Construction method of intelligent manufacturing scene knowledge base based on artificial intelligence

Abstract

The application relates to the technical field of knowledge base construction, in particular to a construction method of an intelligent manufacturing scene knowledge base based on artificial intelligence, which comprises the following steps: synchronously acquiring multi-source data and extracting candidate facts; calculating the extraction confidence of each fact by fusing source internal consistency and relationship similarity, weighting and aggregating the confidence in combination with the long-term reliability of the data source to obtain a global belief value, and screening out reliable facts according to the global belief value; fusing multi-source evidence to generate a relationship comprehensive vector, calculating the similarity between the relationship comprehensive vector and existing semantics of the knowledge base, and finally making a decision in combination with the global belief value to judge whether the fact is admitted or not. The application solves the problem of unreliable facts caused by data conflicts in traditional knowledge base construction, realizes self-dissolution of conflicts and online self-evolution of the knowledge base through quantitative evidence and dynamic decision-making, and improves the consistency and integrity of the knowledge base facts.

Description

Technical Field

[0001] This application relates to the field of knowledge base construction technology, specifically to a method for constructing a knowledge base for intelligent manufacturing scenarios based on artificial intelligence. Background Technology

[0002] Artificial intelligence-based intelligent manufacturing is an advanced factory operation model driven by data, aiming to replace traditional rigid rules. It integrates real-time production data from multiple heterogeneous sources, including equipment sensors, PLCs, MES, CAD systems, and even maintenance logs and visual quality inspection. Utilizing machine learning or deep learning technologies, it intelligently optimizes the entire production process. Its core tasks include real-time fault diagnosis, adaptive adjustment of process parameters, and personalized production scheduling. To effectively integrate this multimodal data, building a unified knowledge base is crucial. This knowledge base not only contains interpretable symbolic rules and graphs but also integrates numerical embeddings that can be directly invoked by deep models. This enables simultaneous support for symbolic reasoning and distributed representation computing, providing a solid knowledge foundation for intelligent decision-making in intelligent manufacturing.

[0003] Traditional methods for constructing knowledge bases in intelligent manufacturing have inherent flaws. Their processes typically rely on regularized templates, CRF, or a single BERT model for information extraction, supplemented by simple string matching or similarity thresholding for deduplication. Finally, triples are filled into a pre-defined graph based on expert-written mapping rules, followed by qualitative updates. However, the fundamental weakness of traditional knowledge base construction methods lies in their complete disregard for the credibility of data sources and the lack of probabilistic detection of conflicting evidence. Therefore, when faced with complex manufacturing scenarios characterized by large differences in heterogeneous data representation, high sensor noise, or incomplete coverage of expert rules, traditional methods are highly prone to generating numerous false positive or false negative triples, ultimately compromising the factual consistency and integrity of the knowledge base. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence, thereby resolving the existing problems.

[0005] The method for constructing an AI-based intelligent manufacturing scenario knowledge base in this application adopts the following technical solution: Obtain all candidate facts and their relational subvectors from various data sources in an AI manufacturing scenario under different time windows; Based on the number of times each candidate fact under each data source was repeatedly counted within the preset time period before the current time window, and the similarity between the relation sub-vector of each candidate fact under each data source and its corresponding prototype vector within the current time window, the extraction confidence of each candidate fact under each data source within the current time window is determined. The long-term reliability of each data source within the current time window is obtained, and combined with the extraction confidence, the evidence strength of each candidate fact under each data source is determined to determine the global belief value of each candidate fact, and then all facts are selected from all candidate facts within the current time window. The evidence strength and embedding vectors of each fact across all data sources are combined to obtain a comprehensive relation vector. The comprehensive average vector of all prototype vectors with the same relation sub-vectors as each fact is obtained from the knowledge base. Based on the similarity between the comprehensive relation vector and the comprehensive average vector, and combined with the global belief value, it is determined whether to write each fact in the current time window into the knowledge base.

[0006] Preferably, the method for determining the extraction confidence of each candidate fact from each data source within the current time window is as follows: Based on the number of times each candidate fact under each data source was repeatedly counted within the preset time period before the current time window, the source consistency factor of each candidate fact under each data source within the current time window is determined. Based on the similarity between the relation subvector and its corresponding prototype vector of each candidate fact under each data source in the current time window, determine the relation consistency factor of each candidate fact under each data source in the current time window. The extraction confidence of each candidate fact from each data source within the current time window is the result of a positive fusion of the intra-source consistency factor and the relational consistency factor of each candidate fact from each data source within the current time window.

[0007] Preferably, the expression for the intra-source consistency factor of each candidate fact under each data source within the current time window is: In the formula, This represents the intra-source consistency factor of candidate facts t under data source s within the current time window; This represents the evidence count of candidate facts t under data source s within a preset time period prior to the current time window; The preset value is represented by ; min[ ] represents the minimum value function.

[0008] Preferably, the expression for the relational consistency factor of each candidate fact under each data source within the current time window is: In the formula, This represents the consistency factor of the relationship between candidate facts t under data source s within the current time window; This represents the similarity between the relational subvector and the corresponding prototype vector of candidate fact t under data source s within the current time window; max[] represents the maximum value function.

[0009] Preferably, obtaining the long-term reliability of each data source within the current time window includes: Long-term reliability of data source s within the current time window The expression is: In the formula, , These represent the number of true cases and the number of false cases in data source s within the preset time period prior to the current time window, respectively. , These represent the preset initial number of true cases and the preset initial number of false cases, respectively.

[0010] Preferably, the strength of evidence for each candidate fact under each data source is the product of the extraction confidence and the corresponding long-term reliability of each candidate fact under each data source.

[0011] Preferably, determining the global belief value for each candidate fact includes: The evidence strength of each candidate fact under each data source within all time windows in the preset period prior to the current time window is used to form the evidence sequence of each candidate fact under each data source in the current time window. Based on the correlation of evidence sequences for each candidate fact among all data sources within the current time window, and in conjunction with the strength of the evidence, the global belief value for each candidate fact within the current window is determined, specifically: Global belief value of candidate fact t within the current time window The expression is: In the formula, This indicates the strength of evidence for candidate fact t under data source s within the current window; This represents the mean of the correlation coefficients of the evidence sequences of candidate facts t among all data sources within the current window; This represents the data source that provides all supporting evidence for candidate fact t.

[0012] Preferably, the step of filtering all facts from all candidate facts within the current time window includes: Based on the global belief value, determine the support level of each candidate fact within the current time window; If the support of the current candidate fact is greater than 0, then the current candidate fact is accepted as a fact; otherwise, the current candidate fact is rejected as a fact.

[0013] Preferably, the fusion of the evidence strength and embedding vectors of each fact across all data sources to obtain the comprehensive relation vector includes: The evidence strength of each fact under each data source within the current time window is used as the weight of the corresponding embedding vector. The embedding vectors of each fact under all data sources that provide supporting evidence are weighted and fused. The fused vector is used as the comprehensive relationship vector of each fact within the current time window.

[0014] Preferably, the determination of whether to write each fact within the current time window into the knowledge base includes: Based on the similarity between the relational composite vector and the composite average vector, and combined with the global belief value, the admission score for each fact within the current time window is determined. If the admission score of the current fact is greater than or equal to the preset first value, the current fact is written into the knowledge base; if the admission score of the current fact is less than or equal to the preset second value, the current fact is rejected from being written into the knowledge base; if the admission score of the current fact is between the preset first value and the preset second value, the current fact is stored in the temporary storage area, where the preset first value is greater than the preset second value.

[0015] One embodiment of this application provides a method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence, the method comprising the following steps: This application has at least the following beneficial effects: This application dynamically quantifies the extraction confidence of candidate facts by fusing source consistency and relational consistency factors. It evaluates the evidence from two dimensions: sustained stability and semantic accuracy. This effectively suppresses "false evidence" caused by transient noise or single false alarms, and filters out semantic conflicts arising from ambiguity or inconsistencies in context. This significantly improves the initial reliability of candidate facts at the source, laying the foundation for building a highly consistent and accurate knowledge base. Furthermore, this application introduces the long-term reliability of the data source as a weight, combined with decorrelation factor correction, to calculate the global belief value of candidate facts using probabilistic aggregation. Conflict resolution is then conducted using the log-likelihood ratio with mutually exclusive facts. This method not only amplifies the evidence contribution of high-confidence sources but also suppresses low-quality data and... The impact of redundant information achieves a comprehensive balance between the quantity and quality of evidence. Furthermore, by actively adjudicating contradictory information, it enables the self-resolution of conflicts, ensuring that the selected facts have a high degree of comprehensive credibility and providing a guarantee for building a consistent and accurate knowledge base. Finally, this application generates a comprehensive relation vector by weighted fusion of multi-source evidence and uses harmonic averaging to combine it with the similarity of the existing semantics of the knowledge base and the global belief value to calculate the admission score. This mechanism constructs a double-insured admission threshold, requiring not only solid multi-source evidence to support the facts but also a high degree of alignment with the existing semantic system of the knowledge base. Through the sensitivity of harmonic averaging to weaknesses, it effectively suppresses situations where the evidence is strong but semantically inconsistent or semantically consistent but the evidence is weak, thereby improving the consistency and completeness of the facts in the knowledge base. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the steps of a method for constructing an AI-based intelligent manufacturing scenario knowledge base, as provided in one embodiment of this application. Figure 2 A fact screening flowchart provided for one embodiment of this application. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by this application to achieve the intended inventive purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of the method for constructing an artificial intelligence-based intelligent manufacturing scenario knowledge base proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] The following section, in conjunction with the accompanying drawings, details the specific scheme of the method for constructing an AI-based intelligent manufacturing scenario knowledge base provided in this application.

[0021] This application provides an embodiment of a method for constructing an AI-based intelligent manufacturing scenario knowledge base. Specifically, the method is described below. Please refer to [link to relevant documentation]. Figure 1 The method includes the following steps: Step S1: Obtain all candidate facts and their relational subvectors from each data source in the AI ​​manufacturing scenario under different time windows.

[0022] The multi-source heterogeneous data stream in intelligent scenarios refers to a real-time data set with different modalities and semantic granularities collected from different data sources during continuous production. This includes, but is not limited to, timing signals from high-frequency sensors, discrete PLC event logs, semi-structured MES work order texts, geometric and process annotations in CAD or AE files, and vision system outputs. These data are processed in parallel in a streaming manner to form a multi-angle observation of the same manufacturing process. The multi-source data is divided into time windows, with each time window being 1 hour in size. In practical applications, implementers can also set the time window size according to specific circumstances. This embodiment does not impose any special restrictions.

[0023] Candidate facts refer to independent knowledge units that may be true and identified from the above data stream. They are represented as "head entity-relationship-tail entity" triples (e.g., main shaft-vibration over-limit-0.8mm / s), with timestamps and data source identifiers. They are the smallest semantic granularity and basic operation objects for building a knowledge base.

[0024] The extraction process is as follows: First, multi-source data from 1Hz to 10kHz is resampled to 1kHz under a unified time baseline to achieve timestamp alignment. A sliding window is then created with a 10-minute step size. Each modal data within the sliding window is encoded into a 256-dimensional context embedding vector. Entity recognition is then performed, using an entity-relation joint extraction model (BERT+PointerNet) to identify candidate triples of "head entity-relation-tail entity" and their start and end indices. Based on these indices, the context embedding vectors are subjected to position masking and average pooling to obtain 256-dimensional head entity sub-vectors, relation sub-vectors, and tail entity sub-vectors. These are then combined through conflict deduplication and splicing projection to form 256-dimensional candidate triple embedding vectors. Further, from the error-free historical data, windows are divided and triples are extracted using the same method as for the candidate facts. The prototype vectors that correspond to the candidate facts in the intelligent manufacturing scenario are selected.

[0025] The acquisition of prototype vectors is a well-known technique, and will not be elaborated further.

[0026] Step S2: Based on the number of times each candidate fact under each data source was repeatedly counted within the preset time period before the current time window, and the similarity between the relation sub-vector of each candidate fact under each data source and its corresponding prototype vector within the current time window, determine the extraction confidence of each candidate fact under each data source within the current time window.

[0027] Because heterogeneous sources express the same fact with lexical differences, measurement uncertainties, and noise, directly adopting all candidate triples may lead to conflicts and errors. Therefore, it is necessary to quantify the semantic matching degree and local duplication evidence strength of each candidate fact under different sources. Based on the number of times each candidate fact under each data source was repeatedly counted within a preset time period before the current time window, and the similarity between the relation subvector and its corresponding prototype vector of each candidate fact under each data source within the current time window, the extraction confidence of each candidate fact under each data source within the current time window is determined. Specifically: First, this embodiment determines the intra-source consistency factor of each candidate fact under each data source within the current time window based on the number of times each candidate fact is repeatedly counted within a preset time period prior to the current time window. Specifically: In this embodiment, the intra-source consistency factor of candidate facts t under data source s within the current time window is... The expression is: In the formula, This represents the evidence count of candidate facts t under data source s within a preset time period prior to the current time window; The preset value is represented by ; min[ ] represents the minimum value function.

[0028] It should be noted that the preset duration is set manually. In this embodiment, the preset duration is 20h. In actual application, as other implementation methods, implementers can also set it according to specific circumstances. This embodiment does not impose any special restrictions.

[0029] It should be noted that the preset value is set manually. The preset value is used to prevent confidence saturation caused by extremely high repetition, to convert local high-frequency consistency into numerical enhancement, and to reduce the weight of occasional noise or single recognition errors. In this embodiment, the preset value ranges from 3 to 5. In this embodiment, the preset value is 5. In actual application, as other implementation methods, implementers can also set it according to specific circumstances. This embodiment does not impose any special restrictions.

[0030] To further clarify, the evidence count acquisition method is as follows: Based on the start and end indices of candidate fact t, the original data segments of the head entity sub-vector, relation sub-vector, and tail entity sub-vector are extracted from the time window of data source s. The three data segments are converted into standardized event codes, that is, strings uniformly encoded from the original signals are concatenated into a unique key. Furthermore, MD5 is calculated on the key. For example, the relation "vibration exceeds the limit" is converted into the event code "VIB_ALM", similar to the HTTP status code 404. Furthermore, the number of times the MD5 value appears up to the current time window is counted in the evidence counters of each data source.

[0031] Based on the source consistency factor, it can be understood that the source consistency factor is used to quantify the repetition intensity of the same fact from the same source. If the source consistency factor of the current candidate fact is larger, it indicates that the current candidate fact has been observed multiple times and independently within the data source. This means that the current candidate fact is a real and continuous event rather than a momentary disturbance. Therefore, the extraction confidence level of the current candidate fact should be increased so that it can obtain higher weight in subsequent belief aggregation and admission decisions, thereby effectively suppressing the risk of single false alarms. Conversely, if the source consistency factor of the current fact is smaller, it indicates that the appearance of the current candidate fact within the data source is isolated and accidental. It is very likely to be momentary noise in the data collection process or a single misjudgment by the extraction model. It lacks continuous and reproducible evidence support. Therefore, the extraction confidence level of the current candidate fact should be significantly reduced or even zeroed out so that it is given extremely low weight in subsequent belief aggregation and admission decisions, thereby effectively filtering out these unreliable "false evidence" and preventing them from polluting the accuracy and consistency of the knowledge base.

[0032] Furthermore, this embodiment determines the relation consistency factor for each candidate fact under each data source within the current time window based on the similarity between the relation sub-vector and the corresponding prototype vector of each candidate fact under each data source within the current time window. Specifically: In this embodiment, the relationship consistency factor of candidate facts t under data source s within the current time window. The expression is: In the formula, This represents the similarity between the relational subvector and the corresponding prototype vector of candidate fact t under data source s within the current time window; max[] represents the maximum value function.

[0033] It should be noted that there are many methods to measure the similarity between vectors. In this embodiment, the cosine similarity between the relation subvector of candidate fact t under data source s and the corresponding prototype vector within the current time window is used as the similarity between the relation subvector of candidate fact t under data source s and the corresponding prototype vector within the current time window. In practical applications, as other implementation methods, implementers may also choose other methods to measure the similarity between vectors, such as the reciprocal of Euclidean distance, depending on the specific circumstances. This embodiment does not impose any special restrictions on the selection of methods to measure the similarity between vectors.

[0034] The method for calculating cosine similarity is a well-known technique, and its specific calculation process will not be elaborated here.

[0035] Based on the relation consistency factor, it can be understood that the relation consistency factor of candidate fact t under data source s within the current time window reflects the degree of fit between the contextual semantics of candidate fact t in a specific data source s and the semantics of the fact. If the similarity between the relation sub-vector and the corresponding prototype vector of candidate fact t under data source s within the current time window is greater, it indicates that the relation expression of fact t extracted from data source s is highly consistent with the standard semantics of the relation. The extraction confidence of candidate fact s should be increased so that it can obtain higher weight in subsequent processing, thereby effectively filtering out "false evidence" caused by ambiguity or irrelevance of context. Conversely, if the similarity between the relational subvector and the corresponding prototype vector of candidate fact t under data source s within the current time window is smaller, it indicates that the relational representation of fact t extracted from data source s deviates from or is completely unrelated to the standard semantics of the relation. This is very likely due to semantic conflicts caused by misunderstanding of context, data noise, or model misjudgment. Therefore, the extraction confidence of this fact should be significantly reduced or even directly cleared to zero, so that it is given extremely low weight or directly excluded in subsequent belief aggregation and admission decisions. This effectively prevents these semantically inconsistent "false evidences" from polluting the knowledge base and ensures the semantic accuracy and consistency of the knowledge in the database.

[0036] Furthermore, this embodiment determines the extraction confidence of each candidate fact under each data source within the current time window based on the relationship consistency factor of each candidate fact under each data source within the current time window, and in combination with the intra-source consistency factor. Specifically: In this embodiment, the result of positively fusing the intra-source consistency factor and relational consistency factor of each candidate fact under each data source within the current time window is used as the extraction confidence of each candidate fact under each data source within the current time window.

[0037] It should be understood that positive fusion refers to combining two or more indicators through addition or multiplication to obtain a comprehensive indicator, thereby more comprehensively and accurately assessing a phenomenon or problem. This fusion method is not limited to simple arithmetic operations, but can also include more complex statistical models and analytical methods. Implementers can choose according to specific circumstances, and this embodiment does not impose any special restrictions.

[0038] Preferably, as one implementation method, in this embodiment, the product of the intra-source consistency factor and the relational consistency factor of each candidate fact under each data source within the current time window is used as the extraction confidence of each candidate fact under each data source within the current time window. In practical applications, as other implementation methods, implementers may also adopt other positive fusion methods such as sum values ​​according to specific circumstances. This embodiment does not impose any special restrictions.

[0039] Based on the extraction confidence of each candidate fact from various data sources within the current time window, it can be understood that the extraction confidence is used to characterize the initial reliability of the candidate fact before entering global fusion. If the intra-source consistency factor of the current candidate fact is larger, it indicates that the current fact is continuously and repeatedly observed within the data source, which directly improves the extraction confidence, reflects that the evidence has high stability and reproducibility, enhances the ability of the current candidate fact to resist transient noise, and enables it to obtain higher basic weight in subsequent decisions. At the same time, if the relation consistency factor of the current candidate fact is larger, it indicates that the semantic expression of the current candidate fact is highly consistent with the standard definition, and the corresponding extraction confidence is also larger, reflecting the accuracy and high quality of the current candidate fact at the semantic level, and increasing the possibility of it being written into the knowledge base. Conversely, if the source consistency factor of the current candidate fact is smaller, it indicates that the fact appears isolated and sporadically within the data source. This drastically lowers the extraction confidence, reflecting the lack of stability of the evidence. It is very likely to be transient noise or a single misjudgment, thus significantly weakening its ability to resist interference and causing it to be given extremely low basic weight or even directly filtered out in subsequent decisions. At the same time, if the relation consistency factor of the current candidate fact is smaller, it indicates that its semantic expression deviates from the standard definition or is even completely unrelated. This also leads to a sharp drop in extraction confidence, reflecting the inaccuracy and low quality of the evidence at the semantic level. This greatly reduces the possibility of it being written into the knowledge base, thereby preventing semantically ambiguous or erroneous knowledge from polluting the entire knowledge base from the source.

[0040] Thus, this embodiment dynamically quantifies the extraction confidence of candidate facts by integrating source consistency and relational consistency factors. At the same time, it evaluates the evidence from two dimensions: the continuous stability of the evidence and the semantic accuracy. This not only effectively suppresses "false evidence" caused by instantaneous noise or single false alarms, but also filters out semantic conflicts caused by ambiguity of expression or inconsistency of context. As a result, it significantly improves the initial reliability of candidate facts at the source, laying the foundation for the subsequent construction of a knowledge base with high consistency and high accuracy.

[0041] Step S3: Obtain the long-term reliability of each data source within the current time window, and combine it with the extraction confidence to determine the evidence strength of each candidate fact under each data source, so as to determine the global belief value of each candidate fact, and then filter out all facts from all candidate facts within the current time window.

[0042] In multi-source heterogeneous data, there are significant differences in the observation accuracy and semantic expression capabilities of each data source. If the extraction confidence of all sources is aggregated with the same weight, it may lead to erroneous facts from low-confidence sources and correct facts from high-confidence sources being treated equally, thereby causing conflicts and errors in the knowledge base.

[0043] Based on the above analysis, this embodiment obtains the long-term reliability of each data source within the current time window, and combines it with the extracted confidence level to determine the evidence strength of each candidate fact under each data source, thereby determining the global belief value of each candidate fact. Then, all facts are selected from all candidate facts within the current time window. That is, by introducing source reliability as a priori weight into probability fusion, the contribution of unreliable sources is suppressed during the evidence aggregation stage, while the influence of high-confidence sources is amplified. This ensures that the final global belief value reflects both the quantity of evidence and the quality of data. The specific process is as follows: First, this embodiment determines the long-term reliability of each data source within the current time window based on the reliability of each data source within a preset period prior to the current time window. Specifically: In this embodiment, as one implementation method, the long-term reliability of data source s within the current time window is... The expression is: In the formula, , These represent the number of true cases and the number of false cases in data source s within the preset time period prior to the current time window, respectively. , These represent the preset initial number of true cases and the preset initial number of false cases, respectively.

[0044] It should be noted that the method for setting the preset initial number of true cases and the preset initial number of false cases is as follows: obtain the number of true cases and the number of false cases of data source s within a preset time period after the knowledge base starts to be built, add 1 to the number of true cases and use the result as the preset initial number of true cases, add 1 to the number of false cases and use the result as the preset initial number of false cases. In this embodiment, the preset time period is 100h. In actual application, the implementer can also set it according to the specific situation. This embodiment does not impose any special restrictions.

[0045] The true and false cases are known techniques, and their specific acquisition process will not be elaborated here.

[0046] Based on long-term reliability, it can be understood that long-term reliability reflects the historical accuracy and stability of the data source. The higher the long-term reliability of the current data source, the more likely it is that the number of true cases verified by the current data source in the past far exceeds the number of false cases, and its output is highly reliable. This reflects that the current data source has high observation accuracy and semantic expression ability. Conversely, the lower the long-term reliability of the current data source, the more likely it is to have been verified as a hypothesis in the past, resulting in low output credibility. This suggests that the current data source may have insufficient observation accuracy, poor data quality, or systematic bias. The evidence it provides will be given extremely low weight in the calculation of global belief values, thus having a weak influence in conflict resolution and effectively preventing its erroneous information from polluting the knowledge base.

[0047] Furthermore, in this embodiment, the long-term reliability of each data source within the current time window, combined with the extraction confidence level, is used to determine the evidence strength of each candidate fact under each data source, thereby determining the global belief value of each candidate fact within the current time window. Specifically: In this embodiment, the product of the extraction confidence level and the corresponding long-term reliability of each candidate fact under each data source is used as the evidence strength of each candidate fact under each data source.

[0048] Based on the strength of evidence for each candidate fact under each data source, it can be understood that the strength of evidence reflects the support of the data source for the candidate fact. If the extraction confidence of the current candidate fact is higher, the strength of evidence is higher, indicating that the current candidate fact not only has a high semantic matching degree under a single data source, but also has stable observation. At the same time, if the long-term credibility corresponding to the current candidate fact is higher, it indicates that the current candidate fact comes from a historically highly credible data source. Conversely, the lower the extraction confidence of the current candidate fact, the weaker the evidence strength, indicating that the current candidate fact may have semantic conflicts or be merely transient noise under a single data source, lacking stable support. At the same time, the lower the long-term credibility of the current candidate fact, it indicates that the current candidate fact comes from a data source with poor historical performance and low credibility. Even if the evidence it provides exists, its influence will be systematically weakened, making it difficult to play an effective role in global decision-making.

[0049] Furthermore, in this embodiment, the evidence strength of each candidate fact under each data source within all time windows within a preset period prior to the current time window is used to form an evidence sequence for each candidate fact under each data source within the current time window. Based on the correlation of evidence sequences for each candidate fact among all data sources within the current time window, and in conjunction with the strength of the evidence, the global belief value for each candidate fact within the current window is determined, specifically: Global belief value of candidate fact t within the current time window The expression is: In the formula, This indicates the strength of evidence for candidate fact t under data source s within the current window; This represents the mean of the correlation coefficients of the evidence sequences of candidate facts t among all data sources within the current window; This represents the data source that provides all supporting evidence for candidate fact t.

[0050] It should be noted that there are many methods for measuring the correlation between sequences. In this embodiment, the Pearson correlation coefficient of the evidence sequences of candidate facts t among all data sources in the current window is used as the evidence sequence correlation coefficient of candidate facts t among all data sources in the current window. In practical applications, as other implementation methods, implementers may also use Spearman correlation coefficient or Kendall rank correlation coefficient, etc., to measure the correlation between sequences, depending on the specific circumstances. This embodiment does not impose any special restrictions on the selection of methods for measuring the correlation between sequences.

[0051] The calculation method for the Pearson correlation coefficient is a well-known technique, and its specific calculation process will not be elaborated here.

[0052] To elaborate further, S represents the set of data sources that provide all supporting evidence for candidate fact t. Supporting evidence refers to the data source output that can directly or indirectly confirm the validity of candidate fact t. For example, if a vibrator sensor collects the event "the amplitude of the spindle vibration is 0.8 mm / s exceeding the threshold" within 1 second, after semantic alignment, a triple "spindle 001 - vibration exceeds the limit - 0.8 mm / s" is generated. That is, each extracted candidate fact constitutes supporting evidence for itself, and it also includes confirmation of t by other candidate facts. For example, the triple "spindle 001 - vibration exceeds the limit - 0.75 mm / s" obtained from other sources also constitutes supporting evidence for t.

[0053] Based on the global belief value, it can be understood that the global belief value reflects the comprehensive credibility of candidate facts under multiple observations. The stronger the evidence of the current candidate fact, the more significant its contribution. When multiple sources of strong evidence support it simultaneously, the global belief value will approach 1, reflecting the synergistic effect of multiple independent evidence. This can exponentially amplify the trust in candidate facts, enabling rapid and accurate acceptance. At the same time, the smaller the mean correlation coefficient of the current candidate facts, the stronger the independence between data sources. The larger the global belief value, the more realistically it reflects the cumulative strength of evidence, reflecting the diversity and complementarity of multiple information, making belief assessment more robust and reliable, and ensuring the accuracy of knowledge base construction. Conversely, the weaker the evidentiary strength of the current candidate fact, the weaker its contribution. Even with support from multiple sources, the global belief value is difficult to improve effectively. This reflects the generally low quality of evidence, making it difficult for candidate facts to meet the entry threshold, thus effectively filtering out unreliable information. At the same time, if the mean correlation coefficient of the current candidate facts is larger, it indicates a high correlation between data sources, and the global belief value will be suppressed and reduced. This reflects the violation of the assumptions of redundancy and independence of evidence, avoiding false confidence inflation caused by repeatedly calculating highly correlated information, thus ensuring the statistical validity of belief assessment.

[0054] Furthermore, this embodiment determines the support level of each candidate fact within the current time window based on the global belief value, specifically: Support of candidate fact t within the current time window The expression is: In the formula, This represents the global belief value of candidate fact t within the current time window; This represents a candidate fact that is mutually exclusive with candidate fact t; Let represent the global belief value of candidate facts t that are mutually exclusive; ln() represents the logarithmic function with the natural constant as the base.

[0055] Based on the support level of each candidate fact within the current time window, it can be understood that the support level reflects the strength of support for candidate fact t relative to its mutually exclusive candidate facts. The net advantage between the strengths of support is when the combined strength of evidence supporting candidate fact t far exceeds that supporting its mutually exclusive candidate facts. When evidence is presented, A value significantly greater than 0 indicates that the candidate fact t is overwhelmingly believed to be true based on global information. Therefore, the candidate fact t is accepted as a fact, thereby maintaining the factual consistency of the knowledge base and achieving the self-resolution of conflicts. Conversely, when the overall strength of evidence supporting candidate fact t is much weaker than that supporting mutually exclusive candidate facts... When evidence is presented, The value will be significantly less than 0, which reflects that the system determines that candidate fact t is very likely to be false information based on global information. Therefore, it rejects candidate fact t from entering the database, thereby avoiding the knowledge base from being polluted by low-quality or contradictory data. It may also trigger a negative update on the reliability of the data source that provides false evidence, thus realizing the online self-evolution and continuous purification of the knowledge base.

[0056] Furthermore, based on the aforementioned support level, facts are filtered. The fact filtering flowchart provided in this embodiment is as follows: Figure 2 As shown, the specific screening process is as follows: if the support of the current candidate fact is greater than 0, then the current candidate fact is accepted as a fact; otherwise, the current candidate fact is rejected as a fact.

[0057] Thus, this embodiment, by introducing the long-term reliability of the data source as a weight and combining it with decorrelation factor correction, calculates the global belief value of candidate facts in a probabilistic aggregation manner, and then adjudicates conflicts by comparing it with the log-likelihood ratio of mutually exclusive facts. This method not only amplifies the evidence contribution of high-credibility sources and suppresses the influence of low-quality data and redundant information, achieving a comprehensive balance between the quantity and quality of evidence, but also achieves the self-resolution of conflicts by actively adjudicating contradictory information, ensuring that the selected facts have a high degree of comprehensive credibility, and providing a guarantee for building a consistent and accurate knowledge base.

[0058] Step S4: Based on the evidence strength of each fact across all data sources, fuse the relation sub-vectors of each fact across all data sources to obtain a relation comprehensive vector; obtain the comprehensive average vector of all prototype vectors with the same relation sub-vectors as each fact in the knowledge base; based on the similarity between the relation comprehensive vector and the comprehensive average vector, and combined with the global belief value, determine whether to write each fact in the current time window into the knowledge base.

[0059] In this embodiment, based on the strength of evidence for each fact across all data sources, the relational sub-vectors of each fact across all data sources are fused to obtain a comprehensive relational vector. Specifically: For fact t, the embedding vectors generated from different data sources (confirmed to be the same fact by evidence counting) are used. The evidence strength of each fact under each data source within the current time window is used as the weight of the corresponding embedding vector. The embedding vectors of each fact under all data sources that provide supporting evidence are weighted and fused. The fused vector is used as the relationship synthesis vector of each fact within the current time window.

[0060] The average vector obtained by merging all prototype vectors with the same corresponding relationship in the relation sub-vector of each fact in the knowledge base is denoted as the comprehensive average vector of each fact.

[0061] Furthermore, this embodiment determines the admission score for each fact within the current time window based on the similarity between the relational composite vector and the composite average vector, combined with the global belief value. Specifically: As one implementation method, in this embodiment, the expression for the admission score of fact x within the current time window is... for: In the formula, This represents the global belief value of fact x within the current time window; This represents the similarity between the composite vector of relationships of fact x within the current time window and the composite average vector; norm[] represents the normalization function; This indicates a preset constant greater than 0, used to prevent the denominator from being 0. The value is set manually, in this embodiment. The value of is 0.01. Under the premise of ensuring that the denominator is not 0 and does not excessively affect the calculation result, the implementer can also set it according to the specific situation. This embodiment does not impose any special restrictions.

[0062] It should be noted that there are many methods to measure the similarity between vectors. In this embodiment, the cosine similarity between the comprehensive relationship vector of fact x within the current time window and the comprehensive average vector is used as the similarity between the comprehensive relationship vector of fact x within the current time window and the comprehensive average vector. In practical applications, as other implementation methods, implementers may also use other methods to measure the similarity between vectors, such as the reciprocal of Euclidean distance, depending on the specific circumstances. This embodiment does not impose any special restrictions on the selection of methods to measure the similarity between vectors.

[0063] Furthermore, based on the admission score, it is determined whether to write the fact into the knowledge base. Specifically: if the admission score of the current fact is greater than or equal to a preset first value, then the current fact is written into the knowledge base; if the admission score of the current fact is less than or equal to a preset second value, then the current fact is rejected from being written into the knowledge base; if the admission score of the current fact is between the preset first value and the preset second value, then the current fact is stored in a temporary storage area, wherein the preset first value is greater than the preset second value.

[0064] It should be noted that the preset first value and the preset second value are set manually. In this embodiment, the preset first value is 0.8 and the preset second value is 0.2. In actual application, as other implementation methods, implementers can also set them according to specific circumstances. This embodiment does not impose any special restrictions.

[0065] Thus, this embodiment generates a comprehensive relation vector by weighted fusion of multi-source evidence, and uses harmonic averaging to combine it with the similarity of existing semantics in the knowledge base and the global belief value to calculate the admission score. This mechanism constructs a double-insured admission threshold, requiring not only solid multi-source evidence to support the facts, but also a high degree of alignment with the existing semantic system of the knowledge base. By using harmonic averaging to address weaknesses, it effectively suppresses situations where the evidence is strong but semantically inconsistent or semantically consistent but the evidence is weak, thereby improving the consistency and completeness of the facts in the knowledge base.

[0066] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0067] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0068] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for constructing a knowledge base for intelligent manufacturing scenarios based on artificial intelligence, characterized in that, The method includes the following steps: Obtain all candidate facts and their relational subvectors from various data sources in an AI manufacturing scenario under different time windows; Based on the number of times each candidate fact under each data source was repeatedly counted within the preset time period before the current time window, and the similarity between the relation sub-vector of each candidate fact under each data source and its corresponding prototype vector within the current time window, the extraction confidence of each candidate fact under each data source within the current time window is determined. The long-term reliability of each data source within the current time window is obtained, and combined with the extraction confidence, the evidence strength of each candidate fact under each data source is determined to determine the global belief value of each candidate fact, and then all facts are selected from all candidate facts within the current time window. The evidence strength and embedding vectors of each fact from all data sources are combined to obtain a comprehensive relation vector; the comprehensive average vector of all prototype vectors with the same relation sub-vectors as each fact is obtained from the knowledge base; based on the similarity between the comprehensive relation vector and the comprehensive average vector, and combined with the global belief value, it is determined whether to write each fact in the current time window into the knowledge base. The strength of evidence for each candidate fact under each data source is the product of the extraction confidence and the corresponding long-term reliability of each candidate fact under each data source. The determination of the global belief value for each candidate fact includes: The evidence strength of each candidate fact under each data source within all time windows in the preset period prior to the current time window is used to form the evidence sequence of each candidate fact under each data source in the current time window. Based on the correlation of evidence sequences for each candidate fact among all data sources within the current time window, and in conjunction with the strength of the evidence, the global belief value for each candidate fact within the current window is determined, specifically: Global belief value of candidate fact t within the current time window The expression is: In the formula, This indicates the strength of evidence for candidate fact t under data source s within the current window; This represents the mean of the correlation coefficients of the evidence sequences of candidate facts t among all data sources within the current window; This represents the data source that provides all supporting evidence for candidate fact t.

2. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 1, characterized in that, The method for determining the extraction confidence of each candidate fact from each data source within the current time window is as follows: Based on the number of times each candidate fact under each data source was repeatedly counted within the preset time period before the current time window, the source consistency factor of each candidate fact under each data source within the current time window is determined. Based on the similarity between the relation subvector and its corresponding prototype vector of each candidate fact under each data source in the current time window, determine the relation consistency factor of each candidate fact under each data source in the current time window. The extraction confidence of each candidate fact from each data source within the current time window is the result of a positive fusion of the intra-source consistency factor and the relational consistency factor of each candidate fact from each data source within the current time window.

3. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 2, characterized in that, The expression for the intra-source consistency factor of each candidate fact under each data source within the current time window is: In the formula, This represents the intra-source consistency factor of candidate facts t under data source s within the current time window; This represents the evidence count of candidate facts t under data source s within a preset time period prior to the current time window; Indicates a preset value; min[ ] represents the minimum value function.

4. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 2, characterized in that, The expression for the consistency factor of the relationship of each candidate fact under each data source within the current time window is: In the formula, This represents the consistency factor of the relationship between candidate facts t under data source s within the current time window; This represents the similarity between the relational subvector and the corresponding prototype vector of candidate fact t under data source s within the current time window; max[] represents the maximum value function.

5. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 1, characterized in that, The process of obtaining the long-term reliability of each data source within the current time window includes: Long-term reliability of data source s within the current time window The expression is: In the formula, , These represent the number of true cases and the number of false cases in data source s within the preset time period prior to the current time window, respectively. , These represent the preset initial number of true cases and the preset initial number of false cases, respectively.

6. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 1, characterized in that, The process of filtering all facts from all candidate facts within the current time window includes: Based on the global belief value, determine the support level of each candidate fact within the current time window; If the support of the current candidate fact is greater than 0, then the current candidate fact is accepted as a fact; otherwise, the current candidate fact is rejected as a fact.

7. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 1, characterized in that, The fusion of the evidence strength and embedding vectors of each fact across all data sources yields a comprehensive relation vector, including: The evidence strength of each fact under each data source within the current time window is used as the weight of the corresponding embedding vector. The embedding vectors of each fact under all data sources that provide supporting evidence are weighted and fused. The fused vector is used as the comprehensive relationship vector of each fact within the current time window.

8. The method for constructing an intelligent manufacturing scenario knowledge base based on artificial intelligence as described in claim 1, characterized in that, The determination of whether to write each fact within the current time window into the knowledge base includes: Based on the similarity between the relational composite vector and the composite average vector, and combined with the global belief value, the admission score for each fact within the current time window is determined. If the admission score of the current fact is greater than or equal to the preset first value, the current fact is written into the knowledge base; if the admission score of the current fact is less than or equal to the preset second value, the current fact is rejected from being written into the knowledge base; if the admission score of the current fact is between the preset first value and the preset second value, the current fact is stored in the temporary storage area, where the preset first value is greater than the preset second value.

Images