Process parameter decision-making system and method based on multi-source information fusion

Abstract

The invention discloses a process parameter decision-making system and method based on multi-source information fusion, and relates to the technical field of conflict detection.The process parameter decision-making method comprises the steps that multi-source heterogeneous data in the process production process is collected, and feature extraction and uncertainty quantification are conducted on the preprocessed multi-source heterogeneous data; constructing a process knowledge graph, calculating distances and conflict entropies among different evidences, calculating evidence credibility weights by using the distances among the evidences, and correcting each evidence by using the credibility weights; carrying out evidence fusion by using an evidence theory to obtain a probability distribution function of comprehensive evidence; determining a current process state and confidence according to the fused evidence, extracting different process parameter adjustment schemes associated with the current process state from the knowledge graph, and selecting an optimal process parameter adjustment scheme by using a neural network algorithm; and executing the optimal process parameter adjustment scheme, collecting multi-source heterogeneous data of the next production cycle, and judging whether the optimal process parameter adjustment scheme is effective or not.

Description

Technical Field

[0001] This invention relates to the field of conflict detection technology, specifically to a process parameter decision-making system and method based on multi-source information fusion. Background Technology

[0002] With the deepening of intelligent manufacturing, the manufacturing industry is transforming towards flexibility, customization, and green practices. Multi-variety, small-batch customized production models are becoming mainstream, requiring process parameter decision systems to quickly respond to the process requirements of different products and achieve "one-click switching" parameter configuration. For example, in 3C electronic product manufacturing, the same production line needs to be compatible with the assembly processes of different products such as mobile phones and tablets; traditional fixed parameter modes cannot meet the need for rapid switching. Consumers have increasingly stringent requirements for product quality, and the industrial sector has imposed mandatory requirements for product lifecycle traceability. This necessitates process parameter decision systems that can link data from the entire production process to achieve precise location of quality problems and reverse optimization of process parameters. With the development of the Industrial Internet and big data technologies propelling the manufacturing industry into a stage of "deep integration of informatization and industrialization," multi-source information fusion technology is being widely applied in process parameter decision-making.

[0003] However, traditional process parameter decisions often rely on a single data source (such as sensor data or empirical formulas) or adopt a simple data overlay and fusion method, resulting in one-sided information utilization, weak ability to process conflicting information, and difficulty in coping with complex working conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a process parameter decision-making system and method based on multi-source information fusion to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A process parameter decision-making method based on multi-source information fusion, characterized in that the method includes the following steps:

[0007] S100: Collect multi-source heterogeneous data during the production process and preprocess the multi-source heterogeneous data;

[0008] Furthermore, the specific steps for preprocessing multi-source heterogeneous data are as follows:

[0009] S101. Collect multi-source heterogeneous data from different sources, in different formats, and at different collection frequencies during the production process. The multi-source heterogeneous data includes physical sensor data, quality inspection data, production management data, and expert experience and knowledge.

[0010] Examples of physical sensing data include time-series data such as temperature, pressure, and vibration; examples of quality inspection data include discrete data such as product dimensions, surface defects, and performance indicators; examples of production management data include structured data such as material batches, equipment IDs, and maintenance records; and examples of expert experience and knowledge include generation rules or language descriptions.

[0011] S102, The preprocessing includes format normalization, time alignment, and spatial alignment;

[0012] The format standardization means that all data are uniformly mapped to a preset standardized data model; the time alignment means that for time-series data, time synchronization is performed by interpolation or resampling based on a unified timestamp; the spatial alignment means that quality inspection data is precisely associated with processing position and equipment unit.

[0013] It covers four major categories of data: physical sensing, quality inspection, production management, and expert experience, avoiding the limitations of a single data source, ensuring the integrity of decision-making basis, and solving the problems of single data source and one-sided information in traditional process parameter decision-making.

[0014] S200. Perform feature extraction and uncertainty quantification on the preprocessed multi-source heterogeneous data;

[0015] Furthermore, the specific steps for feature extraction and uncertainty quantification of the preprocessed multi-source heterogeneous data are as follows:

[0016] S201. Perform time-frequency domain analysis on physical sensing data to extract time-domain features, frequency-domain features, and time-frequency domain features. The time-domain features include mean, standard deviation, peak value, root mean square skewness, etc. The frequency-domain features are obtained through Fourier transform and include spectrum, dominant frequency, average frequency, spectral centroid, and spectral bandwidth. The time-frequency domain features are obtained through wavelet transform and include wavelet coefficients and wavelet entropy.

[0017] Statistical analysis is performed on the quality inspection data to extract statistical and distribution features. The statistical features include batch mean, batch standard deviation, and pass rate. The distribution features include histogram, mode, interquartile range, skewness, and kurtosis.

[0018] Numerical features, categorical feature codes, and derived features are extracted from production management data. The numerical features include directly used numerical data and normalized numerical data. The categorical feature codes include process status, which includes normal, warning, and fault, and are respectively coded as 0, 1, and 2. The derived features include equipment continuous running time, time since last calibration, and material usage batch.

[0019] The method involves rule parsing and fuzzy language quantization based on expert experience and knowledge. The rule parsing involves extracting the judgment threshold from the rule, and the fuzzy language quantization involves representing the fuzzy language using a membership function.

[0020] We employ specialized extraction methods for different types of data to maximize the retention of core data information and avoid feature loss caused by general extraction methods, thus providing high-quality feature input for subsequent decision-making.

[0021] S202. Quantify the uncertainty of different multi-source heterogeneous data and calculate the data quality score;

[0022] The signal-to-noise ratio (SNR) factor, calibration status factor, and process status factor are calculated and weighted for each physical sensor data. Specifically, the SNR factor is calculated by dividing the data standard deviation by the mean to obtain the SNR level, then calculating the reciprocal of the SNR level and normalizing it. The calibration status factor is calculated by subtracting the last calibration date from the current date to obtain the calibration days, then calculating the reciprocal of the calibration days and normalizing it. The process status factor is calculated by dividing the historical number of equipment failures by the historical number of equipment runs to obtain the equipment failure rate, then calculating the reciprocal of the equipment failure rate and normalizing it.

[0023] The sampling coverage score is calculated by weighted summation of the equipment accuracy score and the sampling coverage score based on the quality inspection data. The equipment accuracy score is obtained by normalizing the data after collection using a lookup table method. The sampling coverage score is specifically obtained by dividing the number of inspection samples by the total number of production samples to obtain the sampling ratio, and then normalizing the sampling ratio to obtain the sampling coverage score.

[0024] The error rate is obtained by subtracting the data error rate from 1 for production management data. Specifically, the data error rate is calculated by dividing the number of erroneous data points in the production management data by the total number of data points, and then normalizing the result.

[0025] The expert experience knowledge is calculated by weighted summation of expert rating score, historical accuracy rate, and rule clarity. The expert rating score is assigned by staff and ranges from 0 to 1. The historical accuracy rate is calculated by dividing the number of correct suggestions made by each expert's experience by the total number of suggestions and then normalizing. The rule clarity is assigned by staff after reading the data and ranges from 0 to 1. The weights for weighted fusion in calculating the data quality score are all set by staff experience.

[0026] By calculating data quality scores, the reliability of different data sources is quantified, enabling the decision-making process to identify low-quality data and reduce its interference with the final results. The quality score dimensions of various types of data are aligned with the characteristics of the data, ensuring the objectivity and relevance of the quantitative results.

[0027] S300. Construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph for conflict detection.

[0028] Furthermore, the specific steps for conflict detection using knowledge graphs are as follows:

[0029] S301. Extract the process objects and elements in the production process as entities. The process objects include products, parts, etc., and the process elements include equipment, production parameters, production steps, etc. Extract different production relationships in the production process as edges. The production relationships include equipment-production parameters, production steps-standards, faults-fault causes, etc. For example: "CNC machine tool" → "Needs to be set" → "Speed ​​1500 r / min"; Use entities as knowledge graph nodes and production relationships as knowledge graph edges to construct a knowledge graph; Transform process objects, elements, and relationships into entities and edges to form a structured knowledge network, enabling the dispersed multi-source data features to form connections, intuitively presenting the causal logic between data (such as parameter-state, fault-cause), and providing knowledge support for decision-making.

[0030] Features of multi-source heterogeneous data are mapped onto corresponding knowledge graph nodes, and knowledge graph edges are supplemented based on the causal relationships between different features of multi-source heterogeneous data in historical production records; the causal relationship represents the relationship in which the occurrence of one feature in historical production records leads to the occurrence of another feature; for example, in expert rules, exceeding a threshold leads to the judgment of equipment failure and the issuance of an early warning.

[0031] S302. In a knowledge graph, when the characteristics of multi-source heterogeneous data from different data sources differ based on the judgments of knowledge graph edges for the same process state, calculate the confidence level of each judgment path and the conflict metric. The formula is as follows:

[0032] ;

[0033] In the formula, Conflict(S) represents the conflict measure of the same process state S, and B i and B j These represent the confidence levels of feature judgment paths for multi-source heterogeneous data from different data sources;

[0034] Preset conflict threshold T c When the conflict metric exceeds the conflict threshold, a significant conflict is identified. By calculating the confidence level and conflict metric of decision paths from different data sources, data contradictions are accurately identified, preventing erroneous decisions caused by data conflicts and proactively identifying decision risks.

[0035] S400. Construct a basic probability allocation function for the process status, use the characteristics of different data sources as evidence, calculate the distance and conflict entropy between different pieces of evidence, use the distance between evidence to calculate the credibility weight of the evidence, and use the credibility weight to correct each piece of evidence; use evidence theory to fuse evidence to obtain the probability allocation function of comprehensive evidence.

[0036] Furthermore, the specific steps for using evidence theory to fuse evidence and eliminate conflicts are as follows:

[0037] S401. For the same process state identification problem, define an identification framework, where the identification framework represents the set of all mutually exclusive process states; construct a power set by including all subsets, the empty set, and the framework itself; for each data source, construct a basic probability allocation function m based on the feature values ​​of multi-source heterogeneous data and the knowledge graph, constraining the probability that the basic probability allocation function is an empty set to 0, and the sum of the probabilities of all elements in the power set to 1; the basic probability allocation function is specifically as follows: In the formula, m x (D) represents the probability that data source x infers the process state as D, u D Q represents the fuzzy membership degree of the eigenvalue in state D; s The data quality score represents multi-source heterogeneous data; the fuzzy membership degree is calculated using a Gaussian membership function; an assignment function is constructed based on data feature values, knowledge graphs, and data quality scores to ensure that the probability allocation of evidence aligns with the actual reliability of the data and avoids subjective assumptions.

[0038] S402. Using the feature values ​​of multi-source heterogeneous data from each data source as evidence, calculate the network distance based on the fundamental probability assignment function of the two pieces of evidence, using the following formula:

[0039] ;

[0040] In the formula, d(m1, m2) represents the network distance of the basic probability assignment functions of the two pieces of evidence, m1 and m2 represent the vector representations of the basic probability assignment functions of the two pieces of evidence, and G represents the matrix of power set m1 × power set m2, where the elements are G(A, B) = |A∩B| / |A∪B|; A and B represent subsets of the power set.

[0041] For example: Suppose the identification frame Θ={S,I,O}, then the power set has 8 elements; the BPAs of the two pieces of evidence m1 and m2 are as follows:

[0042] m1:m({S})=0.6,m({I})=0.2,m(Θ)=0.2

[0043] m2:m({S})=0.1,m({I,O})=0.7,m(Θ)=0.2

[0044] Calculate the difference between matrix D and vector to obtain d(m1, m2).

[0045] The formula for calculating conflict entropy is: In the formula, K represents the conflict entropy, ∩Θ x =∅ represents the constraint condition for summation, specifically: only considering the case where the intersection of all evidence identification framework subsets is an empty set; Θ x The framework for identifying evidence x; the greater the conflict entropy, the more severe the conflict.

[0046] S403. Calculate the credibility weight using the network distance and data quality score between different pieces of evidence, using the following formula:

[0047] ;

[0048] In the formula, w x d represents the credibility weight of evidence x. p This represents the average distance between evidence x and the remaining evidence;

[0049] Each piece of evidence is adjusted using a credibility weight, as shown in the formula:

[0050] ;

[0051] In the formula, m x ’ (H) represents the probability that evidence x is D after modification, and Θ represents the identification frame;

[0052] By applying Dempster's evidence theory combination rule to fuse all conflicting and corrected evidence, a probability allocation function m for the comprehensive evidence is obtained. final The method calculates the distance and conflict entropy between pieces of evidence to quantify the degree of evidence conflict and provide a basis for correction. It uses credibility weights to correct evidence, reduce the impact of low-credibility evidence, and improve the accuracy of fusion. Through a fusion method based on Dempster's evidence theory, it effectively integrates multi-source evidence, eliminates conflicts, forms a unified comprehensive evidence probability allocation, solves the decision-making dilemma caused by contradictions in multi-source data, and improves the consistency and reliability of decisions.

[0053] S500: Determine the current process status and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process status from the knowledge graph, and use a neural network algorithm to select the optimal process parameter adjustment scheme.

[0054] Furthermore, the specific steps for selecting the optimal decision using a neural network model are as follows:

[0055] S501. For the same process state, select the state with the highest probability in the probability allocation function of comprehensive evidence as the current process state; extract process parameter adjustment schemes that have a relationship edge with the current process state from the knowledge graph to form a candidate set; use a neural network algorithm to calculate the impact and confidence of the process parameter adjustment schemes in the candidate set on multiple objectives, including production quality, production cost, and efficiency.

[0056] A multi-objective space is constructed, where each process parameter adjustment scheme represents a point in the multi-dimensional space. The point coordinates include the impact of the process parameter adjustment scheme on multiple objectives. The optimal process parameter adjustment scheme is obtained by solving the Pareto optimality algorithm in the multi-objective space. By combining neural network algorithms and Pareto optimality algorithms, multiple objectives such as production quality, cost, and efficiency are taken into account, avoiding the decline in overall benefits caused by optimizing a single objective, and selecting the globally optimal solution rather than a locally optimal one.

[0057] S600: Execute the optimal process parameter adjustment plan and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment plan is effective.

[0058] Furthermore, the specific steps to determine whether the optimal process parameter adjustment scheme is effective are as follows:

[0059] S601. Execute the optimal process parameter adjustment plan and collect new multi-source heterogeneous data in the next production cycle. Calculate the data quality score of the new multi-source heterogeneous data. If the data quality score is higher than that before the optimal process parameter adjustment plan was executed, the plan is deemed effective; otherwise, it is deemed invalid. When the plan is deemed invalid, discard the current adjustment and return to step S400 to recalculate.

[0060] By collecting multi-source data from the next production cycle, the effectiveness of the solution is verified, forming a closed loop of "decision-execution-verification-feedback" to ensure that the solution truly improves production efficiency. When the data quality score does not improve, invalid adjustments are discarded and the calculation is repeated to avoid resource waste and production losses caused by continuously using incorrect solutions, thus ensuring dynamic optimization of decisions.

[0061] A process parameter decision-making system based on multi-source information fusion, comprising a data acquisition module, a feature extraction module, a conflict detection unit, a conflict elimination module, an optimal solution selection module, and a process parameter adjustment module;

[0062] The data acquisition module is used to collect multi-source heterogeneous data during the production process and to preprocess the multi-source heterogeneous data.

[0063] The feature extraction module is used to extract features and quantify uncertainty in the preprocessed multi-source heterogeneous data.

[0064] The conflict detection unit is used to construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph to perform conflict detection.

[0065] The conflict resolution module is used to construct a basic probability allocation function for the process state, using the characteristics of different data sources as evidence, calculating the distance and conflict entropy between different pieces of evidence, calculating the credibility weight of evidence using the distance between evidence, and correcting each piece of evidence using the credibility weight; and using evidence theory to fuse evidence to obtain a probability allocation function for comprehensive evidence.

[0066] The optimal solution screening module is used to determine the current process state and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process state from the knowledge graph, and select the process parameter adjustment scheme using a neural network algorithm.

[0067] The process parameter adjustment module is used to execute the optimal process parameter adjustment scheme and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment scheme is effective.

[0068] The feature extraction module includes a feature extraction unit and a data quality scoring unit;

[0069] The feature extraction unit is used to extract features using different methods for different multi-source heterogeneous data;

[0070] The data quality scoring unit is used to quantify the uncertainty of different multi-source heterogeneous data and calculate the data quality score.

[0071] The conflict resolution module includes a basic probability allocation function unit, an evidence correction unit, and an evidence fusion unit;

[0072] The basic probability allocation function unit is used to define an identification framework for the same process state identification problem, and construct a power set by constructing all subsets, the empty set, and itself of the identification framework; and to construct a basic probability allocation function m for each data source based on the feature values ​​of multi-source heterogeneous data and the knowledge graph.

[0073] The evidence correction unit is used to calculate the credibility weight using the network distance and data quality score between different pieces of evidence, and to correct each piece of evidence using the credibility weight;

[0074] The evidence fusion unit is used to fuse all conflicting and corrected evidence using Dempster's evidence theory combination rules to obtain a probability allocation function m of the comprehensive evidence. final .

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] 1. This invention breaks through the limitations of traditional single data sources, comprehensively integrates four categories of heterogeneous data: physical, quality, management, and expert data, maximizes the mining of data value, and makes decision-making basis more comprehensive and rich.

[0077] 2. This invention filters noisy data, resolves data contradictions, and corrects unreliable evidence through data preprocessing, feature extraction, uncertainty quantification, conflict detection and elimination, thereby ensuring the accuracy of process status judgment and parameter adjustment schemes and reducing decision-making risks. Attached Figure Description

[0078] Figure 1 This is a schematic diagram illustrating the steps of a process parameter decision-making method based on multi-source information fusion according to the present invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] Example: Figure 1 As shown, the present invention provides a technical solution.

[0081] A process parameter decision-making method based on multi-source information fusion, characterized in that the method includes the following steps:

[0082] S100: Collect multi-source heterogeneous data during the production process and preprocess the multi-source heterogeneous data;

[0083] The specific steps for preprocessing multi-source heterogeneous data are as follows:

[0084] S101. Collect multi-source heterogeneous data from different sources, in different formats, and at different collection frequencies during the production process. The multi-source heterogeneous data includes physical sensor data, quality inspection data, production management data, and expert experience and knowledge.

[0085] Examples of physical sensing data include time-series data such as temperature, pressure, and vibration; examples of quality inspection data include discrete data such as product dimensions, surface defects, and performance indicators; examples of production management data include structured data such as material batches, equipment IDs, and maintenance records; and examples of expert experience and knowledge include generation rules or language descriptions.

[0086] S102, The preprocessing includes format normalization, time alignment, and spatial alignment;

[0087] The format standardization means that all data are uniformly mapped to a preset standardized data model; the time alignment means that for time-series data, time synchronization is performed by interpolation or resampling based on a unified timestamp; the spatial alignment means that quality inspection data is precisely associated with processing position and equipment unit.

[0088] It covers four major categories of data: physical sensing, quality inspection, production management, and expert experience, avoiding the limitations of a single data source, ensuring the integrity of decision-making basis, and solving the problems of single data source and one-sided information in traditional process parameter decision-making.

[0089] S200. Perform feature extraction and uncertainty quantification on the preprocessed multi-source heterogeneous data;

[0090] The specific steps for feature extraction and uncertainty quantification of preprocessed multi-source heterogeneous data are as follows:

[0091] S201. Perform time-frequency domain analysis on physical sensing data to extract time-domain features, frequency-domain features, and time-frequency domain features. The time-domain features include mean, standard deviation, peak value, root mean square skewness, etc. The frequency-domain features are obtained through Fourier transform and include spectrum, dominant frequency, average frequency, spectral centroid, and spectral bandwidth. The time-frequency domain features are obtained through wavelet transform and include wavelet coefficients and wavelet entropy.

[0092] Statistical analysis is performed on the quality inspection data to extract statistical and distribution features. The statistical features include batch mean, batch standard deviation, and pass rate. The distribution features include histogram, mode, interquartile range, skewness, and kurtosis.

[0093] Numerical features, categorical feature codes, and derived features are extracted from production management data. The numerical features include directly used numerical data and normalized numerical data. The categorical feature codes include process status, which includes normal, warning, and fault, and are respectively coded as 0, 1, and 2. The derived features include equipment continuous running time, time since last calibration, and material usage batch.

[0094] The method involves rule parsing and fuzzy language quantization based on expert experience and knowledge. The rule parsing involves extracting the judgment threshold from the rule, and the fuzzy language quantization involves representing the fuzzy language using a membership function.

[0095] We employ specialized extraction methods for different types of data to maximize the retention of core data information and avoid feature loss caused by general extraction methods, thus providing high-quality feature input for subsequent decision-making.

[0096] S202. Quantify the uncertainty of different multi-source heterogeneous data and calculate the data quality score;

[0097] The signal-to-noise ratio (SNR) factor, calibration status factor, and process status factor are calculated and weighted for each physical sensor data. Specifically, the SNR factor is calculated by dividing the data standard deviation by the mean to obtain the SNR level, then calculating the reciprocal of the SNR level and normalizing it. The calibration status factor is calculated by subtracting the last calibration date from the current date to obtain the calibration days, then calculating the reciprocal of the calibration days and normalizing it. The process status factor is calculated by dividing the historical number of equipment failures by the historical number of equipment runs to obtain the equipment failure rate, then calculating the reciprocal of the equipment failure rate and normalizing it.

[0098] The sampling coverage score is calculated by weighted summation of the equipment accuracy score and the sampling coverage score based on the quality inspection data. The equipment accuracy score is obtained by normalizing the data after collection using a lookup table method. The sampling coverage score is specifically obtained by dividing the number of inspection samples by the total number of production samples to obtain the sampling ratio, and then normalizing the sampling ratio to obtain the sampling coverage score.

[0099] The error rate is obtained by subtracting the data error rate from 1 for production management data. Specifically, the data error rate is calculated by dividing the number of erroneous data points in the production management data by the total number of data points, and then normalizing the result.

[0100] The expert experience knowledge is calculated by weighted summation of expert rating score, historical accuracy rate, and rule clarity. The expert rating score is assigned by staff and ranges from 0 to 1. The historical accuracy rate is calculated by dividing the number of correct suggestions made by each expert's experience by the total number of suggestions and then normalizing. The rule clarity is assigned by staff after reading the data and ranges from 0 to 1. The weights for weighted fusion in calculating the data quality score are all set by staff experience.

[0101] By calculating data quality scores, the reliability of different data sources is quantified, enabling the decision-making process to identify low-quality data and reduce its interference with the final results. The quality score dimensions of various types of data are aligned with the characteristics of the data, ensuring the objectivity and relevance of the quantitative results.

[0102] S300. Construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph for conflict detection.

[0103] The specific steps for conflict detection using knowledge graphs are as follows:

[0104] S301. Extract the process objects and elements in the production process as entities. The process objects include products, parts, etc., and the process elements include equipment, production parameters, production steps, etc. Extract different production relationships in the production process as edges. The production relationships include equipment-production parameters, production steps-standards, faults-fault causes, etc. For example: "CNC machine tool" → "Needs to be set" → "Speed ​​1500 r / min"; Use entities as knowledge graph nodes and production relationships as knowledge graph edges to construct a knowledge graph; Transform process objects, elements, and relationships into entities and edges to form a structured knowledge network, enabling the dispersed multi-source data features to form connections, intuitively presenting the causal logic between data (such as parameter-state, fault-cause), and providing knowledge support for decision-making.

[0105] Features of multi-source heterogeneous data are mapped onto corresponding knowledge graph nodes, and knowledge graph edges are supplemented based on the causal relationships between different features of multi-source heterogeneous data in historical production records; the causal relationship represents the relationship in which the occurrence of one feature in historical production records leads to the occurrence of another feature; for example, in expert rules, exceeding a threshold leads to the judgment of equipment failure and the issuance of an early warning.

[0106] S302. In a knowledge graph, when the characteristics of multi-source heterogeneous data from different data sources differ based on the judgments of knowledge graph edges for the same process state, calculate the confidence level of each judgment path and the conflict metric. The formula is as follows:

[0107] ;

[0108] In the formula, Conflict(S) represents the conflict measure of the same process state S, and B i and B j These represent the confidence levels of feature judgment paths for multi-source heterogeneous data from different data sources;

[0109] Preset conflict threshold T c When the conflict metric exceeds the conflict threshold, a significant conflict is identified. By calculating the confidence level and conflict metric of decision paths from different data sources, data contradictions are accurately identified, preventing erroneous decisions caused by data conflicts and proactively identifying decision risks.

[0110] S400. Construct a basic probability allocation function for the process status, use the characteristics of different data sources as evidence, calculate the distance and conflict entropy between different pieces of evidence, use the distance between evidence to calculate the credibility weight of the evidence, and use the credibility weight to correct each piece of evidence; use evidence theory to fuse evidence to obtain the probability allocation function of comprehensive evidence.

[0111] The specific steps for using evidence theory to fuse evidence and eliminate conflicts are as follows:

[0112] S401. For the same process state identification problem, define an identification framework, where the identification framework represents the set of all mutually exclusive process states; construct a power set by including all subsets, the empty set, and the framework itself; for each data source, construct a basic probability allocation function m based on the feature values ​​of multi-source heterogeneous data and the knowledge graph, constraining the probability that the basic probability allocation function is an empty set to 0, and the sum of the probabilities of all elements in the power set to 1; the basic probability allocation function is specifically as follows: In the formula, m x (D) represents the probability that data source x infers the process state as D, u D Q represents the fuzzy membership degree of the eigenvalue in state D; s The data quality score represents multi-source heterogeneous data; the fuzzy membership degree is calculated using a Gaussian membership function; an assignment function is constructed based on data feature values, knowledge graphs, and data quality scores to ensure that the probability allocation of evidence aligns with the actual reliability of the data and avoids subjective assumptions.

[0113] Temperature sensor determines whether the heat input is sufficient.

[0114] Assume the identification frame Θ = {sufficient (S), insufficient (I), excessive (O)}

[0115] Calculate initial support based on eigenvalues:

[0116] Temperature characteristic value T=215°C

[0117] Ideal temperature range: 200-220°C (sufficient)

[0118] Fuzzy membership degree calculation:

[0119] μ S (T)=exp(-(T-210)² / (2×5²))=0.78

[0120] μ I (T)=exp(-(T-190)² / (2×8²))=0.15

[0121] μ O (T)=exp(-(T-230)² / (2×6²))=0.07

[0122] Combined with quality score Q s Correction:

[0123] The sensor has a quality rating of Q. s =0.85 (from step 2)

[0124] The revised basic probability assignment function BPA:

[0125] mx ({S}) = 0.78 × 0.85 = 0.663

[0126] m x ({I}) = 0.15 × 0.85 = 0.128

[0127] m x ({O}) = 0.07 × 0.85 = 0.060

[0128] m(Θ) = 1 - Q x =0.15 (representing the uncertain part)

[0129] Σm x =0.663+0.128+0.060+0.15=1.001≈1 (calculation error);

[0130] S402. Using the feature values ​​of multi-source heterogeneous data from each data source as evidence, calculate the network distance based on the fundamental probability assignment function of the two pieces of evidence, using the following formula:

[0131] ;

[0132] In the formula, d(m1, m2) represents the network distance of the basic probability assignment functions of the two pieces of evidence, m1 and m2 represent the vector representations of the basic probability assignment functions of the two pieces of evidence, and G represents the matrix of power set m1 × power set m2, where the elements are G(A, B) = |A∩B| / |A∪B|; A and B represent subsets of the power set.

[0133] For example: Suppose the identification frame Θ={S,I,O}, then the power set has 8 elements; the BPAs of the two pieces of evidence m1 and m2 are as follows:

[0134] m1:m({S})=0.6,m({I})=0.2,m(Θ)=0.2

[0135] m2:m({S})=0.1,m({I,O})=0.7,m(Θ)=0.2

[0136] Calculate the difference between matrix D and vector to obtain d(m1, m2).

[0137] The formula for calculating conflict entropy is: In the formula, K represents the conflict entropy, ∩Θ x =∅ represents the constraint condition for summation, specifically: only considering the case where the intersection of all evidence identification framework subsets is an empty set; Θ x The framework for identifying evidence x; the greater the conflict entropy, the more severe the conflict.

[0138] S403. Calculate the credibility weight using the network distance and data quality score between different pieces of evidence, using the following formula:

[0139] ;

[0140] In the formula, w x d represents the credibility weight of evidence x. p This represents the average distance between evidence x and the remaining evidence;

[0141] Each piece of evidence is adjusted using a credibility weight, as shown in the formula:

[0142] ;

[0143] In the formula, m x ’ (H) represents the probability that evidence x is D after modification, and Θ represents the identification frame;

[0144] By applying Dempster's evidence theory combination rule to fuse all conflicting and corrected evidence, a probability allocation function m for the comprehensive evidence is obtained. final The method calculates the distance and conflict entropy between pieces of evidence to quantify the degree of evidence conflict and provide a basis for correction. It uses credibility weights to correct evidence, reduce the impact of low-credibility evidence, and improve the accuracy of fusion. Through a fusion method based on Dempster's evidence theory, it effectively integrates multi-source evidence, eliminates conflicts, forms a unified comprehensive evidence probability allocation, solves the decision-making dilemma caused by contradictions in multi-source data, and improves the consistency and reliability of decisions.

[0145] S500: Determine the current process status and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process status from the knowledge graph, and use a neural network algorithm to select the optimal process parameter adjustment scheme.

[0146] The specific steps for selecting the optimal decision using a neural network model are as follows:

[0147] S501. For the same process state, select the state with the highest probability in the probability allocation function of comprehensive evidence as the current process state; extract process parameter adjustment schemes that have a relationship edge with the current process state from the knowledge graph to form a candidate set; use a neural network algorithm to calculate the impact and confidence of the process parameter adjustment schemes in the candidate set on multiple objectives, including production quality, production cost, and efficiency.

[0148] A multi-objective space is constructed, where each process parameter adjustment scheme represents a point in the multi-dimensional space. The point coordinates include the impact of the process parameter adjustment scheme on multiple objectives. The optimal process parameter adjustment scheme is obtained by solving the Pareto optimality algorithm in the multi-objective space. By combining neural network algorithms and Pareto optimality algorithms, multiple objectives such as production quality, cost, and efficiency are taken into account, avoiding the decline in overall benefits caused by optimizing a single objective, and selecting the globally optimal solution rather than a locally optimal one.

[0149] S600: Execute the optimal process parameter adjustment plan and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment plan is effective.

[0150] The specific steps to determine whether the optimal process parameter adjustment scheme is effective are as follows:

[0151] S601. Execute the optimal process parameter adjustment plan and collect new multi-source heterogeneous data in the next production cycle. Calculate the data quality score of the new multi-source heterogeneous data. If the data quality score is higher than that before the optimal process parameter adjustment plan was executed, the plan is deemed effective; otherwise, it is deemed invalid. When the plan is deemed invalid, discard the current adjustment and return to step S400 to recalculate.

[0152] By collecting multi-source data from the next production cycle, the effectiveness of the solution is verified, forming a closed loop of "decision-execution-verification-feedback" to ensure that the solution truly improves production efficiency. When the data quality score does not improve, invalid adjustments are discarded and the calculation is repeated to avoid resource waste and production losses caused by continuously using incorrect solutions, thus ensuring dynamic optimization of decisions.

[0153] A process parameter decision-making system based on multi-source information fusion, comprising a data acquisition module, a feature extraction module, a conflict detection unit, a conflict elimination module, an optimal solution selection module, and a process parameter adjustment module;

[0154] The data acquisition module is used to collect multi-source heterogeneous data during the production process and to preprocess the multi-source heterogeneous data.

[0155] The feature extraction module is used to extract features and quantify uncertainty in the preprocessed multi-source heterogeneous data.

[0156] The conflict detection unit is used to construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph to perform conflict detection.

[0157] The conflict resolution module is used to construct a basic probability allocation function for the process state, using the characteristics of different data sources as evidence, calculating the distance and conflict entropy between different pieces of evidence, calculating the credibility weight of evidence using the distance between evidence, and correcting each piece of evidence using the credibility weight; and using evidence theory to fuse evidence to obtain a probability allocation function for comprehensive evidence.

[0158] The optimal solution screening module is used to determine the current process state and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process state from the knowledge graph, and select the process parameter adjustment scheme using a neural network algorithm.

[0159] The process parameter adjustment module is used to execute the optimal process parameter adjustment scheme and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment scheme is effective.

[0160] The feature extraction module includes a feature extraction unit and a data quality scoring unit;

[0161] The feature extraction unit is used to extract features using different methods for different multi-source heterogeneous data;

[0162] The data quality scoring unit is used to quantify the uncertainty of different multi-source heterogeneous data and calculate the data quality score.

[0163] The conflict resolution module includes a basic probability allocation function unit, an evidence correction unit, and an evidence fusion unit;

[0164] The basic probability allocation function unit is used to define an identification framework for the same process state identification problem, and construct a power set by constructing all subsets, the empty set, and itself of the identification framework; and to construct a basic probability allocation function m for each data source based on the feature values ​​of multi-source heterogeneous data and the knowledge graph.

[0165] The evidence correction unit is used to calculate the credibility weight using the network distance and data quality score between different pieces of evidence, and to correct each piece of evidence using the credibility weight;

[0166] The evidence fusion unit is used to fuse all conflicting and corrected evidence using Dempster's evidence theory combination rules to obtain a probability allocation function m of the comprehensive evidence. final .

[0167] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A process parameter decision-making method based on multi-source information fusion, characterized in that: The method includes the following steps: S100: Collect multi-source heterogeneous data during the production process and preprocess the multi-source heterogeneous data; S200. Perform feature extraction and uncertainty quantification on the preprocessed multi-source heterogeneous data; S300. Construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph for conflict detection. S400. Construct a basic probability allocation function for the process status, use the characteristics of different data sources as evidence, calculate the distance and conflict entropy between different pieces of evidence, use the distance between evidence to calculate the credibility weight of the evidence, and use the credibility weight to correct each piece of evidence; use evidence theory to fuse evidence to obtain the probability allocation function of comprehensive evidence. S500: Determine the current process status and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process status from the knowledge graph, and use a neural network algorithm to select the optimal process parameter adjustment scheme. S600: Execute the optimal process parameter adjustment plan and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment plan is effective.

2. The process parameter decision-making method based on multi-source information fusion according to claim 1, characterized in that: The uncertainty quantification in S200 specifically refers to: Uncertainty quantification is performed on different multi-source heterogeneous data to calculate data quality scores.

3. The process parameter decision-making method based on multi-source information fusion according to claim 1, characterized in that: The specific method for conflict detection using knowledge graphs in S300 is as follows: Extract process objects and process elements from the process production as entities, extract different production relations in the process production as edges, and construct a knowledge graph by using entities as knowledge graph nodes and production relations as knowledge graph edges. The features of multi-source heterogeneous data are mapped onto the corresponding knowledge graph nodes, and the knowledge graph edges are supplemented according to the causal relationships between the features of different multi-source heterogeneous data in historical production records. In a knowledge graph, for the same process state, when the characteristics of multi-source heterogeneous data from different data sources differ according to the judgment of the knowledge graph edge, the confidence of each judgment path is calculated and the conflict metric is calculated. Preset conflict threshold T c When the conflict metric is greater than the conflict threshold, a significant conflict is determined to exist.

4. The process parameter decision-making method based on multi-source information fusion according to claim 1, characterized in that: The specific method for constructing the basic probability allocation function for the process state in S400 is as follows: For the same process state identification problem, an identification framework is defined, which represents the set of all mutually exclusive process states. All subsets, the empty set, and the framework itself are constructed into a power set. For each data source, a basic probability assignment function m is constructed based on the feature values ​​of multi-source heterogeneous data and the knowledge graph. The probability that the basic probability assignment function is an empty set is constrained to 0, and the sum of the probabilities of all elements in the power set is 1. The basic probability assignment function is obtained by multiplying the fuzzy membership degree and the data quality score, and the fuzzy membership degree is calculated using a Gaussian membership function.

5. The process parameter decision-making method based on multi-source information fusion according to claim 2, characterized in that: The probability allocation function for obtaining comprehensive evidence through evidence fusion using evidence theory in S400 is as follows: Using the feature values ​​of multi-source heterogeneous data from each data source as evidence, network distance and conflict entropy are calculated for the basic probability allocation function of two pieces of evidence; credibility weights are calculated using the network distance and data quality scores between different pieces of evidence, and each piece of evidence is corrected using the credibility weights; all evidence with significant conflicts and corrections are fused using the Dempster evidence theory combination rule to obtain the probability allocation function mfinal for comprehensive evidence.

6. The process parameter decision-making method based on multi-source information fusion according to claim 1, characterized in that: The specific method for selecting the optimal process parameter adjustment scheme using a neural network algorithm in S500 is as follows: For the same process state, the state with the highest probability in the probability allocation function of comprehensive evidence is selected as the current process state; process parameter adjustment schemes that have edges related to the current process state are extracted from the knowledge graph to form a candidate set, and the impact and confidence of the process parameter adjustment schemes in the candidate set on multiple objectives are calculated using a neural network algorithm; A multi-objective space is constructed, where each process parameter adjustment scheme represents a point in the multi-dimensional space. The coordinates of the point include the impact of the process parameter adjustment scheme on the multiple objectives. The optimal process parameter adjustment scheme is obtained by solving the Pareto optimality algorithm in the multi-objective space.

7. The process parameter decision-making method based on multi-source information fusion according to claim 1, characterized in that: The determination of whether the optimal process parameter adjustment scheme is effective in S600 specifically involves: The optimal process parameter adjustment scheme is executed, and new multi-source heterogeneous data is collected in the next production cycle. The data quality score of the new multi-source heterogeneous data is calculated. If the data quality score is higher than that before the optimal process parameter adjustment scheme was executed, the scheme is deemed effective; otherwise, it is deemed invalid. When the scheme is deemed invalid, the current adjustment is discarded, and the process returns to step S400 to recalculate.

8. A process parameter decision-making system based on multi-source information fusion, characterized in that: The process parameter decision system includes a data acquisition module, a feature extraction module, a conflict detection unit, a conflict elimination module, an optimal solution selection module, and a process parameter adjustment module; The data acquisition module is used to collect multi-source heterogeneous data during the production process and to preprocess the multi-source heterogeneous data. The feature extraction module is used to extract features and quantify uncertainty in the preprocessed multi-source heterogeneous data. The conflict detection unit is used to construct a process knowledge graph, map the extracted features onto knowledge graph nodes, and use the knowledge graph to perform conflict detection. The conflict resolution module is used to construct a basic probability allocation function for the process state, using the characteristics of different data sources as evidence, calculating the distance and conflict entropy between different pieces of evidence, calculating the credibility weight of evidence using the distance between evidence, and correcting each piece of evidence using the credibility weight; and using evidence theory to fuse evidence to obtain a probability allocation function for comprehensive evidence. The optimal solution screening module is used to determine the current process state and confidence level based on the fused evidence, extract different process parameter adjustment schemes associated with the current process state from the knowledge graph, and select the process parameter adjustment scheme using a neural network algorithm. The process parameter adjustment module is used to execute the optimal process parameter adjustment scheme and collect multi-source heterogeneous data for the next production cycle to determine whether the optimal process parameter adjustment scheme is effective.

9. A process parameter decision-making system based on multi-source information fusion according to claim 8, characterized in that: The feature extraction module includes a feature extraction unit and a data quality scoring unit; The feature extraction unit is used to extract features using different methods for different multi-source heterogeneous data; The data quality scoring unit is used to quantify the uncertainty of different multi-source heterogeneous data and calculate the data quality score.

10. A process parameter decision-making system based on multi-source information fusion according to claim 8, characterized in that: The conflict resolution module includes a basic probability allocation function unit, an evidence correction unit, and an evidence fusion unit. The basic probability allocation function unit is used to define an identification framework for the same process state identification problem, and construct a power set by constructing all subsets, the empty set, and itself of the identification framework; and to construct a basic probability allocation function m for each data source based on the feature values ​​of multi-source heterogeneous data and the knowledge graph. The evidence correction unit is used to calculate the credibility weight using the network distance and data quality score between different pieces of evidence, and to correct each piece of evidence using the credibility weight; The evidence fusion unit is used to fuse all the evidence with significant conflicts and corrections using the Dempster evidence theory combination rules to obtain the probability allocation function mfinal of the comprehensive evidence.

Images