered is overcome through the method of obtaining the multivariate data set, more comprehensive working condition input is provided for self-adaption control, a real-time machining track is obtained, and the machining efficiency is improved. The method comprises the steps of obtaining a track deviation vector through a real-time processing track, constructing a reinforcement learning model, inputting a dynamic feature tensor and the track deviation vector into the reinforcement learning model, obtaining an optimal control set, constructing a deformation self-adaptive model according to the optimal control set, and obtaining deformation quantity prediction data through the deformation self-adaptive model. According to the intelligent linkage method, deformation quantity prediction data is obtained, a composite control instruction is obtained according to the deformation quantity prediction data, the composite control instruction is input into a control module for real-time regulation and control, control is conducted through cutting data feedback typesetting paths, a reinforcement learning model and self-adaptive lifting are integrated, and typesetting precision is improved.

The invention discloses a knowledge graph-AI fusion fault diagnosis method for operation and maintenance management of a power distribution network, and relates to the technical field of artificial intelligence. The problems of low initial diagnosis accuracy, insufficient diagnosis adaptation capability in different grounding forms, long fault positioning time and the like caused by external interference of zero-sequence current signals are solved, and the conditions that the operation and maintenance efficiency is reduced and the power supply loss is increased are finally avoided. The method comprises the following steps: acquiring multi-source data and constructing a dynamically updated power distribution network knowledge graph; multi-mode fault feature extraction is carried out; using an adaptive AI model to identify a fault type and performing primary judgment; and accurately positioning a fault section and analyzing a spreading path in combination with a knowledge graph. According to the method, the single-phase earth fault preliminary diagnosis accuracy is improved to reduce invalid troubleshooting of operation and maintenance, cross-scene diagnosis errors are reduced to adapt to diversified operation and maintenance scenes, average positioning time consumption is shortened, and the supporting capacity of fault diagnosis on operation and maintenance management of the power distribution network is remarkably improved.

The invention belongs to the technical field of hypersonic flight vehicle control, and particularly relates to a hypersonic flight vehicle protection control method considering non-starting of an air inlet channel. The method specifically comprises the steps of establishing a quasi-one-dimensional nominal mechanism model to describe flow field changes of an isolation section and a combustion chamber; adopting a physical information neural network to construct a reduced-order model as a real-time agent of the quasi-one-dimensional model; designing a dual extended Kalman filtering algorithm to realize joint online estimation of model parameters and blind area states, and introducing shock wave position pseudo measurement to form multi-loop feedback; a dual-channel fusion early warning system of a model channel and an intelligent sensor channel is constructed, and robust early warning is realized through adaptive weight; and designing a continuous adaptive model prediction controller, and dynamically adjusting a control strategy according to risk indexes and parameter uncertainty to realize active protection control.

The invention relates to a self-adaptive model updating method and device, electronic equipment and a storage medium. The self-adaptive model updating method comprises the following steps: acquiring distributed external sample data newly generated by a large language model; clustering processing is carried out on the out-of-distribution sample data to obtain an out-of-distribution sample cluster, and the out-of-distribution sample cluster at least comprises out-of-distribution sample data of the same structure label; and performing search training on the out-of-distribution sample cluster, determining a target adapter, and sending the target adapter to a database of the large language model to update the large language model. Static model updating is converted into a time-driven learning process by setting a self-adaptive updating mode, so that a large language model can be subjected to self-adaptive updating operation according to failure result feedback during reasoning in a training process, the model is easier to adapt, error accumulation is reduced, stable updating is ensured, accuracy is improved, and the training efficiency is improved. And the precision and robustness of the model are enhanced.

The invention provides a real-time three-dimensional reconstruction method based on online evolutionary learning, and relates to the technical field of three-dimensional reconstruction methods. The real-time three-dimensional reconstruction method based on online evolutionary learning specifically comprises the following steps: S1, online evolutionary learning, S101, pseudo-label generation: adopting a pseudo-label generation algorithm to generate reliable labels for unlabeled data; s102, updating a self-adaptive model, and dynamically adjusting model parameters according to real-time feedback; s103, sensing and optimizing the environmental resources; s104, carrying out continuous optimization and evolution; s105, terminal output and real-time three-dimensional reconstruction output are carried out; s2, self-adaptive three-dimensional reconstruction; and S3, dynamic and static cooperation and heterogeneous model fusion. According to the method, self-adaptive reconstruction can be carried out under an unsupervised learning and reinforcement learning framework, so that a reconstruction model can be automatically optimized in a dynamic environment, dependence on labeled data is reduced, and complexity and precision of a reconstruction algorithm are adjusted in real time according to computing power and network conditions so as to ensure efficient operation of a system under different resource conditions.

The invention relates to the technical field of multi-modal data processing, and discloses a multi-modal data fusion method and system for an AI application all-in-one machine, and the method comprises the steps: receiving an original data set from a private cloud, a public cloud and an edge cloud, and carrying out the cross-modal feature fusion of the original data set, and obtaining a feature data set; executing cloud distribution on the received AI processing task and the feature data set to obtain a task execution strategy; performing resource state monitoring and access control on each cloud platform according to the task execution strategy to obtain a resource scheduling scheme; and performing federal learning coordination processing on the distributed AI model based on the resource scheduling scheme to obtain an adaptive AI model, the method realizes cross-cloud AI model collaborative optimization, enhances the generalization ability of the model in a heterogeneous cloud environment, and provides a technical solution for efficient operation of an AI application all-in-one machine in a hybrid cloud environment.

The invention discloses a three-stage cooperative lane changing method for a vehicle queue in a ramp diversion area in an intelligent network connection environment, and the method comprises the steps: firstly, determining a safety gap of a target lane based on a minimum safety distance and a maximum deceleration, and selecting a vehicle which can change the lane to the safety gap from a CAV queue for preferential lane changing; secondly, establishing a longitudinal trajectory planning model based on self-adaptive model predictive control, performing longitudinal distance adjustment on a lane-changed vehicle by using the model, and reversely calculating an initial position required by the vehicle to smoothly drive into a ramp based on a quintic polynomial trajectory planning method; embedding into adaptive model predictive control as a ramp tail end constraint to form a new safety gap of the target lane; and finally, the vehicles which do not change lanes are controlled to sequentially change lanes to the corresponding safety gaps, and lane changing of the CAV queue is achieved. According to the method, the CAV queue can efficiently and orderly drive into the exit ramp, the vehicle speed fluctuation of a shunting area can be reduced, and the overall passing efficiency of a road section is improved.

The invention discloses an edge calculation collaborative reasoning method for adaptive model segmentation, a medium and equipment. The method comprises the following steps: pre-training a DNN model by adopting a data set in a model application field; quantizing the DNN model by adopting a post-training quantization method; analyzing layering calculation time delay and data output quantity based on layering characteristics of the DNN model, and constructing a directed acyclic graph of the model; generating a feasible segmentation strategy set by adopting a network flow graph mode in a graph theory; and solving an optimal segmentation strategy under the dynamic network quality by adopting a deep Q network. According to the method, the storage pressure of the terminal equipment is relieved by adopting a model compression method, the action space of the DQN is solved by combining the DAG of the DNN and utilizing a graph theory method, the problem that the action space is too large due to too many feasible segmentation strategies is solved, and then the time for solving the optimal segmentation strategy is shortened, and the reasoning efficiency is improved.

An adaptive model partitioning method and system applied to distributed training, which method and system belong to the technical field of deep learning, and aim at solving the technical problem of how to implement, in respect of distributed training, distributed model training by means of combining deep reinforcement learning and a Bayesian optimization algorithm. The method comprises the following steps: constructing a Q network on the basis of a deep neural network, and defining state information, actions and a reward function, wherein the state information comprises feature vectors of partitioned models, and training times, resource utilization rates and inter-node communication overheads of when the partitioned models are subjected to distributed training by means of distributed computing nodes, each action is a partitioning strategy used by an agent under the current state information, the reward function R is used for evaluating the effect of the current partitioning strategy, and the Q network uses the state information as input to predict and output a Q value of each action that the agent may take; and performing multiple iterative training on a deep reinforcement adaptive model, so as to obtain a final partitioning strategy and parameters of the Q network.

The invention discloses a fan nonlinear load self-adaptive control method and device and a storage medium. The fan nonlinear load self-adaptive control method and device are used for efficiently achieving dynamic tracking control over fan nonlinear loads. The method comprises the steps of collecting operation state data of a fan system; constructing a dynamic neural network model containing a long short-term memory network-attention mechanism mixed structure, inputting the operation state data into the dynamic neural network model, and outputting a dynamic predicted value of the fan nonlinear load; constructing an adaptive model prediction controller based on the dynamic prediction value, and generating a control sequence; calculating a prediction residual error of the dynamic neural network model, and when the prediction residual error exceeds a standard threshold value, triggering a model updating mechanism; performing parameter updating on the dynamic neural network model by adopting an incremental learning algorithm with gradient constraint to obtain updated parameters; and the safety control quantity is output to a fan execution mechanism, and dynamic tracking control over the nonlinear load of the fan is achieved.

Embodiments of the present disclosure relate to validation of unsupervised adaptive models. According to example embodiments of the present disclosure, unlike methods validating with the seen target data, the present disclosure synthesizes new samples by mixing the target samples and pseudo labels. The accuracy between model predictions of mixed samples and the mixed labels are measured for model selection, and the accuracy score may be called PseudoMix. PseudoMix enjoys the combined inductive bias of previous methods. Experiments demonstrate that PseudoMix can keep state-of-the-art performance across different validation settings.

The invention discloses a wireless Mesh network flow prediction method based on multi-domain feature fusion deep learning, and belongs to the technical field of communication networks. The method comprises the following steps: preprocessing data according to the collected data; constructing a multi-domain graph convolution model, and predicting network traffic; the multi-domain graph convolution model comprises a TimesNet time-frequency domain multi-periodic feature model, a space-time graph convolution network topology feature model and a Transform multi-domain feature adaptive model; and training the constructed convolution model divided into the multi-domain graphs, and outputting a network traffic identification result. According to the method, precise modeling of complex space-time correlation and non-stationarity in a topological domain is realized through the TE-GRU and the CCW-GCN; the three-domain features are subjected to adaptive alignment, so that the adaptability to a complex dynamic Mesh network scene is improved; the depth and accuracy of cross-domain feature fusion are improved, higher flexibility and robustness are shown in complex non-stationary Mesh network flow modeling, and efficient unified expression of the time-frequency domain and the topological domain is achieved.