42 results about "Applying knowledge" patented technology

The invention relates to the technical field of data fusion, and discloses a multi-modal multi-source heterogeneous data fusion method and system.The multi-modal multi-source heterogeneous data fusion method comprises the steps that a multi-modal prototype network is constructed, different modal features are extracted, and a semantic mapping relation between modals is established; a memory module is constructed, and multi-modal characterization of historical key samples is stored through sample value evaluation; executing diversity sampling to reserve a rare abnormal mode; modal balance processing is implemented, and contribution weights of different modal gradients are dynamically adjusted; knowledge distillation is applied, and general fusion knowledge is extracted from a data-rich scene to assist small sample decision making; the method solves the problems of low fusion efficiency of small sample and cold start scenes, disastrous forgetting in incremental learning, unbalanced modal learning rate and rare mode retention under long-tail distribution, and is suitable for predictive maintenance, personalized recommendation and medical health monitoring in the manufacturing industry and scenes requiring small sample learning and incremental updating.

The invention discloses a music stave sentiment classification method and system based on multi-level distillation, and belongs to the technical field of music information processing. The method comprises the steps of firstly collecting music stave data and converting the data into stave data vectors, then performing feature extraction by using a long short-term memory network, then constructing a teacher network and a student network for knowledge distillation, and realizing multi-level knowledge transmission through temperature scaling, KL divergence loss and mask feature distillation. And finally, training a lightweight classification model to complete sentiment classification. The knowledge distillation technology is creatively applied to staff sentiment classification, the classification accuracy is effectively improved through an online multi-level distillation mode, and the technical problems that a traditional method lacks semantic information and a self-supervised model is not suitable for sentiment tasks are solved. The method has the main advantages of high classification precision, light model weight, capability of effectively capturing music emotion features and the like.

The invention relates to the technical field of resource scheduling, in particular to a digital-intelligent non-accompanying service management scheduling system, which comprises a service baseline construction module for instantiating a service item into a standard plan event flow; the state portrait generation module is used for collecting service object data flow, establishing a dynamic time sequence signal by applying a recurrent neural network model, and quantifying the occurrence probability of a future abnormal event through multi-step prediction; when the triggering is abnormal, executing causal two-way collision positioning reasons by applying the knowledge graph; the task load calculation module is used for integrating the plan event flow into a plan task component, and generating a worker resource list in combination with the predictive risk score and the worker real-time load and space reachability analysis containing the fatigue index; the system further comprises a task distribution module, the optimal staff is determined and a closed-loop scheduling instruction is generated by minimizing a composite cost function containing predictive load cost, and digital intelligent unattended service management from prediction to scheduling is realized.

The invention provides an aerospace application knowledge graph construction method, device, system and equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining multi-modal data; based on the data types, data processing is carried out on the modal data, and lightweight markup language texts and embedded vectors are obtained; performing knowledge extraction on the lightweight markup language text on the basis of a preset cue word and a large language model to obtain metadata information, and constructing an entity relationship triple on the basis of a relationship between entities and the metadata information; and based on the aerospace knowledge file number information, determining an association relationship among the aerospace knowledge file, the lightweight markup language text, the embedded vector and the entity relationship triple, and constructing the aerospace application knowledge graph according to the association relationship and the entity relationship triple. According to the space application knowledge graph constructed by the method, the space application field entities and the association relationship can be accurately presented, and efficient and accurate knowledge support is provided for space application knowledge questions and answers.

The invention provides an automatic construction and semantic understanding enhancement method and device for a large model knowledge base. Internet multi-type data sources are integrated through multi-source data collection, and text data related to the large model application field is obtained; a knowledge extraction and fusion technology is utilized to extract key knowledge elements from the collected data and eliminate redundant contradictory information to form a unified knowledge representation system; based on the unified knowledge representation system, automatically constructing a knowledge base with rich semantic association according to a predefined structure mode, and introducing a data quality evaluation mechanism to guarantee the reliability of the knowledge base; and based on the constructed knowledge base, the deep semantic analysis technology is used to enhance the semantic understanding ability of the large model, including semantic coding, knowledge graph embedding, semantic reasoning and knowledge completion. According to the scheme, efficient, accurate and automatic construction of the knowledge base can be achieved, the semantic understanding ability of a large model is remarkably improved, the large model can understand and apply knowledge more accurately, and powerful support is provided for artificial intelligence application.

Disclosed in the present application are a network policy management method and apparatus, and a device and a storage medium. The method comprises: acquiring at least one target policy instance for network management; on the basis of policy application knowledge, performing policy relationship analysis on the at least one target policy instance and an existing policy instance, so as to obtain a policy relationship analysis result; and on the basis of the policy relationship analysis result, performing optimization processing on the at least one target policy instance and the existing policy instance. The method can realize systematic management of network policies, can satisfy hierarchical and cross-field systematic universal management of operation and maintenance management policies for various network fields, and has the advantages of strong universality, high levels of flexibility and on-demand expandability.

The headings and Abstract of the Disclosure provided herein are for convenience only and do not limit or interpret the scope or meaning of the embodiments. The various embodiments described above can be combined to provide further embodiments. Aspects of the embodiments can be modified, if necessary to employ concepts of the various patents, application and publications to provide yet further embodiments.Based on the teachings contained in this disclosure, it will be apparent to persons skilled in the relevant art(s) how to make and use embodiments of this disclosure using data processing devices, computer systems and / or computer architectures other than that shown in FIG. 4. In particular, embodiments can operate with software, hardware, and / or operating system implementations other than those described herein.Embodiments of the present disclosure have been described above with the aid of functional building blocks illustrating the implementation of specified functions and relationships thereof. The boundaries of these functional building blocks have been arbitrarily defined herein for the convenience of the description. Alternate boundaries can be defined so long as the specified functions and relationships thereof are appropriately performed.The foregoing description of the specific embodiments will so fully reveal the general nature of the disclosure that others can, by applying knowledge within the skill of the art, readily modify and / or adapt for various applications such specific embodiments, without undue experimentation, without departing from the general concept of the present disclosure. Therefore, such adaptations and modifications are intended to be within the meaning and range of equivalents of the disclosed embodiments, based on the teaching and guidance presented herein. It is to be understood that the phraseology or terminology herein is for the purpose of description and not of limitation, such that the terminology or phraseology of the present specification is to be interpreted by the skilled artisan in light of the teachings and guidance.The breadth and scope of the present disclosure should not be limited by any of the above-described exemplary embodiments but should be defined only in accordance with the following claims and their equivalents.

The application discloses a kind of based on IC level EMC data assets measure quantification evaluation method, comprising: constructing read-only measure sensitivity database, extracting the quantification contribution value of target IC from IC level EMC data assets as measure sensitivity data;According to design mode or diagnosis mode, measure sensitivity data is called to predict measure effect and generate optimization scheme;Verification data is stored to user local application knowledge base, and local correction coefficient is generated for subsequent secondary correction.The present application converts EMC optimization from experience trial and error to quantitative calculation, supports pre-design reference and diagnosis rectification, and realizes user EMC design knowledge self-growth.

The invention discloses a drug-target interaction prediction method combining link-perceived cross attention and link-level contrast learning, and introduces a link-perceived cross attention mechanism which allows global features of a sequence (such as protein) to serve as a query to pay attention to internal details of a paired sequence (such as a drug), so as to predict drug-target interaction. Context-aware local features are dynamically generated for each drug-protein pair. Besides, in order to cope with the challenge that information of a knowledge graph is difficult to apply to a cold start scene, a link-based contrast learning strategy (Link-based contrast learning) is designed, and local features generated by a sequence mode through cross attention are aligned with local features obtained through different relation links in a relation mode, so that the probability that the information of the knowledge graph is applied to the cold start scene is lowered, and the probability that the information of the knowledge graph is applied to the cold start scene is lowered. Richer and finer-grained supervision signals are provided for the model, and the sequence model is prompted to learn a'relation mode 'contained in the knowledge graph instead of memorizing a'relation instance' of a specific entity.

The invention provides a dynamic handwriting enhancement method and device based on learning content semantic analysis and electronic equipment, and relates to the technical field of electronic note processing, and the method comprises the steps: obtaining space-time dynamic data corresponding to an original note when a user uses an electronic pen to write; performing stroke unit segmentation on the spatio-temporal dynamic data, and performing feature extraction on the segmented stroke units to obtain a feature tensor of each stroke unit; and performing semantic analysis based on the feature tensor of each stroke unit, determining a target enhancement strategy based on a semantic analysis result, and enhancing the original note based on the target enhancement strategy to obtain an enhanced note. According to the dynamic handwriting enhancement method and device based on learning content semantic analysis and the electronic equipment, by understanding the knowledge essence of the written content, sensing the real-time cognitive state of the user and applying the optimization strategy matched with the knowledge type, the handwriting readability is remarkably improved, and meanwhile the personal writing style is strictly reserved.

The invention relates to an automatic driving and automatic charging fusion application scene evaluation method and system, and the method comprises the steps: searching an association path between an automatic driving entity and an automatic charging entity in a fusion application knowledge graph, and summarizing an entity combination which accords with the logic and is higher in association degree as a candidate fusion application scene; extracting a numerical feature vector of each candidate fusion application scene from the fusion application knowledge graph; obtaining potential scores of each candidate fusion application scene under different evaluation indexes; obtaining an overall potential score according to the potential score of each candidate fusion application scene under different evaluation indexes; and obtaining a posterior mean value of each candidate fusion application scene according to the overall potential score and the potential scores under different evaluation indexes, sorting all the candidate fusion application scenes according to the posterior mean values, and identifying the most potential automatic driving and automatic charging fusion application scene. Compared with the prior art, the fusion application scene recognition accuracy is improved.

The invention provides an image recognition optimization method and system based on deep learning, aims to solve the problems of large algorithm error, low efficiency and the like of a traditional machine learning algorithm, and aims to realize rapid and accurate recognition of disease types of corn leaves and promote healthy development of the corn industry. The method comprises four core parts: 1) adopting MobileNetV3 to replace CSPDarknet53, and optimizing a lightweight backbone network structure; and 2) a Biform module is introduced, and the image processing efficiency is improved through a double-layer routing attention mechanism (BRA). In the coarse granularity stage, key-value matching items irrelevant to the path are removed, only important path segments are reserved for fine granularity operation, the calculation and storage amount is reduced, and a path weight dynamic adjustment mechanism is added to avoid the situation of characteristic connection loss; (3) key information is selected in real time through a double-layer routing attention mechanism, and the operation burden is relieved; and 4) a knowledge distillation technology is applied, the calculation amount is reduced through a teacher-student model compression method, and the prediction precision is maintained.

The present application relates to the technical field of resource scheduling, in particular to a digital intelligent unaccompanied service management scheduling system, comprising: a service baseline construction module, which instantiates service items into standard plan event streams; a state portrait generation module, which collects service object data streams, applies a recurrent neural network model to establish dynamic time series signals, and quantifies the occurrence probability of future abnormal events through multi-step prediction; when an abnormality is triggered, a knowledge graph is applied to perform bidirectional collision positioning of causes; a task load calculation module integrates the plan event streams into plan task components, combines the predictive risk score, and further combines the real-time load of staff containing a fatigue index and spatial accessibility analysis to generate a staff resource list; the system further comprises a task allocation module, which determines the optimal staff and generates closed-loop scheduling instructions by minimizing a composite cost function containing a predictive load cost, thereby realizing digital intelligent unaccompanied service management from prediction to scheduling.

The invention relates to the technical field of intelligent operation and maintenance of mechanical equipment, in particular to an incremental equipment health assessment method based on dynamic perception sensitivity assessment, the core is to construct a dynamic perception sensitivity assessment perception domain expansion framework, the framework integrates a multi-dimensional parallel attention mechanism and a dynamic perception sensitivity assessment mechanism, and the dynamic perception sensitivity assessment perception domain expansion framework is established. And respectively constructing a representation learning and evaluation sub-network and a perception domain expansion sub-network. According to the multi-dimensional parallel attention mechanism, by applying an attention mechanism on four dimensions, the multi-scale feature extraction capability is enhanced; according to the dynamic perception sensitivity evaluation, knowledge anchoring is applied to mean square error loss by calculating parameter perception sensitivity contribution degree and acts on subsequent tasks, so that important knowledge of historical tasks is protected. According to the method, the low-weight parameters can be properly forgotten in the continuous learning process, and the sensitivity is perceived and the model performance is improved under the condition of not carrying out time-consuming de novo training.

The application relates to the technical field of education training, and discloses a training subject management method and system based on knowledge classification. The system comprises a knowledge classification unit, a knowledge evaluation unit and a matching management unit. Through unique algorithms and technologies, knowledge is accurately divided into three types, namely, basic knowledge, advanced knowledge and application knowledge. For example, basic knowledge is compared by using a keyword library, advanced knowledge is constructed to calculate a correlation degree, and application knowledge is analyzed in combination with actual cases, thereby laying a solid foundation for subsequent training. A standardized evaluation process is designed for different types of knowledge, and the knowledge level of a trainer can be comprehensively understood. Different recommendation strategies are formulated according to the level of each knowledge type of the trainer, for example, excellent basic knowledge is focused on advanced and application knowledge, and large projects are recommended for excellent application knowledge, so that the learning needs of different trainers can be met, the problem of mismatched learning content is avoided, the training effect and efficiency are improved, and the knowledge and application ability of the trainer is more efficiently improved.

The present invention relates to the technical field of data fusion, and discloses a multi-modal multi-source heterogeneous data fusion method and system. Among them, a multi-modal multi-source heterogeneous data fusion method includes: constructing a multi-modal prototype network, extracting different modal features and establishing semantic mapping relationships between modalities; constructing a memory module, storing multi-modal representations of historical key examples through sample value evaluation; performing diversity sampling to retain rare abnormal patterns; implementing modal balance processing to dynamically adjust the weight of different modal gradient contributions; applying knowledge distillation to extract general fusion knowledge from data-rich scenarios to assist small-sample decision-making. The present invention solves the problems of low fusion efficiency in small-sample and cold-start scenarios, catastrophic forgetting in incremental learning, imbalance in modal learning rates, and retention of rare patterns under long-tailed distributions, and is applicable to scenarios such as predictive maintenance in manufacturing, personalized recommendation, medical health monitoring, and scenarios that require small-sample learning and incremental updates.

There is provided a method for unlearning to be performed by an unlearning apparatus, the method comprising: receiving to-be-forgotten target data as input; extracting a feature vector from the to-be-forgotten target data using a first layer that is a pre-attack layer of a pre-trained model; setting a second layer, which is a target attack layer of the pre-trained model, as a first model, and duplicating a second layer to generate a duplicated second layer to be set as a second model; generating a noisy feature vector by adding noise to the feature vector; fine-tuning one of the first model and the second model by applying a knowledge distillation technique to the first model and the second model using the feature vector and the noisy feature vector; and generating an unlearned model for the to-be-forgotten target data based on the first layer and the fine-tuned model.

The application discloses a kind of electroencephalogram privacy protection methods based on knowledge distillation, comprising:1, the pre-processing of original data, including data selection, sample selection, data up-sampling and down-sampling;2, establish deep learning model based on convolutional neural network (CNN) and self-attention network (Transformer);3, source model pre-training, the data of subject is used to train source model, and the source model with good classification effect is obtained;4, forgetting training data selection, including the selection of forgetting subject data and the selection of retaining subject data;5, forgetting model training, relying on memory teacher model and forgetting teacher model, the forgetting model is supervised training.The application applies knowledge distillation to electroencephalogram classification model, can make the model effectively forget the electroencephalogram information of specific subject, and provide effective electroencephalogram data privacy protection for subject.

The invention discloses an agent development and safe operation method and system based on a private cloud. The method comprises the following steps: firstly, constructing a multi-level security virtual private cloud architecture; secondly, deploying a containerized agent platform in each security domain, creating an application through a micro-service architecture, binding an exclusive knowledge base, and establishing an application-knowledge base one-to-one security channel based on digital signature; thirdly, implementing a fine-grained permission management and control mechanism, dynamically verifying a data security level label by adopting an access control model, and guaranteeing the security of sensitive information in combination with a privacy technology; then integrating a cue word optimization engine to improve the interaction performance of an agent, and applying a dynamic load balancing algorithm to realize efficient scheduling of a large model cluster; and finally, constructing an end-to-end security protection chain through an AK / SK authentication system. According to the method, the secret-related intelligent agent is managed and controlled from multiple angles, safe development and efficient operation of the intelligent agent in the secret-related environment are achieved, and it is guaranteed that secret-related knowledge information is safe and controllable.

The invention discloses an agent construction method and device based on knowledge base learning, a medium and a product. The method comprises the following steps: acquiring a target data stream through a data stream scheduling unit; a data loading party of the target data stream is determined from the central processing unit and the quantum processing unit through the data stream scheduling unit, the target data stream is sent to the data loading party, and the data loading party is the central processing unit or the quantum processing unit; when it is determined that the data loading party is the central processing unit, the target data stream is sent to the AI processing unit through the central processing unit; when it is determined that the data loading party is the quantum processing unit, the target data stream is sent to the AI processing unit through the quantum processing unit; and performing multi-application knowledge base learning based on the target data stream through the AI processing unit, and constructing a question response agent based on the multi-application knowledge base through the AI processing unit. According to the embodiment of the invention, efficient data transmission can be realized, so that the agent construction efficiency is improved.

The invention discloses a financial data analysis method and system based on a large model, and a medium, mainly relates to the technical field of financial data analysis, and is used for solving the problem that three links, namely data query, rule application and knowledge calling, of the isolation of each module in a framework in the prior art lack a collaborative mechanism. Comprising the following steps: acquiring initial financial problem data, and disassembling the initial financial problem data into a financial problem character string, a rule problem character string and a query problem character string; obtaining corresponding financial knowledge; obtaining a corresponding financial early warning rule; the large model integrates all the corresponding query sub-SQL into a final query SQL; the execution engine executes the final query SQL to obtain a query result; and taking the financial knowledge, the financial early warning rule and the query result as input parameters of a financial data summary function, further obtaining a cue word composed of the financial knowledge, the financial early warning rule and the query result, inputting the cue word into the large model, and obtaining a returned summary character string.

The present invention discloses a long short-term memory compression method based on pruning with a knowledge distillation recovery strategy. In the implementation of knowledge distillation, in addition to introducing distillation loss for the student model to fit the logits output of the teacher model, the cross-entropy loss between the output probability distribution and the true label is also used to ensure that the output of the student model matches the true label of the sample. These two parts of the loss together constitute the objective function, which helps the student model learn from the "hidden knowledge" of the teacher model and optimize the output probability distribution of the model, thereby improving the accuracy of the pruned model. In addition, the present invention applies knowledge distillation to the pruning process of the LSTM model. By rationally transferring knowledge, the pruned model has stronger representation capabilities.

The invention discloses a high-readability industrial drawing AI intelligent identification and generation method, and relates to the technical field of industrial drawings, and the method comprises the steps: S1, constructing a function map and a drawing map; s2, on the basis of the double maps, constructing a three-chain cooperation mechanism of a function chain, a drawing chain and an application knowledge chain; s21, the function chain takes the function map as a core basis, and the dispersed post requirements are converted into coherent collaborative links; s22, the drawing chain constructs a logic chain of the full life cycle of the industrial drawing through the dynamic knowledge graph; S23, the knowledge chain is applied, and according to the demand weight difference of different posts, the generated drawing report is interpreted based on the dynamically updated knowledge graph; according to the method, through the AI technology, the drawing processing efficiency and readability are remarkably improved, and personalized drawings can be generated and interpreted and matched on the basis of post roles.

The present invention relates to the field of automated control technology for microfluidic chips, and discloses an end-side microcarrier sorting system based on improved NanoDet using knowledge distillation, comprising: a microcarrier generation module based on microfluidic chip technology, which controls the fluid to form microcarriers containing cells through flow focusing and guides them to a processing channel; a real-time monitoring module uses a camera and a NanoDet model compressed by knowledge distillation to quickly identify whether the microcarriers contain target microcarriers; a data processing unit reduces the computing resource consumption and energy consumption of the end-side device by compressing the NanoDet model using a knowledge distillation algorithm; and a miniaturized hardware system capable of processing monitoring data in real time and controlling the microcarrier sorting process. The present invention integrates microcarrier sorting, monitoring, and automated operations on a microfluidic chip, effectively improving the throughput of microcarrier sorting. At the same time, the end-side deployment method based on the model improves the integration and usability of the entire system.

The present invention belongs to the field of knowledge graph construction, and discloses a method for constructing a knowledge graph for maritime mobile targets. Based on the dynamic positioning data of maritime mobile targets, the basic motion behavior semantics are defined, and further, the high-order behavior semantics applicable to multiple types of maritime mobile targets in complex environments are defined. A dynamic knowledge graph is constructed by using the dynamic positioning data of maritime mobile targets and the dynamically changing information in the sea area geographical environment information and meteorological information, and a static knowledge graph is constructed by using the static information of the data. The dynamic and static graphs are fused, and knowledge graph reasoning technology is applied to improve the quality of the knowledge graph. A unified knowledge system is used to organize and represent a large amount of heterogeneous AIS data, maritime mobile target motion patterns and behavior knowledge, accurately express the complex behavior semantics of maritime mobile targets, understand different behavior intentions, realize intelligent analysis and decision-making of maritime mobile target behaviors based on dynamic positioning data, and promote the intelligent monitoring and management of maritime mobile targets based on big data.

The application relates to the technical field of image data processing, and provides a chip visual design defect detection method and system based on deep learning, wherein the chip visual design defect detection method based on deep learning comprises the following steps: constructing a special search space and a hardware performance evaluation model; implementing a differentiable structure search optimization network structure; implementing mixed precision quantization optimization model storage and calculation efficiency; performing multi-objective joint optimization to simultaneously optimize the network structure and bit width configuration; and applying a knowledge distillation technology to maintain high detection precision. Through joint optimization of the network structure and precision configuration, the detection speed is significantly improved, the real-time detection requirement of a chip production line is met, the model size is greatly reduced, the storage space and transmission bandwidth are greatly saved, the demand for hardware resources is reduced, the hardware cost is reduced, lower configuration equipment can also be used for efficient detection, and the universality and economy of the method are improved.

The invention provides a customer service request processing method and device, electronic equipment, a computer readable storage medium and a computer program product, and relates to the technical field of computer technologies, in particular to the technical field of man-machine interaction, natural language processing, intelligent robots and other artificial intelligence. The method comprises the following steps: in response to the situation that associated information related to an order cannot be identified from a customer service request initiated by a user through a preset vectorization knowledge base, outputting an interactive interface for interacting with a preset customer service large model; filling an information input field of the interactive interface with the customer service request to obtain recognized initial semantics and initial confidence; in response to the condition that the initial confidence coefficient is lower than a preset confidence coefficient threshold value, initiating at least one round of supplementary inquiry to the user in the interactive interface, and complementally outputting supplementary information returned by the user in the interactive interface until target semantics with the confidence coefficient not lower than the confidence coefficient threshold value are obtained; and determining solution information corresponding to the target semantics from a preset application knowledge base, and outputting the solution information on the interactive interface. According to the method, all user requirements can be completed in the interactive interface in a one-stop manner, and the processing efficiency of the customer service request is improved.

This invention discloses a method, apparatus, electronic device, and storage medium for content detection of a model-associated knowledge base. The method includes: responding to a content detection request for the associated knowledge base of a target model, acquiring a target associated knowledge base for the target model, wherein the target associated knowledge base includes at least a test knowledge base and a scenario application knowledge base, the test knowledge base storing preset security detection knowledge, and the scenario application knowledge base storing industry security content; determining candidate risk content in the target associated knowledge base based on target prompt information and the target model, and generating test data based on the candidate risk content; testing the target associated knowledge base based on the test data, candidate risk content, and the target model, and calibrating the corresponding target risk content in the target associated knowledge base based on the test results, thereby achieving accurate risk mining across industry scenarios and dynamic optimization of knowledge base content, significantly improving the security detection capability and generalization of the target model.

The invention discloses a generative agent system with a long and short term memory mechanism and an implementation method thereof. The system comprises an agent generator, an agent editor, an agent actuator, a data bus, a knowledge base, a tool library and an algorithm library. The agent generator is responsible for generating a workflow, the agent editor is responsible for issuing the workflow, the agent actuator is responsible for executing the workflow, the data bus is responsible for data management in the workflow and ensures data transmission, storage and application in the workflow execution process, and the knowledge base, the tool library and the algorithm library are responsible for data, tool and algorithm storage respectively. The method is implemented based on the system. The method has the advantages that high stability of the workflow and flexibility of intelligent agent execution are both considered, and the execution efficiency and reliability in a severe engineering business scene are ensured while diversified user requirements are met; and the workflow can be quickly generated or adjusted according to business requirements, and can be published as available services by one key, so that the design and maintenance thresholds are remarkably reduced.

The present invention discloses a training method and device for a generative large language model based on knowledge base feedback. The method utilizes a domain knowledge base and supervised fine-tuning training to optimize the large language model; adopts a reinforcement learning method based on knowledge base feedback, utilizes the domain knowledge base, constructs a reward model, scores and provides feedback on the answers generated by the large language model, and constitutes a reinforcement learning process. The innovation of the present invention is to apply knowledge graph technology to reward construction, so that the large language model can be fine-tuned through the automated process of knowledge engineering, replacing the reinforcement learning based on human feedback, which is conducive to saving a large amount of high cost of human feedback annotation. The correct answer to the domain application obtained based on deterministic knowledge reasoning can correct the key defect of the generative large language model in fabricating facts. It can be used in the scenario of constructing a large language model for industry vertical applications based on the domain knowledge graph, and has strong applicability.