Incremental self-evolution method of ontology model for refining and chemical process

Abstract

The application discloses an ontology model incremental self-evolution method for a refining process and relates to the technical field of refining modeling. The technical problems of the prior art are solved, that is, the ontology knowledge update depends on manual work, lacks incremental self-evolution ability, there is a semantic gap in ontology fusion of multi-source heterogeneous data, and the evolution efficiency is bottlenecked in a high-dimensional parameter space. Through a complete closed loop of multi-modal semantic alignment, ontology reasoning, loss construction, low-dimensional subspace evolution, pattern recognition and structure updating, and feedback calibration, the three modal data of continuous measurement variables, discrete operation events and text logs are mapped to a unified ontology semantic embedding space through comparative learning constraints. Through non-negative tensor decomposition of the multi-unit and multi-time state tensor of the refining process, a new mode component with a mode significance exceeding a threshold is automatically extracted and translated into an ontology knowledge unit, which is projected to a low-dimensional subspace and incrementally added to the existing ontology structure, so that the automatic mapping from data to knowledge is realized.

Description

Technical Field

[0001] This invention belongs to the field of refining and chemical modeling, specifically an incremental self-evolution method for ontology models of refining and chemical processes. Background Technology

[0002] Refining processes (such as catalytic cracking, hydrocracking, and distillation separation) are typical complex industrial processes characterized by strong nonlinearity, multivariate coupling, and multiscale characteristics. With the advancement of digital transformation in the refining industry, effectively integrating process data, mechanistic knowledge, and operational experience to construct reusable and evolvable process knowledge models has become a key issue restricting the industry's intelligent upgrading, necessitating an incremental self-evolving ontology model approach for refining processes.

[0003] Researchers have proposed a knowledge graph framework for digital twins of chemical processes, achieving unified organization and automated collaboration among models, data, and process contexts by constructing a process knowledge ontology. In the petrochemical safety field, existing research has focused on HAZOP analysis for knowledge graph construction, referencing the ISO 15926 international standard to construct HAZOP knowledge ontology and defining seven basic elements. However, current refining process ontology modeling still faces the following technical bottlenecks: ontology knowledge updates rely on manual intervention and lack incremental self-evolution capabilities; existing processing architectures typically analyze continuous variables and discrete events through different channels, ignoring the essential semantic differences between heterogeneous data describing the same operational event. While ontology models should unify these two expressions at a higher semantic level, they lack effective cross-modal semantic alignment mechanisms; refining process ontology models involve hundreds or even thousands of interrelated concepts, relationships, and parameters, and traditional optimization methods face severe computational efficiency bottlenecks in high-dimensional parameter spaces, failing to meet the real-time response requirements of industrial scenarios. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes an incremental self-evolution method for ontology models oriented towards the refining process, which is used to solve the technical problems of ontology knowledge updates relying on manual intervention, lack of incremental self-evolution capability, semantic gap in ontology fusion of multi-source heterogeneous data, and evolutionary efficiency bottleneck in high-dimensional parameter space.

[0005] To address the aforementioned problems, a first aspect of the present invention provides an incremental self-evolutionary method for ontology models oriented towards refining processes, comprising: including: The continuous measurement variable sequence, discrete operation event sequence and text log data of the refining process are mapped to a unified ontology semantic embedding space through cross-modal contrastive learning constraints to generate ontology state representation; Based on the ontology state representation, the prediction outputs of each sub-model are weighted and fused by the dynamic gating weights of each sub-model under the current working condition to obtain the ontology inference output; Incremental evolution updates of ontology model parameters are performed in a low-dimensional subspace spanned by the principal component directions of the covariance matrix of historical parameter samples, and incremental principal component analysis is used to update the low-dimensional subspace as the working conditions change dynamically. Non-negative tensor decomposition is performed on the three-dimensional state tensor composed of the state data of several operation units at different sampling times. Based on the saliency of each mode component obtained by decomposition, the mode components with saliency exceeding the preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure. The updated ontology knowledge, with parameter updates constrained by a low-dimensional subspace, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as a positive sample to participate in the next round of cross-modal contrastive learning, thus completing the incremental self-evolution of the ontology model.

[0006] Optionally, in one example of the above aspects, generating an ontology state representation includes the following steps: The continuous measurement variable sequence is encoded by a multilayer perceptron and added to the time location code to obtain the continuous variable semantic vector; The discrete operation event sequence is encoded by a graph neural network based on ontology relation graph to obtain discrete event semantic vectors; Text log data is encoded using a domain-adaptive pre-trained language model to obtain text semantic vectors; By aligning the text semantic vectors of the three modalities in the ontology semantic embedding space using a comparison learning loss function, an ontology state representation is obtained.

[0007] Optionally, in one example of the above aspects, based on the ontology state representation, the predicted outputs of each sub-model are weighted and fused using dynamic gating weights of each sub-model under the current operating condition to obtain the ontology inference output, including the following steps: Based on the similarity between the ontology state representation of the current working condition and the prototype center of each predefined working condition region, as well as the historical performance records of each sub-model in each working condition region, the advantage score of each sub-model under the current working condition is calculated as the final advantage score of the sub-model. The dynamic gating weights of each sub-model are calculated using a temperature-calibrated Softmax function based on the advantage score. The prediction outputs of each sub-model are weighted and fused according to dynamic gating weights to obtain the ontology inference output.

[0008] Optionally, in one example of the above aspects, the ontology model parameters are incrementally updated in a low-dimensional subspace spanned by the principal component directions of the historical parameter sample covariance matrix, and the low-dimensional subspace is updated dynamically with changing operating conditions using incremental principal component analysis, including the following steps: Collect parameter samples of the ontology model under historical operating conditions and calculate the parameter covariance matrix; Principal component analysis was performed on the parameter covariance matrix, and the parameters whose cumulative variance contribution rate reached a preset threshold were retained. The principal components span a low-dimensional subspace, where... Much smaller than the dimension of the original parameter space of the ontology model; The ontology model parameters are incrementally updated in a low-dimensional subspace. After accumulating a preset number of new parameter samples, incremental principal component analysis is used to update the principal component orientation of the low-dimensional subspace.

[0009] Optionally, in one example of the above aspects, a nonnegative tensor decomposition is performed on the three-dimensional state tensor composed of the state data of several operation units at different sampling times. Based on the saliency of each mode component obtained from the decomposition, mode components with saliency exceeding a preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure, including the following steps: A three-dimensional state tensor composed of the state data of several operation units at different sampling times; Nonnegative tensor decomposition is performed on the three-dimensional state tensor to obtain the unit pattern matrix, variable pattern matrix and time pattern matrix; Calculate the significance of the pattern components, and identify the pattern components whose significance exceeds a preset threshold as new pattern components; The newly added schema components are translated into ontology knowledge units and added to the existing ontology structure in the form of triples.

[0010] Optionally, in one example of the above aspects, the updated ontology knowledge, with parameter updates constrained by a low-dimensional subspace, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as a positive sample to participate in the next round of cross-modal contrastive learning, completing the incremental self-evolution of the ontology model, including the following steps: Calculate the update amount of the ontology model parameters in this round; using the principal component direction matrix of the low-dimensional subspace. Perform constraint filtering on it; The triples in the updated ontology knowledge are embedded into a unified ontology semantic embedding space through a translation model; The embedded ontology knowledge is used as a positive sample to participate in the next round of cross-modal comparative learning, thus completing the incremental self-evolution of the ontology model.

[0011] Compared with the prior art, the beneficial effects of the present invention are: This invention overcomes the shortcomings of existing ontology frameworks in automatically identifying and injecting new patterns by performing non-negative tensor decomposition on the multi-unit, multi-time state tensor of the refining process, automatically extracting new pattern components whose pattern saliency exceeds a threshold, translating them into ontology knowledge units, projecting them onto a low-dimensional subspace and incrementally adding them to the existing ontology structure.

[0012] This invention maps three modalities of data—continuous measurement variables, discrete operation events, and text logs—to a unified ontology semantic embedding space through contrastive learning constraints. This brings different modal representations of the same operation event closer together in the embedding space, effectively solving the semantic gap problem in ontology fusion of multi-source heterogeneous data in the refining process. Compared to methods that process different data types separately, this invention achieves unified expression and reasoning of knowledge at the ontology level.

[0013] This invention enables the ontology model to automatically update its knowledge as new data accumulates through a complete closed loop of multimodal semantic alignment, ontology reasoning, loss construction, low-dimensional subspace evolution, pattern recognition and structure update, and feedback calibration. This overcomes the shortcomings of traditional ontology modeling, which relies on manual knowledge engineering and suffers from delayed knowledge updates.

[0014] This invention projects the ontology parameter space to a low-dimensional subspace for evolutionary updates using principal component analysis, and employs incremental principal component analysis to dynamically adjust the subspace orientation according to the operating conditions. Combined with a sparse update strategy, this significantly reduces the computational complexity of high-dimensional optimization problems. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 The first aspect of this invention provides an incremental self-evolutionary method for ontology models oriented towards refining processes, comprising: The continuous measurement variable sequence, discrete operation event sequence and text log data of the refining process are mapped to a unified ontology semantic embedding space through cross-modal contrastive learning constraints to generate ontology state representation; Based on the ontology state representation, the prediction outputs of each sub-model are weighted and fused by the dynamic gating weights of each sub-model under the current working condition to obtain the ontology inference output; Incremental evolution updates of ontology model parameters are performed in a low-dimensional subspace spanned by the principal component directions of the covariance matrix of historical parameter samples, and incremental principal component analysis is used to update the low-dimensional subspace as the working conditions change dynamically. Non-negative tensor decomposition is performed on the three-dimensional state tensor composed of the state data of several operation units at different sampling times. Based on the saliency of each mode component obtained by decomposition, the mode components with saliency exceeding the preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure. The updated ontology knowledge, with parameter updates constrained by a low-dimensional subspace, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as a positive sample to participate in the next round of cross-modal contrastive learning, thus completing the incremental self-evolution of the ontology model.

[0019] In one embodiment of the present invention, generating an ontology state representation includes the following steps: The continuous measurement variable sequence is encoded by a multilayer perceptron and added to the time location code to obtain the continuous variable semantic vector; The discrete operation event sequence is encoded by a graph neural network based on ontology relation graph to obtain discrete event semantic vectors; Text log data is encoded using a domain-adaptive pre-trained language model to obtain text semantic vectors; By aligning the text semantic vectors of the three modalities in the ontology semantic embedding space using a comparison learning loss function, an ontology state representation is obtained.

[0020] In one embodiment of the present invention, a contrastive learning loss function is used. The formula is: in, Cosine similarity; For temperature parameters; Let be a positive sample vector, represented as... Semantically equivalent vectors from another modality; For the j-th sample vector in the batch, j iterates through all the samples involved in the calculation, including 1 positive sample and all the rest negative samples; Anchor vector is the text semantic vector for a specific modality; N is the total number of samples in the batch. It is an exponential function.

[0021] In one embodiment of the present invention, based on the ontology state representation, the predicted outputs of each sub-model are weighted and fused using dynamic gating weights of each sub-model under the current operating condition to obtain the ontology inference output, including the following steps: Based on the similarity between the ontology state representation of the current working condition and the prototype center of each predefined working condition region, as well as the historical performance records of each sub-model in each working condition region, the dominance score of each sub-model under the current working condition is calculated. ;in Working area The prototype center, For sub-model In the operating area Normalized history performance on For the collection of working condition areas, This is the entity state representation vector for the current operating condition; To traverse all working condition regions z, the maximum value of the calculation result within the parentheses is taken as the final advantage score of the sub-model; The dynamic gating weights of each sub-model are calculated using a temperature-calibrated Softmax function based on the advantage score. The prediction outputs of each sub-model are weighted and fused according to dynamic gating weights to obtain the ontology inference output.

[0022] In this embodiment, the dynamic gating weights of each sub-model The calculation formula is: in The advantage score of sub-model s under the current working condition x; To sum the variables, iterate through all sub-models; For temperature parameters; This represents the total number of sub-models. Based on the dynamic gating weights of each sub-model, the ontology inference output is analyzed, and the formula is as follows: ;in For ontology reasoning output, For the first The predicted output of each sub-model.

[0023] In one embodiment of the present invention, the ontology model parameters are incrementally updated in a low-dimensional subspace spanned by the principal component directions of the historical parameter sample covariance matrix, and the low-dimensional subspace is updated dynamically with changing operating conditions using incremental principal component analysis, including the following steps: Collect parameter samples of the ontology model under historical operating conditions and calculate the parameter covariance matrix; Principal component analysis was performed on the parameter covariance matrix, and the parameters whose cumulative variance contribution rate reached a preset threshold were retained. The principal components span a low-dimensional subspace, where... Much smaller than the dimension of the original parameter space of the ontology model; The ontology model parameters are incrementally updated in a low-dimensional subspace. After accumulating a preset number of new parameter samples, incremental principal component analysis is used to update the principal component orientations of the low-dimensional subspace: Incremental principal component analysis is used to update the principal component orientations of the low-dimensional subspace. ;in, This is the new low-dimensional subspace principal component matrix after the (t+1)th step update. This is an incremental PCA function that does not use all historical data, but only new samples, to iteratively update the principal component bases. Let be the principal component matrix at step t. For the newly collected set of m ontology model parameter samples, The threshold number of samples required to trigger incremental updates.

[0024] In one embodiment of the present invention, incremental evolutionary updates of ontology model parameters in a low-dimensional subspace include: The ontology model parameters are represented in the low-dimensional subspace as follows: in This represents the ontology model parameters in a low-dimensional subspace. The principal component matrix V contains the first... A low-dimensional coordinate vector, For the first Low-dimensional coordinate coefficients, The mean of the parameter samples. This represents the total number of low-dimensional coordinates. Based on the representation of ontology model parameters in a low-dimensional subspace, the incremental evolutionary update formula is set as follows: in This represents the ontology model parameters in the low-dimensional subspace after the update is completed at step t+1. For the first The representation of the ontology model parameters in the low-dimensional subspace. For learning rate, For the first Step in Evolutionary increments on low-dimensional coordinates.

[0025] In one embodiment of the present invention, a nonnegative tensor decomposition is performed on the three-dimensional state tensor composed of state data of several operation units at different sampling times. Based on the saliency of each decomposed mode component, mode components with saliency exceeding a preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure, including the following steps: A three-dimensional state tensor composed of the state data of several operation units at different sampling times; Nonnegative tensor decomposition is performed on the three-dimensional state tensor to obtain the unit pattern matrix, variable pattern matrix and time pattern matrix; Calculate mode components Significance ;in, For the first The singular values ​​of each pattern component, For the first The singular values ​​of each pattern component are obtained; pattern components with a significance exceeding a preset threshold are identified as new pattern components; the singular values ​​of the pattern components are obtained by performing SVD decomposition on the centered original data matrix to directly obtain the diagonal matrix and extracting all singular values. The newly added schema components are translated into ontology knowledge units and added to the existing ontology structure in the form of triples.

[0026] In one embodiment of the present invention, nonnegative tensor decomposition of the three-dimensional state tensor includes:

[0027] in, For a three-dimensional state tensor, For the unit pattern matrix, Indicates the first The distribution of each mode component across each operational unit For the variable pattern matrix, Indicates the first The distribution of each mode component across each state variable For time pattern matrix, Indicates the first The distribution of each mode component at each sampling time For the rank of the tensor, For outer product operation, It is the Frobenius norm.

[0028] In one embodiment of the present invention, the rule for translating newly added pattern components into ontology knowledge units is as follows: If the variable pattern matrix Medium variable components If the absolute value exceeds a preset threshold, a collaborative pattern concept containing the corresponding variable is generated. The attributes of the collaborative pattern concept are determined by the variable components. Define the list of variables corresponding to significantly non-zero elements; If the unit pattern matrix Medium unit components If the absolute value exceeds a preset threshold, a cooperative relationship is generated between the corresponding units, and the relationship weight is equal to the unit component weight. ; If the time pattern matrix If a pattern exhibits trending or cyclical characteristics, a time evolution type attribute is added to the corresponding pattern.

[0029] In this embodiment, the input pattern components are filtered by saliency. ,take out Three-column pattern vector.

[0030] Reading the variable pattern matrix No. Columns; Significant variables are filtered using a fixed threshold / column quantile threshold to generate collaborative pattern concept ontology entities, and the set of filtered variables is used as entity attributes; 75th percentile of columns / fixed threshold 0.25; Reading the cell pattern matrix No. Column; threshold filtering of collaborative unit set; establishing collaborative behavior relationship between pairs of units within the set; relationship weight; column 70th percentile / fixed threshold 0.3.

[0031] Read time pattern vector Linear regression fitting yields the slope. A linear regression fit is performed on the sequence to obtain the slope. If the slope is greater than the trend threshold, a trend is observed. Spectral analysis is used to detect the presence of a significant dominant frequency. When the spectral amplitude corresponding to a certain frequency exceeds a preset period threshold and accounts for more than 15% to 25% of the total spectral energy, it is determined to be an effective period, and the pattern is determined to have periodicity. Based on the detection results, the time evolution type ontological attribute is set for the ontological entity of the pattern, and parameters such as slope and period length are added simultaneously. All generated entities, relationships, and attributes are updated in the ontology knowledge graph.

[0032] In one embodiment of the present invention, the updated ontology knowledge, with parameter update amounts constrained by low-dimensional subspace, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as a positive sample to participate in the next round of cross-modal contrastive learning, completing the incremental self-evolution of the ontology model, including the following steps: Calculate the update amount of the ontology model parameters in this round. ,in, For the first time after the update is completed Representation of ontology model parameters in a low-dimensional subspace. For the first The representation of the ontology model parameters in the low-dimensional subspace; using the principal component direction matrix of the low-dimensional subspace. Perform constraint filtering on it: ,in Let be the principal component matrix at step t. For the first The effective parameter update amount after subspace filtering. For the first The original parameters to be filtered are updated incrementally; The triples in the updated ontology knowledge are embedded into a unified ontology semantic embedding space through a translation model; The embedded ontology knowledge is used as a positive sample to participate in the next round of cross-modal comparative learning, thus completing the incremental self-evolution of the ontology model.

[0033] In this embodiment, during the process of embedding the translation model into a unified ontology semantic embedding space, the translation model is TransE, and its optimization objective is:

[0034] in, For boundary distance loss, For the updated set of ontology triples, for One of the positive sample triples. For the head entity, For the relationship, For tail entities, For the set of negative sampled triples, for One of the negative sample triples. For spacing hyperparameters, , , The embedding vectors are, in order, the embedding vectors of the head entity, relation, and tail entity. For head entity The low-dimensional embedding vector, Tail entity The low-dimensional embedding vector; For hinge loss; The embedded ontology knowledge is used as a positive sample to participate in the next round of cross-modal contrastive learning: in, For the first The cross-modal alignment loss of the step, This is the semantic vector of the current modality data. To embed positive samples of ontology knowledge aligned with the current modality data, for the first ontology in the ontology library. Given a set of sample embeddings, we iterate through all the samples involved in the calculation, including one positive sample and all the rest negative samples. This refers to the temperature parameter. The similarity calculation function is cosine similarity in this embodiment; This is for exponential operations.

[0035] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. An incremental self-evolutionary method for ontology models oriented towards refining processes, characterized in that, include: The continuous measurement variable sequence, discrete operation event sequence and text log data of the refining process are mapped to a unified ontology semantic embedding space through cross-modal contrastive learning constraints to generate ontology state representation; Based on the ontology state representation, the prediction outputs of each sub-model are weighted and fused by the dynamic gating weights of each sub-model under the current working condition to obtain the ontology inference output; Incremental evolution updates of ontology model parameters are performed in a low-dimensional subspace spanned by the principal component directions of the covariance matrix of historical parameter samples, and incremental principal component analysis is used to update the low-dimensional subspace as the working conditions change dynamically. Non-negative tensor decomposition is performed on the three-dimensional state tensor composed of the state data of several operation units at different sampling times. Based on the saliency of each mode component obtained by decomposition, the mode components with saliency exceeding the preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure. The updated ontology knowledge, with parameter updates constrained by a low-dimensional subspace, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as a positive sample to participate in the next round of cross-modal contrastive learning, thus completing the incremental self-evolution of the ontology model.

2. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 1, characterized in that, Generating an ontology state representation includes the following steps: The continuous measurement variable sequence is encoded by a multilayer perceptron and added to the time location code to obtain the continuous variable semantic vector; The discrete operation event sequence is encoded by a graph neural network based on ontology relation graph to obtain discrete event semantic vectors; Text log data is encoded using a domain-adaptive pre-trained language model to obtain text semantic vectors; By aligning the text semantic vectors of the three modalities in the ontology semantic embedding space using a comparison learning loss function, an ontology state representation is obtained.

3. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 2, characterized in that, Contrast learning loss function The formula is: in, Cosine similarity; For temperature parameters; Let be a positive sample vector, represented as... Semantically equivalent vectors from another modality; This represents the j-th sample vector within the batch. Anchor vector is the text semantic vector for a specific modality; N is the total number of samples in the batch. It is an exponential function.

4. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 1, characterized in that, Based on the ontology state representation, the predicted outputs of each sub-model are weighted and fused using dynamic gating weights under the current operating conditions to obtain the ontology inference output, including the following steps: Based on the similarity between the ontology state representation of the current working condition and the prototype center of each predefined working condition region, as well as the historical performance records of each sub-model in each working condition region, the dominance score of each sub-model under the current working condition is calculated. ;in Working area The prototype center, For sub-model Normalized historical performance over the operating range, For the collection of working condition areas, is the ontological state representation vector for the current working condition; is the maximum value of the calculated result within the parentheses after traversing all working condition regions z, which is taken as the final advantage score of the sub-model. The dynamic gating weights of each sub-model are calculated using a temperature-calibrated Softmax function based on the advantage score. The prediction outputs of each sub-model are weighted and fused according to dynamic gating weights to obtain the ontology inference output.

5. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 1, characterized in that, The ontology model parameters are incrementally updated in a low-dimensional subspace spanned by the principal component directions of the historical parameter sample covariance matrix. Incremental principal component analysis is then used to update the low-dimensional subspace dynamically according to the changing operating conditions. This includes the following steps: Collect parameter samples of the ontology model under historical operating conditions and calculate the parameter covariance matrix; Principal component analysis was performed on the parameter covariance matrix, and the parameters whose cumulative variance contribution rate reached a preset threshold were retained. Each principal component spans a low-dimensional subspace, in which Much smaller than the dimension of the original parameter space of the ontology model; The ontology model parameters are incrementally updated in a low-dimensional subspace. After accumulating a preset number of new parameter samples, incremental principal component analysis is used to update the principal component orientations of the low-dimensional subspace: Incremental principal component analysis is used to update the principal component orientations of the low-dimensional subspace. ;in, This is the new low-dimensional subspace principal component matrix after the (t+1)th step update. For incremental PCA functions; Let be the principal component matrix at step t. For the newly collected set of m ontology model parameter samples, The threshold number of samples required to trigger incremental updates.

6. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 5, characterized in that, Incremental evolutionary updates of ontology model parameters in a low-dimensional subspace include: The ontology model parameters are represented in the low-dimensional subspace as follows: in This represents the ontology model parameters in a low-dimensional subspace. The principal component matrix V contains the first... A low-dimensional coordinate vector, For the first Low-dimensional coordinate coefficients, The mean of the parameter samples. This represents the total number of low-dimensional coordinates. Based on the representation of ontology model parameters in a low-dimensional subspace, the incremental evolutionary update formula is set as follows: in This represents the ontology model parameters in the low-dimensional subspace after the update is completed at step t+1. For the first The representation of the ontology model parameters in the low-dimensional subspace. For learning rate, For the first Step in Evolutionary increments on low-dimensional coordinates.

7. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 1, characterized in that, Nonnegative tensor decomposition is performed on the three-dimensional state tensor composed of state data of several operation units at different sampling times. Based on the saliency of each mode component obtained from the decomposition, mode components with saliency exceeding a preset threshold are identified as new mode components and translated into ontology knowledge units. The ontology knowledge units are then projected onto a low-dimensional subspace and incrementally added to the existing ontology structure, including the following steps: A three-dimensional state tensor composed of the state data of several operation units at different sampling times; Nonnegative tensor decomposition is performed on the three-dimensional state tensor to obtain the unit pattern matrix, variable pattern matrix and time pattern matrix; Calculate mode components Significance ;in, For the first The singular values ​​of each pattern component, For the first The singular values ​​of each pattern component; pattern components whose significance exceeds a preset threshold are identified as newly added pattern components; The newly added schema components are translated into ontology knowledge units and added to the existing ontology structure in the form of triples.

8. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 7, characterized in that, Nonnegative tensor decomposition of the three-dimensional state tensor includes: in, For a three-dimensional state tensor, For the unit pattern matrix, Indicates the first The distribution of each mode component across each operational unit For the variable pattern matrix, Indicates the first The distribution of each mode component across each state variable For time pattern matrix, Indicates the first The distribution of each mode component at each sampling time For the rank of the tensor, For outer product operation, It is the Frobenius norm.

9. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 7, characterized in that, The rules for translating newly added schema components into ontology knowledge units are as follows: If the variable pattern matrix Medium variable components If the absolute value exceeds the preset threshold, it is considered a significant variable and the collaborative mode concept ontology entity is obtained. The set of selected variables is used as entity attributes. If the unit pattern matrix Medium unit components If the absolute value exceeds a preset threshold, a cooperative relationship is generated between the corresponding units, and the relationship weight is equal to the unit component weight. ; If the time pattern matrix If a pattern exhibits trending or cyclical characteristics, a time evolution type attribute is added to the corresponding pattern.

10. The incremental self-evolutionary method for ontology models oriented towards refining processes according to claim 1, characterized in that, The updated ontology knowledge, constrained by a low-dimensional subspace and filtered for parameter updates, is embedded into a unified ontology semantic embedding space through a translation model. The embedded ontology knowledge is then used as positive samples in the next round of cross-modal contrastive learning, completing the incremental self-evolution of the ontology model. This process includes the following steps: Calculate the update amount of the ontology model parameters in this round. ,in, For the first time after the update is completed Representation of ontology model parameters in a low-dimensional subspace. For the first The representation of the ontology model parameters in the low-dimensional subspace; using the principal component direction matrix of the low-dimensional subspace. Perform constraint filtering on it: ,in Let be the principal component matrix at step t. For the first The effective parameter update amount after subspace filtering. For the first The original parameters to be filtered are updated incrementally; The triples in the updated ontology knowledge are embedded into a unified ontology semantic embedding space through a translation model; The embedded ontology knowledge is used as a positive sample to participate in the next round of cross-modal comparative learning, thus completing the incremental self-evolution of the ontology model.

Images