Corpus-based language model generation method and system

Abstract

The application discloses a corpus-based language model generation method and system, and belongs to the technical field of artificial intelligence, comprising: constructing a digital twin simulation environment and collecting multi-modal physical true value data; inputting the data into a biological perception simulation layer, simulating human sensory threshold drift and attention fluctuation by using a nonlinear transfer function, and outputting a fuzzy perception vector with dynamic confidence; monitoring the confidence, triggering active perception planning when the confidence is lower than a threshold, generating and executing a virtual exploration action that maximizes information gain; fusing initial perception and supplementary observation, updating fault hypothesis posterior probability based on Bayesian inference, and mapping it into a dynamic thinking chain text; and using the text to fine-tune the language model. The application generates structured corpus containing uncertainty and proof logic by simulating human cognitive mechanism, solves the problem that traditional simulation corpus is too perfect, leading to overconfidence of the model, and enables the language model to have human-like uncertain reasoning ability.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more particularly to a corpus-based language model generation method and system. Background Technology

[0002] With the deepening of strategic implementation, intelligent operation and maintenance and human-machine collaboration in the high-end equipment manufacturing field have become crucial for improving production efficiency and safety. In this process, large-scale language models, with their powerful natural language understanding and generation capabilities, are widely used in scenarios such as equipment fault diagnosis, assembly process guidance, and emergency decision support. However, general-purpose large models are mainly trained based on publicly available internet text, lacking a deep understanding of the correlation between multimodal signals and professional operational logic in specific industrial scenarios. To empower models with domain experts, building a high-quality dedicated corpus has become a core prerequisite. Currently, the industry generally adopts a combination of real data acquisition and cleaning with digital twin simulation generation to expand the corpus size. While real data is accurate, its acquisition cost is extremely high, and long-tail fault samples are scarce; digital twin technology, on the other hand, can generate massive amounts of simulated data containing rare faults at low cost through parameterized perturbation, and annotate the ground truth through automated scripts, becoming the mainstream technical means to solve the data scarcity problem.

[0003] However, existing methods for constructing corpora based on digital twins suffer from significant technical bottlenecks. Their core flaw lies in the vast semantic gap between the physical perfection of the simulated environment and the cognitive ambiguity of real-world human perception. Current technologies typically output precise physical truth values ​​calculated by the simulation engine when generating corpora, using these as deterministic labels to train models. This approach ignores sensor noise, environmental interference, and the perceptual uncertainty and hesitation experienced by human experts during decision-making, all prevalent in real-world scenarios. The direct consequence is that models trained on such overly pure corpora exhibit extreme overconfidence and logical rigidity in real-world deployments. Once the input signal deviates slightly from the simulation's standard curve due to noise, the model fails to recognize it or experiences severe hallucinations, lacking the heuristic reasoning ability of human experts to detect suspected anomalies requiring further verification. Summary of the Invention

[0004] This invention overcomes the shortcomings of the prior art and provides a corpus-based language model generation method and system.

[0005] To achieve the above objectives, the technical solution adopted by this invention is: a corpus-based language model generation method, comprising:

[0006] S1. Construct a digital twin simulation environment containing parameterized fault disturbances and run the simulation to collect multimodal physical truth data sequences;

[0007] S2. Input the multimodal physical truth data sequence into the biological perception simulation layer, use the nonlinear transfer function to simulate the human sensory threshold drift and attention fluctuation, and output a fuzzy perception vector with dynamic confidence distribution.

[0008] S3. Real-time monitoring of the dynamic confidence distribution of the fuzzy perception vector. When the confidence is lower than the uncertainty threshold, the active perception planning mechanism is automatically triggered to generate and execute a virtual exploration action sequence that maximizes information gain.

[0009] S4. The posterior probability of the fault hypothesis is iteratively updated based on the Bayesian inference algorithm by fusing the fuzzy perception vector with the supplementary observation data obtained after executing the virtual exploration action sequence, and the probability convergence process is mapped into a dynamic thought chain text containing hypothesis proposal, verification and correction logic.

[0010] S5. The language model is fine-tuned using structured corpus of dynamic thought chain texts to obtain a language model with uncertainty reasoning ability.

[0011] In a preferred embodiment of the present invention, step S2, which uses a nonlinear transfer function to simulate human sensory threshold drift and attention fluctuations, includes:

[0012] Based on the principles of psychophysics, independent biosensory sub-modules were constructed for different sensory channels, namely vision, hearing, and force.

[0013] Within each submodule, a nonlinear transfer function designed based on Weber-Fechner's law is used to transform the input precise physical signal into a neural excitation pattern with inherent fuzziness.

[0014] At the same time, a dynamic sensory threshold drift mechanism is implemented to adjust the perception threshold in real time based on the virtual subject's fatigue accumulation and adaptation history, simulating missed detections due to fatigue or false alarms due to alertness.

[0015] Furthermore, an attention fluctuation model is introduced to generate a time-varying sequence of attention weights, which enhances or suppresses the perception signals of different spatial regions or modalities. Finally, a confidence score representing the reliability of perception is calculated and output synchronously.

[0016] In a preferred embodiment of the present invention, step S3, generating and executing a virtual exploration action sequence that maximizes information gain, includes:

[0017] Define the current fuzzy perception state as the initial belief state, and construct a virtual action space in the digital twin environment that includes multi-angle observation, repetitive operation, switching detection modes, and adjusting environmental parameters;

[0018] Based on the partially observable Markov decision process framework, predict the probability distribution of the observation results that may be obtained after executing each candidate action;

[0019] Calculate the difference in probability distribution of belief states before and after each candidate action to quantify the expected information gain for each action;

[0020] A heuristic search strategy is used to perform multi-step deductions in the virtual action space to select the action combination that minimizes the uncertainty of the expected posterior probability distribution, which is then issued and executed as the virtual exploration action sequence.

[0021] In a preferred embodiment of the present invention, step S4, mapping the probability convergence process to dynamic thought chain text, includes:

[0022] Maintain a logic template engine to parse key node data in the Bayesian update process in real time. The key node data includes the initial distribution of prior probability, feature description of new evidence, calculation results of likelihood ratio, and drift direction and magnitude of posterior probability.

[0023] Based on the initial fuzzy perception vector, generate statements describing the initial suspected objects and the degree of uncertainty;

[0024] By combining the virtual exploration actions performed and the supplementary observation data obtained, action verification statements are generated that explain the rationale for the verification measures and the new phenomena.

[0025] Based on the convergence results of the posterior probability and the pre-defined rules for the correspondence between logical connectors and probability change intervals, a belief update and decision convergence statement is generated to summarize the degree of evidence support and draw a final conclusion.

[0026] Among them, a strong confirmation logic is used when the posterior probability increment exceeds the high threshold, a strong negation logic is used when the posterior probability increment is lower than the negative high threshold, and a weak correction or a euphemistic expression requiring further verification is used when the absolute value of the probability increment is small.

[0027] In a preferred embodiment of the present invention, step S1, constructing a digital twin simulation environment containing parameterized fault disturbances, includes:

[0028] Based on the geometric topology, material physical properties, and kinematic constraints of the target high-end equipment, a full-element digital mapping model is reconstructed in virtual space, and the model is embedded with a set of differential equations describing the dynamic behavior of the equipment.

[0029] Define a parameterized fault disturbance space, and set the fault type, injection location, severity coefficient, and evolution time constant as independently configurable variables;

[0030] By using orthogonal experimental design or Latin hypercube sampling algorithm, a systematic traversal is performed in the parameterized fault disturbance space to generate a full-scenario simulation test case set covering normal operating conditions to extreme failure boundaries;

[0031] The digital twin environment is driven to execute the full-scenario simulation use case set. Using a timestamp alignment algorithm based on maximizing mutual information, heterogeneous multimodal data streams are mapped to a unified time axis to form a multimodal physical truth data sequence with accurate physical truth labels.

[0032] In a preferred embodiment of the present invention, the severity coefficient in the parameterized fault disturbance space is used to linearly or nonlinearly control the magnitude of the fault impact, simulating a continuous change from minor anomalies to catastrophic failures.

[0033] The parameterized fault disturbance space also includes a spatial distribution function, which defines the range and intensity distribution of the fault in the equipment geometric space according to the physical mechanism of the fault.

[0034] It also includes a time evolution function that describes the dynamic characteristics of the fault as it changes over time from the moment it occurs. The time evolution function can take the form of a step function, a ramp function, or an exponential growth function, respectively simulating sudden fracture, progressive wear, or thermal runaway processes.

[0035] In a preferred embodiment of the present invention, in step S3, the uncertainty threshold is a dynamically adjusted threshold whose value is adaptively adjusted according to the safety level, urgency and historical false alarm rate of the current task, so as to ensure that the sensitivity of capturing potential uncertainties in high-risk scenarios is higher than that in low-risk scenarios.

[0036] In a preferred embodiment of the present invention, step S5, fine-tuning the language model includes:

[0037] The dynamic thought chain text is converted into a standard instruction fine-tuning data format. Each data sample includes an instruction part that defines the task as fault diagnosis and reasoning based on multimodal perception information, an input part that includes fuzzy perception vector description and virtual exploration action sequence description, and an output part that includes the complete thought chain text.

[0038] The generated samples are mixed with high-quality cases annotated by real experts in a certain proportion to form a dedicated instruction fine-tuning dataset;

[0039] The pre-trained language model is trained using a loss function that is aware of thought chains. This loss function not only calculates the error in the model's final conclusion but also introduces a logical consistency regularization term to penalize breaks and inconsistencies in the reasoning logic chain and assigns higher weight coefficients to key logical nodes in the thought chain and their subsequent inferences.

[0040] In a preferred embodiment of the present invention, in step S2, the output fuzzy perception vector not only includes signal strength information, but also embeds the reliability assessment of the perception at the current moment, i.e., the confidence score. The confidence score is lowest when the perception intensity is close to the decision boundary, indicating that the virtual subject is in a hesitant state, and highest when the perception intensity tends to the extreme value and the attention is focused.

[0041] This invention provides a corpus-based language model generation system, comprising:

[0042] The digital twin simulation module is used to build a digital twin simulation environment containing parameterized fault disturbances and run simulations to acquire multimodal physical truth data sequences.

[0043] The biosensory simulation module is used to receive the multimodal physical truth data sequence, simulate human sensory threshold drift and attention fluctuation using a nonlinear transfer function, and output a fuzzy perception vector with dynamic confidence distribution.

[0044] The active perception planning module is used to monitor the dynamic confidence distribution of the fuzzy perception vector in real time. When the confidence is lower than the uncertainty threshold, the active perception planning mechanism is automatically triggered to generate and execute a virtual exploration action sequence that maximizes information gain.

[0045] The dynamic thought chain generation module is used to fuse the fuzzy perception vector with the supplementary observation data obtained after executing the virtual exploration action sequence, iteratively update the posterior probability of the fault hypothesis based on the Bayesian inference algorithm, and map the probability convergence process into a dynamic thought chain text containing hypothesis proposal, verification and correction logic.

[0046] The instruction fine-tuning module is used to fine-tune the language model using the structured corpus of the dynamic thought chain text, so as to obtain a language model with uncertainty reasoning ability.

[0047] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0048] (1) This invention constructs a digital twin simulation environment containing parameterized fault disturbances and introduces a biosensory simulation layer to convert physical truth values ​​into fuzzy perceptual vectors with dynamic confidence distributions. This solves the technical drawback of traditional simulation corpora where the physical perfection of the model leads to the model deviating from real human cognitive logic. Existing technologies directly use the precise physical truth values ​​calculated by the simulation engine as deterministic labels, ignoring the sensor noise and perceptual uncertainty of human experts that are common in real working conditions. This invention simulates the drift of human sensory thresholds and fluctuations in attention, so that the generated corpora naturally contain cognitive fuzziness. This allows the model to learn to maintain a necessary cautious attitude under conditions of incomplete information during the training phase. When facing critical faults with weak signals and fuzzy features in real industrial sites, it can effectively suppress the risk of misjudgment caused by overconfidence.

[0049] (2) This invention addresses the technical deficiency of existing corpus construction methods in failing to reproduce the dynamic thinking process of human experts from doubt to verification by real-time monitoring of the confidence distribution of fuzzy perception vectors and automatically triggering an active perception planning mechanism when the confidence level is below the uncertainty threshold. Existing technologies rely on static statistical distributions for data augmentation, which can only introduce perturbations at the numerical level but cannot simulate the human behavior of actively changing observation perspectives and repeating operations to reproduce signals when confidence is low. This invention generates and executes a virtual exploration action sequence that maximizes information gain, enabling the corpus to fully record the interactive reasoning trajectory from initial doubt to active verification, thus giving the language model the ability to actively plan verification paths when perceiving fuzziness.

[0050] (3) This invention integrates fuzzy perception vectors and supplementary observation data, iteratively updates the posterior probability of the fault hypothesis based on a Bayesian inference algorithm, and maps this probability convergence process into a dynamic thought chain text containing the logic of hypothesis proposal, verification, and correction. This solves the deep-seated defect of existing corpora that emphasize results over process. Existing technologies use the final fault label as the training target, resulting in a lack of training samples for progressive reasoning when evidence is insufficient. This invention transforms the numerical Bayesian update process into a structured natural language reasoning sequence, enabling the language model to learn the cognitive paradigm of human experts weighing multiple pieces of evidence and dynamically correcting beliefs. When faced with complex faults involving multiple coupled factors, it can output interpretable diagnostic conclusions with logical evolution characteristics.

[0051] (4) This invention utilizes dynamic thought chain text containing a complete cognitive evolution trajectory to fine-tune the language model, efficiently injecting structured knowledge generated by biological perception simulation and active verification mechanisms into the model parameters, thus solving the fundamental problem of general pre-trained models lacking domain-specific cognitive logic. Existing fine-tuning methods often only focus on the mapping relationship between input and final label, causing the model to learn to quickly give conclusions but lose intermediate reasoning ability. This invention forces the model to pay close attention to the coherence of causal inference through the loss function of thought chain perception, transforming it from a passive text writer into an active reasoner with human-like cognitive mechanisms, realizing a complete cognitive closed loop from perceiving uncertainty to decision certainty. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of the steps of a preferred embodiment of the present invention;

[0054] Figure 2 This is a system architecture and data flow diagram of a preferred embodiment of the present invention;

[0055] Figure 3 This is a biosensory simulation layer diagram of a preferred embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0058] Application Overview:

[0059] In intelligent operation and maintenance within the high-end equipment manufacturing sector, language models trained on existing simulation corpora exhibit significantly lower decision-making accuracy than human experts when handling critical faults with weak signals or ambiguous features. Human experts typically demonstrate hesitation, repeated verification, and proactive adjustments to detection strategies in such scenarios, while models tend to output high-confidence but erroneous definitive conclusions. This difference stems from existing corpus construction methods that directly equate the precise physical truth values ​​generated by simulations with human cognitive input, neglecting the uncertainty of perceived signals in real-world conditions and the dynamic verification mechanisms involved in human decision-making. This bias in data representation leads to a lack of training samples for probabilistic reasoning and risk avoidance under incomplete information conditions, resulting in the risk of misjudgment due to overconfidence in practical applications.

[0060] Existing technologies struggle to address these issues, primarily due to their data augmentation paradigm based on static statistical distributions. Conventional attempts to simulate uncertainty involve superimposing Gaussian noise or applying static signal detection theory formulas; however, these methods only introduce perturbations at the numerical level and cannot reproduce the nonlinear threshold characteristics and dynamic attention allocation mechanisms of the human perceptual system. Directly integrating signal detection theory from psychophysics as a post-processing module with existing simulation systems fails to simulate the dynamic exploratory behaviors humans exhibit under low confidence levels, such as actively changing perspectives or repeating operations. Consequently, the generated corpus still lacks interactivity and logical evolution.

[0061] This invention proposes a corpus-based language model generation method that addresses the aforementioned technical problems by simulating human cognitive mechanisms. A bio-perception simulation layer is constructed within a digital twin simulation engine. Based on psychophysical principles, precise physical signals are transformed into fuzzy perceptual vectors with confidence distributions, achieving the conversion from physical truth to cognitive representation. Building upon this, the invention establishes a dynamic Bayesian thought chain generation mechanism. When the perceptual confidence level falls below a preset threshold, the system automatically plans and executes the virtual subject's proactive perceptual actions, such as changing the observation perspective or repeating operations to reproduce the signal. This generates a structured corpus encompassing the entire process of hypothesis formulation, action verification, belief updating, and decision convergence.

[0062] Exemplary method:

[0063] like Figure 1 As shown, the corpus-based language model generation method includes the following steps:

[0064] S1. Construct a digital twin simulation environment containing parameterized fault disturbances and run the simulation to collect multimodal physical truth data sequences;

[0065] S2. Input the multimodal physical truth data sequence into the biological perception simulation layer, use the nonlinear transfer function to simulate the human sensory threshold drift and attention fluctuation, and output a fuzzy perception vector with dynamic confidence distribution.

[0066] S3. Real-time monitoring of the dynamic confidence distribution of the fuzzy perception vector. When the confidence is lower than the uncertainty threshold, the active perception planning mechanism is automatically triggered to generate and execute a virtual exploration action sequence that maximizes information gain.

[0067] S4. The posterior probability of the fault hypothesis is iteratively updated based on the Bayesian inference algorithm by fusing the fuzzy perception vector with the supplementary observation data obtained after executing the virtual exploration action sequence, and the probability convergence process is mapped into a dynamic thought chain text containing hypothesis proposal, verification and correction logic.

[0068] S5. The language model is fine-tuned using structured corpus of dynamic thought chain texts to obtain a language model with uncertainty reasoning ability.

[0069] When constructing specialized language models for high-end equipment manufacturing, traditional corpus acquisition methods mainly rely on historical work order records or manual expert annotation. This not only limits the data scale but also makes it difficult to cover various rare faults and extreme boundary conditions that may occur throughout the equipment's entire lifecycle. Therefore, introducing digital twin technology for virtual sample generation has become a consensus. This involves reproducing the operational laws of the physical world through a high-fidelity simulation environment, thereby acquiring massive amounts of annotated data at low cost. However, raw data sequences generated solely by a physics engine can only reflect deterministic causal relationships under ideal conditions, lacking the cognitive uncertainties caused by sensor noise, environmental interference, and individual operator differences in real-world operating environments.

[0070] Therefore, constructing a digital twin simulation environment that supports parameterized fault injection and collecting multimodal physical truth data sequences from it is the primary prerequisite and fundamental step for realizing the method of this application.

[0071] In step S1, parameterized fault disturbance refers to no longer treating various possible faults of the equipment as simple discrete labels, but deconstructing them into a set of continuously adjustable physical parameters. By injecting parameter variables that deviate from the standard values ​​into the mathematical expression of the simulation model, a variety of fault symptoms are induced at the output end.

[0072] Multimodal physical truth data sequence refers to the set of multidimensional signals under ideal conditions that are synchronously output by the simulation system within a specific time window and are free from any sensor noise pollution. It covers heterogeneous data such as vibration waveforms, temperature field distribution, current and voltage curves, and equipment operation logs.

[0073] Specifically, step S1 reconstructs a full-element digital mapping model of the target high-end equipment in virtual space based on its geometric topology, material physical properties, and kinematic constraints. This model not only contains static three-dimensional geometric information but also embeds a set of differential equations describing the dynamic behavior of the equipment, so as to accurately reproduce the real-time response of mechanical components under the coupling of force, thermal field, and electromagnetic field.

[0074] Building upon this foundation, a virtual data acquisition array encompassing visible light vision, acoustic signature spectrum, infrared thermal imaging, and multi-dimensional force sensors is deployed. A parameterized fault disturbance space is defined, where fault type, injection location, severity coefficient, and evolution time constant are set as independently configurable variables. Through orthogonal experimental design or Latin hypercube sampling algorithms, a systematic traversal is performed within this high-dimensional variable space to generate a full-scenario simulation test case set covering everything from normal operating conditions to extreme failure boundaries.

[0075] The digital twin environment is driven to execute the full-scenario simulation test case set, synchronously recording the system's internal state vectors and the raw signals from external sensors at each time step. Using a timestamp alignment algorithm based on maximizing mutual information, heterogeneous multimodal data streams are mapped to a unified microsecond-level time axis, eliminating phase delay and asynchronous sampling issues between different sensors, ultimately forming a multimodal physical truth data sequence with precise physical truth labels.

[0076] Faults in high-end equipment manufacturing scenarios often exhibit significant characteristics such as multi-factor coupling, nonlinear evolution, and long-tail distribution. Traditional simulation data generation methods typically rely on pre-set fixed scripts or simple random noise injection, generating only typical fault samples of a single type, which is insufficient to cover complex composite fault scenarios in the real world. For example, in the assembly of aero-engine blades, faults may be caused by the combined effects of small dimensional tolerance accumulation, material fatigue crack propagation, and abnormal thermal stress concentration, and their evolution process is not a linear superposition. If only conventional methods are used, the generated data will lack this deep physical correlation, leading to severe generalization failures in the trained model when facing real complex operating conditions. This application, by decoupling fault parameters and performing combined injection, can reproduce complex fault evolution paths in the real world, ensuring that each generated data sample has a clear physical cause and a complete evolution trajectory.

[0077] To quantify the injection process of fault disturbances and ensure the effectiveness and controllability of its coverage, this step introduces a parameterized fault injection function, which connects the abstract fault definition with the specific physical simulation.

[0078] Parametric fault injection functions are used to describe the corrective effect of fault factors on the physical state of the system at a specific time point and spatial location. Their core lies in precisely controlling the spatiotemporal evolution characteristics of faults through mathematical forms.

[0079] A representative fault injection formula can be expressed as: ,in, It represents the system state vector after the injection of a fault, and includes multi-dimensional physical quantities such as displacement, velocity, acceleration, temperature, and pressure. The reference state vector, representing the normal operating conditions, is obtained by solving the dynamic equations of the digital twin; The severity coefficient is a normalized variable between 0 and 1, used to linearly or nonlinearly control the magnitude of the failure's impact, simulating a continuous change from minor anomalies to catastrophic failures. It is a spatial distribution function related to the fault type and injection location. This function defines the range and intensity distribution of the fault in the equipment geometry space according to the physical mechanism of the fault, ensuring that the fault impact conforms to physical laws. It is a time evolution function that describes the time from the moment the fault occurs. The dynamic characteristics that begin to change over time can take the form of a step function to simulate sudden fracture, a ramp function to simulate progressive wear, or an exponential growth function to simulate thermal runaway.

[0080] By flexibly adjusting the formula , and By combining the function's parameters, this step enables precise control of the simulated fault scenarios, generating diverse data sequences with clear physical meaning and uniform distribution. The parameterized control method based on mathematical models transforms the simulation process from a black-box, random experiment into an interpretable, reproducible, and fully comprehensive systematic data construction process.

[0081] After constructing a multimodal data sequence covering the entire scenario and with accurate physical truth values ​​in step S1, directly using it for model training would result in a corpus lacking the uncertainty and ambiguity characteristics of real human cognitive processes, failing to reproduce the expert's decision-making logic under critical failures. To bridge the gap between the deterministic signals of the physical world and the subjective ambiguity of human cognition, an intermediate processing layer capable of simulating biological perception characteristics is introduced. This layer does not simply add noise, but rather, based on psychophysical principles, nonlinearly reconstructs and maps precise physical signals, transforming them into fuzzy representations that conform to the laws of human sensory response.

[0082] like Figure 3As shown, step S2 first receives time-aligned multimodal physical truth data (visual, auditory, and mechanics) from the previous step. This data represents a continuous, precise, and deterministic numerical sequence at the physical level. The system then sends these ideal signals to biosensory simulation submodules independently constructed for each sensory channel. Within each submodule, instead of traditional linear transformations or random Gaussian noise injection, a series of nonlinear transfer functions designed based on Weber-Fechner's law are deployed. These functions rigorously simulate the nonlinear response characteristics of human sensory organs under different stimulus intensities.

[0083] For example, the visual channel simulates the logarithmic response of the human eye to light intensity, significantly amplifying minute brightness differences in low-light environments while compressing large brightness variations in high-light environments. In the auditory channel, the frequency sensitivity curve and masking effect of the human ear are simulated, causing specific distortions or attenuation of weak abnormal noise signals in the background noise. Through this nonlinear mapping transformation, the originally clear and sharp physical signal is converted into a neural excitation pattern with inherent ambiguity, thus initially realizing the conversion from physical quantity to perceived quantity.

[0084] To concretely implement the above simulation process, this step constructs a core family of nonlinear transfer functions, the most representative of which is the dynamic perceptual response formula used to simulate sensory threshold drift and attention coupling. This formula describes the mathematical relationship between the output perceptual intensity and its accompanying confidence level after the input physical signal is processed by the biological sensory layer.

[0085] A representative calculation formula can be expressed as: ,in, The normalized sensory intensity vector represents the output at time t, and its value ranges from 0 to 1, directly corresponding to the strength of neural excitation. It is the input physical truth signal from step S1, such as sound pressure level, light intensity or torque value; It is a dynamic sensing threshold, which is a variable that changes over time, determined by the base threshold. With fatigue accumulation function and fitness function Joint decision, that is ,in, These are the fatigue coefficient and the adaptation coefficient, respectively. This part accurately simulates the physiological phenomena of decreased sensory sensitivity after long hours of work and adaptation under continuous stimulation. It is a gain control factor that is positively correlated with the current level of attention; when attention is highly focused... Increasing the value of the sigmoid function makes the slope of the sigmoid function steeper, making the system more sensitive to small signal changes near the threshold, and vice versa. It is the attention space weight factor, with a value between 0 and 1. It is generated by the attention fluctuation model based on the current task focus and is used to simulate the signal enhancement or inhibition effect caused by human selective attention.

[0086] In addition, this step also simultaneously calculates the confidence level of the output signal. The formula shows that when the perceived intensity is close to the decision boundary of 0.5, the confidence level is the lowest, indicating that the virtual subject is in a state of extreme hesitation; while when the perceived intensity tends to 0 or 1 and the attention is focused, the confidence level is the highest.

[0087] Building upon this, step S2 further introduces a dynamic sensory threshold drift mechanism and an attention fluctuation model to deeply simulate the evolution of cognitive states in human experts under prolonged work or high-pressure environments. The sensory threshold drift mechanism aims to reproduce the physiological phenomenon that the human perceptual threshold dynamically shifts with changes in fatigue level, adaptation history, and expectation level. The system calculates the perceptual threshold in real time by constructing a threshold dynamic adjustment function based on the time cumulative effect.

[0088] When the virtual subject is in a state of continuous high-intensity operation, the function will automatically raise the perception threshold, causing weak fault signals that could originally be detected to be judged as background noise or not detected, thus simulating the missed detection behavior caused by fatigue; conversely, when in an alert state or after rest and recovery, the threshold is lowered, and the system becomes oversensitive to small signals, simulating false alarms.

[0089] Meanwhile, the attention fluctuation model is used to simulate the finiteness of human cognitive resources and the uneven distribution of these resources. Based on Markov chains or hidden semi-Markov processes, this model generates a time-varying sequence of attention weights and applies these weights to perceptual signals in different modalities or spatial regions. When attention is focused on a specific task area, the signal-to-noise ratio of that area is significantly improved, and detailed features are enhanced; conversely, when attention is diverted or the task is switched, signals in non-focused areas are significantly suppressed or even ignored, simulating cognitive blind spots that humans cannot see or hear.

[0090] Through a triple coupling process involving nonlinear mapping, threshold drift, and attention fluctuations, the originally determined multimodal physical truth data is transformed into a set of fuzzy perception vectors with a dynamic confidence distribution. This vector not only contains signal strength information but also embeds a reliability assessment of the perception at the current moment, i.e., a confidence score. This score directly reflects the degree of certainty of the virtual subject's judgment of the signal, becoming the direct decision-making basis for triggering subsequent active exploration behavior.

[0091] This application conceives and employs a complex biosensory simulation mechanism stemming from a profound insight into the differences in human-machine decision-making during troubleshooting of high-end equipment. Experience shows that seasoned experts can often intuitively detect anomalies before instrument readings trigger alarms, while data-driven models often overreact to minute numerical fluctuations or completely ignore key signs. Human experts' decisions are based on subjective perceptions filtered, distorted, and uncertain through their own sensory systems, rather than absolute physical truths. Existing corpora, derived directly from sensor logs or simulated truths, are too perfect and deterministic, meaning models have never learned how to reason under conditions of information ambiguity and perceptual uncertainty. To compensate for this deficiency, cognitive noise must be explicitly injected into the data generation process. However, simple random noise cannot simulate the structured biases of human perception; therefore, this application cross-disciplinaryly incorporates findings from cognitive neuroscience and psychophysics.

[0092] Weber-Fechner's law reveals the logarithmic relationship between stimulus intensity and sensory quantity, providing a theoretical foundation for the design of nonlinear transfer functions. Meanwhile, the concepts of sensitivity and decision bias in signal detection theory inspire the construction of dynamic confidence distributions and threshold drift mechanisms. By engineering these two theoretical concepts into computable algorithmic modules, this application successfully concretizes the abstract human cognitive mechanism into a data processing flow, achieving a fundamental transformation from physical signals to cognitive representations.

[0093] In step S2, the system transforms the precise physical truth into a fuzzy perceptual vector with a dynamic confidence distribution, simulating the cognitive uncertainty and hesitation experienced by human experts in complex situations. However, merely remaining at the level of perceptual fuzziness is insufficient to reproduce the complete decision-making logic of human experts. In real-world scenarios, experts, when faced with low-confidence signals, will not passively accept ambiguous results or directly offer speculative conclusions. Instead, they will proactively initiate a series of targeted exploratory behaviors to eliminate uncertainty. If this crucial interaction link from doubt to verification is missing from the corpus, the trained model will still lack the ability to handle incomplete information and will be unable to learn how to plan verification paths. Therefore, it is necessary to establish a mechanism that can dynamically trigger and plan the optimal exploration strategy based on perceptual confidence, transforming static perceptual data into a dynamic interactive process.

[0094] In step S3, the system continuously monitors the confidence scores corresponding to each fault hypothesis in the fuzzy perception vector output in step S2 and compares them with a preset dynamic uncertainty threshold. This threshold is not a fixed constant, but is adaptively adjusted according to the safety level, urgency, and historical false alarm rate of the current task to ensure that potential uncertainties are captured more sensitively in high-risk scenarios.

[0095] Once the confidence level of a certain modality or fault hypothesis is detected to be below a certain threshold, indicating that the virtual agent is in a state of high hesitation or cognitive ambiguity, the proactive perception and planning mechanism is immediately activated. At this time, the system defines the current ambiguous perception state as the initial belief state and constructs a high-dimensional virtual action space containing multiple feasible operations based on the digital twin environment. Each potential action is regarded as an operator that can change the observed state of the system and acquire new information.

[0096] For example, this action space encompasses all physically executable exploratory behaviors, such as controlling a virtual camera to perform multi-angle surround shooting to obtain visual features from different perspectives, instructing a virtual robotic arm to repeatedly perform assembly actions to reproduce auditory abnormalities, switching to infrared thermal imaging or ultrasonic detection modes to obtain the distribution of hidden physical quantities, or adjusting ambient lighting and background noise to enhance signal contrast, etc.

[0097] Building upon this, step S3, based on the partially observable Markov decision process framework, filters out action sequences from the vast action space that minimize system entropy, i.e., maximize information gain. The system predicts the probability distribution of possible observations after executing each candidate action, calculates the difference in probability distribution of belief states before and after the action, typically using KL divergence or mutual information as metrics, thereby quantifying the expected information gain for each action.

[0098] To quantify information gain and guide the generation of action sequences, this step employs a mathematical model based on mutual information to evaluate the value of actions. A representative formula for calculating information gain can be expressed as: ,in, This represents the expected information gain from performing action a; Current state of belief That is, the Shannon entropy of the probability distribution corresponding to the fuzzy perception vector output in step S2 is used to measure the current degree of uncertainty. The higher the entropy value, the more hesitant one is. It represents the mathematical expectation, which is the weighted average of all possible observations o; It is the predicted probability distribution of the observed result o after performing action a under the current belief; The posterior belief state is updated after observing the result o and performing the action a. The entropy. The physical meaning of this formula is: the value of an action is equal to the reduction in the uncertainty of the system before and after performing the action. The goal of the planning algorithm is to find the entropy that makes the system more stable. Maximize action A. In multi-step programming, this formula is extended to a sequential form, where the optimal sequence of actions is determined by recursively calculating the cumulative information gain over multiple future steps.

[0099] By introducing the principle of maximizing information gain, this application ensures that every exploratory action generated is purposeful and logical, rather than blind trial and error. This mechanism not only greatly improves the efficiency and relevance of data generation, but more importantly, it records the complete thought process evolution from uncertainty to certainty, providing an indispensable dynamic context for the subsequent generation of thought chain corpora containing rigorous logical chains.

[0100] The virtual exploration action sequence that maximizes information gain refers to a series of ordered operations optimized by the aforementioned planning algorithm, which can eliminate current cognitive ambiguity with the fewest actions and the highest efficiency. This sequence is not generated by random trial and error, but rather by deep reasoning based on prior knowledge of the fault mechanism and the current perceptual state.

[0101] For example, when the system hesitates about the assumption of wear on the bearing inner ring and the initial auditory signal is masked by noise, the planning algorithm calculates that reducing the rotational speed has the highest information gain because this action significantly suppresses background noise and highlights the wear characteristic frequencies. Next, the algorithm might plan to record near the sound source as a follow-up action to further amplify the signal strength. This combination of actions constitutes an optimal exploration sequence for the current specific uncertainty. The specific algorithm that generates this sequence typically employs heuristic search strategies such as Monte Carlo tree search or gradient-based optimization methods, performing multi-step deductions in a virtual action space to simulate virtual observations after executing different action combinations, evaluating the contribution of these results to updating the posterior probability of the fault hypothesis, and finally outputting the action path that makes the expected posterior probability distribution sharpest, i.e., minimizes uncertainty.

[0102] In step S3, the system generates and executes a sequence of virtual exploration actions aimed at maximizing information gain through an active perception planning mechanism, successfully acquiring supplementary observational data that can correct the initial judgment. However, simply viewing this new data independently or simply replacing the old data will not reproduce the deep cognitive logic of human experts in synthesizing multiple pieces of evidence and dynamically correcting their beliefs during the decision-making process. Real fault diagnosis is not an instantaneous result of a single observation, but a continuous thinking process that deeply integrates initial vague impressions with subsequent verification evidence and iteratively updates the probability distribution. If the corpus lacks this evolutionary trajectory of belief updates, the trained model will struggle to grasp how to weigh evidence from different sources and will be unable to generate a thought chain that reflects the logical correction process. Therefore, it is necessary to construct a mechanism that can mathematically integrate initial vague perceptions with newly added observational data and map the probability evolution at the numerical level into natural language logical reasoning.

[0103] In step S4, the system receives the initial fuzzy perception vector from step S2 as the basis for prior beliefs. This vector carries the initial judgment with an uncertain distribution formed by the virtual subject during the first observation. At the same time, it receives supplementary observation data obtained after executing the virtual exploration action sequence in step S3. This data represents new evidence obtained after actively verifying specific doubts.

[0104] The fusion process is not a simple weighted average, but rather a rigorous probabilistic update framework built upon Bayes' theorem. The system transforms the initial fuzzy perception vector into a prior probability distribution of the fault hypothesis and converts supplementary observation data into a likelihood function, representing the probability of observing current evidence given the fault hypothesis. Subsequently, Bayes' formula is used to calculate the posterior probability distribution, which integrates prior knowledge and new evidence, reflecting the system's updated understanding of the true state of the fault after exploratory actions. During this process, if supplementary evidence strongly supports a hypothesis, the posterior probability of that hypothesis will increase significantly; if the evidence contradicts the hypothesis, its probability will decrease accordingly. This fusion mechanism ensures that each new observation scientifically corrects existing beliefs, simulating the thought process of human experts in discerning truth from falsehood.

[0105] To concretely implement probability fusion and logistic mapping, this step constructs a dynamic mapping function based on Bayesian updates. A representative formula for calculating the posterior probability can be expressed as: ,in, Let E represent the i-th fault hypothesis, and let E represent the set of supplementary observational evidence obtained in step S3. The prior probability from step S2 is obtained by normalizing the confidence distribution of the initial fuzzy perception vector; It is the likelihood, calculated by supplementing observation data with fault assumptions. The degree of matching between the theoretical feature models is obtained; the denominator is a normalization constant to ensure that the sum of the posterior probabilities of all hypotheses is 1.

[0106] Based on this, step S4 maps the aforementioned numerical probability iteration process into a dynamic thought chain text in natural language form. The system internally maintains a logic template engine that can parse key node data in the Bayesian update process in real time, including the initial distribution of prior probabilities, feature descriptions of new evidence, the calculation results of the likelihood ratio, and the drift direction and magnitude of the posterior probability.

[0107] The mapping process follows a rigorous logical syntax: First, a hypothesis proposal statement is generated based on the initial fuzzy perception vector, describing the virtual subject's initial suspected object and its degree of uncertainty. For example, an abnormal vibration signal is initially detected, but due to background noise interference, it cannot be confirmed whether it is bearing wear and the current confidence level is low. Second, combined with the virtual exploration actions performed in step S3 and the supplementary data obtained, an action verification statement is generated, explaining the reasons for taking specific verification measures and the new phenomena observed. Finally, based on the convergence result of the posterior probability, a belief update and decision convergence statement is generated, summarizing the degree of support of the evidence for the hypothesis and drawing a final conclusion.

[0108] In the mapping phase, the system defines a set of rules for the correspondence between logical connectors and probability change intervals:

[0109] When the posterior probability increment is greater than the high threshold, the strong confirmation logic template is triggered, using words such as "confirm" and "confirmed".

[0110] When the increment of the posterior probability is less than the negative high threshold, a strong negation logic template is triggered, using words such as exclusion and overturning.

[0111] When the absolute value of the probability increment is small, a weak correction is triggered or the template needs further verification, using euphemistic expressions such as "inclination" or "increased probability".

[0112] In addition, the system will also include likelihood The specific numerical characteristics are transformed into natural language descriptions.

[0113] In step S4, the system constructs a dynamic thought chain text containing a complete cognitive evolution trajectory, transforming the abstract Bayesian probability update process into a natural language sequence with a rigorous logical structure. However, if these high-quality corpora are merely stored in a database without effective model adaptation, the inherent uncertainty handling and proactive verification logic cannot be internalized into the inherent capabilities of the language model. While general pre-trained language models possess a strong foundation in language generation, their parameter space primarily encodes statistical patterns in internet text, lacking specialized knowledge representations for fuzzy perception reasoning and dynamic decision-making planning in industrial scenarios. Directly applying an untuned model to handle such tasks often leads to it ignoring key logical transitions in the corpus, degenerating into shallow responses based on keyword matching. Therefore, S5 employs a targeted instruction fine-tuning strategy to efficiently inject the structured thought chain corpus generated in step S4 into the model parameters, reshaping the model's reasoning path and transforming it from a passive text writer into an active reasoner with human-like cognitive mechanisms.

[0114] Step S5 converts the dynamic thought chain text generated in step S4 into a standard instruction fine-tuning data format. Each data sample consists of three parts: the instruction part clearly defines the current task as fault diagnosis and reasoning based on multimodal perception information; the input part contains the fuzzy perception vector description generated in step S2 and the description of the virtual exploration action sequence executed in step S3; the output part fully includes the dynamic thought chain text generated in step S4, that is, a natural language sequence containing hypothesis proposal, evidence verification, belief update and final conclusion.

[0115] Subsequently, these samples are mixed with a small number of high-quality cases annotated by real experts in a certain proportion to form the final dedicated instruction fine-tuning dataset, in order to prevent the model from overfitting to the distribution characteristics generated by the simulation.

[0116] Building upon this, step S5 employs full-parameter fine-tuning or efficient parameter fine-tuning techniques to deeply optimize the pre-trained language model. A thought chain-aware loss function is used, which not only calculates the error in the model's final conclusion but also penalizes breaks and inconsistencies in the reasoning logic chain. Specifically, during backpropagation, the system assigns higher weight coefficients to key logical nodes in the thought chain and their subsequent inferences, forcing the model to pay close attention to the coherence of causal deduction. During training, the model receives fuzzy perceptions and action descriptions from the input and attempts to autoregressively generate the thought chain text from the output. If the intermediate reasoning steps generated by the model do not match the Bayesian update logic in the standard answer...

[0117] For example, prematurely drawing definitive conclusions when evidence is insufficient, or ignoring crucial falsification steps, will cause the loss function to generate a large gradient signal, driving adjustments to the model parameters. Through tens of thousands of iterations of training, the model gradually learns not to rush to output conclusions in low-confidence states, but rather to simulate the thinking patterns of human experts, first constructing hypotheses internally, planning verification paths, and then revising beliefs based on virtual or real feedback, ultimately outputting a highly reliable diagnostic report.

[0118] To quantify the training effect and guide parameter updates, this step employs a cross-entropy-based loss function and introduces a logistic consistency regularization term. The total loss function can be expressed as follows: ,in, The fuzzy perception and action description sequence representing the input. It is the thought chain text token generated at time t. This represents the conditional probability of the model generating the correct token; the first term is the standard cross-entropy loss, used to ensure the fluency and accuracy of the generated text. It is a logical consistency regularization term used to penalize logical contradictions in the reasoning process. For example, when the model generates a description of insufficient evidence and then immediately generates a conclusion confirming a fault, this term will produce a huge penalty value. This is a balancing coefficient used to adjust the strength of logical constraints. During training, an adaptive learning rate optimizer, combined with a cosine annealing learning rate scheduling strategy, is employed to ensure that the model quickly converges to a general solution space in the early stages of training, and then fine-tunes the logical inference parameters in the later stages.

[0119] This application employs a dynamic thought chain-based instruction fine-tuning strategy, overcoming the drawback of traditional fine-tuning methods that prioritize results over process. Conventional supervised fine-tuning often trains only using input and final label pairs, causing the model to learn to quickly provide answers but lose the intermediate reasoning ability to handle complex uncertainties. When facing critical faults with weak signals and ambiguous features in high-end equipment manufacturing, this fast-thinking mode is highly prone to misjudgment.

[0120] Exemplary system:

[0121] like Figure 2 As shown, the corpus-based language model generation system includes the following modules: digital twin simulation module, biological perception simulation module, active perception planning module, dynamic thought chain generation module, and instruction fine-tuning module.

[0122] The digital twin simulation module is used to construct a digital twin simulation environment containing parameterized fault perturbations and run simulations to acquire multimodal physical truth data sequences. Internally, the module integrates the geometric topology model of the target high-end equipment, a material physical property library, and a set of dynamic differential equations, and deploys a multimodal virtual sensor array covering vision, hearing, thermal imaging, and force sensing. This module defines and executes the parameterized fault perturbation space, systematically traversing high-dimensional variable spaces such as fault type, injection location, severity coefficient, and evolution time constant through orthogonal experimental design or Latin hypercube sampling algorithms. The module runs a full-scenario simulation test case set, synchronously acquiring the system's internal state vectors and raw signals from external sensors at each time step. It utilizes a timestamp alignment algorithm based on maximizing mutual information to eliminate phase delays in heterogeneous data streams, ultimately outputting a multimodal physical truth data sequence with accurate physical truth labels. This data sequence, serving as a clean physical benchmark, is transmitted in real-time to the next-level biosensing simulation module.

[0123] The biosensing simulation module receives multimodal physical truth data sequences from the parametric simulation construction module, aiming to simulate the cognitive uncertainty and perceptual ambiguity experienced by human experts under complex conditions. This module embeds a family of nonlinear transfer functions designed based on Weber-Fechner's law, constructing independent biosensing sub-modules for visual, auditory, and other channels. Based on psychophysical principles, it nonlinearly reconstructs the input precise physical signals. Internally, the module operates a dynamic sensory threshold drift mechanism and an attention fluctuation model, calculating dynamic perception thresholds and attention weighting factors in real time based on the virtual subject's fatigue accumulation, adaptation history, and task focus. It transforms the defined physical signals into neural excitation patterns with inherent ambiguity, simultaneously calculating the confidence score at each moment. After processing by this module, the originally clear physical data stream is transformed into a set of fuzzy perception vectors with a dynamic confidence distribution. These vectors not only contain signal strength information but also embed a reliability assessment of perception, and are subsequently transmitted to the active perception planning module.

[0124] The active perception planning module receives fuzzy perception vectors from the biosensing simulation module and is responsible for addressing the problem that static observation cannot eliminate cognitive uncertainty. This module continuously monitors the confidence scores of each fault hypothesis in the fuzzy perception vectors and compares them to a preset dynamic uncertainty threshold. Once a confidence score is detected to be below the threshold, indicating that the virtual subject is in a state of high hesitation, the module immediately activates a planning algorithm based on a partially observable Markov decision process (POMDP) ​​framework. It constructs a high-dimensional virtual action space within the digital twin environment, incorporating feasible operations such as multi-angle shooting, repetitive operations, and switching detection modes, and quantifies the expected information gain of each candidate action using a mathematical model based on mutual information. Through heuristic strategies such as Monte Carlo tree search, the module selects action combinations that minimize system entropy, generating a virtual exploration action sequence that maximizes information gain. This sequence is then sent to the digital twin environment for execution, and the acquired supplementary observation data, along with the initial fuzzy perception vectors, is transmitted to the dynamic thought chain generation module.

[0125] The dynamic thought chain generation module receives initial fuzzy perception vectors and supplementary observation data, and is responsible for mapping the numerical probability evolution into natural language logical reasoning. Based on Bayes' theorem, this module constructs a rigorous probability update framework, transforming the initial fuzzy perception vectors into prior probability distributions and the supplementary observation data into likelihood functions, iteratively calculating the posterior probability distribution of the fault hypothesis. Internally, the module maintains a logic template engine that parses key node data in the Bayesian update process in real time, including prior distributions, new evidence features, likelihood ratios, and posterior probability drift magnitudes. According to preset rules corresponding to logical connectors and probability change intervals, the module automatically maps the probability convergence process into dynamic thought chain text encompassing the entire process of hypothesis formulation, action verification, belief updating, and decision convergence. This structured corpus fully records the thought evolution trajectory from "doubt" to "verification" and then to "confirmation," and is subsequently fed into the instruction fine-tuning training module.

[0126] The instruction fine-tuning training module receives structured corpora from the dynamic thought chain generation module and is responsible for the construction and optimization of the final language model. This module transforms the dynamic thought chain text into a standard instruction fine-tuning data format. Each sample contains the task instruction, input parts including fuzzy perception and action description, and the output part of the complete thought chain text. The module employs full-parameter fine-tuning or efficient parameter fine-tuning techniques to deeply optimize the pre-trained language model. Its core lies in applying a thought chain perception loss function, which not only calculates the cross-entropy loss of the generated text but also introduces a logical consistency regularization term to penalize logical breaks and contradictions in the reasoning process. Through the backpropagation algorithm, the module assigns higher weights to key logical nodes in the thought chain, forcing the model to learn the coherence of causal reasoning. After tens of thousands of iterations of training, the parameter space of the general language model is reshaped, internalizing higher-order cognitive mechanisms for handling uncertainty and performing proactive verification, ultimately outputting a target language model with uncertainty reasoning capabilities, completing the closed-loop operation of the entire system.

[0127] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A corpus-based language model generation method, characterized in that, include: S1. Construct a digital twin simulation environment containing parameterized fault disturbances and run the simulation to collect multimodal physical truth data sequences; S2. Input the multimodal physical truth data sequence into the biological perception simulation layer, use the nonlinear transfer function to simulate the human sensory threshold drift and attention fluctuation, and output a fuzzy perception vector with dynamic confidence distribution. S3. Real-time monitoring of the dynamic confidence distribution of the fuzzy perception vector. When the confidence is lower than the uncertainty threshold, the active perception planning mechanism is automatically triggered to generate and execute a virtual exploration action sequence that maximizes information gain. S4. The posterior probability of the fault hypothesis is iteratively updated based on the Bayesian inference algorithm by fusing the fuzzy perception vector with the supplementary observation data obtained after executing the virtual exploration action sequence, and the probability convergence process is mapped into a dynamic thought chain text containing hypothesis proposal, verification and correction logic. S5. The language model is fine-tuned using structured corpus of dynamic thought chain texts to obtain a language model with uncertainty reasoning ability.

2. The corpus-based language model generation method according to claim 1, characterized in that: In step S2, a nonlinear transfer function is used to simulate the drift of human sensory thresholds and fluctuations in attention, including: Based on the principles of psychophysics, independent biosensory sub-modules were constructed for different sensory channels, namely vision, hearing, and force. Within each submodule, a nonlinear transfer function designed based on Weber-Fechner's law is used to transform the input precise physical signal into a neural excitation pattern with inherent fuzziness. At the same time, a dynamic sensory threshold drift mechanism is implemented to adjust the perception threshold in real time based on the virtual subject's fatigue accumulation and adaptation history, simulating missed detections due to fatigue or false alarms due to alertness. Furthermore, an attention fluctuation model is introduced to generate a time-varying sequence of attention weights, which enhances or suppresses the perception signals of different spatial regions or modalities. Finally, a confidence score representing the reliability of perception is calculated and output synchronously.

3. The corpus-based language model generation method according to claim 1, characterized in that: In step S3, a virtual exploration action sequence that maximizes information gain is generated and executed, including: Define the current fuzzy perception state as the initial belief state, and construct a virtual action space in the digital twin environment that includes multi-angle observation, repetitive operation, switching detection modes, and adjusting environmental parameters; Based on the partially observable Markov decision process framework, predict the probability distribution of the possible observation results after executing each candidate action; Calculate the difference in probability distribution of belief states before and after executing each candidate action to quantify the expected information gain for each action; A heuristic search strategy is used to perform multi-step deductions in the virtual action space to select the action combination that minimizes the uncertainty of the expected posterior probability distribution, which is then issued and executed as the virtual exploration action sequence.

4. The corpus-based language model generation method according to claim 1, characterized in that: In step S4, the probability convergence process is mapped to dynamic thought chain text, including: Maintain a logic template engine to parse key node data in the Bayesian update process in real time. The key node data includes the initial distribution of prior probability, feature description of new evidence, calculation results of likelihood ratio, and drift direction and magnitude of posterior probability. Based on the initial fuzzy perception vector, generate statements describing the initial suspected objects and the degree of uncertainty; By combining the virtual exploration actions performed and the supplementary observation data obtained, action verification statements are generated that explain the rationale for the verification measures and the new phenomena. Based on the convergence results of the posterior probability and the pre-defined rules for the correspondence between logical connectors and probability change intervals, a belief update and decision convergence statement is generated to summarize the degree of evidence support and draw a final conclusion. Among them, a strong confirmation logic is used when the posterior probability increment exceeds the high threshold, a strong negation logic is used when the posterior probability increment is lower than the negative high threshold, and a weak correction or a euphemistic expression requiring further verification is used when the absolute value of the probability increment is small.

5. The corpus-based language model generation method according to claim 1, characterized in that: In step S1, a digital twin simulation environment containing parameterized fault disturbances is constructed, including: Based on the geometric topology, material physical properties, and kinematic constraints of the target high-end equipment, a full-element digital mapping model is reconstructed in virtual space, and the model is embedded with a set of differential equations describing the dynamic behavior of the equipment. Define a parameterized fault disturbance space, and set the fault type, injection location, severity coefficient, and evolution time constant as independently configurable variables; By using orthogonal experimental design or Latin hypercube sampling algorithm, a systematic traversal is performed in the parameterized fault disturbance space to generate a full-scenario simulation test case set covering normal operating conditions to extreme failure boundaries; The digital twin environment is driven to execute the full-scenario simulation use case set. Using a timestamp alignment algorithm based on maximizing mutual information, heterogeneous multimodal data streams are mapped to a unified time axis to form a multimodal physical truth data sequence with accurate physical truth labels.

6. The corpus-based language model generation method according to claim 5, characterized in that: The severity coefficient in the parameterized fault disturbance space is used to linearly or nonlinearly control the magnitude of the fault effect, simulating a continuous change from minor anomalies to catastrophic failures. The parameterized fault disturbance space also includes a spatial distribution function, which defines the range and intensity distribution of the fault in the equipment geometric space according to the physical mechanism of the fault. It also includes a time evolution function that describes the dynamic characteristics of the fault as it changes over time from the moment it occurs. The time evolution function can take the form of a step function, a ramp function, or an exponential growth function, respectively simulating sudden fracture, progressive wear, or thermal runaway processes.

7. The corpus-based language model generation method according to claim 1, characterized in that: In step S3, the uncertainty threshold is a dynamically adjusted threshold. Its value is adaptively adjusted according to the safety level, urgency and historical false alarm rate of the current task, so as to ensure that the sensitivity of capturing potential uncertainties in high-risk scenarios is higher than that in low-risk scenarios.

8. The corpus-based language model generation method according to claim 1, characterized in that: In step S5, the language model is fine-tuned using instructions, including: The dynamic thought chain text is converted into a standard instruction fine-tuning data format. Each data sample includes an instruction part that defines the task as fault diagnosis and reasoning based on multimodal perception information, an input part that includes fuzzy perception vector description and virtual exploration action sequence description, and an output part that includes the complete thought chain text. The generated samples are mixed with high-quality cases annotated by real experts in a certain proportion to form a dedicated instruction fine-tuning dataset; The pre-trained language model is trained using a loss function that is aware of thought chains. This loss function not only calculates the error in the model's final conclusion but also introduces a logical consistency regularization term to penalize breaks and inconsistencies in the reasoning logic chain and assigns higher weight coefficients to key logical nodes in the thought chain and their subsequent inferences.

9. The corpus-based language model generation method according to claim 1, characterized in that: In step S2, the output fuzzy perception vector not only contains signal strength information, but also embeds the reliability assessment of the perception at the current moment, namely the confidence score. The confidence score is lowest when the perception intensity is close to the decision boundary, indicating that the virtual subject is in a state of hesitation, and highest when the perception intensity tends to the extreme value and the attention is focused.

10. A corpus-based language model generation system, comprising the corpus-based language model generation method according to any one of claims 1-9, characterized in that, include: The digital twin simulation module is used to build a digital twin simulation environment containing parameterized fault disturbances and run simulations to acquire multimodal physical truth data sequences. The biosensory simulation module is used to receive the multimodal physical truth data sequence, simulate human sensory threshold drift and attention fluctuation using a nonlinear transfer function, and output a fuzzy perception vector with dynamic confidence distribution. The active perception planning module is used to monitor the dynamic confidence distribution of the fuzzy perception vector in real time. When the confidence is lower than the uncertainty threshold, the active perception planning mechanism is automatically triggered to generate and execute a virtual exploration action sequence that maximizes information gain. The dynamic thought chain generation module is used to fuse the fuzzy perception vector with the supplementary observation data obtained after executing the virtual exploration action sequence, iteratively update the posterior probability of the fault hypothesis based on the Bayesian inference algorithm, and map the probability convergence process into a dynamic thought chain text containing hypothesis proposal, verification and correction logic. The instruction fine-tuning module is used to fine-tune the language model using the structured corpus of the dynamic thought chain text, so as to obtain a language model with uncertainty reasoning ability.

Images