Enterprise operation index prediction system based on data mining

Abstract

The application relates to the technical field of industrial data mining, in particular to an enterprise operation index prediction system based on data mining, which comprises a heterogeneous data graph construction module, which is used for identifying equipment, materials, personnel and orders as entity nodes, defining interactive operation as a relationship, connecting all entity nodes and establishing an enterprise operation space-time dependence graph according to equipment maintenance records and operator shift log. In the application, by analyzing unstructured data such as equipment maintenance records and operator shift log, production factors such as equipment, materials, personnel and orders are constructed as entity nodes, and the relationship is defined by interactive operation, the time and space attributes of the relationship are given, and thus the space-time dependence graph reflecting the enterprise operation process is established, the complex time sequence and logical correlation among various factors in the production activities are restored, and the limitation that traditional data analysis can only process flat table data is overcome.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of industrial data mining, and in particular to an enterprise operation index prediction system based on data mining. BACKGROUND

[0002] The technical field of industrial data mining mainly uses machine learning, statistical analysis, pattern recognition and other technologies to analyze and discover knowledge from massive and heterogeneous data in the whole process of industrial production such as manufacturing, supply chain and energy. The existing industrial data mining technology breaks the dynamic and time-series interaction relationship between various elements in the production activities in the data preprocessing stage, flattens the complex process data into isolated feature vectors, loses key context information, and causes the analysis result to stay in the surface correlation. For example, a model may find that a certain operator has a strong correlation with product yield reduction, but the actual root cause may be that the operator is on duty when a certain material batch in the upstream has quality defects, and the specific equipment cooperating with it is in a sub-health state. The traditional feature model is difficult to effectively capture the complex transmission effect composed of materials, personnel and equipment, and is easy to draw incorrect attribution conclusions, leading enterprises to take ineffective management measures such as personnel retraining, and ignoring equipment or supply chain problems. Therefore, improvement is needed. SUMMARY

[0003] The application aims to solve the problems in the prior art and provides an enterprise operation index prediction system based on data mining.

[0004] In order to achieve the above-mentioned purpose, the application adopts the following technical scheme: an enterprise operation index prediction system based on data mining comprises:

[0005] A heterogeneous data graph construction module is configured to identify devices, materials, personnel and orders as entity nodes and define interactive operations as relationships according to device maintenance records and operator shift log, connect all entity nodes, and establish an enterprise operation space-time dependency graph.

[0006] A distributed summary quantization storage module is configured to write graph data into data blocks according to the enterprise operation space-time dependency graph, embed a summary structure of quantile and frequency statistics in the metadata area of each data block, generate an embedded distributed summary data block metadata, access the embedded distributed summary data block metadata, and calculate and obtain a quantization anomaly measurement value.

[0007] A space-time dependent path retrieval module is configured to set product yield index nodes and supply chain material batch nodes as start and end points of a path according to the enterprise operation space-time dependent graph and the quantified abnormality metric value, establish a target node pair with constraints, simultaneously start searching from both ends of the target node pair with constraints, eliminate paths, and generate a candidate causal relationship path set;

[0008] A key influence factor extraction module is configured to traverse each path in the path set according to the candidate causal relationship path set, extract node combinations of material batches, equipment, and operators included in the path, form a structured causal chain sequence, calculate an influence weight of each sequence according to the structured causal chain sequence, and obtain an operation index influence factor.

[0009] Preferably, the obtaining step of the enterprise operation space-time dependent graph comprises:

[0010] According to the equipment maintenance record and the operator shift log, the device number, material batch, personnel identification, and order number fields are parsed, the coding format is unified, the labels are added according to the type, the primary keys are de-duplicated, and the entity node set is formed;

[0011] According to the entity node set, the interactive operation items are extracted from the equipment maintenance record and the operator shift log, the corresponding entity nodes are matched according to the initiator and the pointing party, the operation time field and the work site field are read, and the time and coordinate representations are standardized, and the relationship set with time stamp and space location attributes is generated;

[0012] According to the relationship set with time stamp and space location attributes, the entity nodes are connected according to the relationship direction, the time sequence reference chain is established according to the time stamp, the missing pointers are checked and the irrelevant nodes are eliminated, and the enterprise operation space-time dependent graph is generated.

[0013] Preferably, the obtaining step of the metadata of the embedded distribution summary data block comprises:

[0014] According to the enterprise operation space-time dependent graph, the graph data is written into the data block according to the node time sequence, the value range and the count frequency of each interval are counted, the interval width is calculated, and the width, count, and quantile point information are embedded into the metadata area to form the metadata of the embedded distribution summary data block.

[0015] Preferably, the obtaining step of the quantified abnormality metric value comprises:

[0016] According to the metadata of the embedded distribution summary data block, the target data point is determined to be in a histogram interval and a neighborhood interval set, the count and width of each interval are read, the normalized density of each interval is calculated and saved as a density sequence, and a density parameter set is obtained;

[0017] According to the density parameter set, the quantified abnormality metric value is calculated.

[0018] Preferably, the step of obtaining the target node pair with constraints is:

[0019] According to the enterprise operation space-time dependence graph and the quantitative abnormality metric value, the product yield indicator node and the supply chain material batch node are specified as start and end points, a time window and a threshold value of the quantitative abnormality metric value are set, edges with a relationship time stamp falling within the time window and a quantitative abnormality metric value not less than the threshold value are scanned, and the start and end points are registered in pairs to form the target node pair with constraints.

[0020] Preferably, the step of obtaining the candidate causal relationship path set is:

[0021] According to the target node pair with constraints, the search is started from both ends at the same time, a forward queue and a backward queue are established, edges with a time stamp not earlier than a previous edge and not later than an upper limit of the time window are added during the expansion, edges with a quantitative abnormality metric value less than the threshold value are discarded immediately, path segments that can meet at both ends are retained, and an initial path candidate set is obtained.

[0022] According to the initial path candidate set, each path segment is spliced into a complete path from the product yield indicator node to the supply chain material batch node in chronological order, and paths with consistent time window coverage and satisfied logical constraint conditions are checked to generate a candidate causal relationship path set.

[0023] Preferably, the step of obtaining the structured causal chain sequence is:

[0024] According to the candidate causal relationship path set, the triplets of material batch nodes, equipment nodes and operator nodes are extracted in chronological order, the start and end positions of the triplets in the path and the path identifier are recorded, and a structured causal chain sequence is formed.

[0025] Preferably, the step of obtaining the operation indicator influence factor is:

[0026] According to the structured causal chain sequence, the number of occurrences is counted, the product yield indicator value when the sequence occurs and the product yield indicator value when the sequence does not occur are collected, the average path edge number of the sequence occurrence path is calculated, and a frequency and product yield statistical parameter set is obtained.

[0027] According to the frequency and product yield statistical parameter set, the operation indicator influence factor is calculated, and the enterprise operation indicator is judged according to the operation indicator influence factor.

[0028] Compared with the prior art, the application has the advantages and positive effects that:

[0029] In the present application, by analyzing the unstructured data such as equipment maintenance records and operator shift logs, the production factors such as equipment, materials, personnel and orders are constructed as entity nodes, and the relationship is defined by interactive operation, the time and space attributes are given to the relationship, thereby the space-time dependence graph reflecting the enterprise operation process is established, the complex time sequence and logical association among the factors in the production activities are restored, which overcomes the limitation of traditional data analysis that can only process flat table data, at the same time, the data is stored in blocks and the statistical summary such as quantile and frequency is embedded in the metadata, the instant access to the data distribution characteristics is realized, without scanning the original data, the statistical position of any data point in the neighborhood can be quickly estimated, and the key nodes such as product yield and material batch are set as the start and end points of the trace, combined with the time window and the threshold of abnormal value, the bidirectional search is carried out, the irrelevant path is pruned, and the candidate causal relationship chain from a specific input to the final output is locked, thereby the fuzzy correlation problem is converted into the quantitative evaluation of specific production unit combination, which enables the manager to clearly identify the key influencing factors causing the yield fluctuation, and obtain accurate operation optimization decision basis. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The system flowchart of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application, and are not used to limit the present application.

[0032] Please refer to Figure 1 The present application provides a technical scheme: an enterprise operation index prediction system based on data mining includes:

[0033] A heterogeneous data graph construction module is used for identifying equipment, materials, personnel and orders as entity nodes, defining interactive operation as relationship, connecting all entity nodes, and establishing an enterprise operation space-time dependence graph according to equipment maintenance records and operator shift logs.

[0034] A distribution summary quantization storage module is used for writing graph data into data blocks, and embedding the summary structure of quantile and frequency statistics in the metadata area of each data block, generating embedded distribution summary data block metadata, accessing the embedded distribution summary data block metadata, and calculating and obtaining quantized abnormal measurement value according to the enterprise operation space-time dependence graph.

[0035] The space-time dependent path retrieval module is configured to set product yield index nodes and supply chain material batch nodes as start and end points of a path according to the enterprise operation space-time dependent graph and the quantitative abnormality metric value, establish a target node pair with constraints, search from both ends of the target node pair with constraints at the same time, eliminate paths, and generate a candidate causal relationship path set;

[0036] The key influence factor extraction module is configured to traverse each path in the candidate causal relationship path set, extract node combinations of material batches, equipment, and operators included in the path, form a structured causal chain sequence, calculate the influence weight of each sequence according to the structured causal chain sequence, and obtain an operation index influence factor.

[0037] The acquisition steps of the enterprise operation space-time dependent graph are as follows:

[0038] According to the equipment maintenance record and the operator shift log, the device number, material batch, personnel identification, and order number fields are parsed, the coding format is unified, the labels are added according to the type, the primary keys are removed, and the entity node set is formed;

[0039] According to the entity node set, the interactive operation items are extracted from the equipment maintenance record and the operator shift log, the corresponding entity nodes are matched according to the initiator and the pointing party, the operation time field and the work site field are read, and the time and coordinate expressions are standardized to generate a relationship set with time stamp and space location attributes;

[0040] According to the relationship set with time stamp and space location attributes, the entity nodes are connected according to the relationship direction, the time sequence reference chain is established according to the time stamp, the missing pointers are checked and irrelevant nodes are eliminated, and the enterprise operation space-time dependent graph is generated.

[0041] Specifically, according to the equipment maintenance record and the operator shift log, first, the unstructured text is parsed by using a predefined regular expression rule, for example, a rule DEV-[0-9]{4,8} is designed for the equipment number, a rule MAT-[A-Z0-9]{6,12} is designed for the material batch, a rule EMP-[0-9]{5} is designed for the personnel identifier, and a rule ORD-[A-Z]{2}[0-9]{8} is designed for the order number, the log text is scanned piece by piece, and all field contents that match the pattern are extracted, next, in order to ensure the uniqueness and distinguishability of each entity in the graph, the extracted fields are uniformly formatted, and the specific operation is to add a type prefix before the original identifier to form a unified coding format, such as E:DEV-3041 representing equipment, M:MAT-XF7B34 representing material, P:EMP-89721 representing personnel, and O:ORD-BJ20230815 representing order, then, according to the added prefix, a type label is attached to each code, such as E corresponding to the "equipment" label, M corresponding to the "material" label, P corresponding to the "personnel" label, and O corresponding to the "order" label, this step also stores other descriptive information related to the extracted entities, such as equipment model, material name, and personnel name, as attributes of the nodes, subsequently, in order to eliminate data redundancy, the unified coding format is used as the primary key for deduplication processing, and in the specific implementation, a hash table is constructed, all parsed entity codes are traversed, the code is used as the key, if the hash table does not exist, the entity and its associated type label and attribute information are stored in the hash table, if it already exists, the entity is skipped, after the traversal of all log data is completed, all entries stored in the hash table constitute a set of entity nodes without duplication.

[0042] According to the entity node set generated in the preceding step, each record of the device maintenance record and operator shift log is re-traversed to extract the interactive operation entries contained therein, and the extraction process is based on a pre-set verb and entity type association rule library, for example, the rule library defines that the verbs "replace", "repair", and "maintain" are usually associated with "personnel" and "device" entities, and the verbs "take" and "feed" are usually associated with "personnel", "material", and "device" entities. By performing word segmentation and part-of-speech tagging on the log text, the key verbs are identified, and the initiator and the pointing party of the interaction are determined according to the verb association rules, for example, in the record "operator EMP-89721 replaced part A with batch MAT-XF7B34 on DEV-3041", the "replace" operation is identified, it is determined that the initiator is "personnel: EMP-89721", and the pointing party is "device: DEV-3041", and at the same time, a "material: MAT-XF7B34" to "device: DEV-3041" relationship is formed. In the matching process, the original identifiers of each entity are used to search in the entity node set to obtain their unified coding format as the nodes at both ends of the relationship. Then, the operation time field and the work site field are read from each interactive operation entry, the time field is standardized, all different formats of time representations (such as "2023 / 08 / 15 14:30", "Aug 15, 2023, 2:30 PM") are uniformly converted into 13-bit Unix timestamps in UTC time zone to realize accurate sorting and time calculation, and for the work site field, a pre-set geographic information mapping table is used for coordinate standardization, the mapping table is converted from enterprise workshop layout drawings, and the text description of the site (such as "No. 2 workshop No. 3 production line", "A7 warehouse shelf") is mapped to specific two-dimensional or three-dimensional coordinates, for example, {"No. 2 workshop No. 3 production line": (120.345, 30.678), "A7 warehouse shelf": (120.355, 30.690)}. Finally, each parsed interactive operation is integrated into a data structure containing the unified coding of the initiator node, the unified coding of the pointing party node, the standardized timestamp, and the standardized spatial coordinates. All these data structures are collected to generate a relationship set with timestamp and spatial location attributes.

[0043] According to the set of relations between the timestamp and the spatial position attribute, the uniform coding of the initiator and the pointing node recorded in the relation is used to find the corresponding entity node in the set of entity nodes, and a directed connection is established, thereby connecting the discrete nodes into an initial network structure. Subsequently, in order to reflect the dynamic evolution process of the operation activities, all the relations are globally sorted in ascending order according to their standardized timestamp attribute. The sorted relation sequence is used to construct a time sequence reference chain. Specifically, for any entity node in the graph, all the relations related to it (whether as the initiator or the pointing node) are linked in chronological order to form a timeline recording all the interaction history of the entity. Next, the integrity of the graph is checked. Each relation in the set of relations is traversed to check whether the uniform coding of the initiator and the pointing node exists in the set of entity nodes. If it is found that the node coding referred to by a certain relation does not exist in the node set, the relation is regarded as an invalid relation containing a missing pointer and is removed. After the pointer checking is completed, the relevance of the nodes in the graph is checked to remove irrelevant isolated nodes. The judgment standard of this process is whether the degree (the sum of the in-degree and the out-degree) of a node is zero. The specific operation is to calculate the total degree of each node in the graph, identify all the nodes with a degree of zero as irrelevant nodes, and remove them from the set of entity nodes. Through the connection, sorting, checking and removing operations described above, a compact and logically complete graph structure, i.e., the enterprise operation space-time dependency graph, is finally constructed.

[0044] The acquisition step of the metadata of the embedded distribution summary data block is:

[0045] According to the enterprise operation space-time dependency graph, the graph data is written into the data block according to the node time sequence, the value range and the count frequency of each interval are counted, the interval width is calculated, and the width, count and quantile point information are embedded into the metadata area to form the metadata of the embedded distribution summary data block.

[0046] Specifically, based on the enterprise's spatiotemporal dependency graph, all nodes and relationships in the graph are first globally sorted in ascending order according to their associated timestamps. Then, the sorted data stream is sequentially filled into fixed-size data blocks, for example, each data block is set to 64MB. When a data block is full, the next one is automatically created and written. Simultaneously with writing data to each data block, a parallel statistical process is initiated. This process is specifically responsible for analyzing the numerical attribute data written to the current data block, such as equipment operation duration or material consumption. For each attribute to be analyzed, the statistical process first determines its maximum and minimum values ​​within the current data block, and then calculates the width of each interval based on a preset number of intervals, for example, 100 intervals. The width is calculated by dividing the difference between the maximum and minimum values ​​by the number of intervals. Then, all data points within the data block are traversed, each data point is assigned to its corresponding interval, and the frequency of data points falling into each interval is counted. At the same time, key quantiles of the data within the data block are calculated, such as the 10th, 25th, 50th, 75th, and 90th percentiles. These quantiles are obtained by sorting the data within the data block and selecting them according to their positional proportions. Finally, the specific value range of each statistical interval, the counting frequency of the interval, the calculated uniform interval width, and all calculated quantile values ​​and their corresponding quantile values ​​are embedded as a structured summary information set into the metadata area of ​​the data block header, thus forming the embedded distributed summary data block metadata.

[0047] The steps for obtaining quantitative anomaly metrics are as follows:

[0048] Based on the metadata of the embedded distribution summary data block, determine the histogram interval and neighborhood interval set where the target data point is located, read the count and width of each interval, calculate the normalized density of each interval and save it as a density sequence to obtain the density parameter set;

[0049] Based on the density parameter set, the quantization anomaly metric is calculated using the following formula:

[0050] ;

[0051] and:

[0052] ;

[0053] in, To quantify outlier metrics, For the interpolation quantiles of the target data points, The quantiles of the left quantiles. The quantiles of the right-hand quantiles. The values ​​are taken from the left quantile. The values ​​are taken from the right-hand quantile. count of target bin, normal number to prevent denominator zero, small normal number to prevent denominator zero, smoothing constant to prevent division by zero, count of target bin, width of target bin, normalized density of target bin, normalized density of the bin in the neighborhood, neighborhood index set with the bin as the center and the radius of , inverse distance weight of neighborhood bin, neighborhood weighted average density.

[0054] Specifically, according to the embedded distribution summary data block metadata, when a new target data point needs to be evaluated, first, the value range of the bin in the metadata area record is quickly located by the binary search method to determine the specific falling interval index of the target data point, for example, if the interval range is defined as [10.0, 12.5), [12.5, 15.0), [15.0, 17.5), a target data point with a value of 14.8 will be located to the second interval. After determining the target interval, according to a preset neighborhood radius parameter, for example, set the radius to 3, to determine its neighborhood bin set, which includes the target bin itself and the 3 adjacent bins on the left and right, together forming a set of 7 bins. Then, the count frequency and interval width information of the 7 bins are accurately read from the embedded distribution summary data block metadata. Subsequently, for each of the 7 bins, the normalized density of the bin is calculated by dividing its count frequency by its interval width. This calculation converts the original frequency information into a density measure that is not affected by the interval width, allowing for a more fair comparison of the density of different width intervals. All calculated normalized density values are sequentially stored in a sequence to form a density sequence corresponding to the neighborhood bin set. This density sequence, together with the original statistical information such as the count and width of each bin, forms a density parameter set for subsequent anomaly measurement calculation.

[0055] In the formula of the quantified anomaly metric, the first part of the formula evaluates the extremeness of the data point in its overall distribution, i.e. how far it is from the median, while the second part evaluates the isolation of the data point in its local environment by comparing the density of the interval it belongs to with the average density of its neighborhood. Multiplying the two parts together makes a data point only get a high anomaly score when it satisfies both the global position remoteness and the local environment sparsity, avoiding the misjudgment of points in the tail of the distribution but belonging to a dense cluster, or points in the center of the distribution but locally isolated, thus improving the accuracy of anomaly detection;

[0056] is the interpolation quantile of the target data point, representing the relative position of the target data point in the entire data set, with a value range of 0 to 1, which is estimated using the two adjacent quantile point information stored in the embedded distribution summary data block metadata. The specific acquisition steps are as follows: first, determine the value of the target data point , then find the two nearest quantile point values and in the metadata that are just less than and greater than , and read their corresponding quantile numbers and , and calculate by linear interpolation formula. For example, the metadata records that the value of the 25th quantile point ( ) is 115.2 ( ), the value of the 50th quantile point ( ) is 140.8 ( ), and now there is a target data point , then its interpolation quantile is calculated by the formula

[0057] is the value of the target data point, representing the actual observation value that needs to be evaluated for anomaly, which comes from the specific data generated in real time during enterprise operation, such as the duration of a certain equipment operation, a certain quality detection index of a batch of materials, or the delivery period of an order. This value is the starting point for all anomaly metric calculations, and it is obtained by directly reading from the data source (such as a sensor, a log file, or a database record) without any transformation. For example, on a production line, a sensor monitors the processing time of a certain process as 16.3 seconds, so the value of is 16.3, which will be used to determine which histogram interval it belongs to and as the core input for calculating the interpolation quantile ;

[0058] and are the nearest known quantiles on both sides of the target data point, respectively, and and are the specific values corresponding to the two quantiles, all four parameters are read directly from the metadata of the embedded distribution summary data block, which pre-stores a series of quantile point pairs (quantile, quantile value), such as (0.1, 98.5), (0.25, 115.2), (0.5, 140.8), (0.75, 182.4), (0.9, 210.3), when a target data point occurs, it will search among these pre-stored quantile values to find the two quantile values and that just enclose , and simultaneously obtain their corresponding quantiles and , for example, for the target data point , by querying the above metadata, it can be determined that it falls between 115.2 and 140.8, so , , , ;

[0059] , and are very small normal numbers to prevent the denominator from being zero or invalid operation in calculation, for example ;

[0060] is the count of the interval where the target data point is located, is the width of the interval, both parameters are directly read from the metadata of the embedded distribution summary data block, when constructing the metadata, histogram statistics have been performed on the data in the data block, obtaining the specific data point quantity (count) and the width of each interval, when a target data point is located in the th interval, the count value and the width value corresponding to the index are directly extracted from the metadata, for example, the value range of a data block is 0-200, which is divided into 100 equal-width intervals, so the width of each interval is 2.0, if the target data point falls in the 8th interval [16.0, 18.0), and the metadata record shows that there are 5 data points in the interval, then , ;

[0061] is the normalized density of the target interval, is the normalized density of the interval in the neighborhood, The two parameters are calculated from the read count and the width, the formula is This calculation converts the original frequency into the frequency per unit width, i.e. density, thus eliminating the impact of inconsistent interval widths, making the densities of different intervals comparable. For example, according to the parameters obtained earlier, the density of the target interval is Similarly, for an interval in its neighborhood, if its count is and its width is , then its density is ;

[0062] is the neighborhood index set centered on the interval with a radius of , where the radius is a preset integer that defines the size of the "local" range, The choice of needs to balance the sensitivity and stability of the calculation. A smaller (such as 1 or 2) makes the calculation more sensitive to the nearest intervals, while a larger (such as 5) considers a wider context and smooths out some local noise, The value of can be optimized by cross-validation on historical data, the goal is to choose a value that best distinguishes known normal points and abnormal points on the validation set, in this example, set , if the target interval is the 8th interval ( ), then its neighborhood index set includes (excluding the center interval ), a total of 4 intervals;

[0063] is the neighborhood weighted average density, which is a comprehensive assessment of the density around the target interval, obtained by calculating the weighted average of the normalized densities of all intervals in the neighborhood, the weight is the inverse of the distance between the neighborhood interval and the target interval, i.e. This makes the closer neighbors of the target interval contribute more to the average density, more in line with the idea of locality, in the calculation, first get the normalized density of each interval in the neighborhood index set , then calculate its weight according to the distance, finally perform weighted summation and divide by the sum of the weights, for example, for the scenario of , we need to calculate The value of the target data point.

[0064] Calculate according to the parameters:

[0065] Set the target data point .

[0066] Query from the embedded distribution summary data block metadata, get the nearest quantile of : (corresponding quantile ) and (corresponding quantile ), set .

[0067] Calculate the interpolated quantile :

[0068] ;

[0069] The target data point falls in the th interval, which ranges from [16.0, 18.0) with a width , the metadata records its count .

[0070] Calculate the normalized density of the target interval:

[0071] ;

[0072] Set the neighborhood radius , the neighborhood index set is , read the count and width of the neighborhood interval from the metadata and calculate the density:

[0073] Interval 6: .

[0074] Interval 7: .

[0075] Interval 9: .

[0076] Interval 10: .

[0077] Calculate the neighborhood weighted average density :

[0078] ;

[0079] Set , .

[0080] Calculate the quantified anomaly metric value :

[0081] ;

[0082] ;

[0083] ;

[0084] This result indicates that the target data points The quantified anomaly metric was 0.5333, which was compared to a threshold to determine whether it was an anomaly. This threshold was based on all data points in the historical data. The score distribution can be used to set the threshold. For example, the 99.5 percentile of the distribution can be used as the threshold. If the threshold is set to 3.0, since the calculated 0.5333 is much smaller than 3.0, the data point is judged to be a point within the normal operating range, rather than an anomaly.

[0085] The steps for obtaining constrained target node pairs are as follows:

[0086] Based on the enterprise's spatiotemporal dependency graph and quantitative anomaly measurement values, product yield index nodes and supply chain material batch nodes are designated as start and end points. Time windows and thresholds for quantitative anomaly measurement values ​​are set. Edges whose relation timestamps fall within the time window and whose quantitative anomaly measurement values ​​are not less than the threshold are scanned and registered in pairs according to the start and end points to form constrained target node pairs.

[0087] Specifically, according to the enterprise operation space-time dependence graph and the quantitative abnormality metric value, first, from the node set of the graph, according to the type label and attribute information of the node, all nodes representing product yield indicators and nodes representing supply chain material batches are screened out, and they are respectively taken as the starting and ending point candidates of path search. Then, a dynamic time window is set, and the setting of the window is based on the production cycle. For example, for a product with an average production cycle of 7 days, the time window can be set to 10 days before the product yield indicator recording time point, that is, [T-10 days, T], where T is the yield recording time. The window length is slightly longer than the average cycle to cover possible pre-impact factors. At the same time, the threshold value of the quantitative abnormality metric value is set. The threshold value is determined by statistical analysis of all quantitative abnormality metric values calculated from historical data. The mean and standard deviation are calculated, and then the threshold value is set to the mean plus 1.8 times the standard deviation. For example, if the mean of the historical abnormality metric value is 1.2 and the standard deviation is 0.9, the threshold value is set to 1.2 + 1.8 * 0.9 = 2.82. Subsequently, each relationship (i.e. edge) in the enterprise operation space-time dependence graph is traversed to check whether the timestamp carried by the edge falls within the set time window and whether the associated quantitative abnormality metric value is not less than the calculated threshold value 2.82. Only the edges that meet both conditions are considered as "abnormal and related" edges. For each edge that passes the screening, record the starting node and the ending node, and register these node pairs in the format of (starting node, ending node). Finally, all registered node pairs are summarized to form the target node pair with constraints.

[0088] The acquisition step of the candidate causal relationship path set is:

[0089] According to the target node pair with constraints, start searching from both ends at the same time, establish a forward queue and a backward queue, and only add edges with timestamps not earlier than the previous edge and not later than the upper bound of the time window during expansion. Discard edges with quantitative abnormality metric values less than the threshold value immediately. Keep the path segments that can converge at both ends to obtain the initial path candidate set.

[0090] According to the initial path candidate set, each path segment is spliced into a complete path from the product yield indicator node to the supply chain material batch node in chronological order. The paths that meet the consistent time window coverage and logical constraint conditions are checked, and the candidate causal relationship path set is generated.

[0091] Specifically, according to the target node pair with constraints, for each pair of nodes in it, such as (product yield indicator node A, supply chain material batch node B), a bidirectional breadth-first search algorithm is started, and the specific operation is to initialize two queues, one for forward search starting from node A, called forward queue, and the other for backward search starting from node B, called backward queue, at the same time, two access sets are created to record the nodes that have been accessed by forward and backward search respectively, in each step of forward search expansion, a node is taken out from the forward queue, all its outgoing edges are traversed, for each outgoing edge, it is checked whether its timestamp is not earlier than the timestamp of the edge entering the current node and not later than the upper limit of the preset time window, at the same time, it is checked whether the quantization abnormality metric value of the edge is not less than the threshold 2.82, if the quantization abnormality metric value of the edge is less than the threshold, the path branch is pruned immediately and no longer expanded, the neighbor nodes and their path information that meet all the conditions are added to the forward queue, the backward search is carried out in a similar way, but in the opposite direction, that is, a node is taken out from the backward queue, all its incoming edges are traversed, and the same time and abnormality metric value constraints are applied, in the search process, it is continuously checked whether the expanded node of the forward search exists in the accessed set of the backward search, or vice versa, once an intersection point is found, that is, a node is accessed by both forward and backward search, the forward path segment from A to the intersection point and the backward path segment from B to the intersection point are recorded, the two path segments together constitute a potential complete path, when both queues are empty or the search reaches the preset maximum depth, the search process terminates, and the set of all found forward and backward path segment pairs that can meet at some intermediate node, that is, the initial path candidate set.

[0092] According to the initial path candidate set, each pair of path segments in the set is processed, for example, for a pair of segments consisting of a forward path A -> N1 ->... -> M and a backward path B -> Nk ->... -> M, the backward path segment is reversed to obtain M ->... -> Nk -> B, then the forward path segment is spliced with the reversed backward path segment at the intersection point M to form a complete candidate path from the product yield index node A to the supply chain material batch node B, after splicing, the complete path is checked as a whole, first, check whether the timestamps of all edges on the path constitute a monotonically non-decreasing sequence, and whether the time span of the entire path (i.e. the difference between the timestamps of the first edge and the last edge) is within the preset time window coverage range, for example, no more than 10 days, second, the logical constraint condition is checked, these logical constraints are rules defined in advance based on domain knowledge, for example, the "material taking" operation must occur before the "device feeding" operation, or a "device testing" operation must follow a "device maintenance" operation, traverse the complete path after splicing, check whether the operation sequence on the path violates any of these logical rules, only those complete paths that are consistent in time window coverage range and completely satisfy all preset logical constraints are finally retained, all paths that pass the check are collected to generate a candidate causal relationship path set.

[0093] The acquisition step of the structured causal chain sequence is:

[0094] According to the candidate causal relationship path set, the triples of material batch nodes, device nodes and operator nodes are extracted in time sequence, the start and end positions of the triples in the path and the path identifier are recorded, and a structured causal chain sequence is formed.

[0095] Specifically, according to the candidate causal relationship path set, each complete path in the path set is examined one by one, for any path, the internal nodes and the time stamps inherent to the relationship are scanned in order, in the scanning process, a sliding window with a size of 3 is maintained, the window moves on the path node by node, at each window position, the node types contained in the window are checked, if a material batch node, a device node, and an operator node exist in the window at the same time, a potential triad is determined to be found, then the unique identifiers of the three nodes are combined in the predetermined order of (material batch, device, operator) to form a standardized triad, for example (M:MAT-XF7B34, E:DEV-3041, P:EMP-89721), at the same time, the index position of the first node constituting the triad in the path is recorded as the starting position, the index position of the last node is recorded as the end position, and the unique identifier of the path being processed is also recorded, this complete information including the triad itself, the start and end positions, and the path identifier is taken as a record unit, after the traversal of a single path is completed, all record units extracted therefrom are summarized, this operation is repeated for all paths, and finally all record units extracted from all paths are collected to form a structured causal chain sequence.

[0096] The operation index influence factor obtaining step is:

[0097] According to the structured causal chain sequence, the occurrence frequency is counted, the product yield index value when the sequence appears and the product yield index value when the sequence does not appear are collected, the average path edge number of the sequence appearing path is calculated, and the frequency and product yield statistical parameter set is obtained;

[0098] According to the frequency and product yield statistical parameter set, the operation index influence factor is calculated, and the operation index of the enterprise is judged according to the operation index influence factor, and the calculation formula is:

[0099] ;

[0100] Among them, is the operation index influence factor of the sequence is the average value of the product yield index value when the sequence appears, is the average value of the product yield index value when the sequence does not appear, is the standard deviation of the product yield index value when the sequence appears, is the standard deviation of the product yield index value when the sequence does not appear, is the sample number of the product yield index value when the sequence appears, is the sample number of the product yield index value when the sequence does not appear, is the sample number of the product yield index value when the sequence appears, For sequence The number of times it appears, For containing sequences The average number of path edges, To prevent small positive numbers with a denominator of zero, It is a positive number and is used to control the exponential decay strength of the average path edge number.

[0101] Specifically, based on the structured causal chain sequence, all records are first grouped. The grouping is based on the content of the triples themselves. All records with the same combination of material batch, equipment, and operator node are grouped together. For each unique triple sequence, a series of statistical calculations are performed. The first step is to calculate its occurrence frequency, which equals the total number of records in the corresponding group. The second step is to collect the relevant product yield index values. By using the path identifier in each record, the original candidate causal relationship path set is traced back to find the starting point of the path, i.e., the product yield index node, and its corresponding yield value is read. The yield values ​​corresponding to all paths containing the triple sequence are collected to form a yield data set "when the sequence occurs". Simultaneously, a comparative dataset for "when the sequence does not appear" is constructed. This dataset contains the product yield index values ​​corresponding to all paths in the candidate causal relationship path set that do not contain the triple sequence currently being analyzed. The third step is to calculate the average number of edges of the paths where the sequence appears. Specifically, all paths containing the triple sequence are traversed, the number of edges of each path is counted, and then the average number of edges is calculated. The occurrence frequency calculated for each unique triple sequence, the yield dataset for "when the sequence appears", the yield dataset for "when the sequence does not appear", and the average number of path edges are organized into a structured record. After summarizing the records of all unique sequences, a set of frequency and product yield statistical parameters is obtained.

[0102] In the formula for calculating the impact factor of operational indicators, the first part of the formula is a t-test statistic used to measure the confidence level of the difference between the mean yield of two groups of products when the sequence s appears and does not appear. The denominator is the standard error of the difference between the means of the two samples, taking into account the sample size and variance to ensure the statistical reliability of the difference. The second part amplifies the impact of the frequency of sequence occurrence through a logarithmic function, giving higher weight to frequently occurring event patterns to reflect their universality. The third part introduces an exponential decay penalty for path length. A shorter path means that the impact from materials to the final yield is transmitted more directly, and therefore is given higher weight. The three are multiplied together to obtain a comprehensive impact factor score that takes into account the reliability, universality and directness of the impact.

[0103] , and Each represents a specific sequence The mean, standard deviation and sample size of the associated product yield indicator values at the time of occurrence, which are calculated from the "at the time of occurrence" yield data set in the frequency and product yield statistical parameter set obtained in the previous step, specifically, is the total number of yield values in the data set, is the mean of the yield values, is the sample standard deviation of the yield values, for example, for sequence s1, its associated "at the time of occurrence" yield data set is {92.1, 93.5, 91.8, 92.6, 92.0}, then the sample size , the mean , and the sample standard deviation are approximately 0.682 according to the standard deviation calculation formula;

[0104] , and represent the mean, standard deviation and sample size of the product yield indicator values at the time of non-occurrence of a specific sequence The calculation method of these three parameters is similar to , , , but the data source is the "at the time of non-occurrence" yield data set in the frequency and product yield statistical parameter set, which collects the yield values corresponding to all paths that do not contain the sequence , for example, for sequence s1, for example, there are 1000 product yield data points in the entire analysis period, of which 5 are related to the occurrence of s1, then the remaining 995 data points constitute the "at the time of non-occurrence" data set, from which the sample size can be calculated, and the mean and standard deviation of the 995 yield values can be further calculated, for example, , ;

[0105] is the number of occurrences of sequence , which is directly obtained from the frequency and product yield statistical parameter set, it is the count value obtained when counting the occurrence frequency of each unique sequence, reflecting the frequency of the operation sequence in the observation time window, for example, when analyzing s1, it is found that it has occurred 25 times in all paths, then ;

[0106] is the number of occurrences of sequence the average number of edges of the paths of the sequence s1, which is also obtained from the frequency and product yield statistical parameters set, and it reflects the average length or complexity of the causal conduction paths from the supply chain material batch nodes to the product yield indicator nodes, the longer the path, the more intermediate links, the lower the directness of the impact, for example, the number of edges of the 25 paths containing sequence s1 is {5, 6, 5, 7, 6,...} respectively, and the average value of these edge numbers is calculated as ;

[0107] is a small positive number set to prevent the denominator from being zero, and is set to ;

[0108] is a positive number used to control the strength of the path length penalty, and its value determines the speed at which the operating indicator influence factor decays with the increase of the path length, The setting of can be based on prior knowledge, that is, to define the expected path length when the influence factor decays to a certain proportion, for example, set when the path length reaches 6, the value of its penalty term decays to 30% of the initial value (when the path length is 1), then , the solution is Therefore This value will remain the same in the calculation of all sequences.

[0109] According to the parameters, the calculation is as follows:

[0110] For a specific sequence , the following parameters are obtained according to the foregoing steps:

[0111] , , ;

[0112] , , ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] The operating indicator influence factor is calculated :

[0118] ;

[0119] ;

[0120] The results show that the sequence of the operating index impact factor is 13.52, reflecting the statistical significance of the negative impact of the sequence on the product yield (the average from 96.5 to 92.4), the frequency and the directness of the impact path. By calculating the impact factor of all different sequences, a ranking list can be obtained, in which the sequence with higher impact factor value is considered to be the key impact factor leading to the fluctuation of product yield. Enterprise operation managers can prioritize the top-ranked sequences for in-depth analysis and intervention based on this ranking, for example, if the impact factor greater than 10.0 is set as a high-risk factor, then the sequence will be marked as an enterprise operation index impact factor that needs to be focused on.

Claims

1. A data mining-based enterprise operation index prediction system, characterized by, The system comprises: A heterogeneous data graph construction module is configured to identify equipment, materials, personnel, and orders as entity nodes and define interactive operations as relationships based on equipment maintenance records and operator shift logs, connect all entity nodes, and establish an enterprise operation space-time dependency graph; A distribution summary quantization storage module is configured to write graph data into data blocks and embed a summary structure of quantile and frequency statistics in the metadata area of each data block based on the enterprise operation space-time dependency graph, generate embedded distribution summary data block metadata, access the embedded distribution summary data block metadata, and calculate a quantitative anomaly metric value; A space-time dependency path retrieval module is configured to set product yield indicator nodes and supply chain material batch nodes as start and end points of a path, establish a constrained target node pair, simultaneously start searching from both ends of the constrained target node pair, eliminate paths, and generate a candidate causal relationship path set based on the enterprise operation space-time dependency graph and the quantitative anomaly metric value; A key influence factor extraction module is configured to extract node combinations of material batches, equipment, and operators included in each path in the path set based on the candidate causal relationship path set, form a structured causal chain sequence, calculate the influence weight of each sequence based on the structured causal chain sequence, and obtain an operation indicator influence factor. The quantitative anomaly metric value is obtained by: Determining the histogram interval and neighborhood interval set in which the target data point is located based on the embedded distribution summary data block metadata, reading the count and width of each interval, calculating the normalized density of each interval and saving it as a density sequence, and obtaining a density parameter set; Calculating the quantitative anomaly metric value based on the density parameter set.

2. The data mining based business operation indicator forecasting system according to claim 1, wherein, The enterprise operation space-time dependency graph is obtained by: Analyzing the equipment number, material batch, personnel identifier, and order number fields based on equipment maintenance records and operator shift logs, unifying the coding format, adding labels according to types, removing duplicates according to primary keys, and forming an entity node set; Extracting interactive operation items from equipment maintenance records and operator shift logs based on the entity node set, matching corresponding entity nodes according to the initiator and the pointing party, reading the operation time field and the work site field, and standardizing the time and coordinate representation to generate a relationship set with timestamp and spatial location attributes; Connecting entity nodes according to relationship directions based on the relationship set with timestamp and spatial location attributes, sorting by timestamp to establish a time sequence reference chain, checking missing pointers and eliminating irrelevant nodes, and generating an enterprise operation space-time dependency graph.

3. The data mining based business operation indicator forecasting system according to claim 1, wherein, The embedded distribution summary data block metadata is obtained by: Writing graph data into data blocks according to node time order based on the enterprise operation space-time dependency graph, counting the value range and frequency of each interval, calculating the interval width, and embedding the width, count, and quantile point information into the metadata area to form embedded distribution summary data block metadata.

4. The data mining based business operation indicator forecasting system according to claim 1, wherein, The constrained target node pair is obtained by: According to the enterprise operation space-time dependence graph and the quantitative abnormality metric value, a product yield indicator node and a supply chain material batch node are designated as start and end points, a time window and a threshold value of the quantitative abnormality metric value are set, edges with a relationship timestamp falling within the time window and a quantitative abnormality metric value not less than the threshold value are scanned, and the start and end points are registered in pairs to form a target node pair with constraints.

5. The data mining based business operation indicator forecasting system according to claim 1, wherein, The obtaining step of the candidate causal relationship path set is: According to the target node pair with constraints, a forward queue and a backward queue are established by starting searching from both ends, edges with a timestamp not earlier than a previous edge and not later than an upper limit of the time window are added during expansion, edges with a quantitative abnormality metric value less than the threshold value are discarded immediately, path segments that can converge at both ends are retained, and an initial path candidate set is obtained. According to the initial path candidate set, each path segment is spliced into a complete path from the product yield indicator node to the supply chain material batch node in chronological order, paths with consistent time window coverage and satisfied logical constraint conditions are checked, and a candidate causal relationship path set is generated.

6. The data mining based business operation indicator forecasting system according to claim 1, wherein, The obtaining step of the structured causal chain sequence is: According to the candidate causal relationship path set, triplicate paths are traversed to extract a triple of a material batch node, a device node and an operator node in chronological order, record the start and end positions of the triple in the path and the path identifier, and form a structured causal chain sequence.

7. The data mining based business operation indicator forecasting system according to claim 1, wherein, The obtaining step of the operation indicator influence factor is: According to the structured causal chain sequence, the number of occurrences is counted, the product yield indicator value when the sequence occurs and the product yield indicator value when the sequence does not occur are collected, the average number of edges of the sequence occurrence path is calculated and recorded as the average path edge number, and a frequency and product yield statistical parameter set is obtained; According to the frequency and product yield statistical parameter set, an operation indicator influence factor is calculated, and an enterprise operation indicator is determined according to the operation indicator influence factor.

Images