Data standard consensus generation method and platform based on generative AI

Abstract

The application provides a generative AI-based data standard consensus generation method and platform, relates to the technical field of AI and data standard generation, and comprises the following steps: collecting a heterogeneous data set containing reaction principles, electrical signal modulation parameters and hierarchical control logic, and generating a three-domain joint embedding vector; constructing a generative backbone network, inserting a three-domain cross attention layer, and introducing a physical and electrical constraint dynamic mask to generate a standard draft; mapping device, edge and platform layer differentiation views, and iteratively closing the loop based on on-site electrical signal feedback; the platform is correspondingly configured with six modules such as three-domain perception collection and joint embedding preprocessing, supports rapid asset import and source code consistency checking, and breaks the limitation of traditional standards focusing only on text semantics; the generated dynamic executable standard adapts to complex working conditions, solves the problem of cross-domain data interaction landing failure, and improves the reliability of industrial scene data interaction.

Description

Technical Field

[0001] This invention relates to the field of AI and data standard generation technology, and in particular to a data standard consensus generation method and platform based on generative AI. Background Technology

[0002] With the advancement of the Fourth Industrial Revolution, global manufacturing is accelerating its transformation from digitalization to networking and intelligence. The Industrial Internet, as a new type of infrastructure, aims to break down information silos and achieve comprehensive interconnection between people, machines, and things. Under this grand vision, data has become the lifeblood of intelligent manufacturing. However, the long-standing reality in industrial settings is a chaotic mix of equipment from various countries: PLCs (Programmable Logic Controllers) from different brands such as Siemens, Mitsubishi, and Omron coexist; various sensors use heterogeneous protocols such as Modbus, Profinet, and EtherCAT; and even for the same physical quantity (such as temperature, pressure, and voltage), there are significant gaps in definition, accuracy, and semantics across different devices, times, and environments. Therefore, the generation of cross-domain data standard consensus has become a major bottleneck restricting the large-scale implementation of the Industrial Internet. Traditional industrial data standard consensus generation mainly relies on a human-led expert seminar model. Typically, leading companies or standardization organizations take the lead, convening representatives from equipment manufacturers, software developers, and end-users to discuss and align the names, types, units, and ranges of thousands of data fields over months or even years. This model is essentially "textual semantic alignment," producing a static, rigid, and idealized protocol document. However, the complexity of industrial environments far exceeds what text can encompass. Existing technologies reveal fatal flaws when dealing with real physical and electrical environments, primarily manifested in three significant disconnects: First, current standard consensus only reaches the literal matching of field names, completely ignoring the physical generation mechanisms behind the data. Take pH monitoring, the most common issue in industrial environments, as an example. Sensors from different manufacturers are based on different electrochemical principles. Manufacturer A might use a composite glass electrode, whose output follows the Nernst equation, with a logarithmic relationship between potential and hydrogen ion activity, and is affected by temperature; Manufacturer B might use a metal oxide electrode, whose response mechanism and temperature coefficient are drastically different. In the existing standard consensus process, both parties only agree on the field name "pH" and the range "0-14". However, due to the neglect of reaction principles, readings from the two "consensus" standards will exhibit systematic deviations when ambient temperature fluctuates or the ionic strength of the solution changes. This deviation is not a communication error, but rather a "different meanings for the same name" at the physical mechanism level. Although generative AI performs well in natural language processing, its training corpus is mostly text, lacking a deep understanding of the symbolic logic of physicochemical equations. This results in standards generated by it failing to anchor to physical essence, easily producing "illusionary" clauses that violate energy conservation or matter balance. Second, data in standard documents is pure numbers, but in the physical link, it exists in the form of noisy electrical signals. Existing standard setters usually lack a strong hardware background, and electrical characteristics are never specified in standard documents. Taking the BMS (Battery Management System) of an energy storage power station as an example, the acquisition of cell voltage relies on the ΔΣ modulation and analog-to-digital conversion of the AFE (Analog Front-End) chip.In this process, sampling clock jitter, power supply ripple, PCB wiring crosstalk, and the effective number of bits (ENOB) of the ADC all introduce microscopic errors. For example, a certain model of AFE has a minimum waiting window of 420 microseconds from the CONVST trigger signal to the completion of the conversion. If CPU interrupt nesting causes the trigger pulse phase jitter to be ±12 SYSCLK cycles, it will cause the voltage reading of the same cell to jump by ±15 millivolts in the static state. The standards generated by the existing technology cannot perceive these electrical noises. When the power supply ripple in the field is as high as 120mVpp, the standard still requires the data accuracy to reach 1mV, which is not only impossible to achieve, but also causes the edge computing layer to fall into the "crying wolf" dilemma due to frequent false alarms. More importantly, the electrical environment in the industrial field is dynamic. High temperature will cause the crystal oscillator frequency deviation to increase, and sudden load changes will cause power supply fluctuations. Static standard texts are powerless to deal with this. Third, the Industrial Internet is a typical hierarchical architecture, including the device layer (millisecond level), the edge layer (second level), and the platform layer (minute level). Existing standard consensus typically adopts a "one-size-fits-all" approach, generating a single, universal field definition expected to fit all layers. However, each layer has drastically different data requirements. The device layer focuses on the real-time nature and determinism of raw data, the edge layer focuses on statistical characteristics and anomaly filtering, and the platform layer focuses on long-term trends and storage efficiency. For example, if the device layer refreshes data at 100ms intervals and the edge layer aggregates data in 1-second windows, data truncation or overlap during edge layer aggregation will inevitably occur if their intervals are not integer multiples of each other. Existing technologies rely on manually writing complex synchronization code to solve this problem, which is not only inefficient but also prone to errors. Generative AI, when generating standards, often treats the hierarchical architecture as background information rather than constraints on the generation logic, leading to extensive secondary development for hierarchical mapping during standard implementation.

[0003] To overcome these shortcomings, the industry has attempted to introduce federated learning or purely generative AI technologies. Federated learning focuses on privacy-preserving model aggregation, with its optimization goal being the convergence of model parameters rather than the physical compliance and electrical robustness of the data standards themselves. It cannot solve the underlying data drift problem caused by electrical signal jitter. While purely generative AI (such as large language models) can assist in writing standard documents and accelerate text alignment, its lack of analytical capabilities regarding reaction principle equations, perception of electrical signal modulation parameters, and modeling of hierarchical control logic often results in "standards" that are superficial and even contain numerous clauses that violate physical principles. For example, while existing patented technologies include standard matching schemes based on semantic embedding, their embedding vectors are derived solely from text annotations and are not integrated into a symbol tree structure; while there are synchronization methods based on dynamic modulation, they only focus on link-layer timing and do not address upper-layer semantic governance; and while there are hierarchical data processing architectures, their mapping rules are mostly statically configured and not dynamically coupled with the standard generation process.

[0004] In summary, existing data standard consensus technologies, when faced with the complex physical environment, harsh electrical conditions, and stringent hierarchical control requirements of industrial sites, suffer from several challenges. These include the failure of cross-domain data interaction, data bias in model training, and misjudgments in operational decisions. Furthermore, the stability of noise separation and the interpretability of diagnostic conclusions fail to simultaneously meet engineering requirements. Therefore, there is an urgent need for a novel data standard consensus generation method and platform based on generative AI. Summary of the Invention

[0005] The purpose of this invention is to provide a data standard consensus generation method and platform based on generative AI, in order to solve the technical bottlenecks in existing technologies, such as focusing only on text semantic alignment, not anchoring the underlying reaction mechanism of physical quantities, being unable to adapt to the dynamic modulation characteristics of underlying electrical signals, and static solidification of hierarchical control mapping rules. These bottlenecks lead to large voltage jumps in BMS cells of energy storage power stations, high SOC estimation errors, high failure rate of pre-thermal runaway signals, and "same name, different meaning" fields in multi-brand PLC mixed production lines, frequent false alarms at the edge layer. Furthermore, existing solutions also address the lack of analytical capabilities of pure generative AI in physicochemical equations, the tendency to generate illusionary clauses that violate energy conservation, federated learning only optimizing model aggregation parameters and not covering the physical compliance and electrical robustness of the standard itself, the need for manual writing of cross-layer synchronization code for standard implementation, iteration lag, and high failure rate of cross-domain data interaction implementation. The specific technical solution is as follows: This invention provides a data standard consensus generation method based on generative AI, including: S10: Collect heterogeneous data sources from multiple participants to construct a heterogeneous dataset containing reaction principles, electrical signal modulation parameters, and hierarchical control logic. S20, normalize the heterogeneous dataset and generate a three-domain joint embedding vector with uniform dimension; S30, construct a generative backbone network, and input the three-domain joint embedded vector into the network for feature fusion and context modeling; S40 introduces a dynamic masking mechanism based on physical and electrical constraints during the decoding phase to generate a preliminary standard consensus draft. S50 performs hierarchical projection mapping to generate differentiated standard views that adapt to the device layer, edge layer, and platform layer; S60 performs closed-loop iterative optimization based on on-site electrical signal feedback, and outputs the final executable standard consensus.

[0006] Furthermore, S10 includes: extracting the reaction principle equations of the corresponding physical quantities for each data source participating in the standard consensus, and parsing the equations into a symbol tree structure; the symbol tree structure contains operator nodes, variable nodes, and constant nodes, and records the dimensions of each node; synchronously collecting the electrical signal modulation parameters and timing distribution of the physical link layer, and encapsulating them into an electrical parameter package; the electrical signal modulation parameters include the root mean square value of the sampling clock jitter, the peak-to-peak value of the power supply ripple, the effective number of bits of the analog-to-digital converter, and the distribution of the bus frame interval.

[0007] Further, S20 includes: constructing a reaction principle embedder, mapping the symbol tree into a high-dimensional vector, retaining dimensional constraint information as an attention bias during the mapping process; constructing an electrical signal embedder, inputting the normalized scalar parameters in the electrical parameter package into a multilayer perceptron, inputting the shape parameters of the distribution parameters into an independent coding layer, and outputting an electrical feature vector; and performing weighted fusion of text annotation embedding, reaction principle embedding, and electrical signal embedding to generate a three-domain joint embedding vector; wherein, the weight coefficients are preset according to the sensitivity of the physical quantity to the physical mechanism.

[0008] Furthermore, S30 includes: employing a encoder-decoder generative backbone network based on a transformer architecture, setting the hidden layer dimension to 1024, the number of attention heads to 16, and loading weights that have been pre-trained on industry standard corpora during training; at the encoder input, the three-domain joint embedding vector is used as the backbone feature, and state encoding of hierarchical control logic is introduced; the hierarchical control logic includes a device layer refresh cycle of 100ms, an edge layer aggregation window width of 1s, and a platform layer database entry cycle of 60s; before the fully connected layer of the network decoder head, a three-domain cross-attention layer is inserted to force the decoding process to query the corresponding reaction principle embedding and electrical signal embedding when generating each field definition.

[0009] Furthermore, S40 includes: defining a dynamic tolerance interval generation function to calculate the tolerance half-width of each field in real time based on the electrical parameter package; constructing a reaction principle consistency mask: when the value range of the decoded field conflicts with the output range of the reaction principle equation, the generation probability of the corresponding position of the field is set to zero; constructing an electrical signal feasibility mask: when the accuracy required by the generated standard clause is higher than the accuracy supported by the electrical parameter package, the generation of the clause is forcibly blocked.

[0010] Furthermore, S50 includes: generating a device layer projection view: retaining the original data type, byte order, and refresh period of 100ms, and adding a tolerance interval calculated by the dynamic tolerance interval generation function; generating an edge layer projection view: automatically generating an aggregation operator, defined as a truncated mean operator; generating a platform layer projection view: mapping the fields to the storage structure of a time series database, defining a downsampling strategy to retain the mean, maximum, minimum, and standard deviation, with a storage period of 60s.

[0011] Furthermore, in S40, if the peak-to-peak value of the power supply ripple in the electrical parameter package exceeds a preset threshold of 100mV, an enhanced saturation limiting strategy is adopted. This strategy specifically includes: calculating the last bit skipping amplitude caused by ripple noise, wherein the effective number of bits of the analog-to-digital converter is set to 12 bits; temporarily extending the dynamic tolerance half-width to accommodate ripple interference; calling the enhanced saturation limiting function when decoding the generated field value range to allow the value to fluctuate within a larger range; and adding annotation labels to the generated standard consensus draft to indicate the high ripple environment and enable dynamic extended tolerance. If the reaction principle equation is a nonlinear high-order differential equation and involves multivariate coupling, a symbolic tree pruning masking strategy is adopted, specifically including: parsing the symbolic tree, identifying the dominant and minor terms in the equation, and calculating the sensitivity contribution of each variable node; if the field value range in the generated standard draft causes the contribution of the minor term to exceed 20% of the dominant term, it is determined to be a physical mechanism imbalance; generating a mechanism imbalance mask, multiplying the generation probability within the value range by a decay coefficient of 0.01, forcing the model to regenerate a numerical range that conforms to the conservation law; and mapping the corrected numerical range back to the range or alarm threshold field of the standard draft. If there are non-integer multiples of refresh cycles across layers in the hierarchical control logic, an asynchronous mapping correction mask strategy is adopted. Specifically, this includes: detecting the ratio of the device layer refresh cycle to the edge layer aggregation window width; if the ratio is not an integer, it is determined that there is an asynchronous mapping risk; calculating the boundary effect loss of the least common multiple cycle to determine the data truncation error at the edge of the aggregation window; generating an asynchronous correction mask, prohibiting the generation of static mapping rules that may lead to data loss, and forcing the generation of elastic mapping rules that include linear interpolation or previous value preservation logic; and generating a cross-layer synchronization description appendix in the consensus draft to describe in detail the data processing logic under non-integer multiple cycles.

[0012] Furthermore, in S50, if the truncation loss calculated by the edge layer aggregation operator is greater than a preset threshold of 5%, an adaptive truncation coefficient adjustment strategy is adopted. Specifically, this includes: statistically analyzing the proportion of sample points removed by the truncation function when the truncation coefficient is set to 5%; if the proportion is greater than 5%, it is determined that the current truncation coefficient is too strict, and the truncation coefficient is automatically lowered to 3%, and the aggregation operator is recalculated; the quality control indicators in the edge layer projection view are updated synchronously, and the mean deviation before and after the truncation adjustment is recorded; the adjusted truncation coefficient is fed back to the generative backbone network as prior knowledge for the generation of similar scenarios in the future. If the platform layer projection view involves long-term archiving requirements and the data retention period is greater than one year, a hot and cold data tiered storage mapping strategy is adopted. Specifically, this includes: analyzing the access frequency of fields and determining whether a physical quantity is a critical process parameter based on the reaction principle equation; if it is a critical process parameter, defining hot data area, warm data area, and cold data area in the platform layer view, each corresponding to a different downsampling granularity; generating a cross-storage area index mapping table to ensure that query requests can be automatically routed to the corresponding storage area; and adding a data lifecycle management strategy to the consensus draft to clarify the rolling update mechanism of each storage area. If the confidence level of the electrical signal embedding in the three-domain joint embedding vector is lower than the threshold of 0.8, a conservative projection generation strategy is adopted, which specifically includes: retrieving processing cases of similar low-confidence scenarios from the historical database and extracting the corresponding degradation processing operators; in the device layer view, marking the data quality label of the field as pending verification and forcibly enabling the original value upload mode; in the edge layer view, disabling complex aggregation operators, using simple arithmetic average instead, and expanding the uncertainty label; in the platform layer view, generating data completion prompts, and suggesting adding confidence level color warnings to the display interface.

[0013] Further, in step S60, if the feedback signal is negative for three consecutive iterations, a physical constraint-enhanced retraining strategy is adopted, specifically including: pausing the execution of the current standard consensus, locking the mapping relationship between the device layer and the edge layer to prevent data overflow; backtracking the collected original electrical signal data, refitting the jitter distribution parameters, and updating the electrical parameter package; inputting the updated electrical parameter package into the generative backbone network, freezing the text embedding layer weights, and only fine-tuning the reaction principle embedder and the electrical signal embedder; re-executing the steps from generating the preliminary draft to generating the differentiated view, generating the revised standard consensus draft, and conducting comparative tests to verify the revision effect; If the hierarchical control logic changes, for example, the device layer refresh cycle is adjusted from 100ms to 200ms, a hierarchical projection cascade update strategy is adopted, which specifically includes: capturing hierarchical change events and analyzing the parameter differences before and after the change; recalculating the optimal value of the edge layer aggregation window width based on the time constant of the reaction principle equation to maintain the sample number unchanged; cascading update of the platform layer data entry cycle to ensure that the aggregation window width is an integer multiple of the refresh cycle, and updating the sample number value in the mapping operator; generating a change impact analysis report, automatically notifying the affected upstream and downstream systems, and updating the version number of the standard document; If the generative backbone network frequently triggers the reaction principle consistency mask during the decoding phase, a symbol tree-guided contrastive learning optimization strategy is adopted. This strategy includes: recording the high-frequency fields that trigger the reaction principle consistency mask and their corresponding error generation sequences to construct a hard example sample library; adding symbol tree contrastive loss to the training objective of the generative backbone network to narrow the vector distance between correct physical quantity values ​​and the symbol tree output features, while widening the distance between error values ​​and the symbol tree; adjusting the fusion weights in the three-domain joint embedding vector to increase the weight of reaction principle embedding and strengthen the role of physical constraints in feature fusion; and incrementally training the backbone network using the corrected hard example sample library until the frequency of triggering the reaction principle consistency mask drops below 1%.

[0014] This invention also provides a data standard consensus generation platform based on generative AI for implementing the method, the platform comprising: The three-domain sensing and acquisition module is used to acquire heterogeneous data sources that include reaction principles, electrical signal modulation parameters, and hierarchical control logic. The joint embedding preprocessing module is used to transform heterogeneous data sources into three-domain joint embedding vectors of uniform dimension; The generative backbone inference module is used to perform feature fusion and context modeling; The dynamic mask quality control module is used to introduce physical and electrical constraints during the decoding stage to filter out illegal outputs; The hierarchical projection mapping module is used to generate differentiated standard views that adapt to the device layer, edge layer, and platform layer; The closed-loop iterative optimization module is used to continuously optimize the model and standards based on on-site feedback.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method described herein.

[0016] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method.

[0017] The beneficial effects of this invention are as follows: This invention eliminates the semantic bias of "same name, different meaning" cross-domain data fields at the source through three-domain joint representation, solving the problem of implementation failure caused by static standards being unable to adapt to complex industrial conditions; it introduces a dynamic masking mechanism based on physical constraints and electrical noise, effectively avoiding the illusion problem of generative AI, ensuring that the output standard terms conform to basic laws such as energy conservation and matter balance, and adapting to strong interference conditions such as power supply ripple exceeding 120mVpp and clock jitter reaching 12ns; the generated differentiated hierarchical projection view can adapt to the device layer, edge layer, and platform without manually writing cross-layer synchronization code. The heterogeneous control requirements of the layer reduce the voltage jump amplitude of the BMS cells in energy storage power stations, reduce the missed detection rate of thermal runaway pre-signals, and decrease the false alarm rate of the edge layer. The closed-loop iterative mechanism based on field electrical signal feedback enables the standard to have dynamic evolution capabilities, which can adapt to changes in operating conditions such as device aging and environmental temperature changes, and shorten the standard revision cycle. The supporting platform functions can complete the import of standard assets of existing industrial systems within 30 minutes, improve the consistency verification efficiency of design assets and edge computing source code, reduce the workload of secondary development for cross-domain integration, and improve the accuracy, reliability and full life cycle operation and maintenance efficiency of data interaction in the industrial Internet scenario.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram illustrating the steps of the data standard consensus generation method based on generative AI of the present invention; Figure 2 This is a schematic diagram of the data standard consensus generation platform based on generative AI of the present invention. Detailed Implementation

[0020] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0022] It should also be noted that when a component is described as "fixed to" or "set on" another component, it can be directly on the other component or there may be an intervening component present. When a component is described as "connected to" another component, it can be directly connected to the other component or there may be an intervening component present.

[0023] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0024] In embodiments of the present invention, a data standard consensus generation method based on generative AI is provided; please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: S10: Collect heterogeneous data sources from multiple participants to construct a heterogeneous dataset containing reaction principles, electrical signal modulation parameters, and hierarchical control logic. S20, perform normalization preprocessing on the heterogeneous dataset to generate a three-domain joint embedding vector with a unified dimension; S30, construct a generative backbone network, and input the three-domain joint embedded vector into the network for feature fusion and context modeling; S40 introduces a dynamic masking mechanism based on physical and electrical constraints during the decoding phase to generate a preliminary standard consensus draft. S50 performs hierarchical projection mapping to generate differentiated standard views that adapt to the device layer, edge layer, and platform layer; S60 performs closed-loop iterative optimization based on on-site electrical signal feedback, and outputs the final executable standard consensus.

[0025] The principle and effects of the above technical solution are as follows: This invention breaks through the limitation of traditional standard consensus focusing only on textual semantics, and innovatively integrates physical reaction mechanisms, dynamic modulation characteristics of electrical signals, and hierarchical control architecture into the entire process of encoding, decoding, and iteration in generative AI. The method generates not static documents, but executable standards with physical compliance, electrical robustness, and hierarchical adaptability. This solves the problem of standard implementation failure in multi-source heterogeneous systems under complex operating conditions due to underlying electrical noise, differences in physical mechanisms, and disconnection in hierarchical control, thus improving the accuracy and reliability of cross-domain data interaction.

[0026] In one embodiment, S10 includes: S101, for each data source participating in the standard consensus, extract the reaction principle equation of the corresponding physical quantity and parse the equation into a symbol tree structure; specifically, for industrial acid and alkalinity sensors, extract the Nernst equation, parse it into a symbol tree containing operator nodes, variable nodes and constant nodes, and record the dimensions of each node, where the unit of thermodynamic temperature is Kelvin and the hydrogen ion activity is a dimensionless value. S102 synchronously acquires the electrical signal modulation parameters and timing distribution of the physical link layer, including the root mean square value of the sampling clock jitter, the peak-to-peak value of the power supply ripple, the effective number of bits of the analog-to-digital converter, and the distribution of the bus frame interval. Specifically, it acquires the clock jitter distribution of the RS485 bus at 9600 baud rate using a high-precision oscilloscope, fits it to a normal distribution, and records the ripple parameter of the power supply module as a peak-to-peak value of 120mV, and encapsulates these parameters into an electrical parameter package.

[0027] The principle and effect of the above technical solution are as follows: Step S10 constructs a heterogeneous data source that has never been addressed in existing technical solutions. Unlike traditional methods that only collect field names and units, this step incorporates the physical laws that determine the nature of the data and the electrical characteristics that determine the data quality into the scope of data collection. This provides a real anchor point in the physical and electrical worlds for standard consensus, avoiding numerical drift problems caused by different names and electrical noise from the source.

[0028] In one embodiment, S20 includes: S201, Construct a reaction principle embedder to map the symbol tree into a high-dimensional vector, retaining dimensional constraint information as an attention bias during the mapping process; S202, construct an electrical signal embedder, input the normalized scalar parameters in the electrical parameter package into a multilayer perceptron, input the shape parameters of the distributed parameters into an independent coding layer, and output an electrical feature vector; S203, weighted fusion of text annotation embedding, reaction principle embedding and electrical signal embedding to generate a three-domain joint embedding vector; wherein, the weight coefficient is preset according to the sensitivity of the physical quantity to the physical mechanism, and the reaction principle embedding is given a higher weight for fields with strong mechanism dependence such as acidity and alkalinity.

[0029] The principle and effect of the above technical solution are as follows: Step S20 achieves homogeneous representation of heterogeneous data. By introducing a learnable weighted fusion mechanism, the generative model can simultaneously read the textual definition, mathematical structure, and noise level of the underlying electrical signal when processing fields. This fusion method allows subsequent AI models to no longer treat standards as isolated sequences of words, but as engineering entities governed by physical laws and subject to electrical noise interference, laying the foundation for generating physically feasible standards.

[0030] In one embodiment, S30 includes: S301 employs a encoder-decoder generative backbone network based on a transformer architecture, with a hidden layer dimension of 1024 and 16 attention heads. During training, weights that have already been pre-trained on industry-standard corpora are loaded. S302, at the encoder input end, the three-domain joint embedding vector is used as the backbone feature, and the state encoding of the hierarchical control logic is introduced at the same time; the hierarchical control logic includes a device layer refresh cycle of 100ms, an edge layer aggregation window width of 1s, and a platform layer data entry cycle of 60s. S303 inserts a three-domain cross-attention layer before the fully connected layer of the network decoder, forcing the decoding process to query the corresponding reaction principle embedding and electrical signal embedding when generating each field definition, thus achieving deep coupling at the feature level.

[0031] The principle and effect of the above technical solution are as follows: Step S30 ensures the dominant role of three-domain information in the generation process by improving the network structure. Traditional generative models tend to ignore the physical constraints behind standard text when processing it. This step, by inserting a three-domain cross-attention layer, is equivalent to equipping the AI ​​model with a physical electrical calculator. This forces the model to refer to the calculation results of physical equations and the resolution limitations of the analog-to-digital converter when generating key parameters such as range or accuracy, thereby preventing the generation of standard clauses that violate physical common sense or exceed hardware capabilities.

[0032] In one embodiment, S40 includes: S401 defines a dynamic tolerance range generation function that calculates the tolerance half-width of each field in real time based on the electrical parameter package; S402, Construct a reaction principle consistency mask: When the value range of a field generated by decoding conflicts with the output range of the reaction principle equation, set the generation probability of the corresponding position of that field to zero; S403, Construct an electrical signal feasibility mask: When the generated standard clause requires a higher precision than the electrical parameter package supports, force the generation of that clause to be masked.

[0033] The principle and effects of the above technical solution are as follows: Step S40 is the core quality control link in the generation process. By introducing a dual masking mechanism, the generated content is hard-filtered from two dimensions: physical laws and electrical characteristics. The reaction principle consistency mask ensures that the standard clauses do not violate fundamental laws such as energy conservation or matter balance, while the electrical signal feasibility mask prevents setting excessively high precision requirements in environments with strong interference. This mechanism effectively suppresses the illusion problem common in generative models, ensuring the scientific validity and engineering feasibility of the output standard.

[0034] In one embodiment, S50 includes: S501, Generate device layer projection view: retain the original data type, byte order and refresh cycle of 100ms, and attach the tolerance interval calculated by the dynamic tolerance interval generation function; S502, Generate Edge Layer Projected View: Automatically generate aggregation operators, defined as truncated mean operators, with the following formula: ,in This is the edge layer aggregation operator; f is the field index used to distinguish different monitoring quantities: such as f=1 voltage, f=2 temperature, to achieve multi-channel parallel calculation; This is the original sampling sequence of the device; The equations are constrained by physical principles and are in a symbolic tree structure, used to verify the physical rationality of the output results. The electrical signal modulation parameter package includes underlying noise indicators such as clock jitter and power supply ripple, and dynamically generates the normal fluctuation range of the signal; sat(⋅) is the saturation limiting function, defined as sat(x,[L,U])=max(L,min(U,x)), which clamps out-of-range impulse noise and eliminates spike interference; The nominal working value for field f is determined by physical constraints. The derived center value under normal operating conditions is used as the midpoint of the tolerance interval; The dynamic tolerance half-width is determined by the electrical signal parameters. The determined function ensures that the tolerance range can cover the underlying electrical noise; N is the total number of samples within the aggregation window, N=W e / Δt d ;W e For the edge layer aggregation window width, Δt d The device layer refresh cycle is κ; κ is the truncation coefficient, a dimensionless floating-point number (range 0 < κ < 0.5), determined by the generative model based on the reaction principle. The noise sensitivity is dynamically generated to define the proportion of extrema to be removed; S503, Generate Platform Layer Projection View: Map fields to the storage structure of a time series database, define downsampling strategy as retaining mean, maximum, minimum and standard deviation, storage period is 60s.

[0035] The principle and effect of the above technical solution are as follows: Step S50 solves the compatibility problem of standards in control systems at different levels. Existing technologies typically output a general static standard, which leads to an excessive burden on the device layer or overly coarse data at the platform layer. This step, by generating a hierarchical projection view containing specific aggregation operators, achieves lightweight execution of the standard at the device layer, anti-interference processing at the edge layer, and fine-grained storage at the platform layer, perfectly adapting to the hierarchical architecture requirements of the Industrial Internet.

[0036] In one embodiment, S60 includes: After deployment on-site, S601 continuously monitors the actual distribution of electrical parameter packages, especially the actual root mean square value of sampling clock jitter. S602 calculates the deviation between the theoretical tolerance range and the actual electrical noise. When the deviation is negative, it is determined that the current standard tolerance range is too tight and cannot cover the actual noise. S603 uses the deviation as a feedback signal to send back to the generative backbone network, adjusting the weight coefficients embedded in the reaction principle during the next round of consensus generation. If the deviation is negative, the weight coefficient is increased to enhance the weight of physical constraints.

[0037] The principle and effect of the above technical solution are as follows: Step S60 constructs a closed-loop iterative optimization capability for the standard. Traditional standards become fixed once published, making it difficult to adapt to the slow changes in the field environment. This step feeds back the measured electrical noise data from the field to the generation model, dynamically adjusting the tolerance range and model weights in the standard, giving the standard an evolutionary capability similar to biological adaptability, and ensuring the effectiveness of the standard throughout its entire life cycle.

[0038] In one specific embodiment, in S40, if the peak-to-peak value of the power supply ripple in the electrical parameter package is greater than a preset threshold of 100mV, an enhanced saturation limiting strategy is adopted, and the specific steps are as follows: SA41: Calculates the last bit skipping amplitude caused by ripple noise, where the effective number of bits of the analog-to-digital converter is set to 12 bits.

[0039] SA42: Dynamically tolerates half-width temporary extension to accommodate ripple interference.

[0040] SA43: When decoding the generated field value range, an enhanced saturation limiting function is called to allow the value to fluctuate within a wider range.

[0041] SA44: Add annotation labels to the generated standard consensus draft to indicate high ripple environments and enable dynamic extension tolerance.

[0042] The principle and effect of the enhanced saturation limiting strategy are as follows: In industrial environments with strong electromagnetic interference or poor power quality, power supply ripple is one of the main causes of data anomalies. Existing technologies typically solve this through hardware filtering, which is costly. This step intelligently identifies high ripple characteristics during the standard generation stage and dynamically expands the tolerance range, accommodating unavoidable electrical noise at the data standard level. This ensures data authenticity while avoiding numerous false alarms triggered by minor ripples, thus improving the system's robustness.

[0043] In S40, if the reaction principle equation is a nonlinear high-order differential equation and involves multivariate coupling, then a symbolic tree pruning masking strategy is adopted. The specific steps are as follows: SB41: Parse the symbolic tree to identify the dominant and minor terms in the equation and calculate the sensitivity contribution of each variable node.

[0044] SB42: If the field value range in the generated draft standard causes the contribution of the minor item to exceed 20% of the dominant item, it is judged as a physical mechanism imbalance.

[0045] SB43: Generation mechanism imbalance mask, which multiplies the generation probability within the range by a decay coefficient of 0.01, forcing the model to regenerate a numerical range that conforms to the conservation law.

[0046] SB44: Map the revised numerical range back to the range or alarm threshold field in the draft standard.

[0047] The principle and effect of the symbol tree pruning masking strategy are as follows: For complex physical systems, AI models may generate standards that appear numerically reasonable but violate physical intuition. This step introduces symbol tree parsing into a masking mechanism, enabling the AI ​​to possess physical intuition. By quantifying the contribution of each term in the equation, it ensures that the generated standard always remains within the dominant domain of physical principles, effectively preventing the spread of generative model illusion problems in the field of standard setting.

[0048] In S40, if there are non-integer multiple relationships between refresh cycles across levels in the hierarchical control logic, an asynchronous mapping correction mask strategy is adopted. The specific steps are as follows: SC41: Detects the ratio of the device layer refresh cycle to the edge layer aggregation window width. If the ratio is not an integer, it is determined that there is a risk of asynchronous mapping.

[0049] SC42: Calculate the boundary effect loss of the least common multiple period to determine the data truncation error at the edge of the aggregation window.

[0050] SC43: Generates asynchronous correction masks, prohibits the generation of static mapping rules that may lead to data loss, and forces the generation of flexible mapping rules that include linear interpolation or previous value preservation logic.

[0051] SC44: Generate a cross-layer synchronization specification appendix in the consensus draft, detailing the data processing logic under non-integer multiple periods.

[0052] The principle and effect of the asynchronous mapping correction mask strategy are as follows: In industrial settings, the sampling frequencies of different devices are often difficult to be completely consistent, leading to gaps or overlaps during edge layer aggregation. Existing technologies typically ignore this problem or rely on manually writing complex synchronization code. This step identifies the risk of non-integer multiple periods during the standard generation stage and automatically generates flexible mapping rules containing interpolation algorithms, encapsulating the complexity of hierarchical control within the standard, greatly reducing the difficulty and error rate of subsequent system integration.

[0053] In one specific embodiment, in S50, if the truncation loss calculated by the edge layer aggregation operator is greater than a preset threshold of 5%, an adaptive truncation coefficient adjustment strategy is adopted, and the specific steps are as follows: SA51: Statistics on the percentage of sample points removed by the truncation function when the truncation coefficient is set to 5%.

[0054] SA52: If the percentage is greater than 5%, the current truncation coefficient is determined to be too strict. The truncation coefficient is automatically lowered to 3%, and the aggregation operator is recalculated.

[0055] SA53: Synchronously update the quality control metrics in the edge layer projection view and record the mean deviation before and after truncation adjustment.

[0056] SA54: The adjusted truncation coefficients are fed back to the generative backbone network as prior knowledge for the generation of similar scenarios in the future.

[0057] The principle and effect of the adaptive truncation coefficient adjustment strategy are as follows: While truncation of the mean can effectively remove impulse noise, excessive truncation can lead to the loss of effective information. This step establishes a self-evaluation mechanism for the aggregation operator's effect, dynamically fine-tuning the truncation coefficient by monitoring the truncation loss. This mechanism ensures that while meeting anti-interference requirements, it maximizes the preservation of the statistical characteristics of the original data, enabling the generated standard to achieve an expert-level optimization level during the data cleaning process.

[0058] In S50, if the platform layer projection view involves long-term archiving requirements and the data retention period is greater than one year, a cold and hot data tiered storage mapping strategy is adopted. The specific steps are as follows: SB51: Analyze the access frequency of the field and determine whether the physical quantity is a key process parameter based on the reaction principle equation.

[0059] SB52: If it is a critical process parameter, define a hot data area, a warm data area, and a cold data area in the platform layer view, each corresponding to a different downsampling granularity.

[0060] SB53: Generates an index mapping table across storage areas to ensure that query requests can be automatically routed to the corresponding storage area.

[0061] SB54: Add a data lifecycle management strategy to the consensus draft and clarify the rolling update mechanism for each storage area.

[0062] The principle and effect of the described hot and cold data tiered storage mapping strategy are as follows: With the explosive growth of industrial data, a one-size-fits-all storage strategy would lead to extremely high storage costs. This step automatically plans a tiered storage architecture based on the criticality of physical quantities. This not only optimizes the use of storage resources but also solves the persistent problem of difficult cross-period historical data queries by defining a unified index mapping at the standard level, reflecting the forward-looking consideration of standard consensus for future system operation and maintenance.

[0063] In S50, if the confidence level of the electrical signal embedding in the three-domain joint embedding vector is lower than the threshold of 0.8, a conservative projection generation strategy is adopted. The specific steps are as follows: SC51: Retrieve similar low-confidence scenario processing cases from the historical database and extract the corresponding downgrade processing operators.

[0064] SC52: In the device layer view, mark the data quality label of the field as pending verification and force the raw value upload mode to be enabled.

[0065] SC53: In the edge layer view, disable complex aggregation operators, use simple arithmetic averages instead, and expand the uncertainty label.

[0066] SC54: In the platform layer view, generate data completion prompts. It is recommended to add confidence level color warnings to the display interface.

[0067] The principle and effect of the conservative projection generation strategy are as follows: When the underlying electrical signal features are blurry or the noise is excessive, forcibly generating a high-precision standard view is irresponsible. This step empowers the AI ​​model to recognize what it knows and acknowledge what it doesn't. By identifying low-confidence scenarios and automatically generating conservative views that include degradation processing and early warning prompts, the decision-making risks caused by unknown data quality are effectively avoided, demonstrating the method's deep design in terms of security and reliability.

[0068] In one specific embodiment, in S60, if the feedback signal is negative for three consecutive iterations, a physical constraint-enhanced retraining strategy is adopted, and the specific steps are as follows: SA61: Suspends the execution of the current standard consensus, locks the mapping relationship between the device layer and the edge layer, and prevents data overflow.

[0069] SA62: Backtrack the acquired raw electrical signal data, refit the jitter distribution parameters, and update the electrical parameter package.

[0070] SA63: Input the updated electrical parameter package into the generative backbone network, freeze the weights of the text embedding layer, and fine-tune only the reaction principle embedder and the electrical signal embedder, setting the fine-tuning learning rate to 1e. -5 .

[0071] SA64: Re-execute the steps from generating the initial draft to generating the differentiated view, generate the revised standard consensus draft, and conduct comparative tests to verify the effect of the revision.

[0072] The principle and effect of the physical constraint-enhanced retraining strategy are as follows: This constitutes a key link in the closed-loop iteration of the live standard in this invention. When the on-site electrical environment continuously deteriorates, static standards or one-time generation cannot adapt. This step establishes a feedback-based retraining mechanism to enable the standard to adapt to environmental changes. By targeted fine-tuning rather than full retraining, the model's response speed to changes is ensured while maintaining the stability of learned knowledge, thus ensuring the effectiveness of the standard in long-term operation.

[0073] In S60, if the hierarchical control logic changes, for example, the device layer refresh cycle is adjusted from 100ms to 200ms, then a hierarchical projection cascading update strategy is adopted, and the specific steps are as follows: SB61: Capture hierarchy change events and parse the parameter differences before and after the change.

[0074] SB62: Based on the time constant of the reaction principle equation, recalculate the optimal value of the edge layer aggregation window width to maintain a constant sample size.

[0075] SB63: Cascade update platform layer data entry cycle, ensure that the aggregation window width is an integer multiple of the refresh cycle, and update the sample quantity value in the mapping operator.

[0076] SB64: Generates a change impact analysis report, automatically notifies affected upstream and downstream systems, and updates the version number of standard documents.

[0077] The principle and effect of the hierarchical projection cascading update strategy are as follows: Existing revisions are often lagging and lack cross-layer impact analysis. This step incorporates the hierarchical control logic into the dynamic management scope of standard consensus. When the underlying control parameters change, the standard no longer passively waits for manual revision, but actively cascades and updates all relevant upper-layer views and mapping operators. This mechanism greatly shortens the system change cycle, eliminates systemic risks caused by parameter asynchrony, and achieves synchronized operation between the control architecture and data standards.

[0078] In S60, if the generative backbone network frequently triggers the reaction principle consistency mask during the decoding stage, a symbol tree-guided contrastive learning optimization strategy is adopted. The specific steps are as follows: SC61: Record the high-frequency fields of the consistency mask that triggers the reaction principle and their corresponding error generation sequences to build a difficult example sample library.

[0079] SC62: In the training objective of generative backbone networks, a symbol tree contrastive loss is added to narrow the vector distance between the correct physical quantity values ​​and the symbol tree output features, while widening the distance between the incorrect values ​​and the symbol tree.

[0080] SC63: Adjust the fusion weights in the three-domain joint embedding vector, increase the weight of the reaction principle embedding, and strengthen the role of physical constraints in feature fusion.

[0081] SC64: Incremental training of the backbone network is performed using a modified hard example library until the frequency of triggering the reaction principle consistency mask drops below 1%.

[0082] The principle and effect of the symbol tree-guided contrastive learning optimization strategy are as follows: When the AI ​​model frequently violates physical laws, it indicates that its understanding of the reaction principle is biased. This step uses the symbol tree itself as a supervision signal and uses contrastive learning to help the model understand the meaning of the physical formula. This method does not require manual intervention, realizes the self-correction of the AI ​​model, and improves the scientific accuracy of the generated content.

[0083] Accordingly, such as Figure 2 As shown, this invention provides a data standard consensus generation platform based on generative AI, which implements the data standard consensus generation method based on generative AI. The platform comprises six core modules: a three-domain perception acquisition module, a joint embedding preprocessing module, a generative backbone inference module, a dynamic mask quality control module, a hierarchical projection mapping module, and a closed-loop iterative optimization module. The three-domain perception acquisition module provides heterogeneous data sources containing principles, electrical signal modulation parameters, and control logic for subsequent processing; the joint embedding preprocessing module transforms the heterogeneous data sources into vector representations of a unified dimension; the generative backbone inference module performs feature fusion and context modeling; the dynamic mask quality control module introduces physical and electrical constraints during the decoding stage to filter illegal outputs; the hierarchical projection mapping module generates differentiated standard views adapted to different levels; and the closed-loop iterative optimization module continuously optimizes the model and standard based on on-site feedback. This embodiment of the generative AI-based data standard consensus generation platform clarifies the end-to-end processing logic from physical layer data acquisition to cross-domain semantic consensus. It not only defines the model elements and their attributes for each module but also precisely describes the data flow and constraint relationships between these model elements, providing a solid theoretical foundation and guiding framework for the platform's construction, deployment, and implementation. Each module is described in detail below.

[0084] I. Three-Domain Sensing and Acquisition Module The purpose of the three-domain sensing acquisition module is to break through the limitations of traditional data acquisition that only focuses on text and numerical values, and to construct a comprehensive dataset covering physical mechanisms, electrical characteristics, and system architecture. To this end, the three-domain sensing acquisition module is carefully constructed with three major sub-modules: a reaction principle acquisition unit, an electrical signal modulation acquisition unit, and a hierarchical logic acquisition unit.

[0085] The reaction principle acquisition unit encompasses four model elements: physical quantity identification, reaction principle equations, and symbol tree parsing. In this unit, physical quantity identification is a unique classification of the monitored physical quantity, independent of equipment manufacturer and communication protocol. Examples include pH and cell voltage (V) in industrial settings, clearly defining the physical nature of the data. The reaction principle equation is a mathematical expression describing the generation mechanism of the physical quantity, such as the Nernst equation or the Arrhenius equation, accurately reflecting the physicochemical laws of nature. Symbol tree parsing breaks down the reaction principle equation into a tree structure containing operator nodes, variable nodes, and constant nodes, facilitating machine understanding and calculation. The business object here is represented by physical entities, such as reaction vessels or energy storage cells, which constitute the physical basis of data acquisition.

[0086] The electrical signal modulation and acquisition unit focuses on the accurate acquisition of underlying electrical parameters. This unit includes four model elements: sampling clock jitter, peak-to-peak power supply ripple, effective bits of the analog-to-digital converter (ADC), and bus inter-frame interval. Sampling clock jitter records the deviation distribution of the clock edge relative to its ideal position; for example, the root mean square jitter of an RS485 bus at 9600 baud is 3.2 ns, which directly determines the microscopic noise level of the data. Peak-to-peak power supply ripple describes the interference amplitude caused by power supply instability, such as 120 mVpp, which is a key basis for evaluating data accuracy. The effective bits of the ADC determine the resolution of digital quantization, while the bus inter-frame interval reflects the transmission characteristics of the link layer. These elements together constitute the electrical parameter package, providing quantitative indicators for evaluating data quality.

[0087] The hierarchical logic acquisition unit focuses on the temporal architecture of the system operation. This unit includes three model elements: device-level refresh cycle, edge-level aggregation window width, and platform-level data ingestion cycle. The device-level refresh cycle is 100ms, representing the frequency of data acquisition in the physical world; the edge-level aggregation window width is 1s, defining the time span for statistical analysis at the edge; and the platform-level data ingestion cycle is 60s, determining the granularity of cloud storage. These parameters define the spatiotemporal boundaries of data flow and are prerequisites for ensuring cross-layer data consistency.

[0088] Through the collaborative work of the three sub-modules mentioned above, the three-domain perception and acquisition module achieves holographic perception of the data source, providing irrefutable factual evidence for the subsequent generation of standard consensus.

[0089] II. Jointly Embedded Preprocessing Module The joint embedding preprocessing module is further divided into three sub-modules: text embedding unit, reaction principle embedding unit, and electrical signal embedding unit, and includes a key weighted fusion unit.

[0090] The text embedding unit encompasses three model elements: field annotations, data types, and unit dimensions. Field annotations are human-language descriptions of physical quantities, such as "reaction vessel pH"; data types define the data storage format, such as Float or Int; and unit dimensions specify the measurement standard of the physical quantity, such as mol / L or degrees Celsius. These elements are mapped to high-dimensional vectors, preserving semantic information.

[0091] The reaction principle embedding unit includes symbol tree traversal, dimensional constraint encoding, and physical feature mapping. Symbol tree traversal visits each node of the symbol tree according to a depth-first or breadth-first strategy; dimensional constraint encoding transforms the dimensional information in the physical formula into an attention bias matrix to prevent the generation of standards that violate dimensional harmony; physical feature mapping transforms the structural features of the symbol tree into a fixed-length vector representation, enabling the AI ​​model to "understand" the meaning of the physical formula.

[0092] The electrical signal embedding unit involves scalar normalization, distributed parameter encoding, and noise feature extraction. Scalar normalization maps electrical parameters of different magnitudes (such as mV and ns) to a unified range; distributed parameter encoding extracts characteristic parameters such as shape and scale as input for the jitter distribution probability density function; noise feature extraction separates the feature values ​​of interference components such as ripple and spikes from the original waveform.

[0093] The weighted fusion unit linearly combines the embedding vectors from the three dimensions of text, reaction principle, and electrical signal to generate a three-domain joint embedding vector. The weight coefficients are preset based on the sensitivity of the physical quantity to the physical mechanism. For fields with strong mechanism dependence, such as acidity and alkalinity, the reaction principle embedding is given a higher weight (e.g., 0.4). For direct electrical measurements, the weight of the electrical signal embedding is increased (e.g., 0.35). Text embedding typically retains a basic weight (e.g., 0.25). This refined fusion mechanism ensures the complementary advantages of multi-source information.

[0094] III. Generative Main Reasoning Module The generative backbone inference module employs a transformer architecture as its core, with technical modeling consisting of three parts: the encoder, the decoder, and the cross-attention mechanism. First, it establishes a deep understanding of the input features by the encoder. Then, based on this understanding, the decoder generates a draft of the standard consensus, and the cross-attention mechanism ensures the consistency of the generated content with the three-domain constraints. This design philosophy emphasizes the leading role of physical constraints, ensuring the scientific rigor and engineering feasibility of the generated content.

[0095] In other words, the generative backbone inference module includes an encoder unit and a decoder unit. Based on deployment experience in industrial settings, the encoder unit is designed with three model elements: an input layer, a multi-head attention layer, and a feedforward neural network. The input layer receives the joint embedding vector of the three domains; the multi-head attention layer captures long-distance dependencies between fields; and the feedforward neural network performs nonlinear transformations of features. The decoder unit involves a masked multi-head attention layer, an encoder-decoder attention layer, and an output linear layer. The masked multi-head attention layer ensures the autoregressive nature of the generation process; the encoder-decoder attention layer (i.e., the three-domain cross-attention layer) forces the decoding process to query the corresponding reaction principle embedding and electrical signal embedding when generating each field definition, achieving deep coupling at the feature level; and the output linear layer maps the hidden states to probability distributions on the vocabulary.

[0096] The entire generative backbone inference module establishes a connection with the joint embedding preprocessing module through the three-domain joint embedding vector model element to achieve semantic and physical fusion; it also interfaces with the dynamic mask quality control module through the generated hidden states, thus realizing the module's extensibility. All model elements of the generative backbone inference module can be mapped one-to-one with the pre-trained weights of industry standard corpora, laying the foundation for high-quality standard generation.

[0097] IV. Dynamic Mask Quality Control Module The dynamic mask quality control module adopts a "constraint-first" design philosophy, performing quality control modeling from two parts: dynamic tolerance interval generation and dual mask filtering. First, it establishes a tolerance interval calculation method based on electrical parameter packages. Then, based on this, it executes reaction principle consistency masks and electrical signal feasibility masks, physically blocking unreasonable clauses.

[0098] In other words, the dynamic mask quality control module includes a tolerance interval calculation unit and a mask execution unit. Leveraging extensive industry expertise, the tolerance interval calculation unit designed a dynamic tolerance half-width generation function. This function is a mathematical function determined by electrical signal parameters, ensuring that the tolerance interval can cover underlying electrical noise, such as ripple or jitter.

[0099] The mask execution unit of the dynamic mask quality control module involves two model elements: a reaction principle consistency mask and an electrical signal feasibility mask. The reaction principle consistency mask is a filtering mechanism based on the range constraints of the reaction principle equation. When the range of the decoded field conflicts with physical laws (such as energy conservation), the generation probability of the corresponding position of that field is set to zero. The electrical signal feasibility mask is a filtering mechanism based on the precision constraints of electrical parameter packets. When the precision required by the generated standard clause (such as 0.1mV) exceeds the actual hardware support capability (limited by ripple and ENOB), the generation of that clause is forcibly blocked.

[0100] The entire dynamic mask quality control module establishes a relationship with the three-domain perception and acquisition module through the electrical parameter package model element to achieve the matching of quality control rules with the actual situation on site; it also establishes a weak relationship with the generative backbone inference module through the generation probability model element to dynamically intervene in the output during the decoding stage and realize the configurability of the quality control strategy.

[0101] V. Hierarchical Projection Mapping Module The hierarchical projection mapping module is responsible for transforming the unified standard draft into differentiated views adapted to different hierarchical control architectures, modeling from three parts: device layer projection, edge layer projection, and platform layer projection. Specifically, it first establishes the raw data format and byte order required by the device layer, then designs the aggregation operators and anti-interference logic for the edge layer based on this, and finally defines the storage structure and downsampling strategy for the platform layer. This design philosophy emphasizes the hierarchical adaptability of standard implementation, ensuring a smooth data flow from edge to cloud.

[0102] In other words, the hierarchical projection mapping module includes a device layer projection unit, an edge layer projection unit, and a platform layer projection unit. The device layer projection unit is designed with three model elements: data type, byte order, and refresh cycle. It preserves the original data characteristics and adds a dynamic tolerance range to ensure lightweight execution on the device side.

[0103] The edge layer projection unit involves three model elements: aggregation operator, truncation coefficients, and saturation limiting function. The aggregation operator is defined as a truncated mean operator based on physical constraints, and its mathematical expression is: ,in This is the edge layer aggregation operator; f is the field index used to distinguish different monitoring quantities: such as f=1 voltage, f=2 temperature, to achieve multi-channel parallel calculation; This is the original sampling sequence of the device; The equations are constrained by physical principles and are in a symbolic tree structure, used to verify the physical rationality of the output results. The electrical signal modulation parameter package includes underlying noise indicators such as clock jitter and power supply ripple, and dynamically generates the normal fluctuation range of the signal; sat(⋅) is the saturation limiting function, defined as sat(x,[L,U])=max(L,min(U,x)), which clamps out-of-range impulse noise and eliminates spike interference; The nominal working value for field f is determined by physical constraints. The derived center value under normal operating conditions is used as the midpoint of the tolerance interval; The dynamic tolerance half-width is determined by the electrical signal parameters. The determined function ensures that the tolerance range can cover the underlying electrical noise; N is the total number of samples within the aggregation window, N=W e / Δt d ;We For the edge layer aggregation window width, Δt d The device layer refresh cycle is κ; κ is the truncation coefficient, a dimensionless floating-point number (range 0 < κ < 0.5), determined by the generative model based on the reaction principle. The noise sensitivity is dynamically generated and used to define the proportion of extreme values ​​to be removed.

[0104] The platform-level projection unit involves three model elements: storage structure, downsampling strategy, and data lifecycle. The storage structure defines the table structure of the field in the time-series database; the downsampling strategy specifies the retention of statistical characteristics such as mean, maximum, minimum, and standard deviation; and the data lifecycle manages the tiered storage and expiration cleanup of hot and cold data.

[0105] The entire hierarchical projection mapping module establishes a relationship with the hierarchical logic acquisition unit through model elements such as refresh cycle and aggregation window width to achieve accurate matching between the standard view and the control system architecture; and through the core model element of aggregation operator, it achieves deep coupling between data processing logic and physical and electrical constraints.

[0106] VI. Closed-loop iterative optimization module The closed-loop iterative optimization module is responsible for the full lifecycle management after the standard is released, and models are built from three parts: feedback monitoring, deviation analysis, and model fine-tuning. First, a monitoring mechanism based on field electrical signal feedback is established. Then, the deviation between the theoretical standard and the actual working conditions is analyzed based on this. Finally, the deviation data is used to incrementally train the generative model to achieve the dynamic evolution of the standard.

[0107] In other words, the closed-loop iterative optimization module includes a feedback monitoring unit, a deviation analysis unit, and a model fine-tuning unit. The feedback monitoring unit is designed with three model elements: real-time data stream access, electrical signal parameter extraction, and distribution fitting. It continuously monitors the electrical parameter package at the site, especially the actual root mean square value of the sampling clock jitter, capturing minute changes in the environment.

[0108] The deviation analysis unit involves deviation calculation, threshold judgment, and attribution analysis. Deviation calculation is the difference between the theoretical tolerance range and the actual electrical noise; threshold judgment determines whether the deviation exceeds the allowable range (e.g., negative values ​​for three consecutive iterations); attribution analysis identifies the root cause of the deviation, whether it is due to device aging or environmental degradation.

[0109] The model fine-tuning unit involves parameter freezing, low-rank adaptation, and learning rate scheduling. Parameter freezing locks down common parameters such as those in the text embedding layer to prevent catastrophic forgetting; low-rank adaptation injects a small number of trainable parameters into the reaction principle embedding unit and the electrical signal embedding unit to achieve rapid adaptation; and learning rate scheduling sets the fine-tuning learning rate to 1e. -5 This ensures the stability of the training.

[0110] The entire closed-loop iterative optimization module forms a closed loop with the three-domain perception and acquisition module through the model element of feedback signal, realizing the circulation of data flow; and it connects with the generative backbone inference module through the updated model weights to drive the generation of the next round of standard consensus, reflecting the "liveliness" of the standard.

[0111] The data standard consensus generation platform based on generative AI in this embodiment of the invention also has a variety of functional mechanisms, which will be described in detail below.

[0112] Platform fast import mechanism In the initial implementation phase of the standard consensus platform, the largest investment lies in the standardized management of existing systems, a stage requiring substantial manual data entry and processing. To address this, the platform in this embodiment features a specially designed rapid import mechanism, aiming to complete the import of standard assets from an existing industrial system within 30 minutes, thereby improving efficiency and reducing labor costs. In practical applications, a set of standard assets is formed upon completion of platform modeling. If we consider model elements as database table structures, then standard assets are the table data.

[0113] Optionally, the rapid import mechanism of the data standard consensus generation platform based on generative AI in this embodiment can be divided into two major stages: reaction principle import and electrical parameter import.

[0114] The reaction principle import is based on the existing system design documents, PLC annotations, and equipment manuals. By deeply analyzing the mapping relationship between the document structure and platform model elements, and with the help of OCR and large model technology, the reaction principle equations and symbol trees are automatically extracted.

[0115] Electrical parameter import directly targets the historical time-series data (CSV or database backup) of the existing system, analyzes its statistical characteristics and noise distribution, fits an equivalent electrical parameter package (such as jitter distribution and ripple amplitude), and then uses the platform's backend parsing capabilities to store the fitting results in the three-domain sensing acquisition module, thereby achieving rapid and accurate import of electrical features.

[0116] Platform Standard Iteration AI Agent As industrial equipment is upgraded or processes are adjusted, data standards also enter a new iteration cycle. Through the platform's standard iteration AI Agent function, it can not only effectively manage this iteration process but also record change requests in detail for each iteration and deeply analyze their potential impact. The platform's technical modeling can be understood as the top-level design of the standards, while the go-live switch refers to the transition from using the new standard to the iteration cycle during operation and maintenance.

[0117] Users first input their iteration requirements through the platform's web page. The platform then combines existing assets from modules such as three-domain perception, joint embedding, and backbone inference to generate a dedicated standard iteration AI Agent for each version of the standard asset. Upon receiving a change request, the AI ​​Agent can intelligently identify the purpose of the change (such as "improving measurement accuracy" or "adding monitoring points") and automatically analyze the key impact points on the existing reaction principle, electrical signal tolerance range, and hierarchical mapping. It also presents multiple iteration solutions (such as adjusting the truncation coefficient or expanding the tolerance range) for selection. Users can manually confirm and optimize the solutions, enabling collaborative work with the AI ​​Agent to jointly promote efficient and accurate iteration of the standard asset.

[0118] Platform with multiple storage and version designs To ensure the parallel management of multiple factories and production lines in large group enterprises, it is usually necessary to prepare multiple logical environments to cope with the needs of different production lines. When multiple standard iteration requirements are distributed across different stages of R&D, each stage needs to release a corresponding standard version. Addressing this industrial management pattern, the platform has designed and implemented multi-space and multi-version functionality.

[0119] Within the platform, the storage of standard assets is clearly labeled with their respective storage space and version number. A storage space refers to the platform's multi-tenancy capability; one space is allocated to each factory or production line, and data is isolated between spaces. The front-end query and display also provides filtering options based on storage space and version number, ensuring the isolation of standard asset data viewed by the user across different spaces. Users can easily view the corresponding version's standard documents, aggregation operator definitions, and mapping relationships by switching between different versions, greatly improving the flexibility and convenience of version management and asset viewing.

[0120] Design of one-click generative conversion of platform model elements The aggregation operators and tolerance intervals in the platform-level projection mapping module have a certain connection with the physical quantity identifiers and electrical signal parameters of the three-domain sensing and acquisition module. Simultaneously, the device layer, edge layer, and platform layer views within the platform-level projection mapping module have derivative relationships. Based on this design, the platform in this embodiment achieves efficient one-click generation functionality: a corresponding standard draft framework can be generated with one click based on physical quantity identifiers; a dynamic tolerance interval can be generated with one click based on electrical signal parameters; and an edge layer aggregation operator can be generated with one click based on the refresh cycle and aggregation window width. Furthermore, the platform can automatically establish the relationships between these assets.

[0121] The specific conversion rules are as follows: the basic information such as the name and description of the physical quantity identifier is accurately converted into the name and description of the standard field, and its reaction principle equation is synchronously mapped into the physical constraint annotation of the field; the characteristics such as jitter and ripple in the electrical signal parameter package are meticulously converted into the half-width parameters of the tolerance interval; at the same time, the refresh cycle and aggregation window width in the hierarchical control logic are cleverly converted into the number of samples N and the truncation coefficient κ in the aggregation operator, realizing the smooth connection of the data processing logic.

[0122] The system can automatically generate and convert closely related model elements with a single click based on predefined model element conversion rules. Specifically, it can generate standard drafts based on physical quantity identifiers, tolerance ranges based on electrical signal parameters, and aggregation operators based on hierarchical logic. Furthermore, it can automatically establish the relationships between these assets. Users can then perform appropriate verification and fine-tuning of the generated model elements. This process quickly creates new standard assets and significantly reduces the amount of tedious data entry.

[0123] Platform Interface Design As a tool for carrying out data standard consensus processes, the platform has a clear functional division, which is reflected in the platform interface design.

[0124] The platform in this embodiment provides a user interface for user interaction from three dimensions: user interaction experience, standard governance services, and asset management, and the interfaces of each module are carefully designed.

[0125] In terms of user interaction experience, the platform provides capabilities such as a workbench, process guidance, space configuration, and quick access, aiming to create an intuitive and convenient operating environment for users.

[0126] Standard governance services, as the core of achieving three-domain integration, encompass several important aspects. Among them, three-domain perception includes input of reaction principles, configuration of electrical signal parameters, and preset of hierarchical logic; joint embedding and inference involves visualization of embedding vectors, generative model inference logs, and display of cross-attention weights; quality control and projection include mask trigger statistics, hierarchical view preview, and configuration for interfacing with downstream systems, ensuring the comprehensiveness and accuracy of standard services.

[0127] The asset management system features functions such as statistical reports, change request forms, quick asset import, access control, and version control, providing strong support for the effective management and utilization of assets.

[0128] Automatic generation of edge computing scripts based on platform technology modeling of design assets The model elements of the hierarchical projection mapping module can be mapped one-to-one with the configuration elements of edge computing frameworks (such as EdgeX Foundry or KubeEdge), thus automatically generating script code for edge-side data processing based on the design assets of the platform hierarchical projection mapping module.

[0129] The design assets formed by the model elements of the hierarchical projection mapping module mainly include aggregation operator definitions, tolerance interval parameters, data types, and byte order; the configurable assets mainly include refresh cycles, aggregation window widths, and storage strategies. Combining these two parts with the physical constants referenced by the design assets, the system jointly generates the corresponding edge computing framework configuration file and processing function scripts. The specific file mapping relationships are shown in Table 1 below: Design Asset and System Source Code File Mapping Relationship Table.

[0130] Table 1 Mapping Relationship between Design Assets and System Source Code Files

[0131] The system source code is reverse-engineered using AI capabilities to generate design assets modeled by the platform technology. Based on the mapping relationship between design assets and edge computing scripts built on the hierarchical projection mapping module, the platform can access the source code of edge applications on GitLab and parse it sequentially according to categories such as aggregation operators, tolerance intervals, data types, and configuration items. During the parsing process, the platform delves into the file structure and function definitions, and with the help of AIAgent's intelligent interpretation logic based on code interpretation of large models, it converts them into natural language and extracts the essence, ultimately constructing a complete technical modeling design asset.

[0132] Specifically, the platform parses the `processor.go` file to extract the mathematical logic of aggregation operators (such as truncated mean calculation), input and output parameters, and calls to the physical constant library. It uses an AI Agent to interpret the `config.yaml` file, describing the specific values ​​and constraints of configuration parameters such as tolerance intervals and refresh cycles in natural language. Similarly, it parses structure definition files such as `voltage.go` to generate information such as the names, descriptions, and attribute lists of physical quantities. Furthermore, the platform uses the AI ​​Agent to parse Dockerfiles and Deployment files, generating detailed asset information such as resource limitations and environment variables, thereby enabling reverse engineering from running code to designing assets.

[0133] Consistency verification mechanism for platform technology modeling assets and system source code The development of edge computing scripts strictly adheres to the design assets formed by platform technology modeling. To ensure the rigor of this process, this invention provides a precise integration mechanism, namely a consistency check mechanism. This consistency check mechanism can quickly pinpoint the specific inconsistencies when they occur between the technology modeling assets and the system source code. Furthermore, a highly interactive interface provides users with a clear and intuitive view, enabling them to easily and accurately identify and locate inconsistencies, thereby ensuring a smooth and efficient deployment process.

[0134] The current design assets include various asset types such as aggregation operators, tolerance ranges, data types, and configuration strategies. The consistency check interface performs a detailed check and comparison of the design assets and system source code according to these types, highlighting the differences. During the consistency check process, if the system source code exists but the design asset is missing, the corresponding asset will be marked "To be added," indicating that it needs to be added; conversely, if the design asset exists but the system source code is missing, it will be marked "To be deleted," indicating that it needs to be removed. For cases where both exist but the detailed information (such as the truncation coefficient κ or the tolerance half-width Δvtol) is inconsistent, it will be marked "To be modified," indicating that it needs to be updated. Users can use the multi-dimensional filtering function of the consistency check interface to quickly filter by asset type and difference type, facilitating centralized viewing of differences in a specific category and thus enabling efficient asset management.

[0135] Furthermore, the consistency verification system supports direct upload of source code packages for parsing and direct access to source code on GitLab, providing flexible source code access methods. Simultaneously, leveraging the capabilities of scheduled tasks, the consistency comparison triggering method implements three modes: manual immediate triggering, triggering at a specified time point, and periodic triggering at time intervals, to meet the verification needs of different scenarios and further enhance the intelligence and automation level of asset management.

[0136] Platform technology modeling asset and system source code quasi-consistency mechanism Leveraging the consistency check interface, the platform's preservation mechanism quickly ensures a high degree of consistency between technical modeling assets and system source code. For design assets marked "to be added," "to be deleted," and "to be modified" on the interface, the platform intelligently reads the system source code, accurately obtains relevant data, and automatically executes the "add," "delete," and "modify" operations for the design assets, ensuring that the technical modeling assets and system source code remain consistent. Furthermore, the platform supports users selecting multiple difference files in batches and quickly completing consistency updates for multiple files through a one-click synchronization function, further improving the convenience and accuracy of asset management.

[0137] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A data standard consensus generation method based on generative AI, characterized in that, include: S10: Collect heterogeneous data sources from multiple participants to construct a heterogeneous dataset containing reaction principles, electrical signal modulation parameters, and hierarchical control logic. S20, normalize the heterogeneous dataset and generate a three-domain joint embedding vector with uniform dimension; S30, construct a generative backbone network, and jointly embed the three domains into the vector input network for feature fusion and context modeling; S40 introduces a dynamic masking mechanism based on physical and electrical constraints during the decoding phase to generate a preliminary standard consensus draft. S50 performs hierarchical projection mapping to generate differentiated standard views that adapt to the device layer, edge layer, and platform layer; S60 performs closed-loop iterative optimization based on on-site electrical signal feedback, and outputs the final executable standard consensus.

2. The method as described in claim 1, characterized in that, S10 includes: extracting the reaction principle equations of corresponding physical quantities for each data source participating in the standard consensus, and parsing the equations into a symbol tree structure; the symbol tree structure contains operator nodes, variable nodes, and constant nodes, and records the dimensions of each node; synchronously collecting the electrical signal modulation parameters and timing distribution of the physical link layer, and encapsulating them into an electrical parameter package; the electrical signal modulation parameters include the root mean square value of sampling clock jitter, peak-to-peak value of power supply ripple, effective number of bits of the analog-to-digital converter, and the distribution of bus frame intervals.

3. The method as described in claim 1, characterized in that, S20 includes: constructing a reaction principle embedder, mapping the symbol tree to a high-dimensional vector, and retaining dimensional constraint information as attention bias during the mapping process; constructing an electrical signal embedder, inputting the normalized scalar parameters in the electrical parameter package into a multilayer perceptron, inputting the shape parameters of the distribution parameters into an independent coding layer, and outputting an electrical feature vector; and performing weighted fusion of text annotation embedding, reaction principle embedding, and electrical signal embedding to generate a three-domain joint embedding vector; wherein, the weight coefficients are preset according to the sensitivity of the physical quantity to the physical mechanism.

4. The method as described in claim 1, characterized in that, S30 includes: an encoder-decoder generative backbone network based on a transformer architecture, with a hidden layer dimension of 1024 and 16 attention heads. During training, weights that have been pre-trained on industry-standard corpora are loaded. At the encoder input, the three-domain joint embedding vector is used as the backbone feature, and state encoding of hierarchical control logic is introduced. The hierarchical control logic includes a device layer refresh cycle of 100ms, an edge layer aggregation window width of 1s, and a platform layer database entry cycle of 60s. Before the fully connected layer of the network decoder head, a three-domain cross-attention layer is inserted to force the decoding process to query the corresponding reaction principle embedding and electrical signal embedding when generating each field definition.

5. The method as described in claim 1, characterized in that, S40 includes: defining a dynamic tolerance interval generation function to calculate the tolerance half-width of each field in real time based on the electrical parameter package; constructing a reaction principle consistency mask: when the field value range generated by decoding conflicts with the output range of the reaction principle equation, the generation probability of the corresponding position of the field is set to zero; constructing an electrical signal feasibility mask: when the accuracy required by the generated standard clause is higher than the accuracy supported by the electrical parameter package, the generation of the clause is forcibly blocked.

6. The method as described in claim 1, characterized in that, S50 includes: generating a device layer projection view: retaining the original data type, byte order, and refresh period of 100ms, and adding a tolerance interval calculated by the dynamic tolerance interval generation function; generating an edge layer projection view: automatically generating an aggregation operator, defined as a truncated mean operator; generating a platform layer projection view: mapping the fields to the storage structure of a time series database, defining a downsampling strategy to retain the mean, maximum, minimum, and standard deviation, with a storage period of 60s.

7. The method as described in claim 5, characterized in that, In S40, if the peak-to-peak value of the power supply ripple in the electrical parameter package exceeds the preset threshold of 100mV, an enhanced saturation limiting strategy is adopted. Specifically, this includes: calculating the last bit skipping amplitude caused by ripple noise, where the effective number of bits of the analog-to-digital converter is set to 12 bits; temporarily extending the dynamic tolerance half-width to accommodate ripple interference; calling the enhanced saturation limiting function when decoding the generated field value range to allow the value to fluctuate within a larger range; and adding annotation labels to the generated standard consensus draft to indicate the high ripple environment and enable dynamic extended tolerance. If the reaction principle equation is a nonlinear high-order differential equation and involves multivariate coupling, a symbolic tree pruning masking strategy is adopted. Specifically, this includes: parsing the symbolic tree, identifying the dominant and minor terms in the equation, and calculating the sensitivity contribution of each variable node; if the field value range in the generated draft standard causes the contribution of the minor term to exceed 20% of the dominant term, it is determined to be a physical mechanism imbalance; generating a mechanism imbalance mask, multiplying the generation probability within the value range by a decay coefficient of 0.01, forcing the model to regenerate a numerical range that conforms to the conservation law; and mapping the corrected numerical range back to the range or alarm threshold field of the draft standard. If there are non-integer multiples of refresh cycles across layers in the hierarchical control logic, an asynchronous mapping correction mask strategy is adopted. Specifically, this includes: detecting the ratio of the device layer refresh cycle to the edge layer aggregation window width; if the ratio is not an integer, it is determined that there is an asynchronous mapping risk; calculating the boundary effect loss of the least common multiple cycle to determine the data truncation error at the edge of the aggregation window; generating an asynchronous correction mask, prohibiting the generation of static mapping rules that may lead to data loss, and forcing the generation of elastic mapping rules that include linear interpolation or previous value preservation logic; and generating a cross-layer synchronization description appendix in the consensus draft to describe in detail the data processing logic under non-integer multiple cycles.

8. The method as described in claim 5, characterized in that, In S50, if the truncation loss calculated by the edge layer aggregation operator is greater than a preset threshold of 5%, an adaptive truncation coefficient adjustment strategy is adopted. Specifically, this includes: counting the percentage of sample points removed by the truncation function when the truncation coefficient is set to 5%; if the percentage is greater than 5%, it is determined that the current truncation coefficient is too strict, and the truncation coefficient is automatically reduced to 3%, and the aggregation operator is recalculated; the quality control indicators in the edge layer projection view are updated synchronously, and the mean deviation before and after the truncation adjustment is recorded; the adjusted truncation coefficient is fed back to the generative backbone network as prior knowledge for the generation of similar scenarios in the future. If the platform layer projection view involves long-term archiving requirements and the data retention period is greater than one year, a hot and cold data tiered storage mapping strategy will be adopted. Specifically, this includes: analyzing the access frequency of fields and determining whether physical quantities are critical process parameters based on reaction principle equations; if they are critical process parameters, defining hot data areas, warm data areas, and cold data areas in the platform layer view, each corresponding to different downsampling granularities; generating cross-storage area index mapping tables to ensure that query requests can be automatically routed to the corresponding storage area; and adding data lifecycle management strategies to the consensus draft to clarify the rolling update mechanism for each storage area. If the confidence level of the electrical signal embedding in the three-domain joint embedding vector is lower than the threshold of 0.8, a conservative projection generation strategy is adopted, which includes: retrieving similar low-confidence scenario processing cases from the historical database and extracting the corresponding degradation processing operators; marking the data quality label of the field as pending verification in the device layer view and forcibly enabling the original value upload mode; disabling complex aggregation operators in the edge layer view, replacing them with simple arithmetic averages, and expanding the uncertainty label; generating data completion prompts in the platform layer view, and suggesting adding confidence level color warnings to the display interface.

9. The method as described in claim 1, characterized in that, In S60, if the feedback signal is negative for three consecutive iterations, a physical constraint-enhanced retraining strategy is adopted. This strategy includes: pausing the execution of the current standard consensus, locking the mapping relationship between the device layer and the edge layer to prevent data overflow; backtracking the collected original electrical signal data, refitting the jitter distribution parameters, and updating the electrical parameter package; inputting the updated electrical parameter package into the generative backbone network, freezing the text embedding layer weights, and only fine-tuning the reaction principle embedder and the electrical signal embedder; re-executing the steps from generating the initial draft to generating the differentiated view, generating the revised standard consensus draft, and conducting comparative tests to verify the revision effect. If the hierarchical control logic changes, for example, the device layer refresh cycle is adjusted from 100ms to 200ms, a hierarchical projection cascade update strategy is adopted. This strategy includes: capturing hierarchical change events and analyzing the parameter differences before and after the change; recalculating the optimal value of the edge layer aggregation window width based on the time constant of the reaction principle equation to maintain the sample number unchanged; cascading the update of the platform layer's data entry cycle to ensure that the aggregation window width is an integer multiple of the refresh cycle, and updating the sample number value in the mapping operator; generating a change impact analysis report, automatically notifying the affected upstream and downstream systems, and updating the version number of the standard document. If the generative backbone network frequently triggers the reaction principle consistency mask during the decoding phase, a symbol tree-guided contrastive learning optimization strategy is adopted. This strategy includes: recording the high-frequency fields that trigger the reaction principle consistency mask and their corresponding error generation sequences to construct a hard example sample library; adding symbol tree contrastive loss to the training objective of the generative backbone network to narrow the vector distance between correct physical quantity values ​​and the symbol tree output features, while widening the distance between error values ​​and the symbol tree; adjusting the fusion weights in the three-domain joint embedding vector, increasing the weight of reaction principle embedding, and strengthening the role of physical constraints in feature fusion; and incrementally training the backbone network using the corrected hard example sample library until the frequency of triggering the reaction principle consistency mask drops below 1%.

10. A data standard consensus generation platform based on generative AI, used to implement the method described in any one of claims 1-9, characterized in that the platform... include: The three-domain sensing and acquisition module is used to acquire heterogeneous data sources that include reaction principles, electrical signal modulation parameters, and hierarchical control logic. The joint embedding preprocessing module is used to transform heterogeneous data sources into three-domain joint embedding vectors of uniform dimension; The generative backbone inference module is used to perform feature fusion and context modeling; The dynamic mask quality control module is used to introduce physical and electrical constraints during the decoding stage to filter out illegal outputs; The hierarchical projection mapping module is used to generate differentiated standard views that adapt to the device layer, edge layer, and platform layer; The closed-loop iterative optimization module is used to continuously optimize the model and standards based on on-site feedback.

Images