16 results about "Model domain" patented technology

The invention belongs to the related technical field of sketch retrieval, and provides a sketch retrieval three-dimensional model method and system based on multistage joint contrast learning, and the method comprises the steps: constructing a loss function of the consistency of a noise quantized value of a sketch training sample energy distance and a corresponding disturbance vector; accurately quantized noise information is adaptively fused into probability embedding representation, so that the interference problem caused by sketch drawing difference is fundamentally reduced; constraining inter-domain deviation from a sketch training sample to the center of a three-dimensional model domain and from a three-dimensional model training sample to the center of the sketch domain and deviation of category center distribution of the sketch domain and the three-dimensional model domain by adopting a multi-stage joint contrast learning framework; probability embedding of sketch training samples and cross-domain alignment of three-dimensional model multi-view fusion features are achieved, cross-domain distribution differences are eliminated, and retrieval accuracy and robustness are comprehensively improved.

The disclosure provides a three-dimensional model retrieval method and system based on single-image cross-modal fusion, which relates to the technical field of three-dimensional model retrieval, comprising: acquiring a to-be-queried image and a multi-view three-dimensional model set; inputting the to-be-queried image and the multi-view three-dimensional model set rendered into a trained single-image cross-modal fusion network to output a corresponding three-dimensional model retrieved; the single-image cross-modal fusion network introduces a data exchange process, gives image domain data to an additional channel of a three-dimensional model domain with a set probability, gives model domain data to an additional channel of an image domain with a set probability, and inputs the image domain network and the three-dimensional model domain network after domain feature alignment respectively into a cross-modal network, fuses information of different modalities, and uses a contrast learning to solve the problem of mining of difficult negative samples of triple loss.

The invention relates to the technical field of artificial intelligence and software engineering, and discloses a small sample individual fairness testing method based on a domain adaptive flow model. Aiming at the problems of difficulty in obtaining target model training set data, low discrimination sample generation efficiency, insufficient sensitive attribute decoupling and the like in the existing black box fairness test, the invention provides the following technical scheme: firstly, constructing a gating residual domain adaptive flow model, and performing dynamic fusion on a frozen source domain projection matrix and a trainable low-rank adapter to obtain an adaptive flow model; domain adaptation of the flow model is realized under the condition of not depending on a target model training set; secondly, a two-stage anti-fact generation strategy is adopted, a seed sample set is constructed by utilizing a dynamic radius attenuation mechanism, and discriminated samples are efficiently expanded in combination with triple disturbance; finally, gradient orthogonal constraint is applied to the hidden space of the flow model, and decoupling control over the sensitive attributes and the semantic features is achieved. The black box fairness testing method can be used for black box fairness testing of classification models such as tables, texts and images, and has good practical significance and actual effect.

This invention discloses a multimodal large-scale model transfer fine-tuning question answering method based on active learning. The method first groups the unlabeled image dataset using the K-Means clustering algorithm, extracts a predetermined proportion of samples from each cluster for labeling, forming an initial visual question answering dataset for preliminary fine-tuning of the multimodal large-scale model. Then, based on the initial fine-tuned multimodal large-scale model, the uncertainty index of the unlabeled images is calculated, and high-value samples are selected for labeling to expand the dataset. Finally, the model is iteratively fine-tuned and optimized until a predetermined termination condition is met, and the final transferred multimodal large-scale model completes the domain-specific visual question answering task. This invention significantly reduces the data labeling cost in visual question answering tasks, improves data quality and model domain adaptability, and efficiently realizes the transfer application of multimodal large-scale models in professional fields.

This invention discloses a low-cost, accurate flood inundation model construction method based on low-precision hydrodynamics and data-driven principles, comprising the following steps: S1: Constructing a low-fidelity hydrodynamic model; S2: Establishing a mapping module to estimate flood inundation in the model domain based on two relationships; S3: Inputting boundary conditions into the low-fidelity hydrodynamic model to obtain a low-fidelity estimate of the river channel depth, wherein the boundary conditions include inflow, outflow, and initial water level; S4: Processing the low-fidelity estimate of the river channel depth using the mapping module constructed in S2, and outputting the flood inundation range and depth of the entire model domain. The advantages of this invention compared to existing technologies are: providing a low-cost, accurate flood inundation model construction method based on low-precision hydrodynamics and data-driven principles for efficient and accurate simulation of flood inundation range and depth.

The invention discloses a cross-domain target detection method based on high-frequency signals and semantic regularization. The method comprises the following steps: constructing a training set, a test set and a verification set comprising source domain data and target domain data; data preprocessing is carried out; extracting features of high-frequency signals of the original RGB image data in the training set and the verification set, and performing channel type splicing on the two types of data to obtain four-channel image data; constructing a cross-domain target detection model training architecture, and introducing inter-domain alignment branches and regularization branches to assist in detection model training, so that parameter values learned by the detection model have a domain adaptive detection function; designing a loss function, and training a cross-domain target detection model; after training is finished, the weight of the detection model is saved, and model deployment used for actual detection is executed. According to the method, the model domain adaptation difficulty is reduced, the model is guided to pay more attention to a key instance area in the image, interference of irrelevant background noise is effectively suppressed, and therefore the cross-domain detection precision is greatly improved, and the model omission ratio is also obviously reduced.

Methods and computer systems for methods, computer systems, and computer-readable memory media for constructing a system model. A first computer-aided x (CAx) model of a tangible object is received that is described in a first CAx model domain and includes a first model-based definition (MBD) and first CAx product manufacturing information (PMI). The first MBD includes geometric data and topological data. Second CAx PMI is received from a second CAx model domain different from the first CAx model domain. The second CAx PMI is mapped to the topological data of the first MBD using a systems modelling language, and a system model is constructed that includes the first MBD and the mapped second CAx PMI. The system model is stored in a non-transitory computer-readable memory medium.

The invention relates to an interactive zero sample composite fault diagnosis method based on fuzzy semantics and FNN, and the method comprises the following steps: S1, collecting vibration signals of different types of faults of a bearing in a rotating machine under different working conditions, and generating a single fault data set and a composite fault data set; s2, constructing a fault diagnosis framework, wherein the fault diagnosis framework comprises a feature extraction module, a semantic construction module, a semantic embedding module, a reasoning module and an interactive expansion module; the feature extraction module extracts single fault sample features, and the semantic construction module generates composite fault generation semantics; the semantic embedding module adopts an FNN to generate prediction semantics; the reasoning module gives a prediction result; s3, training the fault diagnosis framework; and S4, inputting the composite fault data set into the trained fault diagnosis framework, and outputting a fault diagnosis result. According to the method, the self-adaptive capability of the model domain is considered while accurate recognition of the unseen composite fault is ensured, and the method is suitable for composite fault diagnosis of multiple unknown domains.

The invention provides a large-model domain self-adaptive fine tuning method based on mixed tasks and hierarchical fine tuning, which is characterized in that a method for dynamically adjusting a sampling proportion is adopted, and the sampling proportion of a domain data set and a general data set is dynamically adjusted according to the loss reduction condition in the fine tuning process; freeze, LoRA and other fine adjustment methods are combined to carry out fine adjustment process design, the cosine similarity of a sampled general data set and a constructed field data set is calculated by using Sension Bert, and the difference between the general data set and existing data distribution of a base large model is measured by using the confusion degree; and selecting a proper general data set according to the cosine similarity and the confusion degree. Compared with the prior art, the method has good generalization, large model field self-adaptive fine adjustment is carried out from the aspects of data set construction, fine adjustment process and the like, the field capability of the large model after specific field fine adjustment is improved, the general capability of the large model is kept as much as possible, and the field fine adjustment requirement of the large model in production practice can be met.

The invention discloses a multi-modal large model migration fine-tuning question-answering method based on active learning, and the method comprises the steps: firstly carrying out the grouping of an unlabeled image data set through a K-Means clustering algorithm, extracting a preset proportion of sample labels from each cluster, and forming an initial visual question-answering data set for the preliminary fine tuning of a multi-modal large model; then, based on the primary fine-tuning multi-modal large model, calculating uncertainty indexes of unlabeled images, and screening high-value samples for labeling so as to expand a data set; and finally, the optimization model is finely adjusted through iteration until a preset termination condition is met, and a visual question and answer task of the field is completed through the finally migrated multi-modal large model. According to the method, the data labeling cost in the visual question and answer task is remarkably reduced, the adaptability of the data quality and the model field is improved, and migration application of the multi-modal large model in the professional field is efficiently achieved.

The invention provides a data display method, a program product, electronic equipment and a storage medium, the method is applied to a digital main line, and the method comprises the following steps: sending a data acquisition signal to a domain system, and acquiring a domain data file; the domain system comprises an architecture design domain and a service domain; the business field comprises at least one of a simulation model field, a simulation verification field, a system test data field, an IDS system and a three-dimensional collaborative management system; respectively constructing a model tree corresponding to each field data, and creating an index of a data object in each model tree in the digital main line; associating nodes in the model tree corresponding to the service field with system structure nodes corresponding to the architecture design field to obtain an associated network; mounting a data object corresponding to the domain data to a system structure node based on the data features of the association network and / or the domain data, and generating a converged data model; and displaying the domain data in the domain system. And convergence and display of cross-domain data are realized.

The application provides a distillation method and device of a large language model, and relates to the technical field of model training. The method comprises the following steps: obtaining domain data, general data, a pre-trained teacher model and a student model to be trained; training the student model according to a style loss corresponding to the domain data, a KL divergence loss corresponding to the general data and a supervised fine-tuning loss of the student model; the style loss is calculated by inputting the domain data into the teacher model and the student model respectively; the KL divergence loss is calculated by inputting the general data into the teacher model and the student model respectively; and the supervised fine-tuning loss is calculated by inputting the general data and the domain data into the student model. According to the embodiment, the model obtained by training can balance the professional nature of the model domain knowledge and the general ability, and also has natural (non-written) output.

This invention discloses a method for calculating soil infiltration in coastal saline-alkali land based on temperature tracing. It acquires soil temperature, moisture content, water level, and soil property parameters for different soil layers. For the target depth, a heat transfer model considering soil salinity is established, and the initial thermal parameters and boundary conditions are estimated. The model domain is discretized using a finite difference scheme to determine the total simulation period and divide it into stress periods, each stress period being discretized into multiple time nodes. The temperature and moisture content at the target depth are predicted, and a loss function is established by minimizing the sum of squared residuals between the measured and calculated temperature values. An adaptive optimization algorithm is used to iteratively correct the thermal conductivity and water flux within each stress period. The optimized parameters of the previous stress period are used as initial conditions for subsequent periods to achieve full-cycle recursive calculation. Finally, the water flux at the target depth is obtained through a convergence criterion as a quantitative characterization of the soil infiltration rate.

Methods and computer systems for methods, computer systems, and computer-readable memory media for constructing a system model. A first computer-aided x (CAx) model of a tangible object is received that is described in a first CAx model domain and includes a first model-based definition (MBD) and first CAx product manufacturing information (PMI). The first MBD includes geometric data and topological data. Second CAx PMI is received from a second CAx model domain different from the first CAx model domain. The second CAx PMI is mapped to the topological data of the first MBD using a systems modelling language, and a system model is constructed that includes the first MBD and the mapped second CAx PMI. The system model is stored in a non-transitory computer-readable memory medium.

A method for determining a configuration of a data centre data hall comprises creating model domains having N server cabinets arranged away from edges of the domain, and at least N+1 fan wall units (F

This invention discloses a multimodal basic model domain adaptation method and apparatus for holistic modeling of category association, addressing the problems of insufficient semantic utilization and limited contextual reasoning capabilities in existing methods. The method constructs multi-part textual prompts for the target category, generating multi-part textual features via a text encoder; deep features are extracted from the input image via a visual encoder, and multi-part visual features are generated through a unified attention module; cross-category relationship modeling is performed on the bimodal multi-part features, mining cross-modal associations and generating cross-relationship representations; these are then input into a multilayer perceptron to output category prediction results. This invention reconstructs the model fine-tuning and prediction mechanism, achieving a paradigm shift from one-to-one alignment to one-to-many integration. By combining multi-part prompts and visual encoding strategies, predictions are integrated with category contextual information, improving fine-grained feature recognition and association reasoning capabilities. Furthermore, it requires only minor parameter tuning, resulting in high adaptation efficiency and significantly outperforming existing methods in fine-grained image recognition tasks.