22 results about "Deep processing" patented technology

The invention relates to the technical field of visual questions and answers, in particular to a scene graph question and answer inference model construction method based on reinforcement learning, which comprises the following steps: calling an existing large model interface, performing deep processing on an initial scene data set, and constructing a first batch of training sets; carrying out first-stage reinforcement learning training on a pre-constructed multi-modal large model by utilizing the first batch of training sets; obtaining an LLaVA-CoT data set, screening the LLaVA-CoT data set, and inputting the screened LLaVA-CoT data set into the multi-modal large model trained at the first stage to obtain a reasoning result; calling an existing large model interface to carry out accuracy evaluation and correction on the reasoning result of the multi-modal large model to obtain a second batch of training sets; and performing second-stage reinforcement learning training on the multi-modal large model by using the second batch of training sets to obtain a final scene graph question and answer inference model. According to the method, a staged and differentiated training strategy is adopted, the characteristic advantages of two batches of data are brought into full play, staged training is carried out on the large model, and the model performance is gradually improved.

The invention provides an emotion metaphor recognition method and device based on knowledge extraction and co-evolution reasoning, and relates to the technical field of natural language processing.The method comprises the steps that a high-quality emotion metaphor corpus is constructed, and advanced feature vectorization technology is combined to conduct deep processing on text data; key features are accurately extracted by using a fuzzy multi-granularity knowledge extraction technology, and an optimized feature subset is formed, so that redundant information is effectively removed, and the accuracy of feature selection is improved; furthermore, dynamic recognition and labeling of the emotion metaphor are achieved through a matching network and a co-evolution reasoning mechanism, and metaphor labels and emotion categories can be output at the same time. The method aims at solving the problems that in the prior art, emotion metaphor recognition precision is insufficient, the reasoning ability is limited, and noise and redundant information in large-scale text data are difficult to process.

The invention belongs to the technical field of nuclear power station file management, and particularly relates to an intelligent nuclear power station file classification and retrieval system based on artificial intelligence. The file acquisition module is responsible for uniformly acquiring and accessing multi-modal files in the nuclear power station; the preprocessing and natural language processing module is used for carrying out deep processing and semantic analysis on the collected multi-modal document; the dynamic classification module is used for carrying out multi-dimensional automatic classification on the multi-modal document based on a deep learning model and a nuclear power industry knowledge graph; the dynamic knowledge updating module is used for continuously monitoring new documents and user feedback and keeping adaptability to industry changes; the intelligent retrieval module is used for providing a query function for a user through semantic vectorization and a large-scale rapid retrieval technology; and the intelligent recommendation module is used for providing targeted document recommendation for the user through association analysis and collaborative filtering. According to the method, the problems of low classification efficiency, insufficient retrieval accuracy, weak dynamic adaptive capacity and limited multi-modal data processing capacity are solved.

The invention relates to the technical field of speech sentiment analysis, and discloses an AI speech sentiment analysis method and platform of a deep learning architecture, and the platform comprises a signal collection unit, a noise reduction processing unit, a speech separation unit, a feature extraction unit, and a sentiment classification unit. In the signal processing link, analog-to-digital conversion is carried out according to the Nyquist sampling theorem, and short-time Fourier transform is utilized to respectively process noisy voice and noise, so that background interference is effectively eliminated; in the speaker separation stage, voiceprint features are extracted by means of an LPCC method, and the Euclidean distance clustering technology is combined, so that the voice of a known speaker can be accurately distinguished, and a new sounding individual can be recognized; the emotion feature extraction layer is used for mining emotion information in the speech from multiple dimensions by calculating the speech speed and analyzing intonation changes; in a sentiment classification link, the multi-layer perceptron model carries out deep processing and probability calculation on sentiment features through collaborative operation of an input layer, a hiding layer and an output layer, and it is ensured that a sentiment prediction result is scientific and reliable.

The invention belongs to the technical field of data processing, and provides a risk assessment method and system for a ship classification society and classification society inspection based on a large language model, and the method comprises the steps: S1, pre-training a large language model: pre-training the large language model based on a preset prompt word and a risk assessment data set, and obtaining a ship risk assessment large model; and S2, performing classification society risk and classification society inspection risk assessment on the ship based on the large language model to generate a first risk assessment result, a second risk assessment result and a third risk assessment result. According to the scheme, deep processing is carried out on existing ship operation information in combination with various third-party data according to a preset risk rule mainly by calling a large language model, corresponding risk scores are accurately calculated, and risk assessment suggestions and countermeasures with reference values are generated.

The invention provides a deep processing method and system for a digital delivery result, and relates to the technical field of data processing. A delivery result file set is disassembled and recombined into a unified object main data set and a three-dimensional entity set; each object attribute record and each object relation record have a unified object identifier, a source identifier and evidence positioning information at the same time, and multiple sets of coding calibers of the same object can be gathered to the same unified object identifier; refined model replacement is executed by taking a unified object identifier as an anchor point in three-dimensional treatment, continuity of a composition relationship and a connection relationship is kept, and after project handover and commissioning, online and field inconsistency found by operation and maintenance personnel is converted into a revision chain capable of being rechecked, judged and audited, so that the operation and maintenance efficiency is improved. Therefore, the problems that the unstructured data is difficult to reuse, the coding calibers are different, the object relationship is easy to lose and the data in the operation and maintenance period is not credible are solved.

The embodiment of the invention provides a text detection method and device, electronic equipment and a computer readable storage medium, and relates to the field of natural language processing. The method comprises the steps of dividing a to-be-detected text into at least one to-be-detected fragment, determining vector representation of the to-be-detected fragment, and determining a reference fragment corresponding to the to-be-detected fragment from a retrieval tree based on the vector representation; according to the semantic similarity between the to-be-detected fragment and the corresponding reference fragment and the TF-IDF value of the to-be-detected fragment, determining the target similarity of the to-be-detected fragment; selecting the to-be-detected fragments with the target similarity greater than a preset similarity threshold value from the to-be-detected fragments as target detection fragments; and inputting the target detection fragment and the corresponding reference fragment into a type prediction model, obtaining a plagiarism type corresponding to the output target detection fragment, and generating a detection result. The problems that the plagiarism condition of the text subjected to deep processing cannot be detected and the plagiarism type of the text cannot be accurately distinguished are solved.

The invention provides a Mongolian lip language recognition method based on multi-modal training. The Mongolian lip language recognition method comprises the steps of S1, a preprocessing stage, S2, a training stage and S3, a reasoning stage. The method comprises the following steps: S1, a preprocessing stage: segmenting words in a language model training text into forms of stems, lattice suffixes and verbs except verbs; s2, a training stage: firstly, performing feature extraction on audio and video by using ResNet to obtain feature representation of each mode; the feature representation is then input into a Conformer encoder, and a multi-modal shared representation is extracted through deep fusion; s3, a reasoning stage: firstly, performing fine feature extraction on a video image through ResNet, and then inputting the extracted video features into a Conformer encoder for deep processing; and finally, accurately predicting the text content by using a forward S4D decoder. The audio information is fused in the training process, so that the system can capture more voice features from a multi-mode angle, and the fuzzy and uncertainty possibly caused by only depending on the video information are made up.

This application discloses a method, apparatus, electronic device, and storage medium for constructing a gait recognition enhancement model. The method and apparatus are applied to an electronic device to construct a gait recognition enhancement model capable of multimodal gait recognition of documents. Specifically, in response to a user's model construction request, multimodal data is extracted from multiple documents; the multimodal data undergoes deep processing and standardization to achieve deep alignment, obtaining standardized features; a bio-inspired dynamic fusion mechanism is used to dynamically fuse the standardized features to obtain semantic information; and a generative large language model is used to train the original documents and semantic information to obtain the gait recognition enhancement model. This model can obtain gait information from documents, thereby improving researchers' efficiency in reading documents and the convenience of academic research, assisting researchers in quickly and accurately understanding domain knowledge, and thus helping researchers improve their learning and research capabilities.

The invention discloses a data processing method and device for VR interactive training of a cross-country container forklift. The method comprises the steps that VR interactive training data information of a user is acquired; the user VR interactive training data information comprises an action data set and an expression data set; processing the action data set to obtain a user action evaluation result; processing the expression data set to obtain a user emotion evaluation result; and according to the user action evaluation result and the user emotion evaluation result, performing personalized training suggestion and training content dynamic adjustment on the VR interactive training system. According to the invention, through efficient acquisition, deep processing and intelligent analysis of multi-dimensional data generated in the training process, the forklift driving operation level and safety awareness of trainees are accurately evaluated, the training result is fed back in real time, and the training strategy is dynamically adjusted, so that the quality and efficiency of the forklift driving VR interactive training are improved, and the training experience is improved. And the actual training cost and the safety risk are reduced.

The present invention relates to the field of machine tool processing technology, and discloses a method and system for recovering machine tool processing files from interruption, wherein a method for recovering machine tool processing files from interruption comprises: performing G-code semantic analysis and graph structure conversion to convert a G-code processing program into a structured knowledge graph with semantic association; performing processing intent reasoning based on neural symbolic reasoning to understand the deep processing intent behind the program; realizing self-evolutionary learning of processing experience, and continuously learning and optimizing from historical experience; locating interruption recovery points based on intent graphs to accurately find the optimal recovery position; generating and executing interruption recovery strategies to achieve smooth recovery of the processing process; the present invention improves the success rate and accuracy of interruption recovery by deeply understanding the G-code semantics and processing intent, and is suitable for high-value parts processing scenarios.

The present invention provides a method and system for automatic question labeling and classification, which photographs a target question and converts the photographed image into corresponding question text data, identifies the English text and Chinese text respectively contained in the question text data sample, and then obtains the corresponding subject type according to the vocabulary types respectively contained in the English text and the Chinese text, and performs adaptive labeling. Finally, according to the labeling results, the target question is automatically classified into the corresponding question set. In this way, a large number of target questions of different types can be automatically labeled and classified in a targeted and efficient manner, thereby improving the efficiency of deep processing of the questions and facilitating subsequent analysis and processing of the questions.

The invention discloses a wood veneer sound-absorbing composite material and a preparation method thereof, and belongs to the technical field of building materials. The material comprises a decorative surface layer and a sound-absorbing base material layer, the decorative surface layer is an aluminum foil veneer formed by compounding a natural veneer and an aluminum foil layer; the sound-absorbing base material layer adopts a colored velvet plate; and the aluminum foil wood veneer and the colored velvet plate are hot-pressed and compounded to form an integrated plate body through the aluminum foil layer of the aluminum foil wood veneer. The preparation method comprises the steps that S1, the aluminum foil veneer is prepared, wherein the natural veneer and the aluminum foil layer are compounded to obtain the aluminum foil veneer; s2, preparing a base material: cutting a colored velvet plate; s3, laminating and compounding to form an integrated plate body; s4, waxing the surface; and S5, carrying out integrated deep processing to form the designed sound absorption structure. The invention relates to a novel building material with functions of decoration, sound absorption and noise reduction, in particular to a material prepared by taking natural veneers and recyclable colored velvet plates as main raw materials through processes such as dyeing, compounding, waxing and CNC (computer numerical control) engraving, and application of the material in various public places.

The application provides a glass deep processing process analysis system and method based on AI cloud data, the method comprising: performing semantic analysis and structured packaging on the obtained process query text in combination with a glass industry term library to obtain a structured query instruction; performing database table association, list matching and context enhancement on the structured query instruction based on a bidirectional mode link algorithm, and performing sentence optimization on the matching result after context enhancement based on a binary selection strategy to obtain a primary query sentence; performing logical self-correction on the primary query sentence based on a multi-round syntax tree traversal method to obtain a database query sentence; querying the original data set from the glass processing database by using the database query sentence, and performing multi-dimensional data analysis on the original data set based on a dynamic threshold anomaly detection algorithm and a mutual information algorithm to obtain an analysis result set; and generating an interactive report according to the analysis result set.

This invention discloses a lightweight deep processing method based on BIM, belonging to the field of BIM lightweight technology. It involves constructing the complexity of a target model from the structural data obtained through identification. If the obtained complexity exceeds expectations, a lightweight strategy is provided for the target model based on several model features using a model optimization knowledge graph. The target model is simplified, and the accuracy of the simplified model is analyzed. If the obtained model accuracy exceeds expectations, the data within the target model is compressed. After performance testing of the target model, a performance coefficient is constructed from the test data set. If the obtained performance coefficient is not lower than its previous value, the optimization degree of the target model is constructed. If the optimization degree exceeds expectations, a re-optimization instruction is issued. This completes the lightweight processing of the model, reducing its bloat, maintaining model performance, improving model operating efficiency, and enhancing the efficiency of subsequent work.

This invention relates to the field of quality information traceability technology, and discloses a method and system for tracing the quality information of Epimedium sagittatum. The method includes acquiring basic data and planting parameters of the current batch of Epimedium sagittatum seedlings; processing the basic data and planting parameters using a hash algorithm to generate a basic query code; using the basic query code as a parent node and combining it with growth data to generate a growth code; generating a raw material code for the current batch by combining it with process parameters, and associating the raw material code with the growth code; generating an outbound code using a timestamp and the raw material code; generating auxiliary material codes for auxiliary materials involved in deep processing, associating the auxiliary material codes with the outbound code, and generating a product code by combining it with production parameters. This invention facilitates rapid identification of the source and responsible link of quality fluctuations, sets incremental collection frequencies and differentiated collection indicators for different stages, improves the sufficiency of quality assessment basis, and reduces the risk of tampering or falsification during the process.

The invention provides an event trigger word determination method and equipment, and the method comprises the steps: sequentially carrying out initial candidate trigger word extraction, supplementary candidate trigger word extraction and deep processing of at least one iteration round on a target text based on a pre-trained lightweight large model through a preset event trigger word extraction instruction and a sample set, obtaining a candidate set containing each candidate trigger word corresponding to the target text; according to the semantic similarity between each candidate trigger word in the candidate set and each preset historical statement and each historical event trigger word, selecting a preset number of candidate trigger words from the candidate set as target trigger words; and filtering the target trigger word based on a preset similarity score threshold and event type clustering to obtain an event trigger word corresponding to the target text and an event type corresponding to the event trigger word. According to the method, the event trigger word extraction speed can be ensured, and the event trigger word extraction precision and classification precision can be improved at the same time.

The invention discloses an engineering audit data compliance checking method and system based on a Transform model, and relates to the technical field of engineering audit data checking, and the method comprises the steps: S1, collecting financial data and material consumption data in the whole process of engineering audit; according to the method, structural analysis and deep processing are carried out on project auditing full-process data by relying on the context semantic understanding capability of a Transform model, and accurate semantic matching of auditing data and compliance terms is realized by combining dynamic construction of a compliance rule knowledge base and utilizing a self-attention mechanism, so that the auditing accuracy is improved. The problems that in the prior art, manual checking is low in efficiency and insufficient in accuracy, and a traditional rule matching method is insufficient in semantic understanding and poor in adaptability are solved, intelligence and high efficiency of engineering audit data compliance checking are achieved, the comprehensiveness of checking coverage and the accuracy of anomaly recognition are improved, the scene adaptability of the checking method is enhanced, and the checking efficiency is improved. And reliable technical support is provided for engineering audit.

The invention provides an intelligent red date sorting and shunting method and system, and relates to the technical field of image data analysis and processing, and the method comprises the steps: collecting image data of red dates; extracting size and shape features of the red dates from the image data of the red dates, comparing the extracted features with a preset appearance standard, and preliminarily screening out the red dates meeting the requirements of the first red date processed product as raw materials of the first red date processed product; and the red dates which do not meet the requirements of the first red date processed product are classified as raw materials of the second red date processed product, so that the problems of low sorting efficiency and high labor cost in the prior art are solved, and the problems of large fluctuation of product taste and physical characteristics and unstable quality in the deep processing process of the red dates are solved.

The present invention provides a method and device for sentiment metaphor recognition based on knowledge extraction and co-evolutionary reasoning, which relates to the technical field of natural language processing. The method deeply processes text data by constructing a high-quality sentiment metaphor corpus and combining advanced feature vectorization techniques; uses fuzzy multi-granularity knowledge extraction technology to accurately extract key features and form an optimized feature subset, thereby effectively removing redundant information and improving the accuracy of feature selection; further, through a matching network and a co-evolutionary reasoning mechanism, realizes dynamic recognition and annotation of sentiment metaphors, and can output metaphor labels and sentiment categories simultaneously. It aims to solve the problems of insufficient accuracy of sentiment metaphor recognition, limited reasoning ability, and difficulty in processing noise and redundant information in large-scale text data in the prior art.

The invention relates to the technical field of intelligent patrol assistance, in particular to an intelligent patrol assistance method and system based on large language model fine tuning, and the method comprises the steps: firstly deploying a large language model LLM after instruction fine tuning in a local government affair server; a user uploads a patrol document, and a back end executes preprocessing; constructing a patrol clue vector library, performing semantic comparison on the preprocessed text and the patrol clue vector library, and identifying potential clues; calling a built-in Prompt template to generate a first draft of a patrol report; in combination with a thematic lexicon and LLM situation judgment, labeling type labels and scores for the suspected high-risk paragraphs; cross-text clustering is carried out on the multiple documents, and repeated problem types are recognized; outputting the structured table and the natural language report; according to the method, hundreds of documents can be processed in batches through fusion application of the local deployment instruction fine-tuning large language model and the semantic retrieval technology, automatic deep processing of patrol documents is achieved, and the working efficiency is remarkably improved.

The invention discloses a multi-modal semantic segmentation method and a storage medium, and the method comprises the steps: obtaining a color feature map and a depth feature map of any ith hierarchy through a feature extraction module of a color processing branch and a depth processing branch at the hierarchy based on one or more groups of multi-modal sample images composed of a color image and a depth image; when i is greater than or equal to 2, inputting the color feature map and the depth feature map of the hierarchy into a modal feature selection module corresponding to the hierarchy to obtain a fusion feature map of the hierarchy; inputting the feature map, the color feature map of the level and the depth feature map of the level into a cross-modal fusion module of the level to obtain an enhanced fusion feature map of the level; and obtaining a semantic segmentation result corresponding to each group of multi-modal sample images according to the enhanced fusion feature map of the last level. The complementarity of the multi-modal features is enhanced through the modal feature selection module, the fusion of the multi-modal features is enhanced through the cross-modal fusion module, and high-precision real-time semantic segmentation is realized.