AI interaction intelligent module based on hybrid architecture

Abstract

The invention relates to the technical field of artificial intelligence and Internet of Things, and discloses an AI interaction intelligent module based on a hybrid architecture, comprising a user interaction unit which supports voice, text and image multi-modal input and integrates intention recognition and context understanding algorithms; the data processing unit is used for carrying out structured processing on the electric appliance specification and the historical fault data and constructing a dynamically updated knowledge graph; the hybrid architecture core unit comprises a deep learning subunit for realizing natural language understanding and generation based on a Transform model, and a knowledge reasoning subunit; a fault diagnosis unit; and a feedback optimization unit. According to the method, seamless cooperation of deep learning and symbol logic is realized through a dynamic routing strategy, a high-confidence-coefficient scene generates a response through a Transform model, a medium-confidence-coefficient scene calls a knowledge graph rule for verification, and a low-confidence-coefficient scene supplements information through multiple rounds of interaction, so that the effect of improving balance efficiency and safety is achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and the Internet of Things technology, and in particular to an AI interactive intelligent module based on a hybrid architecture. Background Art

[0002] With the rapid development of artificial intelligence technology, people's demand for intelligent interactive systems is growing, especially in the field of electrical equipment. They expect to communicate with the equipment more conveniently and efficiently to solve problems encountered during use. The traditional single interaction method can no longer meet the diverse needs of users. When describing electrical faults, users may express themselves in various forms such as voice, text, and images. Traditional systems often rely on fixed symbolic logic to understand user intentions. At present, it is difficult to explain the reasoning process by relying solely on a single symbolic logic, which makes the system's decision-making lack of explainability and reliability, and leads to insufficient flexibility and difficulty in taking into account the depth of semantic understanding. Summary of the Invention

[0003] In order to make up for the above shortcomings, the present invention provides an AI interactive intelligent module based on a hybrid architecture, which aims to improve the problem that the reasoning process is difficult to explain, resulting in the lack of explainability and reliability of interactive decision-making.

[0004] In a first aspect, the present invention provides the following technical solution: an AI interactive intelligent module based on a hybrid architecture, comprising: User interaction unit, supporting multimodal input of voice, text, and images, integrating intent recognition and context understanding algorithms; The data processing unit is used to perform structured processing on appliance manuals and historical fault data, and to build a dynamically updated knowledge graph; The core units of the hybrid architecture include: a deep learning sub-unit that implements natural language understanding and generation based on the Transformer model, and a knowledge reasoning sub-unit that combines the knowledge graph with the rule engine for logical reasoning; The fault diagnosis unit integrates sensor data and user feedback to dynamically update the fault probability model through a Bayesian network; Feedback optimization unit uses reinforcement learning to adjust interaction strategies and diagnostic logic in real time.

[0005] Preferably, the user interaction unit includes: The multimodal input fusion subunit maps speech, text, and image features into a unified semantic space through a large language model; The context tracking subunit maintains the user session history through dialogue state management and parses long context dependencies based on a multi-head cross-modal attention mechanism.

[0006] Preferably, the data processing unit includes: The data parsing subunit converts the text data of appliance manuals and historical fault data into structured knowledge through a joint extraction model of OCR (optical character recognition) and entity relationships, and extracts the relationship between faults and solutions from the appliance manuals. The knowledge graph construction sub-unit builds a network of entities and relationships through graph database storage, integrates multi-source data using entity disambiguation and alignment technology, and constructs a hierarchical structure of domain layer, device layer, and user layer; The dynamic update sub-unit is used to update entities and relationships with new appliance manuals and historical fault data using few-sample learning technology, integrating new knowledge into the existing knowledge graph, and using Git-like version control to record graph changes and support rollback to historical versions.

[0007] Preferably, the collaborative logic between the deep learning subunit and the knowledge reasoning subunit of the hybrid architecture core unit includes: If the intent confidence level output by the deep learning subunit is higher than a first threshold, a response is directly generated; If the confidence level is lower than the first threshold but higher than the second threshold, the knowledge reasoning subunit is called to perform rule verification; If the confidence level is lower than the second threshold, multiple rounds of clarification dialogues are initiated to supplement the input information.

[0008] Preferably, the deep learning subunit calculates the probability distribution of each intent category through a Softmax function, and takes the highest probability value as the confidence level; The first preset threshold is 0.85, and the second preset threshold is 0.6, which are obtained through training of historical interaction data.

[0009] Preferably, the Bayesian network of the fault diagnosis unit includes: Dynamic node expansion mechanism, which automatically adds nodes and initializes verification probabilities based on historical data when new undocumented failure modes are detected; The interpretable output mechanism generates a maintenance recommendation report containing the failure probability derivation path.

[0010] Preferably, the feedback optimization unit includes: The reward calculation subunit normalizes the user's feedback score, response time, and fault resolution rate into reward signals; The policy update subunit updates the interaction policy based on the policy gradient algorithm. It is used to select the optimal response action based on the current user intention and device status, including direct answers, counter-questions for clarification, or pushing graphic and text guidance.

[0011] In a second aspect, the present invention provides the following technical solution: an AI interaction method based on a hybrid architecture, the method comprising the following steps: S1. Multimodal intent recognition and fusion: Users input questions about appliances via voice, text, or images. A large language model aligns multimodal semantics, extracts intent, calculates confidence scores, and combines contextual history to understand user needs. S2, dynamic routing and collaborative reasoning, triggers hierarchical decision-making based on intent confidence, including high confidence, where responses are directly generated by deep learning sub-units, medium confidence, combined with knowledge graph rule verification, and low confidence, where information is supplemented through multiple rounds of interaction; S3, multi-source diagnosis and interpretable output, integrates sensor data and user descriptions, calculates fault probability through dynamic Bayesian networks, automatically expands new fault nodes and generates maintenance reports, and annotates the inference logic; S4, multimodal response and automatic control, generates text steps, voice prompts and AR visual guidance. If the device supports automation, it triggers the operation instructions and simultaneously provides feedback to the user on the diagnosis results and operation status; S5, closed-loop feedback and self-optimization, then collects user ratings and fault resolution rates, optimizes interaction strategies through reinforcement learning, updates the knowledge graph and adjusts diagnostic model parameters to achieve continuous iterative upgrades of the system.

[0012] In the third aspect, the invention provides the following technical solution: a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned AI interaction method based on a hybrid architecture when executing the computer program.

[0013] In a fourth aspect, the present invention provides the following technical solution: a readable storage medium having a computer program stored thereon, and the computer program, when executed by a processor, implements the above-mentioned AI interaction method based on a hybrid architecture.

[0014] The present invention has the following beneficial effects: 1. In the present invention, a dynamic routing strategy is used to achieve seamless collaboration between deep learning and symbolic logic. In high-confidence scenarios, responses are generated by the Transformer model. In medium-confidence scenarios, knowledge graph rules are called for verification. In low-confidence scenarios, information is supplemented through multiple rounds of interaction, thereby achieving the effect of improving balance efficiency and security.

[0015] 2. In the present invention, a multi-head cross-modal attention mechanism is used to integrate the temporal correlation of speech, text, and images in long conversations, and dynamically adjust the intent parsing weights to correct the ambiguity of the initial text description, ensure the accuracy of intent recognition in complex scenarios, and effectively resolve reference ambiguity and cross-modal conflicts in multi-round interactions.

[0016] 3. In the present invention, unstructured instructions are converted into structured knowledge through the joint extraction of OCR and entity relationships, and dynamic expansion of the graph is achieved by combining with few-sample learning. It supports automatic addition of nodes for multiple new cases, and records knowledge changes through Git-like version control to achieve rapid structured conversion of unstructured data. It also supports continuous evolution and historical backtracking of knowledge graphs, ensuring the timeliness, reliability and traceability of the diagnostic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is an architectural diagram of an AI interactive intelligent module based on a hybrid architecture proposed in the present invention; Figure 2 This is a user interaction unit architecture diagram of an AI interactive intelligent module based on a hybrid architecture proposed by the present invention; Figure 3 This is a data processing unit architecture diagram of an AI interactive intelligent module based on a hybrid architecture proposed in the present invention; Figure 4 This is a diagram of the feedback optimization unit architecture of the AI interactive intelligent module based on a hybrid architecture proposed by the present invention; Figure 5 This is a flowchart of an AI interaction method based on a hybrid architecture proposed by the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0019] Example 1 Reference Figure 1 - Figure 4 In a first embodiment of the present invention, the present invention provides an AI interactive intelligent module based on a hybrid architecture, comprising: User interaction unit, supporting multimodal input of voice, text, and images, integrating intent recognition and context understanding algorithms; The data processing unit is used to perform structured processing on appliance manuals and historical fault data, and to build a dynamically updated knowledge graph; The core units of the hybrid architecture include: a deep learning sub-unit that implements natural language understanding and generation based on the Transformer model, and a knowledge reasoning sub-unit that combines knowledge graphs and rule engines for logical reasoning; The fault diagnosis unit integrates sensor data and user feedback to dynamically update the fault probability model through a Bayesian network; Feedback optimization unit uses reinforcement learning to adjust interaction strategies and diagnostic logic in real time.

[0020] Specifically, the user interaction unit receives and processes multimodal inputs such as voice, text, and images to parse user intent and maintain contextual coherence. It uses a cross-modal attention mechanism to align the semantics of heterogeneous data, improving the accuracy of fuzzy intent recognition (for example, associating "the machine is loud" with high-frequency sensor noise data). This resolves reference ambiguity in multi-round interactions and makes interactions more consistent with human conversational logic. The data processing unit converts unstructured data such as appliance manuals and historical fault records into structured knowledge, constructing a dynamically scalable knowledge graph. This enables efficient knowledge organization and rapid retrieval (for example, locating the solution corresponding to the "E05 fault code" from a thousand-page manual in seconds). The unit also supports real-time updates and version backtracking of the knowledge graph, facilitating root cause tracing and knowledge base maintenance. This improves knowledge organization efficiency, enables rapid location and reasoning of complex logical relationships, and supports the continuous evolution of the knowledge graph. The core unit of the hybrid architecture handles diverse user needs through the collaboration of data-driven deep learning and symbolic rule-based reasoning. Its Transformer-based natural language understanding model covers long-tail semantic variations (such as the equivalent mapping of "refrigeration failure" and "not cool enough"). The rule engine constrains high-risk operations (for example, prohibiting non-professional users from removing high-voltage components), balancing flexibility and safety. The fault diagnosis unit integrates real-time sensor data and user feedback to calculate fault probabilities and analyze causes. It also uses a dynamic Bayesian network to fuse multi-source heterogeneous data (e.g., sudden temperature rise + user description of "unusual noise") to identify complex fault modes (e.g., "power supply overload combined with capacitor aging"). It also supports automatic expansion and probabilistic deduction of unrecorded faults based on knowledge graphs. The feedback optimization unit continuously optimizes interaction strategies and diagnostic logic based on user behavior and system performance. It uses reinforcement learning to dynamically adjust response strategies, such as prioritizing high-scoring solutions. It also iteratively updates model parameters through online learning mechanisms to improve the system's adaptability and ensure a steady increase in long-term service interaction effects. The user interaction unit includes: The multimodal input fusion subunit maps speech, text, and image features into a unified semantic space through a large language model; The context tracking subunit maintains the user session history through dialogue state management and parses long context dependencies based on a multi-head cross-modal attention mechanism.

[0021] Specifically, the multimodal input fusion subunit maps input data from heterogeneous modalities such as speech, text, and images into a unified semantic space, eliminating the semantic gap between different modalities and deeply correlating and aligning the information of multimodal data. The cross-modal encoding capabilities of large language models (such as the CLIP model and GPT-4V) map speech content (such as "air conditioner noise"), text descriptions (such as user-entered fault codes), and captured images (such as dust accumulation inside the device) into the same semantic vector, addressing the problem of incomplete single-modal information. For example, when a user describes "the machine is noisy," the audio spectrum features collected by the microphone can be combined to accurately locate "fan bearing wear" rather than "circuit failure," thereby improving the accuracy of intent recognition in complex scenarios. This is especially important for reliably analyzing the true intent of ambiguous expressions or cross-modal conflicting input. The context tracking sub-unit maintains the user conversation history, analyzes the context dependencies in long conversations, ensures the coherence and logical consistency of multiple rounds of interactions, and captures the temporal correlation between text, voice, and images in long conversations through a multi-head cross-modal attention mechanism (such as the user first describing "cooling difference" and subsequently uploading a filter image), dynamically adjusts the current intent parsing weight, and improves the accuracy of intent recognition.

[0022] The data processing unit includes: The data parsing sub-unit converts the text data of appliance manuals and historical fault data into structured knowledge through an OCR and entity relationship joint extraction model, and extracts the relationship between faults and solutions from the appliance manuals; The knowledge graph construction sub-unit builds a network of entities and relationships through graph database storage, integrates multi-source data using entity disambiguation and alignment technology, and constructs a hierarchical structure of domain layer, device layer, and user layer; The dynamic update sub-unit is used to update entities and relationships with new appliance manuals and historical fault data using few-sample learning technology, integrating new knowledge into the existing knowledge graph, and using Git-like version control to record graph changes and support rollback to historical versions.

[0023] Specifically, the data parsing sub-unit converts unstructured raw data such as appliance manual text and historical fault records into structured knowledge, extracting the relationship between fault phenomena and solutions. It uses OCR technology to convert tables and diagrams in paper or electronic manuals into processable text data. Combined with the entity-relationship joint extraction model (UIE), it automatically identifies the "fault code-solution" mapping relationship, achieving rapid structural conversion of unstructured data and significantly improving the efficiency of subsequent knowledge retrieval and reasoning. The knowledge graph construction sub-unit organizes the parsed structured knowledge into a multi-level knowledge graph, supporting efficient associative query and reasoning. It uses a graph database (Neo4j) to store entities such as devices, faults, and solutions, as well as their relationships. Entity disambiguation technology eliminates homonyms and constructs a hierarchical structure consisting of the domain layer (general fault logic), the device layer (specific model parameters), and the user layer (personalized historical records), accelerating knowledge query speed. By dynamically updating sub-units, new data is continuously integrated and the timeliness and consistency of the knowledge graph are maintained. Based on few-sample learning technology, only a small amount of labeled data (such as 5 new fault cases) is needed to expand the knowledge graph nodes and relationships. A Git-like version control mechanism is used to record each graph change, supporting one-click rollback to historical versions to ensure the reliability and maintainability of the knowledge base.

[0024] The collaborative logic between the deep learning subunit and the knowledge reasoning subunit of the core unit of the hybrid architecture includes: If the intent confidence level output by the deep learning subunit is higher than a first threshold, a response is directly generated; If the confidence level is lower than the first threshold but higher than the second threshold, the knowledge reasoning subunit is called to perform rule verification; If the confidence level is lower than the second threshold, multiple rounds of clarification dialogues are initiated to supplement the input information.

[0025] Specifically, the collaborative logic uses a dynamic routing mechanism to hierarchically call deep learning and knowledge reasoning capabilities based on the confidence level of intent. It responds quickly to high-confidence scenarios, reduces the misjudgment rate through cross-validation of knowledge graphs and rule engines in medium-confidence scenarios, and improves the accuracy of fuzzy intent resolution through multiple rounds of interaction to supplement information in low-confidence scenarios. Ultimately, it takes into account real-time, security, and fault tolerance in complex interactions.

[0026] The deep learning subunit calculates the probability distribution of each intent category through the Softmax function and takes the highest probability value as the confidence level; The first preset threshold is 0.85, and the second preset threshold is 0.6, which are obtained through training of historical interaction data.

[0027] Specifically, the deep learning sub-unit converts the model output into an interpretable intent probability distribution through the Softmax function, takes the highest probability value as the confidence score, and combines it with the dual threshold optimized based on historical data to achieve quantitative decision-making for intent analysis.

[0028] The Bayesian network of the fault diagnosis unit includes: Dynamic node expansion mechanism, which automatically adds nodes and initializes verification probabilities based on historical data when new undocumented failure modes are detected; The interpretable output mechanism generates a maintenance recommendation report containing the failure probability derivation path.

[0029] Specifically, the system automatically identifies and integrates unrecorded fault modes through a dynamic node expansion mechanism, and initializes prior probabilities based on historical similar data (for example, setting the initial probability to the mean of similar faults, 0.3), enabling rapid adaptation to new faults. Through the interpretable output mechanism, the black-box reasoning process of the Bayesian network is converted into an understandable diagnostic report, which enhances user trust and decision-making basis, makes the diagnosis transparent, enables maintenance personnel to trace the decision basis, and reduces the risk of secondary failures caused by misjudgment.

[0030] The feedback optimization unit includes: The reward calculation subunit normalizes the user's feedback score, response time, and fault resolution rate into reward signals; The policy update subunit updates the interaction policy based on the policy gradient algorithm. It is used to select the optimal response action based on the current user intention and device status, including direct answers, counter-questions for clarification, or pushing graphic and text guidance.

[0031] Specifically, the reward calculation subunit normalizes indicators from different dimensions, such as user feedback scores, response time, and fault resolution rate, and converts them into a unified reward signal. This comprehensively considers multiple key factors in the interaction process, quantifies the effectiveness and quality of user-system interactions, and provides a clear and quantifiable basis for subsequent strategy optimization. The effectiveness of the current interaction strategy is evaluated based on the level of the reward signal, providing precise guidance for strategy updates and continuously optimizing the system towards improving user satisfaction and fault resolution efficiency. The interaction strategy is updated through the policy gradient algorithm in the strategy update subunit. Based on the current user intention and device status, the optimal response method is selected from multiple response actions such as direct answers, counter-questions for clarification, or pushing graphic and text guidance to flexibly respond to different interaction situations. The strategy is dynamically adjusted based on real-time information to provide the response that best meets user needs, improve the success rate of fault resolution, and enhance the fluency and satisfaction of user interaction with the system.

[0032] Example 2: Reference Figure 5 In a second embodiment of the present invention, the present invention provides an AI interaction method based on a hybrid architecture, the method comprising the following steps: S1. Multimodal intent recognition and fusion: Users input questions about appliances via voice, text, or images. A large language model aligns multimodal semantics, extracts intent, calculates confidence scores, and combines contextual history to understand user needs. S2, dynamic routing and collaborative reasoning, triggers hierarchical decision-making based on intent confidence, including high confidence, where responses are directly generated by deep learning sub-units, medium confidence, combined with knowledge graph rule verification, and low confidence, where information is supplemented through multiple rounds of interaction; S3, multi-source diagnosis and interpretable output, integrates sensor data and user descriptions, calculates fault probability through dynamic Bayesian networks, automatically expands new fault nodes and generates maintenance reports, and annotates the inference logic; S4, multimodal response and automatic control, generates text steps, voice prompts and AR visual guidance. If the device supports automation, it triggers the operation instructions and simultaneously provides feedback to the user on the diagnosis results and operation status; S5, closed-loop feedback and self-optimization, then collects user ratings and fault resolution rates, optimizes interaction strategies through reinforcement learning, updates the knowledge graph and adjusts diagnostic model parameters to achieve continuous iterative upgrades of the system.

[0033] Specifically, through five major steps: multimodal intent recognition, dynamic routing collaborative reasoning, multi-source data diagnosis, multimodal response, and closed-loop self-optimization, a full-link intelligent closed loop of "perception-decision-execution-evolution" is constructed. Deeply integrating deep learning with a rule engine, a confidence-driven hierarchical decision-making mechanism (high-confidence direct answer, medium-confidence verification, and low-confidence clarification) balances efficiency and safety. A dynamic Bayesian network supports real-time expansion and explainable reasoning of unknown faults, resolving the cold start problem of traditional diagnosis. Subsequently, reinforcement learning-driven closed-loop optimization continuously improves interaction strategies and knowledge base coverage. In practical applications, the system achieves sub-second responses in scenarios such as smart homes and industrial operations and maintenance, improving the adaptability, transparency, and reliability of human-computer interactions.

[0034] In one embodiment: User voice inquiry: "Why does the washing machine vibrate so loudly during spin drying?" The system converts voice into text through voice recognition and calls the knowledge graph to match the fault tree related to "dehydration abnormality".

[0035] The hybrid architecture core unit determines the problem by combining sensor data (such as load weight and rotation speed) The fault diagnosis unit analyzes real-time data.

[0036] The system will provide feedback on possible causes: "uneven clothing distribution" or "shock absorber failure," and guide the user to perform a balance test or schedule a repair. After the user selects "Balance Detection", the system demonstrates the operation steps through the AR interface and records the solution results to optimize the model.

[0037] Example 3 The third embodiment of the present invention is based on the same inventive concept. The present invention proposes a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of an AI interaction method based on a hybrid architecture of the above embodiment.

[0038] Example 4 The fourth embodiment of the present invention is based on the same inventive concept. The present invention proposes a computer, which includes: a processor and a memory; the processor and the memory communicate with each other; the memory is used to store instructions; the processor is used to execute the instructions in the memory, and execute an AI interaction method based on a hybrid architecture of the above embodiment.

[0039] It should be understood that various components of the present invention may be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods may be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof may be used: a discrete logic circuit having logic gate circuits for implementing logic functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc.

[0040] Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or make equivalent replacements for some of the technical features therein. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An AI interactive intelligent module based on a hybrid architecture, characterized in that: include: User interaction unit, supporting multimodal input of voice, text, and images, integrating intent recognition and context understanding algorithms; The data processing unit is used to perform structured processing on appliance manuals and historical fault data, and to build a dynamically updated knowledge graph; The core units of the hybrid architecture include: a deep learning sub-unit that implements natural language understanding and generation based on the Transformer model, and a knowledge reasoning sub-unit that combines the knowledge graph with the rule engine for logical reasoning; The fault diagnosis unit integrates sensor data and user feedback to dynamically update the fault probability model through a Bayesian network; Feedback optimization unit uses reinforcement learning to adjust interaction strategies and diagnostic logic in real time.

2. The AI interactive intelligent module based on hybrid architecture according to claim 1, characterized in that: The user interaction unit includes: The multimodal input fusion subunit maps speech, text, and image features into a unified semantic space through a large language model; The context tracking subunit maintains the user session history through dialogue state management and parses long context dependencies based on a multi-head cross-modal attention mechanism.

3. The AI interactive intelligent module based on hybrid architecture according to claim 1, characterized in that: The data processing unit includes: The data parsing sub-unit converts the text data of appliance manuals and historical fault data into structured knowledge through an OCR and entity relationship joint extraction model, and extracts the relationship between faults and solutions from the appliance manuals; The knowledge graph construction sub-unit builds a network of entities and relationships through graph database storage, integrates multi-source data using entity disambiguation and alignment technology, and constructs a hierarchical structure of domain layer, device layer, and user layer; The dynamic update sub-unit is used to update entities and relationships with new appliance manuals and historical fault data using few-sample learning technology, integrating new knowledge into the existing knowledge graph, and using Git-like version control to record graph changes and support rollback to historical versions.

4. The AI interactive intelligent module based on hybrid architecture according to claim 1, characterized in that: The collaborative logic between the deep learning subunit and the knowledge reasoning subunit of the hybrid architecture core unit includes: If the intent confidence level output by the deep learning subunit is higher than a first threshold, a response is directly generated; If the confidence level is lower than the first threshold but higher than the second threshold, the knowledge reasoning subunit is called to perform rule verification; If the confidence level is lower than the second threshold, multiple rounds of clarification dialogues are initiated to supplement the input information.

5. The AI interactive intelligent module based on hybrid architecture according to claim 4, characterized in that: The deep learning subunit calculates the probability distribution of each intent category through the Softmax function and takes the highest probability value as the confidence level; The first preset threshold is 0.85, and the second preset threshold is 0.6, which are obtained through training of historical interaction data.

6. The AI interactive intelligent module based on hybrid architecture according to claim 1, characterized in that: The Bayesian network of the fault diagnosis unit includes: Dynamic node expansion mechanism, which automatically adds nodes and initializes verification probabilities based on historical data when new undocumented failure modes are detected; The interpretable output mechanism generates a maintenance recommendation report containing the failure probability derivation path.

7. The AI interactive intelligent module based on hybrid architecture according to claim 1, characterized in that: The feedback optimization unit includes: The reward calculation subunit normalizes the user's feedback score, response time, and fault resolution rate into reward signals; The policy update subunit updates the interaction policy based on the policy gradient algorithm. It is used to select the optimal response action based on the current user intention and device status, including direct answers, counter-questions for clarification, or pushing graphic and text guidance.

8. An AI interaction method based on a hybrid architecture, characterized in that: An AI interactive intelligent module based on a hybrid architecture according to any one of claims 1 to 7, the method comprising the following steps: S1. Multimodal intent recognition and fusion: Users input questions about appliances via voice, text, or images. A large language model aligns multimodal semantics, extracts intent, calculates confidence scores, and combines contextual history to understand user needs. S2, dynamic routing and collaborative reasoning, triggers hierarchical decision-making based on intent confidence, including high confidence, where responses are directly generated by deep learning sub-units, medium confidence, combined with knowledge graph rule verification, and low confidence, where information is supplemented through multiple rounds of interaction; S3, multi-source diagnosis and interpretable output, integrates sensor data and user descriptions, calculates fault probability through dynamic Bayesian networks, automatically expands new fault nodes and generates maintenance reports, and annotates the inference logic; S4, multimodal response and automatic control, generates text steps, voice prompts and AR visual guidance. If the device supports automation, it triggers the operation instructions and simultaneously provides feedback to the user on the diagnosis results and operation status; S5, closed-loop feedback and self-optimization, then collects user ratings and fault resolution rates, optimizes interaction strategies through reinforcement learning, updates the knowledge graph and adjusts diagnostic model parameters to achieve continuous iterative upgrades of the system.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the AI interaction method based on hybrid architecture as described in claim 8.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, and when the computer program is executed by the processor, it implements the AI interaction method based on hybrid architecture as claimed in claim 8.

Images