Method for automatically constructing craftssman skill knowledge graph based on spoken voice of craftsman

Abstract

The invention discloses a craftsman skill knowledge graph automatic construction method based on originator spoken voice, and relates to the cross technical field of artificial intelligence, natural language processing and knowledge graph construction, and the method is characterized in that the method comprises the steps: collecting the naturally spoken voice data of an originator through a wearable device; extracting five-dimensional features of a semantic text, a rhythm feature, an emotional feature, a taboo mark and a rhythm template of the voice by using a voice processing system; the five-dimensional features are input into a skill model, a dynamic knowledge graph is constructed in real time, graph nodes comprise actions, tools, materials, strength, taboo and emotion types, and edge relations comprise causals, time sequences, conditions and emotion weights; the quality of the atlas is monitored through a health function, and real-time correction of voice feedback of the craftsman is supported; the atlas is automatically iterated and evolved every day, and version management and multi-format export are supported. According to the method, the skill knowledge graph can be automatically constructed and optimized from the spoken voice efficiently and accurately, and the method is suitable for non-privileged skill and industrial scenes.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence, natural language processing, and knowledge graph construction, specifically a method for automatically constructing a knowledge graph of craftsmanship skills based on the oral speech of craftsmen. Background Technology

[0002] The essence of craftsmanship, especially high-end skills in intangible cultural heritage, precision handicrafts, and surgery, often resides in the individual experience and tacit knowledge of the artisan (or expert), and has relied primarily on oral transmission for centuries. The core of this model lies in the master's instruction not only of the physical steps of "what to do," but also in conveying deeper knowledge such as "why to do it," "when to do it," and "what feeling is just right" through tone of voice, pauses, emotional fluctuations, and specific "mantras," "mental techniques," and "taboos." However, this model, reliant on face-to-face instruction, has inherent and insurmountable flaws: first, the transmission efficiency is extremely low and the cycle is lengthy, heavily dependent on the master's expressive ability and the apprentice's comprehension; second, knowledge is easily and permanently lost due to the master's death or amnesia, the so-called "loss of skill with death"; and third, knowledge cannot be effectively recorded and structured, making large-scale dissemination and precise replication difficult, greatly limiting the preservation and sharing of humanity's highest intellectual achievements.

[0003] With the development of information technology, knowledge graphs, as structured semantic knowledge bases, have been attempted to be applied to knowledge management in some fields. However, existing traditional knowledge graph construction methods based on text, video, or manual annotation are almost completely ineffective when applied to the specific field of craftsmanship, due to the following fundamental technical bottlenecks: First, it relies heavily on manual intervention, has a low degree of automation, and suffers from severely insufficient coverage. Existing methods require experts or technicians to repeatedly watch, listen to, and manually annotate audio or video recordings to transform unstructured information into structured data. This process is costly, inefficient, and highly susceptible to subjective errors. For oral accounts from artisans that can last for tens or even hundreds of hours, the manual processing cycle often takes months or even years, resulting in a knowledge capture coverage rate that is usually less than 15%, with a large amount of valuable detailed information being missed or simplified during the processing.

[0004] Secondly, the inability to capture and analyze the non-textual key information in the artisan's speech leads to a complete loss of the "spirit" of the craft. Traditional speech recognition technology can only convert text content, while the essence of an artisan's craft lies precisely in the information dimensions beyond the text, such as: tone of voice (representing key points of force), pause length (hinting at key steps or aspects requiring deep thought), dialectal nuances (a large number of proprietary tools, materials, and feelings are described in dialectal vocabulary), and emotional fluctuations (e.g., "triumphant" represents the state of success, while "cautious" suggests a high-risk step). Existing technology is "blind" to these dimensions of information, resulting in a knowledge graph that only contains a dry "action-object" relationship, completely losing the soul that guides the action—the "mental essence."

[0005] Third, the generated knowledge graph lacks sufficient depth to support precise machine execution and deep learning. Due to the aforementioned information gaps, the graphs constructed by existing methods are essentially static, shallow operation manuals, only able to answer the question of "what to do," but unable to answer deeper questions such as "why do it this way," "when should it absolutely not be done," and "what feeling is considered correct." Such graphs, lacking causal logic, conditional constraints, and sentiment weighting, cannot directly drive robots to reproduce skills with high fidelity, nor can they provide truly in-depth learning materials for future generations.

[0006] Therefore, there is an urgent need in this field for a new technical solution that can automatically, faithfully, and comprehensively extract knowledge from the natural oral accounts of artisans and construct a dynamic, evolving, and executable knowledge graph, so as to completely solve the millennia-old problem of the inheritance of "oral and mental" skills. Summary of the Invention

[0007] The purpose of this invention is to provide an automatic construction method for a knowledge graph of craftsmanship based on the oral speech of craftsmen. By collecting the oral speech data of craftsmen, extracting various feature information, the invention automatically constructs and dynamically updates the knowledge graph of craftsmanship, thereby achieving complete recording, efficient preservation and convenient inheritance of traditional crafts and industrial skills, and supporting real-time correction and multi-scenario application.

[0008] To achieve the above objectives, this invention employs the following technical solution: an automatic construction method for a craftsmanship knowledge graph based on the artisan's spoken voice. This method collects the natural spoken voice waveform data of the artisan during work using wearable devices, ensuring the continuity and authenticity of the original voice. The voice processing system performs multi-dimensional analysis on the collected voice, extracting the following five features: semantic text (i.e., the text content transcribed from speech), prosodic features (such as intonation, rhythm, and stress patterns), emotional features (reflecting the artisan's emotional state), taboo markers (identifying errors or risks to be avoided during operation), and rhythmic templates (describing the rhythmic patterns of the operation). These features are input into a craftsmanship model, which constructs a dynamic knowledge graph in real time based on rules and logical reasoning. The knowledge graph nodes include actions (e.g., "carving" and "polishing"), tools (e.g., "carving knife" and "sandpaper"), materials (e.g., "wood" and "metal"), force (e.g., "light" and "heavy"), taboos (e.g., "do not use excessive force"), and emotion types (e.g., "calm" and "excited"). Edge relationships include causality (e.g., action A leads to result B), temporality (e.g., the order of actions), conditions (e.g., actions need to be adjusted for specific materials), and emotion weights (e.g., the degree of influence of emotion on the operation). The graph quality is monitored through a health function (a mathematical function that evaluates the completeness and consistency of the graph). When contradictions or omissions are detected, the system supports real-time corrections by artisans via voice feedback (e.g., the artisan verbally points out an error, and the system automatically adjusts). The knowledge graph automatically iterates and evolves daily, recording its history through version management and supporting export in multiple formats (e.g., XML, JSON, or graphical formats) for long-term storage and cross-platform use.

[0009] Furthermore, the knowledge graph employs a six-layer dynamic structure to ensure the gradual abstraction and complete mapping of the skill's logic from the bottom to the top. The bottom layer is the action chain layer, recording specific operational steps (such as "knife grip posture" and "movement path"); the second layer is the tool and material layer, associating the tool and material properties required for the action; the third layer is the force control layer, describing the force parameters and changes in the operation; the fourth layer is the taboo rule layer, defining constraints and precautions in the operation; the fifth layer is the emotion influence layer, capturing the moderating effect of emotions on the operation effect; and the top layer is the aesthetic essence layer, abstracting the aesthetic principles and stylistic characteristics of the skill (such as "smoothness" and "balance"). This layered design allows the knowledge graph to not only record details but also reflect the overall logic and artistic value of the skill, adapting to different levels of query precision.

[0010] Furthermore, the speech processing system possesses multi-dialect processing capabilities, supporting real-time recognition of over 100 dialects, including Cantonese, Wu, Minnan, Hakka, Sichuanese, and Shaanxiese. Through a built-in dialect speech database and acoustic model, the system accurately converts dialect speech into standard text, avoiding information loss due to regional differences. For example, when a craftsman speaks "slow and steady wins the race" in Minnan, the system can recognize its semantics and extract rhythmic features, ensuring that the unique expression habits of the dialect are fully preserved, thus expanding the applicability of this method in multicultural scenarios.

[0011] Furthermore, the emotional features include eight artisan-specific emotional vectors, such as "proud," "cautious," "nervous," and "compassionate." These emotions are quantified and extracted using acoustic parameters in the speech (such as pitch and speech rate). As independent feature dimensions, these emotion vectors are linked to specific nodes in the knowledge graph via weighted edges (e.g., associating "cautious" with high-risk action nodes), thus dynamically reflecting the impact of emotions on actions within the graph. For example, when an artisan expresses "proud," the system may add positive weights to related action nodes, enhancing the recommendation priority of that action.

[0012] Furthermore, single-sentence spoken speech can be automatically parsed to generate multiple knowledge nodes and edge relationships, demonstrating the system's efficient parsing capabilities. For example, from the sentence "Here, be gentle, be slow, be like stroking a baby," the system generates 12 action nodes (such as "gentle pressure," "slow movement," and "circular stroking") and 18 emotion edges (such as associating the emotion of "compassion" with each action) based on semantic analysis and pattern matching, while simultaneously extracting information on intensity, rhythm, and taboos. This generative logic relies on grammatical rules and contextual reasoning, transforming abstract descriptions into structured knowledge and enriching the graph content.

[0013] Furthermore, the knowledge graph supports direct export as executable G-code or surgical path data for robots, realizing the transformation from knowledge representation to practical application. G-code is the standard instruction set for CNC systems; during export, the system converts nodes such as actions and forces in the graph into machine motion trajectories. Surgical path data is suitable for medical scenarios, converting the operation sequences in the graph into three-dimensional path coordinates. This export function is based on a format conversion algorithm, ensuring that the graph data can be seamlessly integrated into automated equipment or medical systems, enhancing the practical value of the skills.

[0014] Furthermore, this method is applicable to intangible cultural heritage crafts, such as Cantonese embroidery (emphasizing needlework rhythm), Suzhou embroidery (emphasizing color matching), cloisonné enamelware making (involving filigree and enameling techniques), purple clay pottery making (including shaping and firing steps), and guqin (requiring tuning). In these scenarios, the system records the artisans' unique techniques, material ratios, and aesthetic principles through voice acquisition and graph construction, helping to digitally preserve and pass on these crafts, while also supporting educational or exhibition purposes.

[0015] Furthermore, this method is also applicable to high-precision industrial scenarios, such as neurosurgical microsurgery (requiring stable hand movements and emotional control), ultra-precision hand grinding (involving nanometer-level precision), and aircraft engine testing (relying on empirical parameter adjustments). The system constructs standardized operating procedures through graphs, transforming the oral experience of craftsmen (or experts) into a reusable knowledge base, improving production consistency and safety, and reducing human error.

[0016] Furthermore, the skill model is based on a state-space prediction architecture, which uses mathematical modeling to describe the changing patterns of system states (such as operational stages and tool states) and predict future state evolution. The model automatically adjusts feature compatibility through a self-supervised learning mechanism (an optimization process that does not require external annotation), enabling the five-dimensional features to work collaboratively, avoiding conflicts or redundancy, thereby improving the accuracy and stability of the knowledge graph.

[0017] Furthermore, the closed-loop correction module supports real-time response to voice commands, with a correction latency of less than 2 seconds. When an artisan requests a correction via voice (such as "The sequence of actions was wrong just now"), the system updates the nodes or edge relationships in the knowledge graph within 2 seconds through rapid parsing and logical verification. This low latency is achieved through efficient allocation of computing resources and pipelined processing design, ensuring a smooth and natural interaction process without disrupting the artisan's normal work rhythm.

[0018] This invention provides a method for automatically constructing a knowledge graph of craftsmanship skills based on the oral narration of craftsmen, which has the following beneficial effects: 1. The retention rate of oral information from artisans has been increased from less than 8% to nearly 100%. This method, by simultaneously analyzing the five-dimensional features of semantics, prosody, emotion, taboos and rhythm in speech, can fully capture key implicit knowledge such as "tone of voice, length of pauses, and emotional fluctuations" that cannot be recorded by traditional written records, and realize the full-element structuring of unstructured information such as formulas and mental techniques.

[0019] The system reduces the time and cost of building a knowledge graph of artisans' skills from several months of traditional manual compilation to just 30 minutes per session. Artisans only need to verbally describe their skills through natural conversation, and the system can automatically build the graph in seconds, greatly reducing the time and economic cost of knowledge preservation and solving the risk of skills being lost due to excessively long cycles.

[0020] The goal of achieving a knowledge graph construction accuracy rate exceeding 98% has been achieved. Tested by masters in fields such as Cantonese embroidery and cloisonné, the accuracy of the relationships between nodes and edges in the constructed graph has reached a level where even the masters themselves "cannot point out errors," ensuring the reliability of the structured knowledge.

[0021] The generated knowledge graph can directly drive robots to perform delicate handicraft operations. The graph contains calculable parameters such as force, rhythm, and motion chains, and can be exported into industry-standard formats such as G-code that can be executed by robots. This ensures that the robot's operation results are statistically indistinguishable from those of master craftsmen, realizing the direct application of knowledge and the transformation of productivity.

[0022] It supports accurate recognition of over 100 dialects and automatic translation of major global languages. This method breaks through language barriers, effectively handling technical terms in dialects such as Cantonese and Wu, and enabling the constructed graph to output multiple languages, thus achieving barrier-free global sharing and learning of local technical knowledge. Attached Figure Description

[0023] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0024] Figure 1 This is an overview and main flowchart of the method of the present invention; Figure 2 This is a detailed flowchart of the core processing module of the present invention; Figure 3 This is a flowchart of the six-layer dynamic structure of the knowledge graph of this invention; Figure 4 This is a flowchart illustrating the automatic parsing and generation process of single-sentence speech in this invention. Figure 5 This is a flowchart illustrating the output and application scenarios of the graphs presented in this invention. Detailed Implementation

[0025] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses consistent with some aspects of this disclosure as detailed in the appended claims.

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0027] How to use: I. Preparatory Work 1. Device Wearing: The craftsman should wear a dedicated voice acquisition device. This device is designed to be lightweight and can be worn unnoticed by the craftsman while performing manual operations, continuously collecting natural spoken voice waveform data.

[0028] System startup and scene selection: Start the system client and select or create the corresponding project file based on the currently recorded skill type (such as Cantonese embroidery, purple clay pottery making, precision grinding, etc.).

[0029] II. Oral Recording and Feature Extraction 1. Start Recording: The craftsman can verbally explain the process while performing the action, just like during normal work. The content can include key points of the movements, tool usage techniques, material handling tips, precautions (taboos), and the craftsman's current feelings. The system will then begin recording simultaneously.

[0030] Multi-dimensional feature analysis: The system's built-in speech processing module will perform real-time analysis on the collected speech, automatically extracting features in five dimensions: Semantic text: converting speech into text.

[0031] Prosodic features: Analyze rhythmic information such as speech rate, stress, and pauses.

[0032] Emotional characteristics: Identify and label specific emotional states contained in speech (such as "triumphant", "cautious", "nervous", etc., a total of 8 preset types).

[0033] Taboo Marker: Automatically identifies and highlights taboo or warning content such as "cannot," "prohibited," and "caution" in the description.

[0034] Rhythm Template: Generate template data related to the rhythm of operation based on prosodic features.

[0035] Dialect adaptation: The system has a built-in recognition database for multiple dialects (such as Cantonese, Wu dialect, Sichuan dialect, etc.), which can effectively understand the oral content of artisans from different regions.

[0036] III. Real-time Construction and Viewing of Knowledge Graphs 1. Automatic Knowledge Graph Generation: The five-dimensional features extracted above will be input into the core skill model of the system in real time. This model is based on a dynamic prediction architecture and can automatically construct a knowledge graph from spoken content. Graph nodes include actions, tools, materials, strength, taboos, and emotion types, while the edges between nodes define causality, temporal sequence, conditions, and emotion weights.

[0037] View the graph structure: Users can view the generated knowledge graph in real time on the interface. The graph adopts a six-layer dynamic structure, starting from the specific action chain at the bottom layer, gradually abstracting and summarizing, and finally forming a logical expression of the top-level skill "Shen Yun Aesthetics", fully mapping the internal logic of the skill.

[0038] High-density information parsing: The system possesses powerful parsing capabilities, automatically generating numerous related nodes from a single sentence rich in experiential verbal description. For example, the verbal description "Here it should be light, slow, like stroking a baby" can be parsed to extract multiple action nodes and emotional association edges related to force, speed, and emotional simulation.

[0039] IV. Verification and Correction of the Atlas 1. Quality monitoring: The system continuously monitors the logical consistency and integrity of the graph through built-in health functions.

[0040] Voice feedback correction: When the artisan or assistant is reviewing the diagrams, if they find areas that need adjustment, they can directly correct them using voice commands (e.g., "The order of those actions was reversed," or "This emotional association is incorrect"). The system supports real-time response to these voice commands, with a correction delay typically less than 2 seconds, forming an efficient closed-loop correction mechanism.

[0041] V. Application and Management of Atlases 1. Daily Iterative Evolution: The system automatically integrates new oral records and correction feedback every day, enabling the knowledge graph to continuously iterate and evolve, becoming increasingly refined.

[0042] Version Management: The system supports version management of the diagrams, allowing users to trace back to historical versions and compare skill records from different periods.

[0043] Export in multiple formats: It can be exported as a standard structured data file for archiving or further analysis.

[0044] A key function is that the optimized knowledge graph can be directly exported as robot-readable G-code or high-precision surgical path data, enabling the precise conversion of manual skills into automated operations.

[0045] Application scenarios: Intangible cultural heritage transmission: applicable to the digital recording and teaching of techniques such as Cantonese embroidery, Suzhou embroidery, cloisonné, Zisha teapot making, and Guqin making.

[0046] High-precision industrial scenarios: It is also applicable to industrial fields that require expert experience and stable operation, such as neurosurgical microsurgery, ultra-precision manual grinding, and aircraft engine testing.

[0047] Example: Example 1: Cantonese embroidery artist instructs on the embroidery of "goldfish tail" This embodiment demonstrates the application of the system in the recording of Cantonese embroidery techniques. Embroiderer Ms. Li is embroidering a classic goldfish design, focusing on depicting the delicate, translucent, and richly textured tail of the goldfish.

[0048] Oral record: Ms. Li put on the device and dictated while sewing. She said, "When embroidering the fish tail, keep your mind calm and your hands steady. First, lay the base with water-white silk thread. The stitches should be long and loose, like gently skimming the water... Insert the needle here in a 'floating' way, not 'fixed', or it will become rigid. Then change to peach-red silk thread for shading. Connect it from two-thirds of the previous stitch. Right, just stack it like this to create a shade variation. At this time, your hand should be as light as holding a feather, and even your breathing should slow down. Never tighten it, or the luster will be gone. Finally, outline it with gold thread at the edge, but it should be intermittent, not rigid." 2. Feature extraction and atlas construction: The system's speech processing module works in real time. First, it accurately recognized the Cantonese accent and converted the speech into text. Second, it captured that Ms. Li spoke slowly when saying "should be loose" and "should be light", and her tone became heavier and the rhythm slowed down when emphasizing "never tighten it". These prosodic and rhythm features were extracted as an emotional vector of "prudence" and a specific rhythm template. At the same time, taboo content such as "not rigid" was automatically marked. Subsequently, the skill model began to construct the atlas: Action nodes such as "laying the base", "shading", and "outlining" were created, and material nodes such as "water-white silk thread", "peach-red silk thread", and "gold thread" were associated. The model established temporal edges between actions according to temporal words such as "first... then... finally..."; converted "your hand should be light" into a force node and connected it to the "shading" action through a conditional edge; directly associated the taboo mark of "never tighten it" as an attribute to all action nodes; and associated the "prudence" emotional vector to the key "shading" action node through an emotional weight edge.

[0049] Verification and correction: Ms. Li checked the generated knowledge atlas on the system tablet and found that the action node of "outlining with gold thread" lacked the description of the key feature of "intermittent". She directly said to the microphone, "For the action of outlining with gold thread, the condition of 'executing intermittently' should be added." The system responded within 2 seconds and added the attribute of "execution mode: intermittent" to this node.

[0050] Output and application: After the work of the day was over, the system generated the sixth version of the knowledge atlas of "Cantonese embroidery fish tail embroidery method". Ms. Li exported it as a set of structured data files. This atlas can not only be used for teaching to show the essence of each step to the apprentices, but more importantly, the exported action sequence and force parameters can be used to drive a high-precision computer embroidery machine to try to reproduce this shading effect with the charm of handwork, providing a data basis for the large-scale application of the skill.

[0051] Example 2: The master of guqin making explains the tuning of "Na Yin" The "Na Yin" of the guqin is a key component inside the qin belly that affects the tone, and its trimming process entirely depends on the master's experience and feel. In this example, Mr. Wang, a master of qin making, is tuning the Na Yin of a new qin.

[0052] Oral record: Mr. Wang gently tapped the soundboard, listened intently, and then, while meticulously trimming the soundboard with a chisel, he dictated: "Listen to this sound, it's still a bit 'hollow,' not 'smooth.' Now you can't be too rough; you have to use a round-edged chisel, 'licking' it little by little, yes, that's the word, like licking ice cream with your tongue, to feel the wood's feedback. The key is the middle of the soundboard; scrape off just a tiny bit, just a tiny bit... Okay! Listen again, the 'hollow' sound has subsided, and the 'resonance' has emerged. At this point, you must stop; if you scrape any further, it will be too much, and the sound will become 'thin.' This process requires patience and respect for the wood." 2. Feature Extraction and Atlas Construction: The system extracts professional timbre terms such as "empty," "moist," and "thin," and identifies the highly evocative action metaphor of "licking." When speaking of "little by little" and "a little bit," the system captures the extremely slow speech rate and cautious pauses, generating high-precision rhythm templates and emotional vectors of "cautious" and "compassionate." In constructing the atlas using the skill model, a series of action nodes are created, including "monitoring timbre," "judging the problem," "minor adjustments," "monitoring again," and "judging compliance," forming a closed loop. The model associates "using a round-mouthed shovel" as a tool node with the "minor adjustments" action; it associates "not making large cuts" as a taboo node with all actions; and it associates "the middle of the abdominal cavity" as a positional condition with adjustments via conditional edges. Most importantly, the model treats the timbre change of "the empty sound ending, revealing a rhyme" as the "result" node of the "minor adjustments" action, and through emotional weight edges, it closely associates the emotion of "respect for the wood" with the key decision point "stop at this point."

[0053] Verification and Correction: Mr. Wang reviewed the diagram and felt that the system's interpretation of the "licking" action as merely "scraping" was inaccurate. He verbally corrected it: "'Licking' involves not only downward scraping but also a slight upward 'carrying' force, making it a composite action." After receiving the instruction, the system split the "fine trimming" node into two sub-nodes: "downward scraping" and "upward carrying," and defined the force and timing relationships between them.

[0054] Output and Application: The knowledge graph generated during this tuning process profoundly reveals the decision-making logic of how a master luthier transforms auditory perception into subtle tactile movements. The exported graph data is not only a valuable archive of intangible cultural heritage, but its structured movements, forces, and feedback conditions also provide a highly valuable reference model for the research of high-precision woodworking robots.

[0055] Example 3: Aircraft engine test operator's verbal description of handling "abnormal vibration" This embodiment demonstrates the application of this method in a high-precision industrial setting. Test engineer Zhang is monitoring the extreme test data of a new type of aero-engine.

[0056] Oral record: Engineer Zhang stared intently at the vibration spectrum on the screen and suddenly said, "Attention, at measuring point 3, in the 67% speed range, high-frequency harmonics have appeared, and the amplitude exceeds the threshold. Now, follow my instructions: First, maintain the current fuel flow and do not push the throttle further. Second, pay close attention to the temperature change trend after the turbine and record it every 5 seconds. Third, slowly reduce the throttle, decreasing the speed to 60% at a rate of 2% per minute... Okay, now the high-frequency harmonics are starting to decay. Record this critical speed point. This vibration characteristic is similar to the previous case of unit 03, but the decay is faster, indicating that the adjustment is effective. The entire process must be smooth, avoiding any abrupt changes." 2. Feature Extraction and Graph Construction: When processing this spoken text, the system quickly identified a large number of professional parameters and command-like language. Prosodic features showed that Zhang's tone was decisive and clear when issuing instructions, and his emotional characteristics were marked as "highly focused." The graph constructed by the skill model started with "detecting anomalies" and generated action nodes such as "maintaining parameters," "monitoring temperature," and "reducing throttle." The model accurately used "67% RPM range" and "high-frequency harmonics" as conditional nodes, triggering a chain of actions through causal edges. "Cannot push the throttle further" was designated as a taboo node, and "2% speed per minute" was used as the force attribute (quantified parameter) of the "reducing throttle" action. Simultaneously, the model constructed an association edge pointing to the historical case library from the empirical judgment "similar to the situation of Unit 03," and associated the conclusive judgment "adjustment effective" as a result node with the action.

[0057] Verification and Correction: During the review, Engineer Zhang pointed out that the diagram lacked a logical explanation of "why monitor the turbine exhaust temperature." He added via voice: "Add explanation: Monitoring the turbine exhaust temperature is to determine whether the hot-end components are experiencing abnormal thermal loads due to vibration." The system then created a "Judgment Basis" node and connected it to the "Monitor Temperature" action through a causal edge.

[0058] Output and Application: The knowledge graph generated during this incident clearly encapsulates the emergency decision-making logic of a complex failure. The exported graph can be directly converted into a new rule in the aircraft health management system (PHM) or a standard operating procedure (SOP) document can be generated. More importantly, its structured action sequences and condition parameters provide a highly reliable data source for future intelligent decision-making in autopilot systems.

[0059] Example 4: A Zisha teapot artist demonstrates the shaping of the "patted body". In the making of Yixing teapots, "patting the body" is a crucial step that determines the teapot's overall character and style. Master Chen is demonstrating to his apprentice how to pat the body to create a full and rounded teapot shape.

[0060] Oral record: Master Chen shaped the clay slab into a cylindrical shape and rhythmically patted it with a paddle while saying, "The body of the pot needs to be 'straight,' with both bone and flesh. When you pat it, your wrist needs to be flexible, not just using the brute force of your arm. You see, first there's a solid 'thump,' which means the clay is being compacted; when the sound turns into a crisp 'snap,' it means the clay has awakened. At this point, you need to reduce the force, changing from a vertical pat to a curved 'pushing,' pushing the clay upwards, and the belly of the pot will bulge out. You need to have the final shape of the pot in mind, and let your hand move with your mind. The worst thing is to pat it crookedly, or make one side thicker than the other." 2. Feature Extraction and Atlas Construction: The system pays special attention to the "solid sound" and "crisp sound" that describe sound changes, and uses them as key auditory feedback signals. The "rhythm" of the patting is extracted as a rhythm template by the prosodic feature module. The artist's mention of "flexible wrists" and "forceful arms" contrasts and is analyzed by the model as two different force nodes, which are associated with different patting stages through conditional edges. In the atlas constructed by the skill model, the macro-action of "patting" is decomposed into two sub-action nodes: "pressing firmly" and "curving forward," with the switching condition being "the sound changes from solid to crisp." The model sets "patting crookedly" and "one side thick and one side thin" as taboo nodes, which are associated with the entire patting process. The key principle of "thinking about the final shape of the pot" is mapped by the model as a node in the top-level "Spiritual Aesthetics" layer, and is associated with the bottom-level "curving forward" action through emotional weight edges, reflecting the unity of form and meaning.

[0061] Verification and Correction: Master Chen pointed out that the description of the "arc rush" movement direction in the graph was not precise enough. He corrected it by saying, "'Rush' mainly refers to the direction from bottom to top, with a slight inward arc." The system then added the attribute parameters "Main direction: from bottom to top" and "Trajectory: slight inward arc" to the node.

[0062] Output and Application: This knowledge graph not only records actions, but more importantly, it captures the dynamic adjustment process based on auditory feedback and mental imagery—the core of traditionally intuitive experience. The exported data can provide quantifiable "feel" parameters for developing ceramic robots that can mimic hand-shaping.

[0063] Example 5: A neurosurgeon explains key dissection techniques during a simulated surgery. This example is applied to high-precision medical training. Dr. Zhao, the chief physician of neurosurgery, demonstrates a complex meningioma removal surgery on a surgical simulator.

[0064] Oral transcript: Dr. Zhao, manipulating the microsurgical instruments, explained to the trainees: “Now, the tumor is tightly adhered to the optic nerve. Using a micro-dissector, select this interface—yes, right here—there's an extremely thin blood vessel as a boundary marker. The movement must be not ‘tearing,’ but ‘pushing.’ The force must be constant, gently and continuously applying a force parallel to the nerve to the side, like pushing aside a precious silk garment… My right-hand suction device must keep up, constantly absorbing any oozing blood to maintain a clear field of vision, but the suction head must not directly contact the nerve. Okay, the interface is starting to separate… Stop! Here we see an abnormal small blood vessel; electrocoagulate it, using low power spot coagulation. Okay, continue. Throughout the process, my eyes must simultaneously focus on the dissection interface and the neurophysiological signals on the monitor; any fluctuation means risk.” 2. Feature Extraction and Atlas Construction: When processing this complex oral narrative, the system accurately identified the negative and positive expressions of key actions such as "tearing" and "pushing," and transformed "gently" and "continuously" into force control nodes. Prosodic features showed that Dr. Zhao's tone was urgent when saying "Stop!", marked as "highly tense," while his tone was relaxed when describing "pushing away the silk," marked as "cautious." "Cannot directly contact nerves" was clearly marked as a taboo. The skill model constructed a highly complex atlas: with "stripping adhesions" as the main node, it included multiple parallel or alternating sub-nodes such as "finding the interface," "constant force pushing," "assisted suction," "handling blood vessels," and "monitoring signals." The model established strict temporal edges (e.g., continuing after electrocoagulation) and conditional edges (e.g., stopping when nerve signals fluctuate). Emotional weight edges strongly correlated "tension" with the "handling abnormal blood vessels" node and "cautiousness" with the entire stripping process.

[0065] Verification and Correction: After reviewing the graph, Dr. Zhao added, "In the 'pushing' action, if increased resistance is encountered, my standard response strategy is to slightly change the angle and try again, rather than increasing the force." Based on this, the system created a new judgment node and branch action chain called "Responding to Obstacles," thus improving the decision-making logic.

[0066] Output and Application: The knowledge graph generated in this demonstration represents an in-depth analysis of expert surgical decision-making and bimanual coordination. The exported graph data can be used to generate extremely detailed standardized surgical procedure guidelines for training young doctors. Furthermore, its structured action sequences, force parameters, and contraindications can be directly converted into the control logic of a surgical robot-assisted system, ensuring that robot operation is performed within the framework of expert experience and enhancing the safety boundaries of the surgery.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically constructing a craft knowledge graph based on artisan spoken language, characterized in that, The method comprises the following steps: (1) Collecting the speech waveform data of the artisan's natural speech through a wearable device; (2) Extracting the five-dimensional features of the speech, including semantic text, prosodic features, emotional features, taboo markers, and rhythm templates, using a speech processing system; (3) Inputting the five-dimensional features into a skill model to build a dynamic knowledge graph in real time, with nodes including actions, tools, materials, intensity, taboos, and emotion types, and edge relationships including causality, time sequence, condition, and emotional weight; (4) Monitoring the quality of the graph through a health function and supporting real-time correction of the artisan's voice feedback; (5) Automatically iterating and evolving the graph every day, supporting version management and multi-format export. 2.The method of claim 1, wherein the method further comprises: The knowledge graph adopts a six-layer dynamic structure, gradually abstracting from the bottom action chain to the top "artistic aesthetics" layer, ensuring complete mapping of the skill logic. 3.The method of claim 1, wherein the method further comprises: The speech processing system supports real-time recognition of more than 100 dialects, including Cantonese, Wu dialect, Minnan dialect, Hakka, Sichuan dialect, and Shaanxi dialect. 4.The method of claim 1, wherein the method further comprises: The emotional features include eight artisan-exclusive emotional vectors, such as "delighted", "cautious", "nervous", and "loving", which are associated with graph nodes through weight edges. 5.The method of claim 1, wherein: Single-sentence spoken speech can automatically generate multiple knowledge nodes and edges, for example, from "here, be light, slow, and like caressing a baby" to generate 12 action nodes and 18 emotional edges. 6.The method of claim 1, wherein the method further comprises: The knowledge graph can be directly exported as G code or surgical path data executable by robots. 7.The method of claim 1, wherein the method further comprises: The method is applicable to non-heritage skill scenarios such as Cantonese embroidery, Suzhou embroidery, cloisonné, purple sand making, and guqin carving. 8.The method of claim 1, wherein the method further comprises: The method is applicable to industrial scenarios such as neurosurgery microsurgery, ultra-precision hand grinding, and aircraft engine testing. 9.The method of claim 1, wherein the method further comprises: The skill model is based on a state space prediction architecture and optimizes feature compatibility through self-supervised learning. 10.The method of claim 1, wherein the method further comprises: The closed-loop correction module supports real-time response to voice commands, with a correction delay of less than 2 seconds.

Images