428 results about "Training phase" patented technology

The invention discloses a three-dimensional scene reconstruction method and device based on large model geometric prior, and a medium, and aims to solve the problems that a conventional 3DGS is liable to have artifacts and detail loss in geometric discontinuity, data redundancy and illumination variation scenes, and predicts a dense depth map and a normal map from a monocular image by using a pre-trained large model. The position and form of the Gaussian kernel are constrained as additional geometric priori; a primitive adjustment strategy based on kernel density estimation is introduced in the training stage, small Gaussian primitives with similar structures and adjacent spaces are combined into a large Gaussian primitive, the rendering quality is kept, redundancy is reduced, and the volume of the model is reduced; an exposure coefficient is adaptively estimated for each input image, an exposure compensation image loss function is constructed, and floating artifacts caused by illumination differences at shooting moments are eliminated. Experiments show that compared with the prior art, the method improves the three-dimensional reconstruction precision and real-time rendering quality of complex illumination and less-texture areas in a public data set and an unmanned aerial vehicle aerial photography scene.

The invention relates to the field of intelligent automatic driving, in particular to a multi-vehicle cooperative controllable confrontation test method based on a diffusion model. Comprising the following steps: S1, a diffusion model training stage: training a diffusion model based on real driving data, and learning vehicle behavior distribution through forward noise addition and reverse denoising to obtain fixed model parameters; s2, scene and multi-vehicle space-time modeling: fusing the road map, the lane topology and the vehicle state information, and constructing a space-time scene representation of multi-vehicle interaction as input of diffusion generation; s3, a diffusion generation mechanism based on adversarial guidance; s4, performing partial diffusion control and diversified confrontation generation; and S5, evaluating the authenticity and controllability of the multi-agent confrontation scene. Compared with the prior art, the method has the advantage that the authenticity, controllability and closed-loop consistency of multi-vehicle cooperative confrontation scene generation are obviously improved.

The invention provides a VAE-ALSTM-based bridge structure state anomaly detection method and system, and the method comprises the steps: carrying out the standardization and sliding window segmentation of vibration monitoring data, extracting potential features through a variational auto-encoder, and carrying out the time sequence prediction and reconstruction in combination with an attention-enhanced long-short-term memory network, thereby achieving the detection of the abnormal state of a bridge structure. Fusing the prediction error and the reconstruction error to generate an abnormal score; a threshold value is automatically set through quartile distance statistics, and self-adaptive judgment under different bridge types is achieved; when the score crosses the boundary, real-time alarm is triggered; the system correspondingly comprises a data preprocessing and sample construction module, a VAE-ALSTM model construction module, a training stage prediction and reconstruction module, an error calculation and anomaly scoring module, a threshold setting module and an anomaly judgment and alarm module. The method and the system do not need manual threshold parameter adjustment, are high in precision and good in real-time performance, and can be widely applied to the fields of bridge health monitoring, operation and maintenance early warning and the like.

The invention discloses a three-dimensional point cloud geometric information compression method based on implicit neural representation, and the method comprises the steps: constructing a trunk structure of an implicit neural network through a plurality of sine representation network layers which are connected in series, and introducing a variable-scale position coding mechanism on this basis, the method enables a network to obtain higher geometric reduction precision while keeping a compression ratio, and comprises the following steps: (1) inputting space coordinates of divided voxels into a position coding module with adjustable scale parameters; (2) feeding a coding result into an implicit neural network constructed by a network layer based on sine representation, and outputting the occupancy probability of the voxel through an activation function; (3) in a training stage, the model continuously optimizes parameters, so that the output probability distribution is highly consistent with a real occupied label; (4) after model training is completed, a method of combining an AdaRound second-order quantization optimization strategy and quantization perception training is introduced, network weight is finely adjusted, and quantization errors are reduced; (5) in a reasoning stage, judging whether the voxel is occupied or not according to a preset threshold value, and when the prediction probability exceeds the threshold value, regarding the voxel as occupied; and (6) all voxels judged to be occupied are aggregated, and reconstruction of the geometric structure of the point cloud is completed.

The invention relates to the field of artificial intelligence deep learning model optimization, and discloses an ultra-low parameter efficient fine tuning method based on sparse frequency domain projection. Comprising the steps of setting an upper projection matrix which is frozen after random initialization in an attention mechanism layer or a feedforward mapping layer of a pre-training model for mapping a low-dimensional reconstruction result to an original weight dimension; constructing a sparse spectrum matrix in a frequency domain, generating a non-zero index set according to a frequency priority and a random sparse mixed strategy, and keeping the whole training process fixed; performing inverse transformation on the sparse spectrum matrix to obtain a time domain reconstruction matrix, and generating a weight increment through linear mapping of an upper projection matrix; only updating the non-zero spectrum coefficient under the limited parameter budget, and keeping the pre-training weight and the upper projection matrix frozen; optionally, spectrum regularization, energy constraint and gradient clipping are applied in the training stage, and quantification and distillation joint optimization is performed in the reasoning stage.

Systems and methods for object tracking are described. One or more aspects of the systems and methods include receiving a video depicting an object; generating object tracking information for the object using a student network, wherein the student network is trained in a second training phase based on a teacher network using an object tracking training set and a knowledge distillation loss that is based on an output of the student network and the teacher network, and wherein the teacher network is trained in a first training phase using an object detection training set that is augmented with object tracking supervision data; and transmitting the object tracking information in response to receiving the video.

The invention discloses an unmanned cluster dynamic task game confrontation method based on MAPPO deep reinforcement learning of Actor-Critic. An MAPPO deep reinforcement learning strategy of Actor centralized training and Critic distributed execution is adopted as a training algorithm of unmanned cluster dynamic task game confrontation. According to the method, centralized training is carried out by using global information in a training stage, and distributed execution is carried out by only using local observation information and historical information in an execution stage, so that the task can be completed through deep cooperation among multiple agents; the problem that the task completion efficiency is low due to the fact that the flight path drifts and jitters and is prone to falling into a local optimal solution during training of a traditional algorithm is solved, multi-aircraft cooperative path planning can be effectively achieved, and the learning efficiency and stability are improved in a complex multi-agent scene.

The invention provides a fan fault diagnosis method for recognizing a vibration map based on a convolutional neural network, and relates to the technical field of neural networks, and the method comprises the steps: obtaining a multi-dimensional vibration signal in the operation process of a fan, and generating a two-dimensional vibration map through time-frequency transformation; and a convolutional neural network is utilized to automatically extract multi-layer time-frequency features and realize fault category discrimination. In the training stage, parameter optimization is carried out based on known fault samples, in the reasoning stage, real-time signals are input into a trained model to obtain fault type probability distribution, the fault type is determined according to the maximum probability, a fault evolution result is generated in combination with the historical operation trend, and therefore automatic, intelligent and rapid diagnosis of fan faults is achieved.

The invention discloses an aircraft cluster multi-task scheduling system based on a cooperative game and a working method thereof. The system comprises a simulation environment module, a hierarchical strategy network module, a centralized value network module and a training and execution module. The simulation environment module is used for constructing a multi-agent air combat confrontation environment and generating states, rewards and interaction data required by training; the hierarchical strategy network comprises a shared space-time representation encoder f theta (.), a high-layer strategy network pi H (aHz) and a low-layer strategy network pi L (aLz; aH); the centralized value network is used for receiving global state information in a training stage, estimating the overall return of our agent cluster, calculating a dominant function, and realizing the optimization of the network by minimizing the value loss; and the training and execution module optimizes a centralized value network parameter and hierarchical strategy network parameters theta H and theta L by using a global state St in a centralized training stage, and outputs an air combat decision ai according to local observation independent decisions of each agent in a distributed execution stage.

There is provided a system and method of runtime defect detection in a semiconductor specimen. The method includes obtaining a plurality of runtime images acquired for a plurality of dies on the specimen, feeding the plurality of runtime images to a plurality of input channels of a neural network (NN) in an input order, wherein the NN is previously trained in a training phase, and processing, by the NN, the plurality of runtime images simultaneously, to obtain a plurality of defect maps, each corresponding to a respective runtime image and indicating probabilities of defect candidate presence thereof. Each given runtime image is processed as a target image using remaining images in the plurality of runtime images as reference images of the target image, and the defect map of the target image remains invariant, irrespective of changes to the input order.

The invention relates to the technical field of artificial intelligence and deep learning, and discloses an artificial intelligence-combined real-time attitude estimation and action recognition deep learning system, which comprises a sensor module, an edge computing platform and a deep learning system. According to the real-time attitude estimation and action recognition deep learning system combined with artificial intelligence, the problem of calculation efficiency is solved through layered pruning and cross-layer reconstruction design of a lightweight deep learning model, multi-task loss is synchronously optimized in a pre-training stage, LASSO regularization is applied to a convolutional layer through structure sparsification to retain a high-response channel, and the real-time attitude estimation and action recognition deep learning system can be used for realizing real-time attitude estimation and action recognition. BN layer parameter re-parameterization is combined to eliminate reasoning overhead, a deformable convolution block is inserted in a cross-layer reconstruction mode to dynamically adjust sampling points to compensate feature expression loss, then the pruned student network approaches the teacher network feature expression ability through knowledge distillation, the calculation complexity is reduced, meanwhile, the model precision is maintained, and the real-time reasoning requirement of edge equipment is met.

The invention discloses an answer generation method based on a thinking chain, a product, equipment and a medium, and relates to the technical field of calculation data processing, and the method comprises the steps: constructing a self-adaptive thinking chain generation mechanism based on reward driving based on the matching degree between the complexity score of question data and the quantized value of the length of the thinking chain, therefore, the strategy model can autonomously learn a strategy that a high-complexity problem generates a long thinking chain and a low-complexity problem generates a short thinking chain in a training stage, so that the technical problems of waste of model computing resources and aggravation of system response delay caused by adoption of an undifferentiated thinking chain generation strategy in related technologies are solved; the beneficial effects of reducing the redundancy calculation overhead, improving the system response efficiency and realizing on-demand allocation of the calculation resources according to the problem difficulty are achieved.

The invention discloses a DRDoS attack detection method oriented to a high-speed network. The method is divided into an offline training stage and an online detection stage. In the off-line training stage, high-speed network traffic and DRDoS attack traffic are mixed, and the traffic processing scale is greatly reduced while space-time correlation is reserved by adopting a window sampling technology based on a time interval. And then, deeply analyzing the difference between the DRDoS attack traffic and the legal traffic of the high-speed network, and constructing a one-way traffic feature set which has high characterization and is suitable for a real asymmetric routing scene based on three traffic modes of the DRDoS attack. In order to efficiently extract a traffic mode in a high-speed network, a DRD-Sketch data structure with fine-grained spatial-temporal feature perception capability is designed to realize multi-dimensional traffic feature extraction. And finally, training a lightweight machine learning model to realize DRDoS attack traffic discrimination. In the online detection stage, traffic is captured from key nodes of a backbone network, features are extracted by adopting a sampling method and a data structure which are the same as those in the offline training stage, and the features are sent into a trained detection model to realize DRDoS attack low-delay alarm in a high-speed network environment.

Systems and methods for automatically extracting one or more salient images from a surgical video stream are described. A plurality of records including annotated images from recorded surgical procedures are used as training data to generate an image extraction machine learning model. Features, extracted from the training data, are used as inputs to the image extraction machine learning model in a training phase, which outputs salient images. After training, features extracted from the surgical video stream are input into the trained image extraction machine learning model to output the one or more salient images from the surgical video stream.

Systems and methods for identifying anomalous activities in a cloud computing environment are provided. According to one embodiment, a customer's infrastructure may be fortified by leveraging deep learning technology (e.g., an encoder-decoder machine-learning (ML) model) to predict events in the cloud environment. During a training phase, the ML model may be trained to make a prediction regarding a next event based on a predetermined or configurable length of a sequence of contextual events. For example, historical events (e.g., cloud application programming interface (API) events logged to a cloud activity trace) observed within the customer's cloud infrastructure over the course of a particular date range may be split into appropriate event / context pairs and fed to the ML model. Subsequently, during a run-time anomaly detection phase, the ML model may be used to predict a next event based on a sequence of immediately preceding events to facilitate identification of anomalous activity.

A method for a first apparatus, the method comprising: obtaining (713) at least one matching factor, wherein the at least one matching factor indicates a degree matching between inputs used during a machine learning training phase and inputs used during a machine learning inference phase; obtaining at least one threshold matching factor from a second apparatus; comparing (715) the at least one threshold matching factor and the at least one matching factor; performing at least a machine learning model switching based on the monitoring decision request from the second apparatus, the monitoring decision request based at least on: the obtained at least one matching factor; and at least one channel characteristic.

In some example embodiments, there may be provided a method that includes receiving, by a machine learning encoder as part of a training phase, channel state information as data samples; generating, by the machine learning encoder, a latent variable comprising a log likelihood ratio value representation for the channel state information, wherein the latent variable provides a lower dimension binary representation when compared to the received channel state information to enable compression of the received channel state information; generating, by the binary sampler, a binary coding value representation of the latent variable, wherein the binary coding value converts the latent variable to a binary form; and generating, by the machine learning decoder, a reconstructed channel state information, wherein the generating is based in part on the binary coding value representation of the latent variable generated by the binary sampler. Related systems, methods, and articles of manufacture are also disclosed.

The invention discloses a crack detection model training method and device based on dynamic multi-strategy active learning, and the method comprises the steps: collecting and carrying out the preprocessing of a crack image, and constructing a data set containing a labeled subset and an unlabeled subset; performing preliminary training on the double-branch deep learning model by using the labeled subset; then, through an iterative active learning framework, comprehensively evaluating the value of an unlabeled sample from three dimensions of uncertainty, difficulty and representativeness, and dynamically adjusting the weight of each dimension according to a training stage to perform sample screening; and a mixed domain attention module fusing space and frequency domain information is introduced to realize more accurate difficulty perception, and the segmentation precision of the crack edge is improved by including boundary optimization loss. According to the method, the problems of high data labeling cost and low efficiency in deep learning crack detection can be solved, a detection model with better performance can be obtained with less labeling quantity, and the training efficiency and the detection accuracy are remarkably improved.

This invention discloses a method and apparatus for single-channel blind source separation of wireless signals based on deep neural networks. Specifically, in the training phase, a signal dataset is created, including a training set and a validation set; a dual-path complex time-domain convolutional network is constructed; a custom network training objective function is created; the obtained signal dataset is used to train and validate the dual-path complex time-domain convolutional network, resulting in a trained dual-path complex time-domain convolutional network; in the application phase, the time-frequency aliased signal to be separated is acquired through a single antenna; the time-frequency aliased signal is segmented and processed to obtain a fixed-length complex-form aliased signal; using the trained dual-path complex time-domain convolutional network, multiple signals are separated from the complex-form aliased signal and sent to a demodulator for demodulation to recover the bit data of the separated signal. This invention can effectively separate simultaneously and in-frequency mixed signals under single-antenna reception conditions, improving the anti-interference capability of wireless communication.

The invention discloses a multi-modal intelligent garbage detection method based on dynamic adaptive meta-learning, and the method comprises the steps: constructing a framework based on a Faster R-CNN two-stage detector fusing uncertainty estimation and multi-modal features, fusing a dynamic adaptive meta-learning mechanism, and supporting and querying a set to achieve robust adaptation through environment interaction; an enhanced garbage proposal module is designed, multi-modal feature fusion is adopted, and deep interaction with query features is supported; and constructing an advanced garbage classification module, and introducing a dynamic soft attention mechanism to realize space-semantic alignment. In the training stage, mixed meta-learning is adopted, an extended FSOD data set is combined with self-supervised pre-training, N-way K-shot subtask training is performed, meta-knowledge is acquired, and dynamic fine tuning is performed by using a labeled junk image in combination with real-time environment data. In the optimization stage, real-time detection of various types of garbage is realized in the test stage through a multi-task joint loss function supervision framework.

The invention requests to protect a wisdom education scene crowd counting method based on meta-learning and a multi-modal large model. The method comprises the following steps: positioning individual positions through a point regression mode by branches from top to bottom, and outputting a first people number predicted value; a large language model is called through a space prior analysis module from bottom to top to generate regionalized semantic description, vision-language feature alignment and fusion are achieved through a regionalized multi-head cross-attention module, and then a second people number predicted value is obtained through density map regression; the double-branch features are further fused to form a third path, and a third people number predicted value is output; and finally, adaptive weighting is carried out on the three-path prediction result through a learnable weight, and a final crowd counting result is obtained. In the model training stage, a meta-learning strategy is adopted, and supervised optimization is carried out through a multi-task loss function in combination with point regression loss, counting loss and optimal transmission loss, so that the model has rapid scene adaptation capability under the condition of few samples.

The invention discloses an electroencephalogram motion decoding method based on cross-modal semantic alignment, which comprises the following steps: fusing electroencephalogram (EEG) coding and video coding, mapping EEG signals and motion video features to a shared semantic space by utilizing comparative learning, and training a classifier which only depends on EEG input on the basis of the mapping, so as to obtain an electroencephalogram motion decoding result. Two innovative loss functions are adopted in the comparison learning stage and the training stage, the accuracy of the corresponding relation between the EEG and video action semantics and the EEG single-mode discrimination capability can be improved at the same time, and accurate discrimination of the corresponding action category of the EEG signal is achieved. According to the invention, high identification precision and robustness can be maintained under the cross-subject condition, and the demand for large-scale labeling and calibration of new subjects is reduced. The modular design facilitates expansion and integration, the system can be widely applied to rehabilitation training, virtual reality interaction, exoskeleton control and other scenes, and an efficient and reliable solution is provided for real-time electroencephalogram motion decoding and cross-modal retrieval.

The invention relates to a soil compaction system for compacting soil (2), comprising: a soil compaction machine (1); a data processing unit (8); and a determination device (10) for determining at least one input variable which is a criterion for the behaviour of at least one component of the soil compaction machine during a working phase; wherein the data processing unit (8) has an AI (artificial intelligence) device (9); the data processing unit (8) has been trained during a training phase; the data processing unit (8) is trained to derive at least one soil compaction value from the at least one input variable with the aid of the AI device (9); and the soil compaction value is an absolute soil compaction value.

The invention discloses a generative evaluation task-based artificial intelligence model performance test method and system, and the method comprises the steps: determining at least one to-be-executed code outputted by an artificial intelligence model in response to at least one generative evaluation task, the artificial intelligence model being an artificial intelligence model in a training stage or completing training, the to-be-executed codes are in one-to-one correspondence with the generative evaluation tasks; loading the to-be-executed code into a target container matched with the to-be-executed code in the sandbox system to run, and obtaining a code execution result of the target container; and based on the code execution result and the expected effect of the corresponding generative evaluation task, evaluating the performance of the artificial intelligence model to obtain a performance test result of the artificial intelligence model. According to the method, decoupling between artificial intelligence model performance verification and artificial intelligence model training and reasoning can be realized, and the use efficiency and the use safety of the artificial intelligence model are improved while the artificial intelligence model performance verification is realized.

The invention relates to the technical field of artificial intelligence and medical health, in particular to an intelligent medical auxiliary system based on multi-model driving and multi-knowledge-base collaboration, and the system comprises an input layer which is used for receiving a query request submitted by a user in the forms of texts, voices and images; the data storage layer comprises a structured knowledge base, and the data storage layer is used for storing disease knowledge data; the processing core layer is used for receiving a query request, selecting a processing flow according to the query request, and generating a response text and interaction feedback in combination with the data stored in the data storage layer; and the output layer is used for outputting a query result which comprises a response text, an original text source of a retrieval basis and interactive feedback of a training link. The system has the capability of integrating texts, voices and images, creatively adopts a multi-knowledge-base RAG architecture oriented to special injury treatment and daily diseases, can provide professional medical auxiliary services in an off-network environment, and meets high-safety scene requirements.

This invention relates to the field of artificial intelligence technology and discloses a training method, apparatus, device, and storage medium for an emotional speech synthesis model, which can be applied to intelligent voice dialogue scenarios in finance, insurance, and medical fields. This invention selects at least one layer of a pre-trained speech synthesis model as the target layer, loads a VB-LoRA fine-tuning module onto the target layer, and then fine-tunes it using emotional speech data. This enables the model to achieve emotional speech synthesis and output. No emotional information is added during the training phase; emotional information is only added during fine-tuning. This allows for fine-tuning by adding emotional information of different emotion categories, giving the model the ability to express different emotion categories, thus enhancing the model's scalability and flexibility. Furthermore, during fine-tuning, only the parameters of the target layer with the VB-LoRA fine-tuning module are adjusted, eliminating the need for full parameter fine-tuning of the entire model, reducing the workload of model fine-tuning and lowering computational costs.

This invention relates to the field of segmentation algorithm technology, and in particular to a lung nodule segmentation algorithm based on TransUNet. The steps include: replacing the Transformer layers in the TransUNet network with MaxViT networks to form an MM-TransUNet network as the segmentation model; training the segmentation model using the MM-TransUNet network; after each training round, saving the segmentation model parameters for that round; comparing the detection results obtained by the segmentation model with the actual results to calculate the loss; adjusting the parameters of the segmentation model based on the loss; reading the optimal parameters saved during the training phase; substituting the optimal parameters into the segmentation model to obtain the optimal lung nodule segmentation model; applying the MM-TransUNet network as the segmentation model can effectively improve the segmentation accuracy of lung nodules; MM-TransUNet combines residual mechanisms, CNNs, and multi-axial self-attention mechanisms as an encoder module, playing a role in more completely extracting lung nodule features.

The invention discloses a robot AI control system fused with physical constraints, and the system comprises the following steps: S1, employing a differentiable constraint loss layer which comprises kinematics constraint loss, dynamics constraint loss, collision-free constraint loss and contact constraint loss, S2, calculating a total loss function as the weighted sum of task loss and each constraint loss, and S3, estimating the gradient of the non-micro constraint by adopting a finite difference method. According to the method, a differential constraint loss layer is adopted in a training stage by adopting a double-level constraint fusion mechanism, a loss item based on physical constraint is introduced, gradient updating is directly guided, a constraint layer based on classifier guidance is matched, and for a controller based on a diffusion model and other iteration generation models, in the sampling process of denoising, the noise is reduced, and the noise is reduced. And a constraint function gradient is utilized to guide the generation direction, so that the device has high safety guarantee, the output quality is improved, the generalization ability is enhanced, and the modularization and flexibility are improved.

This invention discloses a method, self-sensing system, and medium for predicting rotor displacement of a magnetic levitation bearing, belonging to the interdisciplinary field of artificial intelligence and magnetic levitation systems. The method includes: a training phase: collecting waveform data of the winding current and actual rotor displacement of the magnetic levitation bearing under different operating conditions; using the winding current values ​​and operating condition categories at times (t-1), (t-2), ..., (t-n) as input and the rotor displacement value at time t as output, training a rotor displacement prediction model; where n is an integer greater than or equal to 2; and an application phase: acquiring the winding current values ​​and operating condition categories from the previous n times in real time, inputting them into the trained rotor displacement prediction model, and predicting the rotor displacement value at the current time. This invention can predict rotor displacement through real-time winding current information of the magnetic bearing, thereby enabling fault response to displacement sensors in magnetic levitation systems, reducing dependence on displacement sensor hardware, and even replacing displacement sensors, thus reducing costs.