A multi-layer nested AI self-correction general architecture containing local and global tensor double normalization and a control method

Abstract

This invention discloses a multi-layered nested AI self-correction general architecture and control method with dual local and global tensor normalization. The architecture logic is divided into three control domains: local, global, and AI iteration, which perform local and global two-level standardized operations respectively, unifying data dimensions and layering to prevent error accumulation. AI is broken down into three modules: preprocessing, raw decision-making, and global optimization. The upper layer only fine-tunes the parameters of abnormal units based on error indicators, eliminating blind global training. The architecture can add a top-level compliance constraint logic control domain, quantifying legal and ethical regulations into rule factors and adding ethical standardization operations. Ethical verification has the highest priority, achieving end-to-end compliance interception. This solution can be externally adapted to systems such as autonomous driving, industrial control, large-scale models, and medical AI, suppressing AI illusions and numerical divergence, and reducing computing power and compliance costs.

Description

Technical Field

[0001] This invention belongs to the cross-technical fields of intelligent control algorithms, layered collaboration of artificial intelligence, complex system software architecture, multi-source signal correction, and dynamic system self-iterative optimization. Relying on a logical multi-level hierarchical nested architecture combined with a two-level standardized constraint mechanism, it can be widely used in various dynamic intelligent systems such as autonomous driving, EEG biosignal denoising and analysis, industrial measurement and control, power dynamic regulation, intelligent sensing and detection, general large models, AI agent value alignment and ethical and safety management, etc. It belongs to the category of general underlying infrastructure technologies. Background Technology

[0002] End-to-end pure AI relies on neural network data fitting to make decisions, but lacks deterministic mathematical constraints and a unified convergence boundary. Its output is highly random, prone to creating the illusion of intelligence, and parameter overflow and divergence. It depends on training with all global parameters, making it extremely prone to overfitting and operating condition oscillations. Its safety and reliability cannot meet the requirements of high-level deployment. Traditional modular control systems adopt pipeline layout, with inconsistent data dimensions among modules. They can only perform single-point local corrections, lack unified constraints across the entire domain, and continuously accumulate coupling errors. They are not intelligent enough, highly dependent on repeated manual calibration and debugging, and lack autonomous adaptive iteration capabilities. Existing tensor and AI fusion solutions only treat multidimensional numerical carriers as conventional numerical calculation tools, without establishing a logical hierarchical nested control system, without distinguishing between local and global differentiated standardization constraints, without an AI layered functional collaboration mechanism, and lack a control architecture where the upper layer relies on error indicators for cross-layer directional parameter tuning and the lower layer has no right to reverse intervention. At the same time, generative AI generally suffers from pain points such as difficulty in value alignment, generation of illegal content, and bypassing of security rules by jailbreaking; existing ethical control measures are mostly superficial interception methods such as post-blacklists and prompt word restrictions, which lack underlying mathematical rigid constraints and cannot achieve irreversible ethical control across the entire chain. In summary, the industry urgently needs a general intelligent underlying architecture that balances convergence certainty, intelligent generalization capability, error layering and isolation, autonomous iterative optimization, and can be embedded with rigid ethical constraints at the underlying level. Summary of the Invention

[0003] 3.1 Overall Design Principles This invention uses multidimensional ordered numerical carriers (tensors, multidimensional matrices, multidimensional arrays, structured feature matrices, and other similar mathematical carriers) as its underlying mathematical foundation. It pioneers a basic general architecture with multi-level hierarchical nesting of logic, two-level standardized constraints (local and global), layered AI functional collaboration, upper-level cross-domain control, and targeted local parameter tuning of error indicators. This architecture can selectively add external or internal value logic constraint domains, quantifying laws and regulations, public order and good customs, and industry ethics into hierarchical rule factors and conducting ethical standardization calculations. This forms a two-layer bottom-line control system of physical constraints and ethical constraints, solving three major industry pain points: uncontrolled divergence of pure AI, insufficient intelligence of traditional control, and difficulty in underlying ethical constraints for AI. 3.1.1 Definitions of key terms (all equivalent replacements fall under protection, avoiding renaming and rebranding) Multidimensional numerical carriers: These include multidimensional parameter carriers such as tensors, multidimensional matrices, multidimensional arrays, and feature tensor matrices, and are not restricted by naming conventions. Standardization constraint (normalization): realizes the operation of data unit unification, parameter weight regularization, and output value range anchoring. It is also known as value range limit, global standardization, and parameter boundary constraint. It achieves the same function and falls into the protection. Logical multi-level hierarchical nesting: not limited to physical wrapping and layering, it covers architectures such as logical subordination, parent-child control, distributed layering, and plug-in mounting, satisfying the feature that the upper layer can control the lower layer across layers, which is the nesting feature; Error metrics include quantitative metrics such as the difference between the AI ​​output and the corrected true value, residual, error weight, and loss weight. Parameter tuning of abnormal modules can be protected by using any of these metrics for screening. Value logic constraint domain: The highest logical control level of the system. It can be deployed in any form, and can be independently added externally, embedded in the global control domain, or distributed and embedded in units at all levels. The protection boundary is defined only by the constraint priority and is independent of the physical layout. Rule factors: Quantitative expression of ethical and legal constraints. The storage carrier includes a multi-dimensional matrix, rule database, configuration form, and penalty weight table. The carrier transformation does not circumvent protection. Note: The technical solutions in the industry that are simply nested, without dual standardization constraints, and without upper-level cross-domain directional parameter tuning are existing public technologies. Even if the architecture layout is changed, as long as the two key technical means of two-level standardization and upper-level directional parameter tuning are missing, all the beneficial effects of this invention cannot be achieved, and it does not constitute an equivalent substitution. 3.2 Seven Core Technical Features of the Infrastructure The system logic is divided into local control domains, global control domains, and AI iteration control domains. Each control domain can be physically merged or split into multiple units. The number of levels is not fixed. The upper layer has unidirectional control authority over the lower layer. Multidimensional carriers and AI functions are layered and coordinated: Multidimensional carriers are responsible for error correction and boundary constraints, while AI is divided into three categories according to function: preprocessing, raw decision-making, and global optimization. These can be deployed independently as modules or integrated into the same model for logical partitioning. Two-level standardization constraint mechanism: It is divided into local standardized operation exclusive to local control domain and global standardized operation of the whole system. The two-level operation is independent, constrained at each level, and the division of labor is fixed and cannot be interchanged. The upper layer relies on error indicators for targeted local parameter tuning: the upper layer control domain filters out abnormal sub-units by using either the difference or error weight as error indicators, fine-tunes the parameters of abnormal units only, freezes the parameters of stable units throughout the process, and prohibits global indiscriminate training of the entire model. Standardized dual priority constraints: the time priority of standardized operations is higher than that of all AI operations; regardless of the order of operations, all AI outputs must be intercepted by the standardized value range and cannot exceed the constraint boundaries. The architecture is infinitely expandable: adding any subsystem, perception dimension, or control module only requires supplementing the corresponding multi-dimensional factors and locally nested units, while the underlying rules and mathematical logic remain unchanged. One-way cross-domain control between upper and lower levels: The global control domain and the AI ​​iterative control domain can collect error indicators, monitor the lower level, trace anomalies, and issue parameter adjustment instructions across layers. The lower level has no right to modify the upper level's constraint logic in reverse. 3.3 Logical Control Domain Division Description Local control domain: The system is divided into several independent sub-units, each of which performs local standardized calculations independently, unifies the internal parameter dimensions, eliminates hardware drift, environmental disturbances, and acquisition noise locally, and prevents single-point errors from propagating upwards; Global control domain: It aggregates all locally standardized unit data, constructs a global multidimensional numerical carrier, and performs global standardization operations to eliminate cross-module coupling errors, adaptive weight allocation, and rigidly lock the output value domain of the entire system; AI Iterative Control Domain: Divided into three levels of AI functions: preprocessing, raw decision-making, and global optimization. The bottom-level functions complete the filtering of clutter and invalid data, the middle-level functions realize scenario generalization and raw decision generation, and the top-level global optimization function realizes cross-domain control and parameter tuning based on error indicators. 3.4 Two-tiered standardized operation mechanism Localized computation: Constructing multi-dimensional numerical carriers based on unit hardware, environment, disturbance, and state parameters, unifying the internal dimensions, weights, and output range of the unit, and closing the loop locally for single-point disturbances; Global standardized computation: The output of all system units is aggregated to construct a global multi-dimensional carrier, unify the data benchmark of the entire system, eliminate cross-domain coupling interference, lock the upper limit of system output, and avoid numerical overflow and divergence. 3.5 Rigid Criteria for Collaboration between Multidimensional Carriers and AI Deterministic fallback priority: Two-level standardized computation takes precedence over AI end-to-end computation, defining the convergence boundary for AI; AI functions are assigned to specific roles and are separated: the responsibilities of the three types of AI functions—preprocessing, decision-making, and global optimization—are fixed and cannot be used interchangeably. Targeted convergence optimization: Only upper-level units with excessive error are selected and corrected, stable module parameters are fixed, and overfitting caused by global blind training is eliminated; Non-intrusive adaptation principle: Standardized constraints are an external enhancement mode that does not require modification of the original system, the underlying structure of the AI ​​model and weight parameters, and is compatible with various native intelligent solutions. 3.6 Seven Rigid Operating Rules for the Infrastructure Two-level standardization with fixed division of labor, local standardization to eliminate single-point errors, and global standardization to lock the boundaries of the entire domain, with functions that are not interchangeable or replaceable; The standardization adopts a dual constraint of time priority + value range interception to control the range of AI output throughout the process; The upper layer controls the lower layer in one direction, and the lower layer cannot reverse and tamper with the upper layer's constraint logic; The three types of AI functions have fixed divisions of labor, with no overlap or overstepping of responsibilities; Only the upper-level trigger parameters are fine-tuned, while the parameters of the stable module are permanently frozen; Local error domain closed-loop elimination and cross-domain coupled error global unified correction achieve error hierarchical zero accumulation; New business modules only supplement local units and multi-dimensional factors, while the underlying architecture rules remain constant. 3.7 Advanced Extension: Compliance Constraint Logic Control Domain + Hierarchical Rule Factor System 3.7.1 Extension Principle Based on the existing three-level logical control domain, a compliance constraint logical control domain can be added as the highest logical level of the entire system; compliance constraint rules, laws and regulations, and industry standards are decomposed into hierarchical rule factors and quantitatively stored, and a new ethical dimension is added for standardized calculation, with full-link execution: local standardization → global physical standardization → AI original decision → ethical standardization verification → compliance output, forming a closed loop of physical + ethical dual constraints. 3.7.2 Expanding Six Technical Features Top-level logical constraint domain: The compliance constraint logic control domain is the highest control level in the entire system. Its deployment form is arbitrary, and it unidirectionally constrains all control domains below it. Hierarchical and quantitative storage of rule factors: Divided into unmodifiable core factors, authorizable scenario factors, and customizable business factors, with no limit on storage medium; Ethical-specific standardized computation: Conduct ethical value range verification on all AI decision and iteration parameters to block illegal outputs; Targeted correction of ethical error indicators: Violating units are screened based on ethical difference and ethical weight, and parameters are finely adjusted by the upper level. Multi-level computation priority: Ethical normalization computation > Global physical normalization computation > Local normalization computation > All AI computations; Rule encryption and solidification mechanism: Core rule factors are stored in encryption, and only the top layer has one-way incremental update permissions, while the lower layer cannot tamper with or bypass permissions. 3.7.3 Five new rigid ethical operating rules have been added. Ethical standardization has the highest priority; no AI operation or output must exceed the ethical boundaries. The three-level rule factor hierarchy is fixed, and the core factor is permanently locked and cannot be tampered with. If ethical indicators are exceeded, the output will be immediately blocked, and operation can only continue after correction to meet the standards. AI outputs must pass both physical and ethical standardization verifications simultaneously. The entire ethical process generates tamper-proof encrypted audit logs, enabling full-process traceability and auditing. Attached Figure Description Figure 1 is a schematic diagram of the overall architecture of this invention. The diagram is arranged in layers from top to bottom: the top dashed box is the optional extended compliance constraint logic control domain; below, the solid boxes are arranged in order of AI iteration control domain, global control domain, and local control domain; the bottom solid box contains the perception sub-unit, control sub-unit, transmission sub-unit, and feedback sub-unit; data flows forward from top to bottom, and the AI ​​iteration control domain controls all lower-level modules unidirectionally across layers through the left vertical arrow; the operation priority is: ethical standardization > global standardization > local standardization > AI operation. Detailed Implementation 5.1 Infrastructure Implementation Examples The autonomous driving system is divided into multiple local control units, including visual perception, radar detection, vehicle dynamics, and positioning. Each unit is locally standardized to eliminate sensor distortion and temperature drift noise. The entire vehicle is globally standardized to lock the output range of steering, acceleration, deceleration, and braking. AI is layered to achieve data cleaning and driving decision-making. The upper layer relies on error indicators to optimize abnormal sub-units in a targeted manner, solving the problems of large model illusion and control quantity overflow divergence. EEG signal analysis system: decomposes power frequency interference, electromyography artifacts, baseline drift, and extracts features from local units; local standardization removes various clutter; global standardization unifies the amplitude and phase of multiple channels; hierarchical AI completes clutter screening, EEG decoding, and sub-unit self-calibration. Industrial multidimensional measurement and control system: It separates temperature, pressure, vibration and electrical sensing units, and eliminates single-point drift error of equipment through local standardization; it unifies the threshold values ​​of whole machine parameters through global standardization, and completes the working condition assessment and control output through AI, relying on error indicators for long-term self-calibration. 5.2 Extended Examples of Compliance Constraints General large-scale model security control: Text generation, knowledge retrieval, and logical reasoning units complete two-level physical standardization; compliance constraint logic control domain loads legal regulations, privacy protection, and prohibited content rule factors; AI-generated content undergoes ethical standardization verification; illegal content generates ethical error indicators and fine-tunes the generation module accordingly; serious anomalies trigger system circuit breakers. Ethical decision-making for autonomous driving: Incorporating ethical factors such as prioritizing life and vulnerable traffic participants, multiple decisions under extreme conditions are unified and optimized to resolve contradictions in extreme ethical decisions. AI-assisted diagnosis in healthcare: It incorporates medical rules and factors that prioritize patient rights, prohibit overtreatment, and protect privacy. The AI ​​diagnosis and treatment plan undergoes ethical review, eliminating unreasonable medication and high-risk treatment suggestions. 6. Beneficial effects of the invention The patent barrier is strong and extremely difficult to circumvent: This invention is the first to create a basic paradigm of multi-level hierarchical nesting logic + two-level standardization + layered AI function collaboration + cross-domain unidirectional control + local parameter tuning of error indicators, and is equipped with a top-level value and ethical constraint architecture; all technical features are written in a functional upper-level style, and competitors can fall into the scope of equivalent infringement by renaming, merging levels, changing physical layout, post-verification, module embedding and other variations, which cannot be easily circumvented. The boundaries of infringement determination are clear: all core technology rules are legally fixed, and the implementation of functions is used as the standard for infringement determination, without being restricted by the form of product software and hardware, model architecture, or application field. Full track coverage: The basic architecture covers hardware control systems such as autonomous driving, industrial measurement and control, and biological brain-computer signals; the value extension architecture fully covers generative large models and AI Agent ethics and risk control tracks, simultaneously locking in two high-value tracks. It offers a dual solution to industry pain points: the infrastructure addresses the uncontrolled divergence of pure AI and the insufficient intelligence of traditional control; the value constraints tackle the common challenges of global AI value alignment, rule generation, and rule overreach from a mathematical perspective. High versatility and ample room for patent expansion: The upper-level claims are not limited to any particular scenario and can be used to derive product patents for various sub-fields, thus constructing a complete patent protection matrix of parent patent + series of dependent patents. Excellent practicality: non-intrusive external deployment, compatible with various existing AI and control systems, low software and hardware deployment costs, stable system iteration, low failure rate, and outstanding industrialization value. Supplement: If only a nested structure is adopted, the local and global two-level standardized constraints are removed, and the cross-layer directional parameter tuning architecture relying on the upper-level error index is removed, it is only a simple modification of the existing conventional nested architecture. It cannot achieve the core technical effects such as local isolation of single-point errors, rigid locking of global output, suppression of AI illusion, layered error clearing, and local fine-grained iterative cost reduction. It cannot solve the inherent pain points of the industry that this invention aims to overcome, such as AI output divergence, layer-by-layer error accumulation, high global blind training computing power, and ethical non-control. This reduction scheme does not have the technical superiority of this invention and is not an equivalent technical solution.

Claims

1. A multi-layered nested AI self-correction general architecture and control method with dual local and global tensor normalization, characterized in that: S1. Build a logically multi-layered hierarchical nested architecture, which is logically divided into local control domain, global control domain, and AI iteration control domain. Each control domain can be physically split or merged. The overall architecture supports unlimited expansion and addition of sub-units. S2. Each local control domain constructs a multi-dimensional numerical carrier and performs local standardized constraint calculations to unify the parameter dimensions within the unit, eliminate single-point errors, and prevent errors from accumulating to the upper layer. S3. Summarize the data of each local unit to construct a global multidimensional numerical carrier and perform global standardization constraint operations to correct cross-module coupling errors and lock the output value range of the entire system; the two-level standardization operations have fixed division of labor and cannot be interchanged. S4 and AI are divided into three categories of functions: preprocessing, raw decision-making, and global optimization. These three functions can be separated into independent modules or integrated into the same logical partition within the same model, each performing its own function. S5. Standardized operation is subject to dual constraints: timing priority is higher than AI operation, and all AI outputs must be intercepted and verified by the standardized value range. S6. The upper-level control domain actively collects full-link error indicators across layers. The error indicators include the difference and error weight, enabling lower-level unit status monitoring, anomaly tracing, and targeted parameter tuning. Lower-level units cannot reverse-modify the upper-level constraint logic and are not authorized to perform large-scale autonomous parameter tuning. S7. The upper layer filters out abnormal units based on the error index threshold or the error weight ratio of the sub-module. Only the parameters of the filtered sub-units that exceed the standard are locally fine-tuned, while the parameters of the other modules are frozen, and global blind training of the entire model is prohibited. S8. The newly added functional modules only supplement the corresponding multi-dimensional factors and local units, while the underlying architecture operation rules remain unchanged.

2. The method according to claim 1, characterized in that: Local errors are eliminated in the corresponding control domain through closed-loop control, and cross-domain coupling errors are uniformly corrected by the global control domain. The responsibilities of the three types of AI functions are isolated, achieving zero error accumulation in the system.

3. The method according to claim 1, characterized in that: The global optimization function serves as the top-level control entity, where abnormal units are screened and optimized in a targeted manner by the upper layer using methods such as difference judgment and error weight ratio. All AI operations are locked in the value range by two levels of standardization to avoid parameter oscillation, overfitting, numerical divergence, and AI illusion.

4. The method according to claim 1, characterized in that: The overall architecture is a non-intrusive add-on enhancement architecture, which does not require modification of the original system and AI underlying structure, and can be adapted to all categories of intelligent control systems such as autonomous driving, biosignals, industry, power, and large models.

5. The method according to claim 1, characterized in that: The architecture can add a compliance constraint logic control domain as the highest level of logic control for the entire system; the deployment form of this constraint domain is not limited, and it can be independently attached externally, embedded in the global control domain, or distributed and embedded in units at all levels to form a dual closed-loop control system with physical constraints superimposed on ethical constraints.

6. The method according to claim 5, characterized in that: Laws, regulations, and ethical guidelines are quantified into hierarchical, multi-dimensional rule factors. The storage carriers for these rule factors include multi-dimensional numerical matrices, rule databases, and configuration forms. Ethical standardization calculations are performed based on these rule factors, and AI outputs must pass ethical verification before they become effective.

7. The method according to claim 6, characterized in that: The computational priorities, from highest to lowest, are: ethical standardization computation, global standardization computation, local standardization computation, and AI end-to-end computation; violating units are screened based on ethical differences or ethical weights and parameters are fine-tuned in a targeted manner; core rule factors are encrypted and solidified, and only the top level supports one-way authorized incremental updates.

8. The method according to claim 5, characterized in that: The entire process of ethical verification, parameter correction, and rule change operations generates tamper-proof audit data; the access of the value logic constraint domain does not alter the underlying operating rules of the original multi-layered nested architecture.

Images