A white-box AI cognitive reasoning method and system based on concept tensor

Abstract

The application provides a white-box AI cognitive reasoning method and system based on a concept tensor, which is executed based on a three-dimensional concept tensor, the three-dimensional concept tensor comprising a concept dimension, an original feature dimension and a knowledge hierarchy dimension adopting a tree-shaped hierarchical structure, following a point-line-surface-body four-order dimension progressive structure, and being capable of extracting common features from feature information of low-level concepts to generate high-level concepts through an upward abstraction mechanism, and obtaining feature information of corresponding low-level concepts from high-level concepts through a downward search mechanism, realizing bidirectional conversion of geometric features and topological structures, and forming unified knowledge representation; the method comprises the following steps: in response to a reasoning request, extracting feature representation of a to-be-reasoned concept on a specified knowledge hierarchy; performing explainable matching analysis on the feature representation to obtain a matching attribute set comprising specific attribute pairs leading to semantic relationships between concepts; and performing deterministic state transition according to the matching attribute set and a preset meta-rule to generate a traceable reasoning result.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a white-box AI cognitive reasoning method and system based on concept tensors, which is particularly suitable for safety early warning and fault root cause tracing in industrial scenarios. Background Technology

[0002] In the long-term practical application of industrial AI, existing AI technologies have consistently faced a fundamental contradiction. On the one hand, data-driven models, represented by deep neural networks and large language models, possess powerful perception and generalization capabilities, automatically learning complex features and patterns from massive amounts of labeled or unlabeled data, demonstrating excellent performance in tasks such as image recognition and natural language processing. However, these models are essentially black-box systems; their internal weight distribution and multi-layered nonlinear transformations make the final decision-making logic difficult for humans to understand. Even with post-analysis using interpretability techniques such as attention mechanisms and gradient visualization, only local, statistically significant correlations are often provided, failing to reconstruct a clear, definite, and causally consistent reasoning path. In safety-sensitive scenarios such as chemical process control, power system fault diagnosis, and rail transit scheduling, this lack of interpretability poses a fatal risk: operators cannot confirm the reliability of the model's judgment basis, and it is even more difficult to trace which input features and which step of reasoning led to the erroneous decision when an anomaly occurs, thus hindering its actual deployment in critical tasks.

[0003] On the other hand, knowledge systems based on symbolic logic, such as knowledge graphs, expert systems, and logic programming frameworks, inherently possess interpretability and reasoning transparency. By explicitly defining entities, attributes, relationships, and logical rules, they can provide step-by-step derivation processes and support the traceability and verification of conclusions. These methods have unique advantages in industrial environments requiring strict compliance, auditing, or causal analysis. However, the construction of symbolic systems is highly dependent on manual intervention: the schema design of knowledge graphs, the annotation of entities and relationships, and the writing of logical rules all require significant participation from domain experts, and maintenance costs rise sharply with changes in business. More importantly, traditional symbolic systems lack the ability to automatically extract new concepts and discover new relationships from unstructured data (such as text, sensor time-series signals, and images), making it difficult to adapt to the needs of rapid business iteration or dynamic environmental changes; their knowledge bases are often prone to lag or even obsolescence.

[0004] In recent years, academia and industry have attempted to integrate the advantages of both approaches through neural-symbolic computation, such as using neural networks for symbolic knowledge extraction and representation, or injecting symbolic constraints into the training process of neural networks. However, these attempts still face many challenges: the semantic gap between neural representation and symbolic reasoning has not been fully bridged, the mechanisms for dynamic knowledge updates and continuous learning are not mature enough, and the overall system often struggles to achieve a balance between flexibility, interpretability, and automation. Therefore, in industrial AI scenarios that require both "autonomous learning capability" and "full interpretability," a unified, efficient, and practical technical framework is still lacking. This situation has prompted researchers to further explore new paradigms for knowledge representation and reasoning to fundamentally alleviate the aforementioned contradictions. Summary of the Invention

[0005] To address the shortcomings and deficiencies of existing technologies, this invention provides a white-box AI cognitive reasoning method and system based on concept tensors. This invention employs a three-dimensional concept tensor as a unified knowledge representation carrier, strictly adhering to the four-dimensional progression principle of point-line-surface-volume to construct a complete knowledge dimensionality enhancement system. It stores various concepts in the system through the concept dimension and supports dynamic expansion; quantifies the atomic attributes of concepts through the original feature dimension and supports dynamic expansion; organizes the generalization and specialization relationships between concepts through a tree-like hierarchical knowledge hierarchy dimension; and achieves strict completeness of the non-full-rank tensor space by reserving illusory spaces. This realizes the bidirectional conversion and organic integration of geometric features and topological structures, breaking the industry pain point of the separation of geometric and topological representations in existing technologies. Building upon this foundation, this invention employs a feature resonance mechanism to perform interpretable concept similarity calculations. This not only yields the degree of similarity between concepts but also outputs specific matching attribute pairs that lead to semantic relationships between concepts, sorted by contribution. Through meta-rule-driven deterministic state transitions and reverse chain derivations, a complete and traceable causal reasoning chain from the outcome state to the root cause state is generated. Simultaneously, a hybrid mechanism combining weak-layer buffer unsupervised incremental learning and single-definition supervised learning enables the automatic accumulation of new concepts and the rapid injection of artificial knowledge. The reasoning process of this invention is completely transparent and traceable. All algorithms are implemented based on linear algebra and deterministic logic, allowing for efficient operation on conventional computing devices without complex deep learning training. Furthermore, it supports incremental knowledge updates without retraining the entire model, making it particularly suitable for edge computing scenarios with high requirements for interpretability, real-time performance, and reliability, such as industrial safety early warning and fault root cause tracing.

[0006] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0007] A white-box AI cognitive reasoning method based on concept tensors is disclosed. This method executes based on a three-dimensional concept tensor, which includes a concept dimension, a raw feature dimension, and a knowledge hierarchy dimension employing a tree-like hierarchical structure. It follows a four-order dimensional progression structure of point-line-plane-volume and can extract common features from the feature information of lower-level concepts through an upward abstraction mechanism to generate higher-level concepts, and retrieve the feature information of corresponding lower-level concepts from higher-level concepts through a downward retrieval mechanism. This achieves bidirectional conversion between geometric features and topological structures, forming a unified knowledge representation. The method includes:

[0008] In response to a reasoning request, extract the feature representation of the concept to be reasoned at the specified knowledge level;

[0009] Perform interpretable matching analysis on the feature representation to obtain a set of matching attributes containing specific attribute pairs that lead to semantic relationships between concepts;

[0010] Deterministic state transitions are performed based on the set of matching attributes and preset meta-rules to generate traceable reasoning results;

[0011] The method supports incrementally adding new concepts and corresponding feature information to the three-dimensional concept tensor.

[0012] Furthermore, the interpretable matching analysis is implemented through a feature resonance mechanism, which calculates the similarity between the feature representations of the concepts to be reasoned, filters attribute pairs that meet the preset feature threshold, calculates the contribution of each matching attribute pair based on the attribute value, and sorts the matching results based on the contribution. After each entity definition or feature state change, the system automatically triggers a global resonance refresh to update the entity resonance relationship at the corresponding level.

[0013] Furthermore, the preset meta-rules are the opposing relationships of qualitative labels, including predefined core meta-rules and derived meta-rules automatically generated through a recursive guessing mechanism; the generation of traceable reasoning results specifically involves recursively tracing from the result state to the root cause state through reverse chain deduction to generate a complete causal reasoning chain.

[0014] Furthermore, the incremental addition of new concepts and corresponding feature information to the three-dimensional concept tensor is achieved by combining unsupervised incremental learning in a weak-layer buffer with supervised learning based on a single definition.

[0015] Furthermore, the knowledge hierarchy dimension adopts a bottom-up initialization method, and the hierarchy affiliation of the new concept is determined by calculating the cosine similarity between the new concept and all existing concepts, with each concept corresponding to a unique parent node; the dimensionality upgrade sequence terminal of the knowledge hierarchy dimension is set with reserved dummy positions to achieve complete regularization of the dimensional space.

[0016] Furthermore, the three-dimensional concept tensor is stored using a sparse matrix combined with a hash table, storing only non-zero data.

[0017] Furthermore, the method is applied to an industrial safety early warning system to perform conflict detection and fault root cause tracing based on multi-sensor data.

[0018] And, a white-box AI cognitive reasoning system based on concept tensors, comprising:

[0019] The tensor construction module is used to construct a three-dimensional concept tensor. The three-dimensional concept tensor includes a concept dimension, a primitive feature dimension, and a knowledge hierarchy dimension with a tree-like hierarchical structure, and follows a four-order dimensional progression structure of point-line-plane-volume.

[0020] The interactive conversion module is used to execute upward abstraction and downward retrieval mechanisms, realizing bidirectional conversion between geometric features and topological structures to form a unified knowledge representation.

[0021] The request processing module is used to respond to reasoning requests and extract the feature representation of the concept to be reasoned at a specified knowledge level.

[0022] The matching analysis module is used to perform interpretable matching analysis on the feature representation to obtain a set of matching attributes containing specific attribute pairs;

[0023] The rule reasoning module is used to perform deterministic state transitions based on the set of matching attributes and preset meta-rules, and generate traceable reasoning results.

[0024] The incremental learning module is used to incrementally add new concepts and their corresponding feature information to the 3D concept tensor.

[0025] Furthermore, the matching analysis module performs interpretable matching analysis through a feature resonance mechanism and supports global resonance refresh; the rule reasoning module generates causal reasoning chains through reverse chain derivation and can automatically generate derived meta-rules through a recursive word guessing mechanism; the incremental learning module integrates weak-layer buffer learning units and single-definition learning units; and the tensor construction module achieves dimensional space completeness by reserving illusory spaces.

[0026] Furthermore, the system is deployed on industrial edge computing devices to perform industrial security early warning and fault root cause tracing.

[0027] And a computer device including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0028] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0029] Compared to existing technologies, this invention and its preferred embodiment unify the geometric features and topological hierarchy of concepts within a three-dimensional tensor representation, fundamentally integrating data-driven autonomous learning capabilities with the complete interpretability of symbolic systems. This allows the system to automatically summarize new concepts from data streams and provide clear, definite, and traceable reasoning paths for each step. Its interpretable matching analysis mechanism not only outputs similarity values ​​but also explicitly reveals the specific attribute pairs that lead to similarities or conflicts between concepts, achieving complete transparency in the decision-making process. The meta-rule-based reasoning framework requires only a small number of manually defined core rules to automatically derive a large number of specific reasoning rules and supports tracing the complete causal chain backward from the result, significantly reducing the cost of knowledge construction and maintenance. Simultaneously, by integrating unsupervised incremental learning and imperative supervised learning, the system flexibly adapts to different needs, from automated data mining to rapid injection of manual knowledge, ensuring the dynamic expansion and updating of the knowledge base.

[0030] The four-order structure of point-line-plane-volume achieves rigorous formalization of knowledge representation, further enhancing the determinism and white-box nature of the system; the reserved dummy space mechanism completes the tensor space, improving the system's adaptability to irregular data; and the global resonance refresh mechanism ensures the real-time self-consistency of the system state, avoiding inference errors caused by delayed consistency.

[0031] Furthermore, thanks to engineering optimizations such as sparse storage, the entire solution has low computational resource requirements and can be deployed on industrial edge devices. It can be directly applied to scenarios such as industrial monitoring, fault diagnosis and early warning that have stringent requirements for security and interpretability, effectively solving the core contradiction between the untrustworthiness of black-box models and the rigidity of symbolic systems. Attached Figure Description

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0033] Figure 1 This is a schematic diagram of the overall implementation framework of the embodiment of the present invention. Detailed Implementation

[0034] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0035] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] To address the industry pain point that existing technologies have consistently failed to simultaneously meet the two key requirements of autonomous learning capability and full interpretability, this invention proposes a novel knowledge representation and reasoning framework that can automatically learn concepts like a black box model, and ensure the complete transparency and traceability of the decision-making process like a symbolic system.

[0038] After comparing and verifying multiple technical approaches, this invention finally determined the overall architecture centered on the "concept tensor". This architecture unifies geometric features and topological structures in the same three-dimensional tensor, using the geometric dimension to quantify the attributes of concepts and the topological dimension to organize the hierarchical relationships of concepts. The two interact through a specific mechanism to jointly constitute a complete cognitive reasoning system.

[0039] like Figure 1 As shown, the following is a detailed demonstration and description of the implementation of the solution of the present invention:

[0040] (1) Concept Tensor Data Structure

[0041] This invention constructs a multilayer three-dimensional binary tensor Each of the three dimensions has its own function and is interconnected, and the whole follows the principle of a four-dimensional progression from point to line to plane to volume:

[0042] C-dimensional is the concept dimension, used to store all concepts in the system (such as water, fire, and temperature sensors), and supports dynamic expansion, allowing new concept nodes to be added at any time;

[0043] The F dimension is the original feature dimension, used to store the atomic properties of concepts (such as liquid, cooling, high temperature), and its values ​​are strictly limited to {0,1}. It also supports dynamic expansion, and new properties can be added according to business needs.

[0044] The L dimension represents the knowledge hierarchy, employing a tree-like hierarchical structure. The root node (the most abstract). The leaf node (the most specific) corresponds to a coarser cognitive granularity, while a larger value corresponds to a finer cognitive granularity.

[0045] The four-dimensional progressive structure of point-line-plane-volume represents a complete dimensional upgrade process of knowledge from atomic storage to a three-dimensional associative space, with each dimension serving as a building block for the next dimension.

[0046] Zero-dimensional point: The feature vector of a single entity in a specific context, which is the smallest knowledge storage and operation unit of the system;

[0047] One-dimensional line: The resonance relationship between two entities in the same layer and context, defined by cosine similarity and matching attribute set;

[0048] Two-dimensional surface: a resonance network formed by all resonance lines in the same layer and context, covering the relationships between all activated entities in this layer;

[0049] Three-dimensional volume: A complete three-dimensional associated space formed by stacking two-dimensional surfaces of all levels through multi-branch tree parent-child paths, which is the unified storage form of all knowledge in the system.

[0050] In practical implementation, this invention uses a sparse matrix combined with a hash table to store the tensor, storing only non-zero values. This not only saves a significant amount of memory space but also ensures the system's operating efficiency.

[0051] The hierarchical dimensional ascending chain of the cognitive architecture of this invention is a strictly unidirectionally increasing sequence of dimensional mappings. Each dimensional structure generates a unitized ascending ray along the ascending direction. The ray terminates with a unique reserved vacancy, serving as a standard interface for cross-dimensional mapping and a spatial completeness constraint unit. The reserved vacancy only performs the functions of dimensional closure, dimensional ascending guidance, and spatial regularization; it does not carry semantic information, participate in feature resonance calculation, or similarity iteration, and maintains zero-value constancy in form. This allows a non-full-rank, non-cubic information point set to achieve strict completeness in its dimensional space without introducing additional semantics or compromising system determinism and white-box nature.

[0052] Regarding the construction of the L-dimensional hierarchy, after repeated testing, a bottom-up initialization strategy and a dynamic hierarchy determination method were ultimately adopted. Initially, the system contains only one root node (L=0). When a new concept is added, the system automatically calculates the cosine similarity between that concept and all existing concepts, finding the most similar node as its parent node. If the similarity is higher than a preset threshold (0.5 was found to be a reasonable value in experiments), the concept becomes a child node of the parent node; if the similarity is lower than the threshold, a new first-level node is created as the parent node of the concept. Each concept has only one unique position in the L-dimensional hierarchy, and the parent-child relationship naturally defines the generalization and specialization relationships of concepts.

[0053] For example, when the system inputs the three specific concepts "water, ice, and steam" in sequence, it will automatically cluster them and find that they share the feature of "material form", thus creating the parent node "water substance" at the l=1 level; it will continue to cluster and find that "water substance" and "fire substance" share the more abstract feature of "matter", thus creating the root node "matter" at the l=0 level.

[0054] The most significant innovation of this solution lies in its pioneering achievement of a unified representation of geometry and topology:

[0055] F-dimensionality, as a geometric dimension, transforms each concept into a point in geometric space. Cosine similarity can be used to accurately calculate the degree of similarity between two concepts.

[0056] The L-dimensional dimension, as a topological dimension, organizes all concepts into a tree-like connected graph, and the hierarchical relationship between concepts can be clearly traced through parent-child relationships.

[0057] The two interact through two mechanisms:

[0058] Upward abstraction (induction): Extracting common features from the feature vectors of lower-level concrete concepts to generate feature vectors of higher-level abstract concepts;

[0059] Downward retrieval (deduction): Starting from a high-level abstract concept, automatically retrieve the specific concepts of all child nodes based on parent-child relationships.

[0060] This unified representation completely breaks the limitation of the separation of geometry and topology in existing technologies, and realizes a complete cognitive ability that can both calculate similarity and trace logical relationships.

[0061] (2) Characteristic resonance mechanism

[0062] Traditional similarity calculation methods (such as word vectors and knowledge graph embeddings) can only output a single scalar value, failing to explain "why these two concepts are similar." To address this issue, this invention designs a feature resonance mechanism that not only calculates the similarity value but also outputs specific matching attribute pairs.

[0063] Given two concepts A and B, the system performs the following steps at the specified knowledge level dimension l:

[0064] Extract the feature vectors a=T[A,:,l] and b=T[B,:,l] of two concepts at this level from the concept tensor;

[0065] Calculate the cosine similarity between two vectors: sim = (a·b) / (||a||||b||).

[0066] Iterate through all feature dimensions, if a i >θ and b i If the threshold value is greater than θ (where θ is a preset feature activation threshold), then record the similar matching pairs (f). i ,f i ) to set P, and simultaneously record the values ​​of two vectors (a i ,b i The contribution of a similar matching pair is defined as (a i +bi ) / 2 is used to sort the similarity matching results. In addition, the system predefines the mapping relationship between features and meta-rule keywords, traversing all feature dimension combinations, if a i >θ and b j If >θ, and the meta-rule keywords mapped by features i and j have an opposing relationship in the meta-rule table, then record the conflict matching pair (f i ,f j (i) to set P. Usually, i ≠ j in this case.

[0067] The final output is the similarity score sim and the set of matching attribute pairs P.

[0068] This mechanism makes similarity calculations fully interpretable, allowing users to clearly see which attributes make two concepts similar.

[0069] This invention also includes a global resonance refresh mechanism as a closed-loop operation to complement the feature resonance mechanism: after each entity definition operation or feature state transition operation is completed, the system automatically triggers a global resonance refresh, locates the changed level, recalculates the cosine similarity and matching attribute set of all entity pairs in each context within that level, updates the weights and matching attributes of the corresponding edges in the resonance network, and ensures that the similarity query results at any time are the latest real-time values, ensuring that the system is always in a completely self-consistent state.

[0070] (3) Weak layer buffer

[0071] To achieve unsupervised incremental learning, this invention designs a weak-layer buffer for temporarily storing unidentified unknown terms in the system. The buffer structure is W={(w,c(w))|w Let T,c(w)∈N}, where w is an unknown term and c(w) is the number of times the term appears.

[0072] In experiments, this invention discovered that when an unknown term appears more than three times, it usually indicates that it is a meaningful concept. Therefore, this invention sets the promotion threshold to three times; when c(w) ≥ 3, the system automatically promotes the term to a formal concept without manual intervention. This mechanism enables the system to automatically learn new concepts from daily interactions, significantly reducing the maintenance cost of the knowledge base.

[0073] (4) Single-definition learning

[0074] In addition to unsupervised incremental learning, the system also supports deterministic concept definition via natural language commands. When receiving a command in the form of "Define X is Y", the system will perform the following operations:

[0075] If Y is the original feature, then set T[X][Y][l]=1 and perform normalization;

[0076] If Y is an existing concept, then perform vector superposition T[X]=Norm(T[X]+T[Y]);

[0077] If Y is a new feature, then create a new dimension in dimension F and perform the above operations again.

[0078] This one-time definition method is very much in line with human learning habits, allowing users to quickly impart new knowledge to the system using natural language.

[0079] (5) Chain derivation and state transition mechanism

[0080] The reasoning capability of this scheme is based on deterministic state transitions and reverse chain derivation driven by meta-rules. This invention defines the basis of all reasoning as a set of meta-rules, which are the opposite relationships of qualitative labels, such as "high" and "low", "rise" and "fall", "positive" and "negative", etc.

[0081] The meta-rule table is stored independently in the system as a dictionary, supporting the dynamic addition of new meta-rules. It is obtained in two ways: first, by manually predefining core meta-rules, which form the axiomatic basis of the system's reasoning; and second, by automatically generating derived meta-rules through a recursive word-guessing mechanism. The recursive word-guessing mechanism maintains a "root-derived word" mapping table. When a user inputs a root word, the system automatically generates all related derived words and applies existing meta-rules to these derived words. In practical implementation, this invention limits the recursion depth to two levels, thus covering most commonly used derived words while avoiding excessive computation.

[0082] The complete logic for conflict detection and state transition is as follows:

[0083] The characteristic resonance mechanism outputs a set of matching pairs P;

[0084] For each matching pair (f) a ,f b Extract its keywords (such as "cooling" in "cooling down" and "high" in "high temperature");

[0085] If the meta-rule table is queried and M["decrease"] = "high", then it is determined that there is a conflict between these two attributes.

[0086] Trigger a state transition and reset conflicting attributes to zero;

[0087] The system continues to check other attributes until there are no conflicts.

[0088] Building upon this foundation, the system supports reverse chain derivation. Starting from the resulting state E, it searches for all state transition rules that can lead to E, obtaining a set of preceding states. This process is then recursively executed until the root cause state, which cannot be further decomposed, is found. Finally, the complete causal chain E←S1←S2←…←S is output. n .

[0089] Just as Euclidean geometry is built upon five axioms, the entire reasoning system of this scheme is built upon a small number of core meta-rules. The core value of this scheme lies not in "automatically discovering conflicts," but in "defining first-dimensional rules, with the system automatically generating a massive amount of reasoning results." For example, by simply defining the conflict between "decrease" and "high," the system can automatically deduce the conflict between "cooling down" and "high temperature," then the conflict between "water" and "fire," and finally, "water extinguishes fire." This is fundamentally different from traditional symbolic systems that require manual definition of every specific conflict rule, significantly reducing the maintenance cost of the knowledge base.

[0090] (6) Three-dimensional spatial state reasoning mechanism

[0091] In a preferred embodiment of the present invention, the system adds a three-dimensional spatial state reasoning module decoupled from the concept tensor kernel, used to synthesize multi-source input information into a deterministic spatial state and perform decision reasoning. The module constructs a three-dimensional Cartesian coordinate system reasoning space, with each of the three dimensions corresponding to an independent evaluation axis. The origin of the coordinate system is a zero-neutral point, the positive half-axis represents positive evaluation, and the negative half-axis represents negative evaluation. The module maps the concept to be evaluated to coordinate points in the reasoning space, receives structured messages from external input and parses them into force vectors acting on the coordinate points, calculates the resultant force of all historical force vectors and updates the coordinate positions, and finally outputs a deterministic reasoning result based on the spatial region where the coordinate point is located.

[0092] The force vector and its component force values ​​for each message are recorded in the system log. When a historical message is determined to be false or erroneous, the system can delete the force vector record and recalculate the resultant force, thus achieving reversible correction of the reasoning results.

[0093] Compared with existing technologies, this solution has the following eight core innovations. These innovations are interconnected and mutually supportive, together forming a complete white-box AI cognitive reasoning system:

[0094] First, this approach achieves a unified knowledge representation of geometry and topology for the first time. Existing technologies either focus only on geometric representations (such as word vectors and knowledge graph embeddings), which can only calculate similarity but cannot trace logical relationships; or they focus only on topological representations (such as knowledge graphs), which can only represent logical relationships but cannot accurately quantify attributes. This scheme simultaneously encodes geometric features (F-dimensional) and topological levels (L-dimensional) in the same three-dimensional tensor, and achieves organic interaction between the two through upward abstraction and downward retrieval mechanisms. This is the fundamental feature that distinguishes this scheme from all existing technologies.

[0095] Second, an L-dimensional design based on a tree-like hierarchical structure is proposed. In existing technologies, the third dimension of tensors is usually used to store different relation types or time steps, and has never been used to represent the abstract hierarchy of concepts. The L-dimensional structure of this scheme adopts a tree-like structure, which clearly depicts the generalization and specialization relationships between concepts, providing a solid foundation for logical reasoning and causal tracing.

[0096] Third, an interpretable feature resonance similarity calculation mechanism was invented. Traditional methods can only output a single similarity value, while the feature resonance mechanism of this scheme simultaneously outputs a specific set of matching attribute pairs, making the similarity calculation completely interpretable, and allowing users to clearly see the basis of the reasoning.

[0097] Fourth, an automated reasoning paradigm based on meta-rules is proposed. Traditional symbolic systems require manual definition of all specific rules, resulting in extremely high maintenance costs. This solution only requires defining a small number of core meta-rules, and the system can automatically derive a massive number of conflict detection and state transition rules, significantly reducing the construction and maintenance costs of the knowledge base.

[0098] Fifth, a hybrid learning mechanism combining weak-layer buffering and single-definition was designed. Weak-layer buffering enables unsupervised incremental learning, while single-definition enables deterministic supervised learning. The combination of the two allows the system to automatically learn new concepts from unstructured data and quickly receive knowledge imparted by humans.

[0099] Sixth, a four-dimensional progressive knowledge representation system of point-line-plane-volume was proposed, which clarified the complete construction hierarchy from atomic storage unit to three-dimensional associated space, and realized the unity of structure and formalization of knowledge representation;

[0100] Seventh, a dimension completion mechanism with reserved illusory spaces was designed, which solves the regularization problem of non-full-rank tensor space without destroying the determinism and white-box nature of the system.

[0101] Eighth, a global resonance refresh mechanism was invented to ensure that the similarity query results at any time in the system are the latest values ​​in real time, thus achieving complete self-consistency of the system state.

[0102] Based on the above innovations, this solution achieves the following significant technical effects:

[0103] The reasoning process is fully traceable: users can check the matching attributes and state transitions at each step, clearly see the entire reasoning process, and completely solve the problem of the unexplainable nature of black-box AI.

[0104] Low-power operation: All algorithms are based on linear algebra and deterministic logic, requiring no complex deep learning training, and can be deployed and run on regular CPUs, making them particularly suitable for edge computing scenarios;

[0105] Incremental concept learning: New concepts can be automatically added to the knowledge base through weak-layer buffering or single definition without retraining the entire model;

[0106] High-accuracy interpretable reasoning: In tasks such as conflict detection, the accuracy of this scheme is significantly higher than that of traditional word vector methods, and it can also output interpretable reasoning.

[0107] The following describes the specific implementation and application of the present invention through several more specific embodiments:

[0108] Example 1: Extinguishing a fire with water

[0109] This is a basic reasoning example that demonstrates the fundamental processes of feature resonance, conflict detection, and state transition.

[0110] Initial settings:

[0111] The feature vector of the concept "water" at level l=2 contains the attribute "cooling down", which is associated with the meta-rule "decrease".

[0112] The feature vector of the concept "fire" at level l=2 contains the attribute of "high temperature", which is associated with the meta-rule "high".

[0113] The meta-rule table predefines conflict relationships ("decrease", "high").

[0114] Input: "Water is poured on fire"

[0115] Execution steps:

[0116] The system uses natural language processing tools to identify the two core concepts "water" and "fire" in the input;

[0117] Extract the feature vectors of the two concepts at level l=2;

[0118] Perform feature resonance calculations to obtain similarity values ​​and identify matching pairs ("cooling down", "high temperature");

[0119] Extract the keywords "decrease" and "high" from the matching pair, query the meta-rule table, and determine that there is a conflict;

[0120] Trigger a state transition, resetting the "high temperature" attribute of "fire" to zero;

[0121] Output the inference result: "fire extinguished".

[0122] In actual testing, this simple example proved to be a good validation of the core logic of this approach. While traditional word vector methods can calculate the similarity between "water" and "fire," they cannot explain why they interact, nor can they automatically deduce the result of "fire going out."

[0123] Example 2: Weak Layer Learning

[0124] This example demonstrates how a system can achieve unsupervised incremental learning through a weak-layer buffer.

[0125] Initial state: The concept of "cyber" does not exist in the concept tensor.

[0126] Execution process:

[0127] The first time text containing "cyber" is input, the system adds it to the weak buffer and counts it as c(cyber) = 1.

[0128] The second input contains text with the word "cyber", and the buffer count c(cyber) = 2;

[0129] The third input contains text with the word "cyber", and the buffer count c(cyber) = 3, reaching the promotion threshold;

[0130] The system automatically creates a new concept "cyber" in the concept tensor, initializing its feature vector to an all-zero vector. Users can later add attribute features to it using the "Define cyber is..." command.

[0131] This mechanism allows the system to automatically accumulate new concepts during daily use, without the need for manual addition.

[0132] Example 3: Single Definition

[0133] This example demonstrates how to quickly impart new knowledge to a system using natural language commands.

[0134] Enter: "Define Administrator is manager"

[0135] Execution steps:

[0136] The system parses natural language commands and identifies the defined concept "Administrator" and the defining concept "manager".

[0137] Check if "manager" is an existing concept. If it is confirmed to exist, extract its feature vector.

[0138] Create a new concept "Administrator" and copy the feature vector of "manager" to "Administrator";

[0139] Perform L2 norm normalization on the eigenvector of "Administrator";

[0140] Output: "Administrator defined successfully".

[0141] Example 4: Quantitative Comparison Experiment

[0142] To verify the performance of this scheme in the conflict detection task, this invention selected 10 pairs of typical concepts and compared the performance of this scheme with Word2Vec.

[0143] Experimental environment: Intel Core i5-10400 CPU, 16GB RAM, Windows 10 operating system. Experimental parameters: The similarity threshold for Word2Vec was set to 0.5; values ​​greater than 0.5 were considered similar, and values ​​less than 0.2 were considered irrelevant. The feature threshold θ for this scheme was set to 0.5.

[0144] The experimental results are shown in the table below:

[0145]

[0146] Experimental Results Analysis: Our proposed solution achieved 100% accuracy (10 / 10) on the test set, with an average inference time of 0.2ms per result and memory usage of less than 100MB. In contrast, Word2Vec's accuracy was only 70% (7 / 10), with an average inference time of 1.5ms per result. More importantly, our solution not only accurately determines the results but also outputs specific matching attribute pairs, explaining the basis for the judgment—something black-box methods like Word2Vec cannot do.

[0147] Example 5: Industrial Scenario Example (Chemical Plant Fire Early Warning System)

[0148] This example demonstrates the application of this solution in a real-world industrial scenario, showcasing its advantages of low computational power and traceable causal reasoning.

[0149] Scenario Description: This solution was deployed as the inference engine for a fire early warning system in a chemical plant warehouse. Traditional fire early warning systems can only trigger alarms based on the threshold of a single sensor, and cannot integrate information from multiple sensors or explain the cause of the alarm, making them prone to false alarms and missed alarms.

[0150] Initial knowledge base:

[0151] The concept of a "temperature sensor" includes a feature vector containing "monitored temperature" and "output value".

[0152] The concept of a "smoke sensor" includes a feature vector containing "monitored particulate matter" and "output concentration".

[0153] Meta-rule tables: ("Elevation", "Danger"), ("Danger", "Alarm").

[0154] Input sequence:

[0155] The temperature sensor outputs a value of 85°C. The system detects that this value is significantly higher than the historical average, determines it as "temperature rise", and adds it as a temporary attribute to the feature vector of the "temperature sensor".

[0156] When the smoke sensor outputs a high concentration value, the system determines that "smoke concentration has increased" and adds it as a temporary attribute to the feature vector of the "smoke sensor".

[0157] Chain derivation process:

[0158] The performance was evaluated using a characteristic resonance mechanism, which detected the simultaneous occurrence of both "increased temperature" and "increased smoke concentration".

[0159] Query the meta-rule table and trigger the ("Elevated", "Dangerous") rule, changing the system state from "Safe" to "Dangerous".

[0160] Continue querying the meta-rule table, trigger the ("Danger", "Alarm") rule, and the system status changes from "Danger" to "Alarm";

[0161] Output a complete and traceable causal chain: Alarm ← Dangerous state ← Temperature rise ∧ Smoke concentration rise.

[0162] Example 6: Verification of the fourth-order structure and dimensionality completeness of the concept tensor

[0163] This example demonstrates the complete execution flow of a four-order structure of point-line-plane-volume, reserved vacancy space, and global resonance refresh.

[0164] Initial settings:

[0165] The initial root node level l=0 contains the concept "matter", with a reserved vacant space in a constant zero value state; the l=1 level contains two sub-concepts "water matter" and "fire matter", which are respectively bound to the attributes "cooling" and "high temperature"; the l=2 level contains three sub-concepts "water", "ice" and "steam", all of which inherit the characteristics of the parent node "water matter".

[0166] Execution steps:

[0167] The system performs an entity definition operation, adding the concept "dry ice" at level l=2, binding the "cooling" attribute, and setting the parent node to "water substance";

[0168] After the operation is completed, a global resonance refresh is automatically triggered, recalculating the resonance relationship for all entity pairs at level l=2, generating similarity values ​​and matching attribute sets between "dry ice" and "water", "ice", and "steam", and updating the two-dimensional surface of the resonance network at this level.

[0169] The system detected that the number of entities at level l=2 is not cubic. It completes the dimensional space regularization by reserving illusory spaces to maintain the strict completeness of the tensor structure. The reserved illusory spaces remain at zero throughout the process and do not participate in any resonance calculations.

[0170] Output the updated complete 3D volume knowledge structure, as well as interpretable resonance results of "dry ice" and other concepts at the same level.

[0171] This example verifies the complete dimensionality upgrade logic of the fourth-order dimensional structure, the self-consistency guarantee of global resonance refresh, and the ability of reserved illusory spaces to complete non-full-rank tensors. The entire operation is traceable and no additional semantics are introduced, maintaining the white-box nature and determinism of the system.

[0172] In summary, this solution integrates information from multiple sensors for inference, significantly reducing the false alarm rate. Furthermore, it outputs a complete causal chain, allowing maintenance personnel to clearly see the cause of the alarm and quickly locate and resolve the problem. In addition, the low computational cost of this solution allows it to be deployed directly on edge computing devices in the warehouse, without relying on cloud GPUs, ensuring the system's real-time performance and reliability.

[0173] Based on the same inventive concept, this invention also provides a computer device, comprising: one or more processors, and a memory for storing one or more computer programs; the programs include program instructions, and the processor executes the program instructions stored in the memory. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, used to implement one or more instructions, specifically for loading and executing one or more instructions stored in a computer storage medium to implement the above-described method.

[0174] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium storing a computer program, which, when executed by a processor, performs the above-described method. This storage medium can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0175] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0176] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

[0177] This invention is not limited to the preferred embodiment described above. Anyone inspired by this invention can derive various other forms of white-box AI cognitive reasoning methods and systems based on concept tensors. All equivalent variations and modifications made within the scope of the claims of this invention should be included in the scope of this invention.

Claims

1. A white-box AI cognitive reasoning method based on concept tensors, characterized in that, The method is based on a three-dimensional concept tensor, which includes a concept dimension, a primitive feature dimension, and a knowledge hierarchy dimension with a tree-like hierarchical structure. It follows a four-dimensional progressive structure of point-line-plane-volume and can extract common features from the feature information of lower-level concepts through an upward abstraction mechanism to generate higher-level concepts, and retrieve the feature information of corresponding lower-level concepts from higher-level concepts through a downward retrieval mechanism. This achieves a bidirectional conversion between geometric features and topological structures, forming a unified knowledge representation. The method includes: In response to a reasoning request, extract the feature representation of the concept to be reasoned at the specified knowledge level; Perform interpretable matching analysis on the feature representation to obtain a set of matching attributes containing specific attribute pairs that lead to semantic relationships between concepts; Deterministic state transitions are performed based on the set of matching attributes and preset meta-rules to generate traceable reasoning results; The method supports incrementally adding new concepts and corresponding feature information to the three-dimensional concept tensor.

2. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The interpretable matching analysis is implemented through a feature resonance mechanism, which calculates the similarity between the feature representations of the concepts to be reasoned, filters attribute pairs that meet the preset feature threshold, calculates the contribution of each matching attribute pair based on the attribute value, and sorts the matching results based on the contribution. After each entity definition or feature state change, the system automatically triggers a global resonance refresh to update the entity resonance relationship at the corresponding level.

3. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The preset meta-rules are the opposing relationships of qualitative labels, including predefined core meta-rules and derived meta-rules automatically generated through a recursive guessing mechanism; the generation of traceable reasoning results specifically involves recursively tracing from the result state to the root cause state through reverse chain deduction to generate a complete causal reasoning chain.

4. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The incremental addition of new concepts and corresponding feature information to the three-dimensional concept tensor is achieved by combining unsupervised incremental learning in a weak-layer buffer with supervised learning based on a single definition.

5. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The knowledge hierarchy dimension adopts a bottom-up initialization method. The hierarchy of the new concept is determined by calculating the cosine similarity between the new concept and all existing concepts, and each concept corresponds to a unique parent node. The dimensionality upgrade sequence terminal of the knowledge hierarchy dimension is set with reserved dummy positions to achieve complete regularization of the dimensional space.

6. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The three-dimensional concept tensor is stored using a combination of sparse matrix and hash table, storing only non-zero data.

7. The white-box AI cognitive reasoning method based on concept tensors according to claim 1, characterized in that: The method is applied to an industrial safety early warning system to perform conflict detection and fault root cause tracing based on multi-sensor data.

8. A white-box AI cognitive reasoning system based on concept tensors, characterized in that, include: The tensor construction module is used to construct a three-dimensional concept tensor. The three-dimensional concept tensor includes a concept dimension, a primitive feature dimension, and a knowledge hierarchy dimension with a tree-like hierarchical structure, and follows a four-order dimensional progression structure of point-line-plane-volume. The interactive conversion module is used to execute upward abstraction and downward retrieval mechanisms, realizing bidirectional conversion between geometric features and topological structures to form a unified knowledge representation. The request processing module is used to respond to reasoning requests and extract the feature representation of the concept to be reasoned at a specified knowledge level. The matching analysis module is used to perform interpretable matching analysis on the feature representation to obtain a set of matching attributes containing specific attribute pairs; The rule reasoning module is used to perform deterministic state transitions based on the set of matching attributes and preset meta-rules, and generate traceable reasoning results. The incremental learning module is used to incrementally add new concepts and their corresponding feature information to the 3D concept tensor.

9. A white-box AI cognitive reasoning system based on concept tensors according to claim 8, characterized in that, The matching analysis module performs interpretable matching analysis through a feature resonance mechanism and supports global resonance refresh; the rule reasoning module generates causal reasoning chains through reverse chain derivation and can automatically generate derived meta-rules through a recursive word guessing mechanism; the incremental learning module integrates weak-layer buffer learning units and single-definition learning units; the tensor construction module achieves dimensional space completeness by reserving illusory spaces.

10. A white-box AI cognitive reasoning system based on concept tensors according to claim 8, characterized in that, The system is deployed on industrial edge computing devices to perform industrial security early warning and fault root cause tracing.

Images