A data integration classification method and system for a data center

Abstract

The application relates to the technical field of electronic digital data processing, in particular to a data integration classification method and system of a data center. The data center is used to acquire runtime sequence data of industrial equipment and labeled labels corresponding to the runtime sequence data. The occurrence frequency of a label pair formed by any two labels is used to determine the weight of a connection line between the two labels in the label pair. The runtime sequence data and the labels are used to determine a candidate cause-effect label pair formed by a candidate cause label and a candidate effect label, and a data-driven label cause-effect relationship is considered. A cause-effect comprehensive confidence of each candidate cause-effect label pair is determined according to the candidate cause-effect label pair and standardized cause-effect rules of an expert experience database. A double cause-effect relationship and a target label relationship graph of fault propagation which are fused with data driving and expert knowledge driving are determined. The missing labels and the target predefined categories are inferred, and the reliability of label inference and classification is improved.

Description

A data integration and classification method and system for data centers Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and more specifically to a data integration and classification method and system for data centers. Background Technology

[0002] In current industrial data centers, the classification of multimodal unstructured data mainly relies on pre-defined labeling systems or rule / statistical model-based automatic labeling methods. These methods perform well when labels are complete and data is standardized. However, in real-world industrial scenarios, due to diverse data sources, high labeling costs, and human oversights, labels are often incomplete or ambiguous, limiting the effectiveness of traditional classification methods. Currently, researchers are attempting to address the label missing problem using graph neural networks or knowledge graphs, but none of these methods consider causal relationships or fault propagation mechanisms between labels, making it impossible to reliably complete labels and achieve proper classification. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention aims to provide a data integration and classification method and system for data centers, the specific technical solution of which is as follows:

[0004] In a first aspect, embodiments of this application provide a data integration and classification method for a data center, including:

[0005] Equipment data of industrial equipment is obtained through a data center. The equipment data includes runtime sequence data and the corresponding labeled tags of the runtime sequence data.

[0006] Based on the occurrence frequency of any two labels forming a label pair, determine the weight of the connection line between the two labels in the label pair, and based on the runtime sequence data and the labels, determine the corresponding candidate causal label pair consisting of candidate cause labels and candidate effect labels.

[0007] Based on the candidate causal label pairs and the standardized causal rules of the expert experience database, the comprehensive causal confidence of each candidate causal label pair is determined, and the weights of the corresponding connectors are updated according to the comprehensive causal confidence to determine the target label relationship graph.

[0008] Based on the target label relationship graph, determine the causal pointing to the missing label, and based on the causal pointing to the missing label and the predefined category, infer the target pointing to the missing label and the target predefined category.

[0009] In one implementation, determining the weight of the connecting line between two tags in a tag pair based on the number of occurrences of the tag pair formed by any two tags includes:

[0010] Determine the conditional probability of each label pair based on the number of occurrences of any two labels.

[0011] Using each of the labels as nodes, the conditional probabilities are mapped to determine the weight of the connection line between the two labels in each label pair.

[0012] In one implementation, determining the candidate causal label pair consisting of the candidate cause label and the candidate effect label based on the runtime sequence data and the label includes:

[0013] By using a preset time window, the time corresponding to the runtime sequence data is divided to obtain several target time windows arranged in chronological order, and the target time window in which each of the tags is located is determined.

[0014] If a first label is determined from the labels, and a second label other than the first label appears within a preset time lag of the target time window after the first label appears, the first label is determined as a candidate cause label and the second label is determined as a candidate effect label. The process of determining the first label from the labels is repeated until all candidate cause labels and candidate effect labels are determined. Each candidate cause label and its corresponding candidate effect label constitute a candidate cause-effect label pair.

[0015] In one implementation, determining the overall causal confidence level of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database includes:

[0016] In the runtime sequence data, the first data of the candidate causal label and the second data of the corresponding candidate causal label for each candidate causal label pair are determined respectively. The first data and the corresponding second data are processed by the mutual information entropy algorithm to determine the first confidence level of each candidate causal label pair at the data level.

[0017] Each candidate cause label and its corresponding candidate effect label are matched with the expert experience database to determine the corresponding matching label pair. Then, based on the matching label pair and the standardized causal rule, the second confidence level corresponding to each matching label pair is determined.

[0018] The collection time of each first data or second data is determined based on the runtime sequence data, the rule update time of the standardized causal rule is obtained, and the first weight corresponding to each first confidence level and the second weight corresponding to each second confidence level are determined based on the collection time, the rule update time and the current time, respectively.

[0019] The causal comprehensive confidence level of each candidate causal label pair is determined by weighted summation based on the first confidence level, the first weight, the second confidence level, and the second weight.

[0020] In one implementation, determining the second confidence level corresponding to each matching label pair based on the matching label pair and the standardized causal rule includes:

[0021] Based on the matching tag pairs and the standardized causal rules, determine the tag rules corresponding to each matching tag pair;

[0022] When the labeling rule indicates a direct causal relationship, the second confidence level is determined to be within the first preset range.

[0023] When the labeling rule indicates that there is no causal relationship, the second confidence level is determined to be the specified minimum value;

[0024] When the labeling rule is a parent-child hierarchical classification, the second confidence level is determined to be the second preset range;

[0025] When the labeling rule is applicable to a specified device model, if the candidate causal label pair corresponding to the matching label pair is determined by the device data of industrial equipment of the specified device model, the second confidence level is determined to be the maximum specified value; otherwise, the second confidence level is determined to be the third preset range.

[0026] Otherwise, based on the confidence levels determined by several experts, a balanced confidence level is determined as the corresponding second confidence level.

[0027] In one implementation, determining the first weight corresponding to each first confidence level and the second weight corresponding to each second confidence level based on the collection time, the rule update time, and the current time includes:

[0028] Determine the first time difference between the current time and the collection time, and the second time difference between the current time and the rule update time;

[0029] The data timeliness of each candidate causal label pair is determined based on the reciprocal of the first time difference, and the rule timeliness of each matching label pair is determined based on the reciprocal of the second time difference.

[0030] The sum of the timeliness of the data for each candidate causal label pair and the timeliness of the rule for the corresponding matching label pair is determined.

[0031] A first weight corresponding to each first confidence level is determined based on a first ratio of the data timeliness of each candidate causal label pair to the sum of the timeliness. A second weight corresponding to each second confidence level is determined based on a second ratio of the rule timeliness of each matching label pair to the sum of the timeliness.

[0032] In one implementation, updating the weights of the corresponding connectors based on the causal comprehensive confidence level to determine the target label relationship graph includes:

[0033] Using a graph database, an initial label relationship graph is determined based on all labels and the connecting lines between them. In the initial label relationship graph, the edge attributes of each connecting line include the weight of the connecting line.

[0034] From the initial label relationship graph, determine the target connection line associated with each candidate causal label pair, and update the weight of the target connection line to the corresponding causal comprehensive confidence level to obtain the updated label relationship graph;

[0035] Candidate label pairs with direct causal relationship are determined from the updated label relationship graph. For each candidate label pair, a directed path set of all labels pointing to the result is found starting from the label of the cause using a directed graph traversal algorithm. The directed path set includes several directed paths.

[0036] The path strength of each directed path is determined based on the weights of the connecting lines of the labels traversed in each directed path. The final corrected confidence is determined based on the path strength of each directed path and the weights of the connecting lines between the candidate label pairs. The weights of the connecting lines of the candidate label pairs in the updated label relationship graph are then updated to the final corrected confidence, thus obtaining the target label relationship graph.

[0037] In one implementation, determining the final correction confidence based on the path strength of each of the directed paths and the weights of the connectors between the candidate label pairs includes:

[0038] The path length of each directed path is determined, and the path weight of each directed path is determined based on the path length.

[0039] The overall path impact of each set of directed paths is determined by weighted summation based on the path strength and path weight of each directed path.

[0040] The final corrected confidence level of the candidate label pair is determined based on the combined influence of the path and the weight of the connecting lines between the candidate label pairs.

[0041] In one implementation, the step of determining the causal pointing to the missing label based on the target label relationship graph, and inferring the target pointing to the missing label and the target predefined category based on the causal pointing to the missing label and the predefined category, includes:

[0042] Acquire new device data, and determine all known labels from the target label relationship graph based on the new device data;

[0043] Starting from each of the known labels, a directed graph traversal algorithm is used to search for causal paths pointing to the missing labels and the set of causal paths ending at the missing labels in the target label relationship graph.

[0044] In each set of causal paths, the comprehensive impact of the target path corresponding to each causal path is determined. The final corrected confidence of each connection line on each causal path is updated based on the comprehensive impact of the target path. All causal pointing missing labels on causal paths with updated final corrected confidence greater than the confidence threshold are taken as valid inference labels.

[0045] For each known label, at least one total label set corresponding to each known label is constructed based on a valid inference label corresponding to the known label, and the matching score of each total label set with the standard category label set of each predefined category is determined;

[0046] The effective inference label corresponding to the total label set with the highest matching score is used as the target pointing to the missing label, the predefined category corresponding to the highest matching score is used as the target predefined category, and the weight of the connection line between the known label and the target pointing to the missing label is updated to the corresponding highest matching score.

[0047] Secondly, embodiments of this application provide a data integration and classification system for a data center, comprising:

[0048] The acquisition module is used to acquire equipment data of industrial equipment through a data center. The equipment data includes runtime sequence data and the corresponding labeled tags of the runtime sequence data.

[0049] The first determining module is used to determine the weight of the connection line between the two labels in the label pair based on the number of occurrences of the label pair formed by any two labels, and to determine the candidate causal label pair formed by the corresponding candidate cause label and candidate effect label based on the runtime sequence data and the labels.

[0050] The second determining module is used to determine the comprehensive causal confidence of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database, and update the weight of the corresponding connection line based on the comprehensive causal confidence to determine the target label relationship graph.

[0051] The reasoning module is used to determine the causal pointing to the missing label based on the target label relationship graph, and to infer the target pointing to the missing label and the target predefined category based on the causal pointing to the missing label and the predefined category.

[0052] The present invention has the following beneficial effects:

[0053] By acquiring runtime sequence data of industrial equipment and corresponding labeled tags from the data center, the weight of the connection line between the two tags in any label pair is determined based on the frequency of occurrence of the pair. Based on the runtime sequence data and the labels, candidate causal label pairs (comprising candidate cause and candidate effect labels) are identified, considering data-driven label causal relationships. Based on the candidate causal label pairs and standardized causal rules from the expert experience database, the comprehensive causal confidence score of each candidate causal label pair is determined. The weight of the corresponding connection line is updated based on the comprehensive causal confidence score, thus determining a dual causal relationship integrating data-driven and expert knowledge-driven approaches, as well as a target label relationship graph for fault propagation. Based on the target label relationship graph, the missing causal label is identified, and based on the missing causal label and predefined categories, the target missing label and target predefined categories are inferred, which helps improve the reliability of label inference and classification. Attached Figure Description

[0054] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 is a flowchart illustrating the steps of a data integration and classification method for a data center according to an embodiment of the present invention.

[0056] Figure 2 is a structural block diagram of a data integration and classification system for a data center provided in an embodiment of the present invention. Detailed Implementation

[0057] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a data integration and classification method and system for a data center proposed according to the present invention. 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.

[0058] 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 invention pertains.

[0059] It should be noted that "exemplary" in the embodiments of this application refers to examples listed for ease of explanation, and other embodiments are not limited to the listed examples.

[0060] The following description, in conjunction with the accompanying drawings, details a specific scheme for a data integration and classification method and system for a data center provided by the present invention.

[0061] Please refer to Figure 1, which shows a flowchart of a data integration and classification method for a data center according to an embodiment of the present invention. This data integration and classification method for a data center may include at least steps S100-S400:

[0062] S100. Obtain equipment data of industrial equipment through the data center. The equipment data includes runtime sequence data and the corresponding labeled tags.

[0063] S200. Based on the number of occurrences of any two labels forming a label pair, determine the weight of the connection line between the two labels in the label pair, and based on the runtime sequence data and the labels, determine the candidate causal label pair formed by the corresponding candidate cause label and candidate effect label.

[0064] S300. Based on the candidate causal label pairs and the standardized causal rules of the expert experience database, determine the comprehensive causal confidence of each candidate causal label pair, and update the weights of the corresponding connectors based on the comprehensive causal confidence to determine the target label relationship graph.

[0065] S400. Based on the target label relationship diagram, determine the causal pointing to the missing label, and based on the causal pointing to the missing label and the predefined category, infer the target pointing to the missing label and the target predefined category.

[0066] The technical solution of this application embodiment obtains the runtime sequence data of industrial equipment and the corresponding labeled tags through a data center. Based on the frequency of occurrence of any two tags forming a tag pair, the weight of the connection line between the two tags in the tag pair is determined. Based on the runtime sequence data and the tags, candidate causal tag pairs consisting of candidate cause tags and candidate effect tags are determined, considering data-driven tag causal relationships. Based on the candidate causal tag pairs and the standardized causal rules of the expert experience database, the causal comprehensive confidence level of each candidate causal tag pair is determined. The weight of the corresponding connection line is updated based on the causal comprehensive confidence level, determining a dual causal relationship that integrates data-driven and expert knowledge-driven approaches, as well as a target tag relationship graph for fault propagation. Based on the target tag relationship graph, the causal pointing to the missing tag is determined. Based on the causal pointing to the missing tag and the predefined category, the target pointing to the missing tag and the target predefined category are inferred, which helps improve the reliability of tag inference and classification.

[0067] In one implementation, within an industrial data center, the data center is situated within a multimodal unstructured data integration and classification scenario. For example, in equipment fault diagnosis, the data may be categorized as "normal state," "mechanical fault," "electrical fault," or "sensor fault." Each category is predefined for subsequent analysis to determine which predefined category it belongs to, thus completing the classification. The equipment data collected by the data center is multimodal unstructured data from diverse sources (such as sensor time-series data, maintenance logs, and images). Tags are often missing due to data collection omissions or incomplete labeling. Therefore, the method in this embodiment uses reasoning to determine the target and fill in the missing tags, and identifies the predefined category to which it belongs, thus overcoming the shortcomings of traditional classification methods.

[0068] In one implementation, operational data of industrial equipment is acquired through a data center. Since this operational data is multi-source and heterogeneous, it may include monitoring data such as temperature, vibration, and pressure acquired from various sensors. Therefore, appropriate preprocessing is performed, including but not limited to:

[0069] 1. Perform format standardization and noise filtering to remove abnormal jumps in sensor data and irrelevant characters in text data;

[0070] 2. Perform data alignment and sampling to ensure that data from different modalities remain consistent across time or entity dimensions;

[0071] 3. Extract representative features from data from different sources, such as word embedding or topic features for text, convolutional features for images, and statistical and frequency domain features for time series data.

[0072] Then, the preprocessed runtime data (with timestamps) is used as runtime sequence data, and this runtime sequence data is associated and labeled with pre-defined tags to determine the labeled tags corresponding to the runtime sequence data. This association and labeling can be automatically performed by the system based on preset rules, such as identifying keywords. For example, if a piece of data meets a certain condition, the corresponding tag is associated and labeled, ultimately forming a feature-tag matrix, i.e., a matrix of runtime sequence data and tags, which serves as the basis for constructing the initial tag relationship graph. It should be noted that the runtime sequence data and its corresponding labeled tags constitute the equipment data of the industrial equipment, which can be obtained through the data center.

[0073] For example, the labels can be "abnormal vibration", "excessive temperature", "bearing wear", "gear damage", etc., and can be defined based on the actual situation without specific limitations.

[0074] It should be noted that label-based data classification technology is a commonly used data classification technique. It uses preset or dynamically generated labels as the basis for classification, structuring and categorizing unordered data. It is one of the fundamental technologies for data management and data analysis. The core of label classification is that labels serve as index keys. Classification is achieved by establishing a data-label correspondence, pre-planning a structured label system, and automatically matching labels using rules or algorithms. Label classification involves three key technical steps: preprocessing data, assigning labels according to rules or models, and finally classifying and storing data based on labels and building an index to ensure accuracy and efficiency. The method in this application embodiment establishes a label relationship graph that reflects the causal relationship between labels, and an intelligent reasoning network that can deeply understand and quantify the inherent propagation mechanism of faults. When a label is missing in industrial unstructured data (i.e., there is a causal link to the missing label), it can reasonably deduce the missing label, determine the target pointing to the missing label, and achieve more accurate unstructured data classification.

[0075] The correlation between industrial data tags stems primarily from the fact that multimodal data in industrial scenarios (such as text, sensor readings, and images) often describe different aspects of the same entity or event, exhibiting inherent statistical correlations or causal relationships in the real world. For instance, equipment leaks often co-occur with tags like abnormal vibrations, forming a co-occurrence pattern. Therefore, to capture the correlation between industrial data tags, resolve tag missingness and multimodal ambiguity, and provide a foundation for subsequent optimization of knowledge networks, inference of missing tags, and classification, it is necessary to quantify tag relationships through statistical co-occurrence and conditional probability calculation, and construct a corresponding tag relationship graph through structured storage in a graph database.

[0076] In one implementation, step S200 determines the weight of the connecting line between two labels in a label pair based on the number of occurrences of the label pair formed by any two labels, including steps S201-S202:

[0077] S201. Determine the conditional probability of each label pair based on the number of occurrences of any two labels.

[0078] Optionally, using device data as a data sample, statistics are compiled on any two labels (e.g., the first label). The first tag and the first Tag pairs consisting of individual tags Number of times This can also be referred to as the number of co-occurrences, generating a label co-occurrence matrix, and determining the number of co-occurrences. Total number of times each tag appears Then, determine the conditional probabilities for each label pair. :

[0079]

[0080] S202. Using each label as a node, map the conditional probability to determine the weight of the connection line between the two labels in each label pair.

[0081] Optionally, using all labels as nodes, the conditional probabilities are mapped (referring to a mathematical transformation performed on the linear fractions of the input system to generate the conditional probability), and the weight C of the connection line between the two labels in each label pair is determined. The range of the weight C is [0,1]. In this embodiment of the invention, the linear mapping can be specifically, for example, a maximum-minimum value normalization process.

[0082] It should be noted that in industrial scenarios, the propagation of faults follows a strict temporal sequence. One fault (cause) occurs first and then triggers another fault (effect) after a period of time. Therefore, by analyzing whether "the past values ​​of known label X can significantly improve the prediction accuracy of the future values ​​of label Y", this temporal dependency can be quantified, and the pure statistical co-occurrence can be distinguished from the potential causal relationship with time orientation.

[0083] In one implementation, step S200 determines candidate causal label pairs consisting of candidate cause labels and candidate effect labels based on runtime sequence data and labels, including steps S203-S204:

[0084] S203. By using a preset time window, the time corresponding to the runtime sequence data is divided to obtain several target time windows arranged in chronological order, and the target time window of each label is determined.

[0085] For example, the preset time window is 5 minutes long. Different device types can be designed with different time lengths. By using the preset time window to divide the time, several target time windows are obtained in chronological order. Based on the data corresponding to each target time window in the runtime sequence data, it is possible to determine which target time window the corresponding tag of these data is located in, and thus determine the target time window where each tag is located.

[0086] S204. Determine the first label from the labels. If the first label is within the target time window... After it appears, if a second label other than the first label appears within the preset time lag of the target time window, the first label is determined as a candidate cause label and the second label is determined as a candidate effect label. The process returns to the step of determining the first label from the labels until all candidate cause labels and candidate effect labels are determined. Each candidate cause label and its corresponding candidate effect label constitute a candidate cause-effect label pair.

[0087] Optionally, the preset time lag can be, for example, 1 minute. The Granger causality test algorithm can be used to analyze the temporal order and dependency relationship between labels. Specifically, the first label (e.g., label X) is determined from the labels. If the first label is within the target time window... After it appears, within the target time window Preset timing lag within (i.e. If a second label (e.g., label Y) appears other than the first label, the first label is determined as a candidate cause label and the second label is determined as a candidate effect label. The process of determining the first label from the labels is repeated until all candidate cause labels and candidate effect labels are determined. Each candidate cause label and its corresponding candidate effect label constitute a candidate cause-effect label pair.

[0088] In one implementation, step S300 determines the comprehensive causal confidence level of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database, including steps S301-S304:

[0089] S301. Determine the first data of the candidate causal label and the second data of the corresponding candidate causal label for each candidate causal label pair in the runtime sequence data. Process the first data and the corresponding second data respectively using the mutual information entropy algorithm to determine the first confidence level of each candidate causal label pair at the data level.

[0090] Optionally, the first data of the candidate cause label and the second data of the corresponding candidate effect label for each candidate causal label pair in the runtime sequence data are determined. Then, using the mutual information entropy algorithm, the first data of the candidate cause label and the second data of the corresponding candidate effect label in each candidate causal label pair are processed individually to obtain the first confidence level of a candidate causal label pair at the data level. Therefore, multiple processing steps can yield the first confidence level of each candidate causal label pair at the data level. The value range is [0,1]. A value closer to 1 indicates higher causal reliability at the data level. It's important to note that each candidate cause label and candidate effect label is configured with a corresponding ID in the data center, based on the candidate cause label ID, candidate effect label ID, and time-series lag. First confidence level The timestamps (or time windows) corresponding to the first and second data can constitute a set of data-driven candidate causal relationship labels.

[0091] S302. Match each candidate cause label with the corresponding candidate effect label and the expert experience database to determine the corresponding matching label pair, and determine the second confidence level for each matching label pair according to the matching label pair and the standardized causal rule.

[0092] It should be noted that the expert experience database contains expert knowledge such as fault mechanisms, maintenance cases, and physical laws. According to the causal relationship (causal certainty) determined by the fault mechanism, the corresponding expert knowledge is converted into standardized causal rules (a set of knowledge-driven causal rules). At the same time, the expert experience database contains all standard labels, standard label IDs, applicable equipment types / scenarios, and update timestamps involved in the field of industrial equipment.

[0093] Optionally, each candidate cause label and its corresponding candidate effect label are matched with standard labels from an expert experience database to determine corresponding matching label pairs. Then, based on the matching label pairs and standardized causal rules, the second confidence level for each matching label pair is determined. Specifically, based on the matching label pairs and standardized causal rules, the label rules corresponding to each matching label pair are determined.

[0094] When the labeling rule indicates a direct causal relationship (a mandatory causal rule), such as "bearing wear → excessive temperature," then the second confidence level is determined. The first preset range is [0.7, 1.0].

[0095] When the labeling rule indicates that there is no causal relationship (prohibited causal rule), such as "temperature exceeds standard → vibration abnormality", the second confidence level is set to a specified minimum value, such as 0, to directly exclude spurious associations.

[0096] When the labeling rule is a parent-child hierarchical classification, such as "mechanical failure → bearing wear", the second confidence level is determined to be the second preset range [0.8, 1.0].

[0097] When the labeling rule is applicable to a specified device model, if the candidate causal label pair corresponding to the matching label pair is determined by the device data of the industrial equipment of the specified device model, that is, the current candidate causal label pair labels the runtime sequence data of the specified device model, then the second confidence level is determined to be the maximum specified value, such as 1. Otherwise, the second confidence level is determined to be the attenuation coefficient λ. The third preset range of the attenuation coefficient λ is [0.2, 0.8], which can be configured according to the device difference.

[0098] Otherwise, if the determination of causal relationships is inconsistent when defining the labeling rules, and it is impossible to determine a unique case falling under any of the above-mentioned circumstances, a balanced confidence level is determined based on the confidence values ​​determined by several experts, and this balanced confidence level is used as the corresponding second confidence level. For example, each expert provides a confidence value, and then the weights corresponding to each expert are summed to obtain the balanced confidence level as the corresponding second confidence level. : In the formula, M represents the total number of experts. This represents the weight corresponding to the z-th expert. denoted by z, the confidence level given by the z-th expert, and norm represents the maximum and minimum value normalization.

[0099] S303. Determine the collection time of each first or second data based on the runtime sequence data, obtain the rule update time of the standardized causal rule, and determine the first weight corresponding to each first confidence level and the second weight corresponding to each second confidence level based on the collection time, rule update time and current time respectively.

[0100] It should be noted that since runtime sequence data may become outdated as equipment conditions and environment change, its value diminishes over time. Although standardized causal rules contain stable physical laws, their specific applications may need to be adjusted due to equipment iteration or cognitive updates. Therefore, the "freshness" of both is quantified by timeliness scores, and weights are assigned accordingly, so that the fusion process can automatically favor information sources that are more timely and reflect the current state.

[0101] First, determine the current time. With collection time First time difference It should be noted that the time difference in this embodiment of the invention specifically refers to the time interval and the current time. With the rule update time Second time difference Rule update time The timing of this rule update depends on the labeling rules corresponding to the matching label pairs. For example, if the second confidence level is determined based on "the existence of a direct causal relationship," then this rule update time... It refers to the update time of the label rule that indicates "a direct causal relationship exists".

[0102] Secondly, the data timeliness of each candidate causal label pair is determined based on the reciprocal of the first time difference. The rule timeliness of each matching tag pair is determined based on the reciprocal of the second time difference. .

[0103]

[0104]

[0105] Where the time unit is days, and exp represents an exponential function with the natural constant as its base.

[0106] It should be noted that, since the input to the natural exponential function (i.e., the exp function) must satisfy the dimensionless principle, the embodiments of this application have limitations in calculating the timeliness of the aforementioned data. and the timeliness of rules Previously, it was possible to set the current time. Collection time and rule update time Perform the operation to eliminate dimensions.

[0107] Then, the sum of the timeliness of the data for each candidate causal label pair and the timeliness of the rule for the corresponding matching label pair is determined. .

[0108] Finally, based on the timeliness of the data for each candidate causal label pair... The first ratio to the sum of timeliness is used to determine the first weight corresponding to each first confidence level. And the timeliness of the rules based on each matching tag pair. The second ratio to the sum of timeliness determines the second weight corresponding to each second confidence level. .

[0109]

[0110]

[0111] S304. Calculate the weighted sum based on the first confidence level, the first weight, the second confidence level, and the second weight respectively to determine the overall causal confidence level for each candidate causal label pair. (Value range [0,1]).

[0112] Specifically, the formula is:

[0113]

[0114] In one implementation, step S300 updates the weights of the corresponding connectors based on the causal comprehensive confidence level to determine the target label relationship graph, including steps S305-S308:

[0115] S305. Using a graph database, determine an initial label relationship graph based on all labels and the connecting lines between them. In the initial label relationship graph, the edge attributes of each connecting line include the weight of the connecting line.

[0116] Alternatively, an initial label relationship graph can be determined using a graph database (such as Neo4j), with labels as nodes, based on all labels and the connections between them. Each tag / node has a corresponding tag ID and tag type attributes (phenomenon type: abnormal vibration, excessive temperature, etc.; attribute type: equipment model, data collection time; attribute type is used to assist in the analysis of phenomenon type). Each connecting line has edge attributes, such as the weight of the connecting line. And the type of association (e.g., initial statistical association, i.e., the first determination of weights). (The association type will change accordingly when updating the weights later).

[0117] S306. From the initial label relationship graph, determine the target connection line associated with each candidate causal label pair, and update the weight of the target connection line to the corresponding causal comprehensive confidence level to obtain the updated label relationship graph.

[0118] Optionally, from the initial label relationship graph, the target connection line associated with each candidate causal label pair is determined, that is, the target connection line involved in the aforementioned candidate causal label pair is determined, and then the weights of the target connection lines are assigned. Updated to the corresponding causal comprehensive confidence level Update the weights of the target connectors to obtain the updated label relationship graph. Among them, updating the label relationship graph Connectors other than the target connector maintain their original weights. The weights used in subsequent calculations are all from the updated label relationship graph. The weights in the equation.

[0119] S307. From the updated label relationship graph, determine the candidate label pairs with direct causal relationship. For each candidate label pair, use a directed graph traversal algorithm to find the set of directed paths that point to the result labels starting from the cause label. The set of directed paths includes several directed paths.

[0120] Optionally, the weights of all connections in the label relationship graph will be updated. If the connection is greater than or equal to a confidence threshold, the label pair connected by the link is considered a candidate label pair with a direct causal relationship. , To update the label relationship graph One label (the label of the reason) To update the label relationship graph There are 10 labels (labels of the results). Using a directed graph traversal algorithm, based on a set maximum path depth k (e.g., k=4, limiting the search range and reducing computational complexity), we find the set of all directed paths pointing to the labels of the results, starting from the labels of the causes. The set of directed paths includes several directed paths. , Representing the There are ... , For the labels / nodes visited:

[0121] In this context, each path in the directed path is an existing directed edge in the updated label relationship graph, and its direction is consistent with the causal propagation direction.

[0122] S308. Determine the path strength of each directed path based on the weights of the connecting lines of the labels passed through in each directed path. Determine the final corrected confidence based on the path strength of each directed path and the weights of the connecting lines between candidate label pairs. Update the weights of the connecting lines of candidate label pairs in the updated label relationship graph to the final corrected confidence, and obtain the target label relationship graph.

[0123] Alternatively, the chain multiplication model can be used to determine the weights of the connectors that traverse each label in the directed path. Determine the path strength of each directed path.

[0124]

[0125] in, For the first The path strength of a directed path The cumulative multiplication symbol is used. For the first The path length of a directed path. For the first The adjacent th on the directed path The, the Weights between tags / nodes .

[0126] First, determine the path length of each directed path (i.e., The path weight of each directed path is determined based on its length. .

[0127] Secondly, the overall path impact of each set of directed paths is determined by weighted summation based on the path strength and path weight of each directed path. :

[0128]

[0129] Then, based on the comprehensive impact of the path and the weights of the connections between candidate label pairs Determine the final corrected confidence level of the candidate label pairs. Specifically, it uses a nonlinear function for calculation, and the formula is as follows:

[0130] in, This is a configurable influence coefficient, ranging from [0,1], used to control the adjustment magnitude of path influence on confidence (weight). The term means that if the current weight When the confidence level is already high, the incremental information space that indirect paths can provide is very small, preventing over-correction of high-confidence (weighted) relationships; conversely, when... At lower levels, indirect paths It can play a greater corrective role, which helps to enhance the credibility of causal relationships where direct evidence is weak but indirect evidence is strong.

[0131] Finally, the label relationship graph will be updated. Weights of the connectors between candidate label pairs Updated to final corrected confidence level Obtain the target label relationship graph Among them, the target label relationship diagram The connecting lines other than the connecting lines of the candidate tag pairs remain unchanged. Medium weight The weights used in subsequent calculations are all from the target label relationship graph. The weights in the graph have now been updated, completing the update of the weights (confidence) of the two connecting lines. This includes the target label relationship graph. Persistent storage is performed in a graph database; in some implementations, false association edges can be removed based on the prohibition of causation rule in the expert rules to improve the target label relationship graph. The credibility of causality.

[0132] In one embodiment, step S400 includes steps S401-S405:

[0133] S401. Obtain new device data and determine all known labels from the target label relationship graph based on the new device data.

[0134] Optionally, a target label relationship graph representing the global relationship of labels is constructed based on existing device data. Subsequently, upon acquiring new device data, new labels can be determined within the new device data based on the same principle. Then, these new labels are used to define the target label relationship graph. Find the same tags and treat them as known tags. If there are multiple known tags Then all known labels This can yield a set of known labels.

[0135] S402. Starting from each known label, use a directed graph traversal algorithm to find the causal path set that points to the missing label and ends at the causal path set in the target label relationship graph.

[0136] Optionally, each known label can be used separately. Starting from the target label relationship graph, a directed graph traversal algorithm is used, with a maximum path depth of k=4. Searching for missing causal labels And causally pointing to missing labels The set of causal paths ending at a certain point That is, to determine the known labels Starting point and causal relationship with missing label All causal paths ending at the destination The set of causal paths constituted It should be noted that the causal relationship points to a missing label. This means, for example, that new equipment data may contain a "temperature exceeded" label, but no subsequent label causally related to "temperature exceeded" appears. If the label corresponding to "excessive temperature" is "abnormal vibration" or another label with a causal relationship, then "abnormal vibration" or another label will be used as the missing causal label. .

[0137] S403. Determine the comprehensive impact of the target path corresponding to each causal path in each causal path set, update the final corrected confidence of each connection line on each causal path according to the comprehensive impact of the target path, and take all causal pointing missing labels on the causal path with the updated final corrected confidence greater than the confidence threshold as valid inference labels.

[0138] Optionally, the combined impact of the target path corresponding to each causal path in each set of causal paths can be determined separately. (Based on the above) (The calculation principle is determined), and the final corrected confidence of each connecting line on each causal path is updated based on the comprehensive influence of the target path, that is, based on the above calculation formula. The principle, the weights in the formula during calculation for Weights in Thus, the new That is, the updated final corrected confidence level. Then, at this time Updated to new According to the new The updated final corrected confidence levels (new) will be used respectively. (greater than the confidence threshold) All causal points along the causal path are missing labels. As a valid inference label .

[0139] S404. For each known label, construct at least one total label set corresponding to each known label based on a valid inference label corresponding to the known label, and determine the matching score between each total label set and the standard category label set of each predefined category.

[0140] It should be noted that predefined categories refer to data classification categories pre-defined in industrial data center scenarios. These categories are typically defined based on business requirements or domain knowledge. For example, in equipment fault diagnosis, predefined categories may include specific fault types such as "normal state," "mechanical fault," "electrical fault," and "sensor fault." Each predefined category has a corresponding set of standard category labels. (No. A set of standard category labels for a predefined category, that is, all standard category labels belonging to that predefined category. For example, under the predefined category of "mechanical failure", the set of standard category labels may include "bearing wear" and "abnormal vibration".

[0141] Optionally, for each known label Based on the known labels A corresponding valid inference label Construct at least one total tag set corresponding to each known tag. By using the Jaccard similarity calculation method, the matching score of each total label set and the standard category label set of each predefined category is determined. :

[0142]

[0143] The numerator represents the number of identical labels after taking the intersection, and the denominator represents the total number of labels after taking the union.

[0144] S405. Take the valid inference label corresponding to the total label set with the highest matching score as the target pointing to the missing label, take the predefined category corresponding to the highest matching score as the target predefined category, and update the weight of the connection line between the known label and the target pointing to the missing label to the corresponding highest matching score.

[0145] Optionally, the highest matching score will be selected. ( It needs to be greater than or equal to The valid inference labels corresponding to the total set of labels) The system targets the missing label, thus enabling complete and automatic reasoning and completion of the missing label; it selects the highest matching score. The corresponding predefined category is used as the target predefined category to complete the automatic classification. In this embodiment, the highest matching score is determined. After that, the matching score needs to be output. Update the indexes (such as classification results, connector weights, and other indices related to inference and classification), and update the weights of the connectors between known labels and target labels pointing to missing labels to the corresponding highest matching scores. That is, to apply the new Update the weights of the corresponding connectors to determine the latest target label relationship graph. The model iteration is then completed. In some embodiments, if the inferred missing label is new, it is added as a node to the graph, and a new connection line is created based on the causal path. In addition, each time the label relationship graph is updated, the corresponding update time when the weights are updated is recorded to reflect the timeliness.

[0146] It should be noted that the embodiments of this application construct a target label relationship graph reflecting fault propagation by fusing confidence levels of dual causal relationships driven by data and expert knowledge. The inference target points to the missing label, making the label set more complete, which is conducive to improving the classification accuracy and robustness in scenarios with missing labels. The weights of the connecting lines in the target label relationship graph can be continuously iterated and updated, with an adaptive update mechanism. For example, if a causal path is verified in multiple inferences, its weight (confidence) is increased; otherwise, it is decreased, which is suitable for complex and ever-changing industrial data environments.

[0147] Referring to Figure 2, a structural block diagram of a data integration and classification system for a data center according to an embodiment of this application is shown. The system may include:

[0148] The acquisition module is used to acquire equipment data of industrial equipment through the data center. The equipment data includes runtime sequence data and the corresponding labeled tags of the runtime sequence data.

[0149] The first determining module is used to determine the weight of the connection line between the two labels in a label pair based on the number of occurrences of the label pair formed by any two labels, and to determine the candidate causal label pair formed by the corresponding candidate cause label and candidate effect label based on the runtime sequence data and the labels.

[0150] The second determination module is used to determine the comprehensive causal confidence of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database, and update the weight of the corresponding connection line based on the comprehensive causal confidence to determine the target label relationship graph.

[0151] The reasoning module is used to determine the causal link to the missing label based on the target label relationship graph, and to infer the target link to the missing label and the target predefined category based on the causal link to the missing label and the predefined category.

[0152] In this embodiment of the application, the functions of each module in the system can be found in the corresponding descriptions in the above methods, and will not be repeated here.

[0153] 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. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0154] 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.

Claims

1. A data integration and classification method for a data center, characterized in that, The method includes: acquiring equipment data of industrial equipment through a data center, the equipment data including runtime sequence data and corresponding labeled tags; determining the weight of the connection line between two tags in a label pair based on the frequency of occurrence of any two tags, and determining candidate causal label pairs consisting of candidate causal labels and candidate effect labels based on the runtime sequence data and the labels; determining the comprehensive causal confidence of each candidate causal label pair based on the candidate causal label pairs and standardized causal rules of an expert experience database, and updating the weight of the corresponding connection line based on the comprehensive causal confidence to determine a target label relationship graph; determining the causal pointing to the missing label based on the target label relationship graph, and inferring the target pointing to the missing label and the target predefined category based on the causal pointing to the missing label and a predefined category.

2. The data integration and classification method for data centers according to claim 1, characterized in that: The step of determining the weight of the connection line between two tags in a tag pair based on the occurrence frequency of any two tags includes: determining the conditional probability corresponding to each tag pair based on the occurrence frequency of any two tags; mapping the conditional probability to each tag as a node to determine the weight of the connection line between two tags in each tag pair.

3. The data integration and classification method for data centers according to claim 1, characterized in that: The step of determining the candidate causal label pair consisting of the corresponding candidate cause label and candidate effect label based on the runtime sequence data and the label includes: dividing the time corresponding to the runtime sequence data into several target time windows arranged in chronological order through a preset time window, and determining the target time window in which each label is located; determining a first label from the labels; if a second label other than the first label appears within a preset time lag of the target time window after the first label appears, determining the first label as a candidate cause label and the second label as a candidate effect label, returning to the step of determining the first label from the labels, until all candidate cause labels and candidate effect labels are determined, and each candidate cause label and the corresponding candidate effect label constitute a candidate causal label pair.

4. The data integration and classification method for data centers according to any one of claims 1-3, characterized in that: The step of determining the comprehensive causal confidence of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database includes: determining the first data of the candidate cause label and the second data of the corresponding candidate effect label for each candidate causal label pair in the runtime sequence data; processing the first data and the corresponding second data using the mutual information entropy algorithm to determine the first confidence of each candidate causal label pair at the data level; matching each candidate cause label and the corresponding candidate effect label with the expert experience database to determine the corresponding matching label pair; and determining the second confidence of each matching label pair based on the matching label pair and the standardized causal rules; determining the collection time of each first data or second data based on the runtime sequence data; obtaining the rule update time of the standardized causal rules; determining the first weight corresponding to each first confidence and the second weight corresponding to each second confidence based on the collection time, the rule update time, and the current time; and performing a weighted sum based on the first confidence, the first weight, the second confidence, and the second weight to determine the comprehensive causal confidence of each candidate causal label pair.

5. The data integration and classification method for data centers according to claim 4, characterized in that: The step of determining the second confidence level corresponding to each matching label pair based on the matching label pair and the standardized causal rule includes: determining the label rule corresponding to each matching label pair based on the matching label pair and the standardized causal rule; when the label rule indicates a direct causal relationship, determining the second confidence level to be a first preset range; when the label rule indicates no causal relationship, determining the second confidence level to be a specified minimum value; when the label rule indicates a parent-child hierarchical relationship, determining the second confidence level to be a second preset range; when the label rule applies to a specified equipment model, if the candidate causal label pair corresponding to the matching label pair is determined through equipment data of industrial equipment of the specified equipment model, determining the second confidence level to be a maximum specified value, otherwise determining the second confidence level to be a third preset range; otherwise, based on the confidence level values ​​determined by several experts, a balanced confidence level is comprehensively determined as the corresponding second confidence level.

6. The data integration and classification method for data centers according to claim 4, characterized in that: The step of determining the first weight corresponding to each first confidence level and the second weight corresponding to each second confidence level based on the collection time, the rule update time, and the current time includes: determining the first time difference between the current time and the collection time and the second time difference between the current time and the rule update time; determining the data timeliness of each candidate causal label pair based on the reciprocal of the first time difference, and determining the rule timeliness of each matching label pair based on the reciprocal of the second time difference; determining the sum of the timeliness of the data timeliness of each candidate causal label pair and the rule timeliness of the corresponding matching label pair; determining the first weight corresponding to each first confidence level based on the first ratio of the data timeliness of each candidate causal label pair to the sum of timeliness, and determining the second weight corresponding to each second confidence level based on the second ratio of the rule timeliness of each matching label pair to the sum of timeliness.

7. The data integration and classification method for data centers according to claim 1, characterized in that: The step of updating the weights of the corresponding connectors based on the causal comprehensive confidence level to determine the target label relationship graph includes: determining an initial label relationship graph using a graph database based on all labels and the connectors between them, wherein the edge attribute of each connector in the initial label relationship graph includes the weight of the connector; determining the target connector associated with each candidate causal label pair from the initial label relationship graph, and updating the weight of the target connector to the corresponding causal comprehensive confidence level to obtain an updated label relationship graph; determining candidate label pairs with direct causal relationships from the updated label relationship graph, and for each candidate label pair, using a directed graph traversal algorithm to find a set of directed paths pointing to the result labels starting from the causal label, wherein the set of directed paths includes several directed paths; determining the path strength of each directed path based on the weights of the connectors of the labels traversed in each directed path, and determining the final corrected confidence level based on the path strength of each directed path and the weights of the connectors between the candidate label pairs; updating the weights of the connectors of the candidate label pairs in the updated label relationship graph to the final corrected confidence level to obtain the target label relationship graph.

8. The data integration and classification method for data centers according to claim 7, characterized in that: The step of determining the final corrected confidence level based on the path strength of each directed path and the weight of the connection line between the candidate label pairs includes: determining the path length of each directed path and determining the path weight of each directed path based on the path length; performing a weighted summation based on the path strength and the path weight of each directed path to determine the overall path influence of each set of directed paths; and determining the final corrected confidence level of the candidate label pairs based on the overall path influence and the weight of the connection line between the candidate label pairs.

9. The data integration and classification method for data centers according to claim 7, characterized in that: The step of determining the causal pointing to the missing label based on the target label relationship graph, and inferring the target pointing to the missing label and the target predefined category based on the causal pointing to the missing label and the predefined category, includes: acquiring new device data, and determining all known labels from the target label relationship graph based on the new device data; using each known label as a starting point, searching for the causal pointing to the missing label and the set of causal paths ending at the causal pointing to the missing label in the target label relationship graph using a directed graph traversal algorithm; determining the comprehensive influence of the target path corresponding to each causal path in each causal path set, and updating the final corrected confidence of each connection line on each causal path based on the comprehensive influence of the target path. Each known label is assigned a causal missing label along a causal path where the updated final corrected confidence score is greater than the confidence threshold. For each known label, at least one total label set is constructed based on a single valid inference label. The matching score between each total label set and the standard category label sets of each predefined category is determined. The valid inference label corresponding to the total label set with the highest matching score is taken as the target missing label, the predefined category corresponding to the highest matching score is taken as the target predefined category, and the weight of the connection line between the known label and the target missing label is updated to the corresponding highest matching score.

10. A data integration and classification system for a data center, characterized in that, include: The acquisition module is used to acquire equipment data of industrial equipment through a data center. The equipment data includes runtime sequence data and the corresponding labeled tags of the runtime sequence data. The first determining module is used to determine the weight of the connection line between the two labels in the label pair based on the number of occurrences of the label pair formed by any two labels, and to determine the candidate causal label pair formed by the corresponding candidate cause label and candidate effect label based on the runtime sequence data and the labels. The second determining module is used to determine the comprehensive causal confidence of each candidate causal label pair based on the candidate causal label pairs and the standardized causal rules of the expert experience database, and update the weight of the corresponding connection line based on the comprehensive causal confidence to determine the target label relationship graph; the inference module is used to determine the causal pointing to the missing label based on the target label relationship graph, and infer the target pointing to the missing label and the target predefined category based on the causal pointing to the missing label and the predefined category.

Images