Production line knowledge graph generation method and server based on industrial internet

Abstract

This invention provides a method and server for generating a production line knowledge graph based on the Industrial Internet. It obtains continuous time series data from multiple measurement points from the same data source on the production line. For each measurement point sequence, it performs adaptive segmentation and symbolic aggregation based on trend inflection points, extracting the positions of trend direction changes, segmenting them, and mapping them to discrete symbol sequences. For any two discrete symbol sequences of measurement points, it calculates multi-delay propagation entropy to obtain directed causal strength, filters measurement point pairs with significant causal relationships and clear directions, and establishes a set of measurement point-level causal relationship edges carrying the measurement point-level effect delay. Then, it uses a pre-constructed static structural constraint model to perform equipment-level semantic reduction, aggregating the measurement point-level causal edges of the same equipment pair into inter-equipment semantic relationship edges, and assigning relationship type labels according to the mechanical connection type to generate a production line knowledge graph. This invention achieves automatic extraction of causal semantic relationships between equipment from massive time-series data, making the knowledge graph both data-driven and physically interpretable.

Description

Technical Field

[0001] This invention relates to the field of industrial knowledge graphs, and more specifically, to a method and server for generating production line knowledge graphs based on the Industrial Internet. Background Technology

[0002] Production line knowledge graphs can describe entities such as equipment and measuring points in industrial production lines and their interactions, which is of great value for condition monitoring and fault tracing. Current technologies typically rely primarily on the physical topology and mechanical connections between production line equipment, combined with human experience to extract features from measuring point data, thereby determining node attributes and edge relationships. However, relying solely on physical topology and human experience makes it difficult to automatically uncover the potential causal logic and information transmission directions between measuring points from massive continuous time-series data. This results in the relationship representation in the graph being limited to structural connections and lacking dynamic causal semantics, thus restricting the knowledge graph's ability to provide a deep description of the production line's operational mechanisms. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide at least one method and server for generating production line knowledge graphs based on the Industrial Internet.

[0004] According to one aspect of the present invention, a method for generating a production line knowledge graph based on the Industrial Internet is provided, comprising: Obtain a continuous time series data set consisting of multiple measurement point variables from the same data source on the production line; For each measurement point variable, an adaptive segmented symbolic aggregation based on trend inflection points is performed on the continuous time series data set. By extracting key locations in the time series where the trend direction changes, adaptive segmentation is performed, and each segment is mapped to a preset symbol set to obtain the corresponding discrete symbol sequence. For any two discrete symbol sequences of measurement point variables, perform multi-delay transfer entropy calculation to obtain the directed causal intensity value that characterizes the direction and driving intensity of information flow between the two measurement point variables. Traverse all pairs of measurement point variables and summarize all directed causal intensity values ​​to obtain the directed causal intensity set. Extract measurement point variable pairs from the directed causal intensity set whose intensity values ​​satisfy the preset causal significance condition and whose direction is clear from cause to effect. For each extracted measurement point variable pair, instantiate measurement point nodes and device nodes of the devices to which the measurement point variables belong. Establish causal relationship edges from the cause measurement point node to the effect measurement point node. Based on the transmission entropy time delay analysis, assign measurement point-level action delay attributes to each causal relationship edge to form a set of causal relationship edges at the measurement point node level. A static structural constraint model is pre-constructed based on the fixed mechanical transmission topology and physical adjacency relationships between production line equipment. The causal relationship edge set at the measurement point node level is semantically reduced at the equipment level using the static structural constraint model. Causal relationship edges at the measurement point node level belonging to the same equipment pair are aggregated and deduced into semantic relationship edges between equipment nodes. Relationship type labels are assigned to each semantic relationship edge according to the mechanical connection type of the equipment at both ends of the edge. A production line knowledge graph is generated based on the equipment nodes and the set of semantic relationship edges with assigned relationship type labels.

[0005] According to another aspect of the present invention, a server is provided, comprising: a processor; and a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method described above.

[0006] This invention acquires continuous time series of multiple measurement point variables from the same data source on the production line and performs adaptive segmented symbolic aggregation based on trend inflection points. It automatically identifies key locations of trend changes and maps the sequences to discrete symbols, effectively compressing data size and reducing noise interference, providing structured input for causal analysis. For the discrete symbol sequences of any two measurement point variables, multi-time-delay propagation entropy calculation is performed to obtain directional causal strength values ​​that quantify the direction and driving intensity of information flow. This overcomes the limitation of traditional correlation analysis in determining directionality, significantly improving the accuracy and reliability of identifying causal relationships between measurement points. After extracting measurement point pairs with significant causal relationships and clear directions, causal relationship edges with measurement point-level effect delay attributes are constructed. This achieves both structured storage of causal relationships and preservation of dynamic transmission response information, enabling the graph to characterize the causal transmission time sequence. By introducing a static structural constraint model pre-constructed based on the mechanical transmission topology and physical adjacency relationships between production line equipment, semantic reduction of causal edges at the measurement point layer is performed at the equipment level. Edges belonging to the same equipment pair are aggregated into semantic relationship edges between equipment nodes, and relationship labels are assigned according to the equipment connection type. This achieves a deep integration of data-driven causal discovery and physical prior knowledge, ensuring strict consistency between the knowledge graph and the physical configuration, while also endowing equipment relationships with causal explanatory power. The resulting production line knowledge graph can simultaneously express the structural associations between equipment and the dynamic causal transmission logic, providing a high-quality graph structure data foundation for process optimization and fault tracing. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram of an application scenario provided by the present invention; Figure 2This is a schematic diagram illustrating the principle of the production line knowledge graph generation method based on the Industrial Internet provided by this invention. Figure 3 This is a flowchart illustrating a method for generating a production line knowledge graph based on the Industrial Internet, provided by this invention. Figure 4 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] To facilitate a clearer understanding of this invention, we will first introduce the application scenarios of the industrial internet-based production line knowledge graph generation method, such as... Figure 1 As shown, the application scenario of this invention includes a server 10 and one or more measurement points; the number of measurement points is not limited here. Figure 1 As shown, the measurement points are connected to the server 10 via a network so that each measurement point can interact with the server 10 via the network.

[0011] It is understood that server 10 can refer to a device that executes the industrial internet-based production line knowledge graph generation method provided in the embodiments of the present invention. Server 10 can be, for example, a single physical server, or a server cluster or distributed system consisting of at least two physical servers. Measurement points can be different sensors, such as displacement sensors, pressure sensors, etc.

[0012] Furthermore, please combine Figure 2 and Figure 3 This is a schematic diagram illustrating the principle and process of a production line knowledge graph generation method based on the Industrial Internet, provided in an embodiment of the present invention. This method can be... Figure 1 The method for generating a production line knowledge graph based on the Industrial Internet is executed on server 10, and may include the following steps: Step S100: Obtain a continuous time series data set consisting of multiple measurement point variables from the same data source on the production line.

[0013] In this invention, the production line can be a complete production system in a manufacturing environment, composed of a series of devices, tooling, and material handling equipment with clear physical connections or process linkages, capable of completing specific product processing or assembly tasks. The common data source is a single data infrastructure that centrally aggregates, uniformly encapsulates, and provides access interfaces for all relevant sensor signals and equipment status data on the production line. Its physical form can be a real-time historical database server deployed at the factory level, or a time-series data storage engine integrated on edge computing nodes. This data source maintains a globally unified and strictly increasing time base, and all accessed data is timestamped according to this time base. Measurement point variables are specific physical quantities or state quantities measured and reported by individual sensors installed in the production line equipment, process media, or environmental space. Each measurement point variable is marked by a globally unique identifier, which typically uses a hierarchical naming convention to concatenate the workshop number, production line number, equipment number, and the physical meaning of the measurement point. The continuous time series data set is a two-dimensional numerical matrix formed by sampling multiple measurement point variables at equal intervals within the same time interval according to a unified sampling period. Each row of the matrix corresponds to a sampling time, and each column corresponds to a floating-point record of a measurement point variable at that time after analog-to-digital conversion and engineering value calibration. The time steps between each row are completely consistent, and there are no missing intervals due to communication interruption or sensor failure.

[0014] For example, when monitoring an automated production line consisting of a high-speed stamping machine, a feeding robot, and a finished product conveyor belt in real time, structured time-series data blocks that have been protocol-adapted and time-aligned are read from an edge intelligent data gateway deployed in the control cabinet. This edge intelligent data gateway communicates with the programmable logic controllers of each device via the industrial Ethernet protocol, synchronously collecting five measurement point variables at a fixed sampling period: the value of the clamping force sensor, the reading of the slider displacement encoder, the hydraulic system oil supply pressure, the vibration amplitude of the robot's end effector, and the temperature of the thermocouple on the mold surface. The clock synchronization module inside the edge intelligent data gateway obtains the standard time from a network time protocol server and uses this standard time as the timestamp reference for all sampled data. After each sampling period, the gateway combines the current value of the five measurement point variables with the corresponding timestamp to form a record, which is then appended to a memory-mapped time-series data circular buffer. All records constitute a continuous time-series data set containing five data streams, covering the entire history from the start of the current batch production to the current moment. The time span of all data columns is completely consistent, and the time interval between adjacent records is always equal to the preset sampling period.

[0015] Step S200: Perform adaptive segmented symbolic aggregation based on trend turning points on the continuous time series data set of each measurement point variable. By extracting key positions in the time series where the trend direction changes, adaptive segmentation is performed, and each segment is mapped to a preset symbol set to obtain the corresponding discrete symbol sequence.

[0016] Trend inflection points are local morphological abrupt changes on a continuous time series curve. At these points, the sign of the local first derivative of the sequence changes from positive to negative or vice versa, signifying a shift from a continuous upward trend to a continuous downward trend, or vice versa. Adaptive segmentation based on trend inflection points does not rely on a preset fixed segment length. Instead, it uses the detected trend inflection points as dividing boundaries to segment the original continuous time series into several segments. Each segment maintains a single trend direction, and the length of segments for different measurement variables or even the same measurement variable varies depending on the rhythm of data change, exhibiting adaptive characteristics. Symbolic aggregation is the process of summarizing the numerical distribution characteristics of each adaptive segment into a discrete symbol, thereby transforming the entire numerical sequence into a string composed of a finite number of symbols. The preset symbol set is an alphabet composed of a finite number of distinct characters. Each character represents a prototype pattern of the numerical distribution of the sequence segments, such as stable and in a high-value region, stable and in a low-value region, slowly rising with large fluctuations, or rapidly falling with small fluctuations. These prototype patterns are defined by statistical feature centers obtained beforehand through clustering a large number of sequence segments.

[0017] In one implementation, step S200 specifically includes the following steps S210 to S260: Step S210: Perform sequential adjacent point slope trend analysis on the continuous time series data set of the measurement point variable, generate local direction vector markers between each adjacent point, and detect the data sampling positions where the local direction vector markers are flipped. Record the data sampling positions as the initial selection inflection points to form the initial selection inflection point sequence.

[0018] Sequential adjacent point slope trend analysis is a process that proceeds point-by-point along the time axis, determining the direction of numerical change for each pair of immediately preceding and following sampling points. During execution, starting from the second sampling point in the sequence, the value of the current sampling point is subtracted from the value of the immediately preceding sampling point to obtain the first-order difference of that adjacent point pair. The local direction vector label is a ternary direction indicator assigned to each pair of adjacent points based on the sign of the first-order difference. A positive label is assigned when the first-order difference is greater than a preset noise tolerance threshold, and a negative label is assigned when the first-order difference is less than a negative noise tolerance threshold. When the absolute value of the first-order difference does not exceed the noise tolerance threshold, to avoid misjudgments triggered by minor fluctuations, the local direction vector label of that adjacent point pair directly adopts the direction label of the previous pair of adjacent points. The noise tolerance threshold is set based on a comprehensive consideration of the sensor's measurement noise level and the natural fluctuation amplitude of the process; its function is to filter out false direction changes caused by quantization errors or minor disturbances. Positive-negative reversal occurs in a local direction vector label sequence when a positive label immediately adjacent to another positive label changes to a negative label, or vice versa. Consecutive identical direction labels or a direction label that remains constant with a corresponding label do not constitute a reversal. The data sampling position is the index number or timestamp corresponding to the sampling point in the original continuous time series where a positive-negative reversal occurs. Initial inflection points are data sampling points identified by the positive-negative reversal detection mechanism as locations of trend direction transitions. The initial inflection point sequence is a list obtained by arranging all initial inflection points in their original chronological order.

[0019] When performing this step on the time series of clamping force sensor values, firstly, a one-dimensional floating-point array of clamping force values ​​is extracted and treated as a sequence arranged in sampling order. Starting from the second element, the difference between the current element and the previous element is calculated one by one. Each difference is compared with a preset noise tolerance threshold: if the difference is greater than the positive tolerance, a positive label is assigned to the pair of adjacent points; if the difference is less than the negative tolerance, a negative label is assigned to the pair of adjacent points; if the absolute value of the difference is not greater than the positive tolerance, the direction label of the previous pair of adjacent points is carried over as the direction label of the current pair. After traversing the entire array, a direction label sequence with a length one label less than the original sequence length is obtained. Next, positive-to-negative reversals are detected in this direction marker sequence: when the k-th marker is found to be positive and the k-plus-one marker is found to be negative, a reversal from positive to negative is confirmed, and the data sampling position of the k-plus-one sampling point in the original sequence is determined as the initial selected inflection point; when the k-th marker is found to be negative and the k-plus-one marker is found to be positive, a reversal from negative to positive is confirmed, and the data sampling position is also determined as the initial selected inflection point. All data sampling positions that meet the reversal conditions are recorded in chronological order to form the initial selected inflection point sequence.

[0020] Step S220: For each initial inflection point in the initial inflection point sequence, define a neighborhood box with the point as the center and a fixed time span. Extract the directional attributes of the initial inflection points contained in each neighborhood box and analyze the concentrated performance of the directional attributes within the box. Retain the center point of the neighborhood box whose directional attribute concentrated performance meets the preset stable aggregation conditions, remove the remaining scattered points, and generate an aggregation strengthening inflection point sequence.

[0021] A neighborhood box is a local time interval enclosed on the timeline with a predetermined initial inflection point as its center of symmetry, extending forward and backward by equal fixed time spans. The fixed time span is a pre-defined constant duration parameter, which specifies a uniform value for the total length of the neighborhood box across the entire sequence of initial inflection points to ensure consistent scale in the center selection. Directional attributes are binary classification labels for the type of trend direction change represented by the initial inflection point, derived from the changes in local directional vector labels before and after the inflection point: if the local directional vector label before the inflection point is positive and the label after the inflection point is negative, then the directional attribute of that point is peak-shaped; if the label before the inflection point is negative and the label after the inflection point is positive, then the directional attribute of that point is valley-shaped. The concentrated representation of directional attributes within the box is the distribution pattern of the directional attributes of all initial inflection points falling within the local time range covered by the current neighborhood box. This is mainly characterized by the majority of directional attribute categories and their proportion within the box, aiming to determine whether the directional characteristics of the inflection points within the box exhibit statistical convergence. The preset stable clustering condition is a set of multi-dimensional judgment rules used to evaluate whether the concentration of directional attributes has reached a sufficient degree of clustering. This rule also examines whether the absolute number of initially selected inflection points within the bounding box meets the minimum density requirement, and whether the proportion of the majority directional attributes exceeds a preset proportion threshold. The clustering strengthening inflection point sequence is an ordered set of the center points of the neighboring boxes that satisfy the stable clustering condition after filtering by neighboring boxes, arranged in chronological order.

[0022] Based on the obtained sequence of initial selection inflection points for mold clamping force, a neighborhood box is defined for each initial selection inflection point in the sequence, centered on the timestamp of that point and extending before and after it for a fixed duration. For each neighborhood box, all initial selection inflection points contained within the box are scanned, and the directional attribute of each initial selection inflection point is read: if the point corresponds to a reversal from positive to negative, its directional attribute is recorded as a peak; if it corresponds to a reversal from negative to positive, it is recorded as a valley. The number of occurrences of peak directional attributes and valley directional attributes within the box are counted, and the sum of the number of peaks and valleys is calculated as the total number of points within the box. The ratio of the number of peaks to the total number of points or the ratio of the number of valleys to the total number of points is then calculated. The total number of points within the box is compared with the preset minimum density requirement, and the proportion of the majority directional attributes is compared with the preset proportion threshold. If the total number of points within a bounding box exceeds the minimum density requirement, and the proportion of the majority of directional attributes exceeds a threshold, then the directional attribute concentration of the neighboring bounding box is determined to meet the preset stable clustering conditions, and the center point of the box is retained. If the total number of points within the bounding box is insufficient, or the directional attribute distribution is too even, causing either condition to be unmet, then the center point of the box is determined to be a scattered point and is removed. After performing this evaluation and screening operation on the neighboring bounding boxes corresponding to all initially selected inflection points, all retained center points are assigned to the clustering strengthening inflection point sequence in chronological order.

[0023] In one implementation, step S220 specifically includes the following steps S221 to S226: Step S221: For the initial selection of turning point sequence, establish a sliding processing window in the time dimension and set the window length. Move the sliding processing window along the time axis in one step. After each move, extract the initial selection of turning point subset contained in the window and the directional attribute distribution of each subset member.

[0024] The sliding processing window in the time dimension is a fixed-length truncation interval that slides sequentially along the time axis. Its length is measured in time units and remains the same as the fixed time span of the aforementioned neighborhood box to ensure logical consistency in subsequent processing. In its implementation, the sliding processing window is jointly defined by the timestamps of its left and right boundaries, and the duration between the left and right boundaries is always equal to the set window length. The sliding processing window moves along the time axis in single steps, shifting both the left and right boundaries backward by one step unit each time. This step unit can be set to a granularity equal to the sampling period to perform a dense, comprehensive scan of the initial selection of inflection point sequences. After each window movement, all initial selection inflection points within the current window's time range are extracted to form a subset of initial selection inflection points. The peak or valley markers of each member in the subset are then read to form the directional attribute distribution at that window position.

[0025] In this embodiment, a sliding processing window is established for the initial selection inflection point sequence of clamping force. The window length is set to the same duration as the fixed time span of the neighborhood box. The left boundary of the sliding processing window is initially aligned to the earliest timestamp of the initial selection inflection point sequence. All initial selection inflection points contained within the window span are extracted to form a subset of initial selection inflection points. Each initial selection inflection point in this subset is traversed and labeled as a peak-shaped directional attribute or a valley-shaped directional attribute according to its corresponding flip type. The number of individuals in each of the two directional attribute categories is counted to obtain the directional attribute distribution at the window position. Then, the sliding processing window is slid backward along the time axis by one step, and a new subset of initial selection inflection points and directional attribute distribution are extracted again. This sliding is gradually performed until the right boundary of the window exceeds the timestamp of the latest initial selection inflection point. The entire scanning process generates a series of snapshots of the subset of initial selection inflection points and their directional attribute distributions corresponding to the window position.

[0026] Step S222: For each window position, the distribution of directional attributes within the initially selected inflection point subset is statistically analyzed, and the consistency representation of directional attributes is calculated. The consistency representation reflects the degree of aggregation of the directional directions of the inflection points within the window.

[0027] The dispersion of directional attributes refers to the uniformity or concentration of directional attribute values ​​at each initial selection turning point within the current window across the entire window. This can be characterized by comparing the relative number of individuals in the majority directional attribute category within the window with the total number of turning points within the window. The consistency measure is a scalar index used to quantify the convergence of directional attributes among all initial selection turning points within the window. It is calculated by counting the number of individuals in the dominant directional attribute category within the window, and then normalizing this count by combining twice the count with the sum of the total number of initial selection turning points and the total number of directional attribute categories within the window. This ensures that the consistency measure approaches its upper limit when directional attributes are completely identical and approaches its lower limit when directional attributes are completely dispersed.

[0028] In implementation, for each subset of initial inflection points obtained at each window position, the number of individuals with peak-shaped directional attributes and the number of individuals with valley-shaped directional attributes in that subset are first calculated, and the larger of the two is recorded as the maximum category count. Simultaneously, the total number of initial inflection points in the subset is counted. Then, a modified consistency measure is used to construct the consistency representation: twice the maximum category count is divided by the sum of the total number of initial inflection points and the total number of directional attribute categories; the quotient is the consistency representation of the directional attributes for that window. If the directional attributes of all initial inflection points within the window are either all peak-shaped or all valley-shaped, the consistency representation equals the theoretical maximum value, indicating extremely strong aggregation; if the number of peak-shaped and valley-shaped individuals is roughly equal, the consistency representation is at a moderate level, indicating significant divergence in directional orientation.

[0029] Step S223: Compare the consistency representation quantity with the preset consistency threshold, mark the window position where the consistency representation quantity exceeds the preset consistency threshold as the effective aggregation window, and determine the initial inflection point corresponding to the center time of the effective aggregation window as the candidate aggregation center point.

[0030] The preset consistency threshold is a pre-defined judgment threshold used to distinguish between windows with significant clustering of directional attribute distributions and windows with no obvious tendency in directional attribute distributions. The effective clustering window is a sliding processing window whose consistency representation exceeds the preset consistency threshold. Within the local time range covered by these windows, the initially selected inflection points exhibit a high degree of consistency in directional attributes. The center time of the effective clustering window is calculated by adding half the window length to the time of the window's left boundary. The candidate clustering center point is the initially selected inflection point whose timestamp is closest to the center time of the effective clustering window. If this center time falls exactly between two initially selected inflection points, the one with the closer distance is selected.

[0031] Following the previous example, after calculating the consistency representation value at each sliding window position window by window, the value of each consistency representation value is compared with a preset consistency threshold. For windows with a consistency representation value greater than the preset consistency threshold, they are marked as valid clustering windows, and the center time of the window is calculated. Then, the initial selection inflection point with the timestamp closest to the center time is retrieved from the initial selection inflection point sequence of clamping force and extracted as a candidate clustering center point. For windows with a consistency representation value not exceeding the preset consistency threshold, the extraction of candidate clustering center points is not performed.

[0032] Step S224: Perform overlap analysis on adjacent effective clustering windows on the time axis. If the coverage areas of adjacent windows overlap and the time interval between their respective candidate clustering centers is less than the preset deduplication interval, then only the candidate clustering centers with larger consistency characterization values ​​are retained, and the remaining candidate clustering centers are filtered out to obtain a set of clustering centers without redundancy.

[0033] Overlap analysis determines the overlap of time periods between the coverage areas of two adjacent valid clustering windows on the time axis. If the right boundary time of the preceding window is later than the left boundary time of the following window, the two windows are considered to overlap in their time coverage areas. The preset deduplication interval is a time length threshold used to define whether the candidate clustering centers of two adjacent valid clustering windows are too close on the time axis to be considered duplicate representations of the same physical turning point event. If adjacent windows overlap and the time interval between the two candidate clustering centers is less than the preset deduplication interval, the redundancy elimination rule is invoked to select one of them to retain. The selection is based on the higher consistency representation value, which is considered to have a stronger representativeness of the turning point event.

[0034] In this embodiment, all valid clustering windows are arranged in chronological order of their center times. Starting with the first valid clustering window, overlap analysis is performed on each of its immediately following valid clustering windows. If the time coverage intervals of two windows overlap, and the time difference between the candidate clustering centers corresponding to the two windows is less than the preset deduplication interval, the consistency representation of the windows to which these two candidate clustering centers belong is compared. The candidate clustering center with the larger consistency representation is retained, and the other candidate clustering center is deleted from the result set. If the time coverage intervals of two windows do not overlap, or if they overlap but the time difference between the two candidate clustering centers is not less than the preset deduplication interval, both candidate clustering centers are retained. After this deduplication process, all remaining candidate clustering centers are assigned to a set of non-redundant clustering centers.

[0035] Step S225: Using each cluster center point in the set of non-redundant cluster center points as a reference, back-reference the direction attributes of all initially selected turning points within its corresponding valid cluster window, and reposition the precise representative time coordinates of the cluster center point based on the time boundary of the initially selected turning points with consistent direction attributes, to obtain the time-corrected cluster center point.

[0036] Backlinking uses the current cluster center point as an index to reverse-locate to the effective cluster window that determined its generation, and re-extracts information from all initially selected inflection points within that window. Initially selected inflection points with consistent direction attributes are those within the effective cluster window whose direction attributes are the same as the majority of those in the window. The time boundary is the time range defined by the earliest and latest timestamps among these initially selected inflection points with consistent direction attributes. The precise time coordinates are obtained by finding the midpoint between the earliest and latest timestamps, i.e., taking the average of the earliest and latest time boundaries.

[0037] After obtaining the set of non-redundant cluster centers, for each cluster center, we first backtrack to the valid cluster window that generated it, enumerate all the initial inflection points contained within the window, and identify the majority directional attribute category within the window, denoted as the dominant directional attribute. Next, we filter out all initial inflection points in the window whose directional attribute is exactly the same as the dominant directional attribute, and determine the earliest and latest timestamps among these initial inflection points. The arithmetic mean of the earliest and latest timestamps is used as the precise representative time coordinate of this cluster center. The original candidate cluster center timestamps are replaced with this precise representative time coordinate, completing the time refinement of this cluster center. This process is repeated for all non-redundant cluster centers, resulting in a batch of cluster centers with refined time coordinates.

[0038] Step S226: Arrange all the time-corrected cluster centers in chronological order and collect them into a cluster strengthening inflection point sequence.

[0039] All the time-corrected cluster center points generated in the previous step are arranged in ascending order according to their precise representative time coordinates, forming an ordered set of points that strictly increases on the time axis. This set of points is the cluster strengthening inflection point sequence.

[0040] Step S230: Calculate the time interval between adjacent inflection points in the clustered strengthening inflection point sequence, compare the time interval with the preset shortest duration constraint of the trend segment, and for short-term trend segments with time intervals lower than the shortest duration constraint, merge and eliminate the inflection points at both ends, and reorganize the eliminated inflection points into a key inflection point sequence.

[0041] The time interval between adjacent inflection points is the absolute value of the difference between the time coordinates of two inflection points that are immediately adjacent in time order within the clustering and strengthening inflection point sequence. The preset shortest duration constraint for a trend segment is a preset minimum time length threshold used to determine whether the trend segment sandwiched between two adjacent inflection points has sufficient physical meaning. If the time interval between two adjacent inflection points is less than this constraint value, the trend segment is considered too short and may be a pseudo-trend caused by noise or instantaneous process disturbances, and should be eliminated. Merging and elimination involves merging the two inflection points that define the short-lived trend segment. Specifically, the inflection point with the lower consistency value during the clustering and strengthening screening process is discarded, while the one with the higher consistency value is retained. Alternatively, when the consistency values ​​of the two inflection points are close, the inflection point with the time coordinate closer to the middle of the sequence is retained. The key inflection point sequence is a sequence reconstructed from the remaining inflection points after the short-lived trend segment merging and elimination. Each inflection point in this sequence corresponds to a trend segment boundary with sufficient duration.

[0042] In the application scenario of clamping force as a measurement variable, for the previously generated clustering reinforcement inflection point sequence, the time difference of each pair of adjacent inflection points is calculated sequentially from beginning to end. Each time difference is compared with the preset shortest duration constraint of the trend segment. When it is found that the time interval between two adjacent inflection points is lower than the preset shortest duration constraint of the trend segment, the left and right inflection points at both ends of the short-term trend segment are located, and the consistency representation values ​​recorded by these two inflection points in the effective clustering window screening stage are retrieved. The magnitude of the two consistency representation values ​​is compared, and the inflection point with the smaller consistency representation value is discarded, while the inflection point with the larger consistency representation value is retained. If the two consistency representation values ​​are equal, the time distance between their precise representative time coordinates and the points in the sequence is further compared, and the closer one is retained. After completing the merging and elimination of all short-term trend segments, all the remaining inflection points are arranged in chronological order to obtain the key inflection point sequence.

[0043] Step S240: Using each key inflection point in the key inflection point sequence as a dividing boundary, the continuous time series data set is truncated with boundary alignment to obtain multiple non-overlapping subsequence segments arranged in chronological order. The central tendency measure and dispersion measure of the internal numerical distribution of each subsequence segment are extracted to obtain the statistical description pair of the subsequence segment.

[0044] The segmentation boundary is the precise time point used to physically divide a time series. When a key inflection point is used as the segmentation boundary, the time interval between two adjacent key inflection points constitutes an independent subsequence segment. Boundary-aligned interval truncation ensures that the right endpoint of the preceding subsequence segment and the left endpoint of the following subsequence segment strictly coincide on the time axis without gaps or temporal overlap, thus ensuring that all subsequence segments completely cover the time range of the original sequence. The propensity to centralize measure is a statistical measure that summarizes the overall value level of all sampled points within a subsequence segment, such as the arithmetic mean or median, to characterize the central location of the numerical distribution within that segment. The dispersion measure is a statistical measure that measures the deviation of all sampled points within a subsequence segment from the propensity to centralize measure, such as the standard deviation or interquartile range, to characterize the spread of the numerical distribution within that segment. A statistical descriptive pair is a binary pair consisting of the propensity to centralize measure and the dispersion measure, comprehensively describing the numerical distribution characteristics of the corresponding subsequence segment.

[0045] In this embodiment, the time coordinates of each critical inflection point in the clamping force critical inflection point sequence are extracted sequentially. A first sub-sequence segment is extracted with the first critical inflection point as the left boundary and the second critical inflection point as the right boundary; a second sub-sequence segment is extracted with the second critical inflection point as the left boundary and the third critical inflection point as the right boundary, and so on, until the entire time range is covered. The resulting sub-sequence segments are connected end-to-end on the time axis and do not overlap. For each extracted sub-sequence segment, the arithmetic mean of all sampled values ​​within that sub-sequence segment is extracted as a measure of central tendency, and the standard deviation of all sampled values ​​within that sub-sequence segment is extracted as a measure of dispersion. These two statistics are paired to obtain a statistical descriptive pair for that sub-sequence segment.

[0046] Step S250: Spatial position comparison of the statistical description pair with the symbol reference representation of each prototype symbol in the preset symbol dictionary, calculation of the offset distance of the statistical description pair relative to each symbol reference representation, selection of the symbol character corresponding to the prototype symbol with the smallest offset distance as the substitute symbol for the subsequence segment, and concatenation of the substitute symbols of all subsequence segments in chronological order to generate a discrete symbol sequence of the measurement point variable.

[0047] The pre-constructed symbol dictionary is a pre-built symbol knowledge base, where each entry contains a unique symbol character and its associated symbol reference representation. The symbol reference representation is a two-dimensional reference coordinate system consisting of a reference setpoint measure and a reference dispersion measure, representing the standard position of the prototype symbol in the numerical distribution feature space. Spatial position comparison involves measuring the distance in a two-dimensional plane between the statistical description pair of a subsequence segment (i.e., the coordinate point formed by the setpoint measure and dispersion measure) and the symbol reference representation coordinates of all prototype symbols in the pre-constructed symbol dictionary. Offset distance is a scalar measure of the difference between the coordinates of the statistical description pair and the symbol reference representation coordinates. Euclidean distance can be used, which involves calculating the difference in the setpoint measure dimension and the difference in the dispersion measure dimension, taking the square root of the sum of the squares of the two differences. The substitute symbol is the symbol character associated with the prototype symbol selected from the pre-constructed symbol dictionary, which will occupy the corresponding position in the discrete symbol sequence.

[0048] In one implementation, step S250 specifically includes the following steps S251 to S256: Step S251: Perform cross-measurement dimensional mapping on the central tendency measure and dispersion measure in the statistical description pair. Based on the global distribution benchmark of all measurement point variable sequences, perform positional alignment and amplitude scaling on the central tendency measure and dispersion measure to obtain the standardized description coordinates of the subsequence segment.

[0049] Cross-measurement-point dimensionless mapping aims to eliminate the problem of incompatible symbol dictionaries caused by differences in physical units and numerical magnitudes between variables at different measurement points. By transforming the central tendency and dispersion measures of each measurement point variable into a dimensionless standardized space, it allows for fair comparison and symbol mapping of sequence fragments from different measurement points within the same feature space. The global distribution baseline is the overall statistical feature framework extracted from all continuous time series data of all measurement point variables. It includes the mean of the global central tendency and the mean of the global dispersion measures, serving as the origin reference for position alignment; it also includes the standard deviation of the global central tendency and the standard deviation of the global dispersion measures, serving as scaling factors. The standardized descriptive coordinates are the two-dimensional coordinates of the central tendency and dispersion measures in the standardized space after position alignment and scaling.

[0050] For example, first, iterate through all continuous time series data of all measurement point variables on the production line, aggregating the sampled values ​​of all measurement point variables at all times into a global data pool. Calculate the arithmetic mean of this global data pool as the global central tendency measure mean, and calculate the standard deviation of this global data pool as the global dispersion measure mean. Simultaneously, calculate the central tendency measure and dispersion measure within the sequence of each measurement point variable individually, then aggregate these central tendency measure values ​​across measurement points and calculate their standard deviations to obtain the global central tendency measure standard deviation. Similarly, aggregate the dispersion measure values ​​of each measurement point and calculate their standard deviations to obtain the global dispersion measure standard deviation. For the statistical description pair of the current subsequence segment to be processed, subtract the global central tendency measure mean from its central tendency measure, and then divide by the global central tendency measure standard deviation to obtain the standardized central tendency measure; subtract the global dispersion measure mean from its dispersion measure, and then divide by the global dispersion measure standard deviation to obtain the standardized dispersion measure. The standardized central tendency measure and the standardized dispersion measure together constitute the standardized descriptive coordinates of this subsequence segment.

[0051] Step S252: Read the preset symbol normalization coordinates of each prototype symbol from the preset symbol dictionary, calculate the vector end distance between the normalization description coordinates and the normalization coordinates of each symbol, and obtain the initial separation degree sorting of the subsequence segment and each prototype symbol.

[0052] Symbol normalized coordinates are the reference positions of each prototype symbol in the predefined symbol dictionary within the normalized space. Their compilation method is completely consistent with that of normalized descriptive coordinates; both are normalized two-dimensional coordinates obtained by aligning positions and scaling amplitudes based on the same global distribution reference to ensure comparability. Vector endpoint distance calculation treats the normalized descriptive coordinates as the endpoints of a two-dimensional vector originating from the origin, and the symbol normalized coordinates as the endpoints of another two-dimensional vector, calculating the straight-line distance between these two endpoints. Initial separation sorting arranges the vector endpoint distances calculated for the current subsequence segment with respect to all prototype symbols in ascending order; the smaller the distance, the closer the numerical distribution of the subsequence segment is to the prototype symbol.

[0053] In this embodiment, the standardized coordinates of the first, second, third, fourth, fifth, and sixth characters of a symbol are read sequentially from a preset symbol dictionary. For the standardized descriptive coordinates of the current sub-sequence segment of the clamping force, the vector terminal distance is calculated between them and the standardized coordinates of the first character of the symbol: the difference is calculated and squared on the central tendency metric dimension of the standardized space, and the difference is also calculated and squared on the dispersion metric dimension. The two squared values ​​are added together and the square root is taken to obtain the separation degree from the first character of the symbol. The separation degree from the remaining five prototype symbols is calculated in the same way. The six separation degree values ​​are arranged in ascending order to generate the initial separation degree ranking of the sub-sequence segment. The prototype symbols ranked higher represent better mapping candidates.

[0054] Step S253: Identify the pre-defined substitute symbol in the time order of the current subsequence segment to form the pre-order symbol background, and extract the set of legal subsequent symbols that are allowed to appear under the pre-order symbol background according to the preset symbol conversion protocol table.

[0055] The leading substitution symbol is the final substitution symbol assigned to the immediately preceding processed subsequence segment. If the current subsequence segment is the first segment in chronological order, the leading substitution symbol takes a special starting mark. The preceding symbol context is an ordered context consisting of the leading substitution symbol and its possible preceding symbol. Its length is determined by the memory order upon which the predefined symbol transition protocol table depends. The predefined symbol transition protocol table is a finite set of state transition rules used to describe the legal transition relationships between consecutive symbols in a symbol sequence. This table is stored in the form of a matrix or state transition diagram. Each record in the table indicates which symbols can appear as subsequent symbols under a given preceding symbol or combination of preceding symbols. The set of legal subsequent symbols is the set of all symbol characters allowed by the protocol table under the constraints of the current preceding symbol context.

[0056] When constructing the preset symbol conversion protocol table, a method can be used to statistically analyze the symbol transfer frequency from historical normal operating data of the production line and combine this with expert experience to eliminate physically impossible symbol transfer paths. For example, if the process logic of the clamping force process determines that the clamping force cannot immediately jump to a continuously decreasing segment after the rapid rise segment, then the transfer record from the symbol representing the rapid rise to the symbol representing the continuously decreasing segment in the protocol table is marked as invalid. When mapping the current subsequence segment, the finally determined replacement symbol of the previous segment is first extracted to form the preceding symbol background. The valid transfer records in the preset symbol conversion protocol table based on the preceding symbol background are searched, and all allowed subsequent symbols in the records are extracted to form a set of valid subsequent symbols.

[0057] Step S254: Apply transformation constraints to the initial separation degree ranking by using the set of legal subsequent symbols, and expand the separation degree values ​​corresponding to prototype symbols that do not belong to the set of legal subsequent symbols to a preset penalty value, thereby obtaining a modified separation degree ranking that takes into account context constraints.

[0058] Transition constraints are based on the contextual legal transition rules of the symbol sequence, imposing restrictions on candidate symbols. Only prototype symbols that fall into the set of legal subsequent symbols are allowed as replacement symbol candidates for the current subsequence segment. The preset penalty value is a value much larger than all normal separation metrics. When a prototype symbol is not in the set of legal subsequent symbols, its corresponding separation degree is forcibly set to this preset penalty value, placing the prototype symbol at a severe disadvantage in the ranking and thus excluding it from the valid candidates. Corrected separation degree ranking is the result of re-ranking after applying penalties to some separation degree values.

[0059] For the current sub-sequence segment of the clamping force, after obtaining the initial separation degree sort and the set of legal subsequent symbols, the six prototype symbols involved in the initial separation degree sort are traversed. Each prototype symbol is checked to see if it exists in the set of legal subsequent symbols. If a prototype symbol is not in the set of legal subsequent symbols, its corresponding separation degree value is replaced with a preset penalty value. This preset penalty value is determined by taking a multiple of the maximum normal separation degree value to ensure that the penalized prototype symbol is at the bottom of the corrected sort and cannot be selected. For prototype symbols that belong to the set of legal subsequent symbols, their separation degree values ​​remain unchanged. Then, all prototype symbols are reordered in ascending order according to the adjusted separation degree values ​​to form the corrected separation degree sort.

[0060] Step S255: Select the prototype symbol with the smallest separation in the corrected separation sorting, and extract the symbol character bound to the prototype symbol as the final replacement symbol for the subsequence segment.

[0061] The first position in the sorted sequence based on separation is the one with the lowest separation. This prototype symbol is most similar to the current subsequence in terms of numerical distribution and also satisfies the legal transformation constraints under the background of the preceding symbol. Therefore, it is determined as the final replacement symbol. The symbol characters bound to this prototype symbol in the preset symbol dictionary are extracted and assigned to the current subsequence, completing the symbolization of this segment.

[0062] Step S256: Append the final substitution symbol to the end of the generated symbol sequence. After all subsequence segments have been processed, output the complete symbol sequence as the discrete symbol sequence of the measurement point variables.

[0063] When processing the first subsequence segment, the generated symbol sequence is empty, and the final substitution symbol is added as the first element. After processing each subsequent subsequence segment, the final substitution symbol of that segment is appended to the end of the symbol sequence. Once all subsequence segments have been symbol mapped, a symbol sequence with a length equal to the total number of subsequence segments is obtained. This symbol sequence is the discrete symbol sequence of the measured variable.

[0064] Step S260: Summarize the discrete symbol sequences generated by all measurement point variables in the production line, catalog and arrange them using the measurement point variable identifier as the index, and obtain a discrete symbol sequence concatenation table. The discrete symbol sequence concatenation table is used to provide target input data for calculating the transfer entropy of the discrete symbol sequences corresponding to any two measurement point variables.

[0065] The measurement point variable identifier is a unique identifier for each measurement point variable registered in the production line monitoring system. It is usually composed of workshop code, production line code, equipment code, and measurement point physical quantity description, using a hierarchical naming rule. The cataloging arrangement uses the measurement point variable identifier as the key and the corresponding discrete symbol sequence as the value to construct an ordered directory with a key-value structure. All discrete symbol sequences of measurement point variables are neatly listed in lexicographical order or registration order according to the measurement point variable identifier.

[0066] After generating discrete symbol sequences for five measurement point variables—the clamping force sensor value, the slider displacement encoder reading, the hydraulic system oil supply pressure value, the vibration amplitude of the robot end effector, and the thermocouple temperature value on the mold surface—each discrete symbol sequence is assigned a corresponding measurement point variable identifier. These five identifiers are then paired with their respective discrete symbol sequences to form a table-structured data structure containing five entries. This table-structured data is the discrete symbol sequence concatenation table. Subsequent calculations of the transfer entropy between any two measurement point variables will directly extract the corresponding discrete symbol sequence from this concatenation table according to the measurement point variable identifier as input.

[0067] Step S300: For any two discrete symbol sequences of measurement point variables, perform multi-delay propagation entropy calculation to obtain the directed causal intensity value that characterizes the direction and driving intensity of information flow between the two measurement point variables. Traverse all pairs of measurement point variables and summarize all directed causal intensity values ​​to obtain the directed causal intensity set.

[0068] Multi-delay transfer entropy calculation, based on the classical transfer entropy framework, introduces multiple different candidate time delay values ​​to examine the changes in information transfer from one discrete symbol sequence to another under different time delay settings, in order to capture the inherent response lag characteristics in process dynamics. The core idea of ​​transfer entropy is to measure the contribution of the past state of the source sequence to the reduction of uncertainty in the current state of the target sequence, given the past state of the target sequence. When significant information transfer exists, the transfer entropy value will be greater than the noise level under random fluctuations. The directed causality strength value is the amount of information transfer when the transfer entropy reaches its maximum value among a series of candidate time delays. This maximum value reflects the strength of the information driving force of the source measurement variable on the target measurement variable. Traversing all measurement variable pairs involves randomly selecting two different measurement variables from the discrete symbol sequence concatenation list, forming a pair as the source and target sequences respectively, and calculating for each possible pair.

[0069] When calculating multi-delay transfer entropy on a discrete symbol sequence linked list consisting of five measurement point variables, the discrete symbol sequence of the clamping force sensor values ​​is first selected as the source sequence, and the discrete symbol sequence of the slider displacement encoder readings is selected as the target sequence. A candidate time delay list is set, which contains multiple integer time delay steps from zero to a preset maximum time delay. For each candidate time delay value, the source sequence is shifted backward by the number of time delay steps, and realigned with the target sequence according to a common effective time range. The transfer entropy from the source sequence to the target sequence under that time delay is calculated. Then, the roles of source and target are swapped, and the transfer entropy from the slider displacement encoder readings to the clamping force sensor values ​​is calculated. This process is repeated until all five measurement point variable pairs, totaling twenty directed relations, are processed. Each directed relation corresponds to a maximum transfer entropy value and a corresponding optimal time delay value. The strength values ​​of all directed relations are summarized to form a directed causal strength set.

[0070] In one implementation, step S300 specifically includes the following steps S310 to S360: Step S310: Read the first discrete symbol sequence and the second discrete symbol sequence of the target measurement point variable pair from the discrete symbol sequence concatenation table. Under the condition of no time delay, count the frequency of co-occurrence of each symbol pair in the first discrete symbol sequence and the second discrete symbol sequence to generate the joint symbol occurrence distribution structure.

[0071] The first and second discrete symbol sequences are the discrete symbol sequences of the two measurement point variables in the target measurement point variable pair, respectively. Before specifying a direction, they are equal in status. Counting the frequency of co-occurrence under no time delay involves aligning the two sequences one-to-one according to the same time position, recording the symbol pairs formed by the symbols of the first and second discrete symbol sequences at each time position, and counting the occurrence of symbol pairs at all time positions. The joint symbol occurrence distribution structure is a structured probability distribution description containing all possible symbol pair combinations and their occurrence probabilities or frequencies.

[0072] In this embodiment, an example of a target measurement point variable pair is the clamping force sensor value and the slider displacement encoder reading. The discrete symbol sequence corresponding to the clamping force sensor value is extracted from the discrete symbol sequence concatenation table as the first discrete symbol sequence, and the discrete symbol sequence corresponding to the slider displacement encoder reading is extracted as the second discrete symbol sequence. At the same time index position of the two sequences, the symbol characters of the first and second discrete symbol sequences are concatenated to form an ordered symbol pair. After traversing all time positions, the total number of occurrences of each possible symbol pair combination in all time positions is counted, and the frequency is converted into a proportion to form a joint probability distribution structure. This structure comprehensively reflects the symbol co-occurrence relationship of the two measurement point variables under time alignment.

[0073] In one implementation, step S310 specifically includes the following steps S311 to S316: Step S311: Obtain the row and column dimensions of the preset symbol frequency statistics container, where the rows correspond to all possible symbols of the first discrete symbol sequence and the columns correspond to all possible symbols of the second discrete symbol sequence. Initialize the count value of each row and column intersection unit in the symbol frequency statistics container.

[0074] The preset symbol frequency statistics container is a two-dimensional counting matrix. Its rows are formed by sequentially arranging the elements of the complete set of symbols allowed to appear in the first discrete symbol sequence, and its columns are formed by sequentially arranging the elements of the complete set of symbols allowed to appear in the second discrete symbol sequence. The two complete sets of symbols are obtained by taking values ​​from their respective discrete symbol sequences, removing duplicates, and then sorting them. Each row and column intersection unit is a counting slot corresponding to the intersection of row and column symbols. During initialization, the count values ​​of all slots are set to zero.

[0075] In this step, the complete set of possible symbols for the first discrete symbol sequence is obtained by extracting all unique symbol characters that have appeared in its own sequence and sorting them in character order; the complete set of possible symbols for the second discrete symbol sequence is obtained in the same way. Then, a two-dimensional integer array with the number of rows equal to the number of elements in the complete set of symbols for the first discrete symbol sequence and the number of columns equal to the number of elements in the complete set of symbols for the second discrete symbol sequence is created as a symbol frequency counting container. All elements of this two-dimensional array are traversed and assigned the value of zero.

[0076] Step S312: Pair the first discrete symbol sequence with the second discrete symbol sequence one by one according to the same time position to generate a symbol pairing sequence. Each pairing element in the symbol pairing sequence consists of the first discrete symbol sequence symbol and the second discrete symbol sequence symbol at the same time position.

[0077] Pairing symbols one by one at the same time position uses a common time index sequence as a reference. At the k-th time position, the k-th symbol of the first discrete symbol sequence and the k-th symbol of the second discrete symbol sequence are extracted and encapsulated into an ordered symbol pair. When the two discrete symbol sequences are exactly the same length, the pairing operation traverses all positions; when the lengths are unequal, the common time range is truncated for pairing. The symbol pairing sequence is a new sequence formed sequentially by these ordered symbol pairs, and its length is equal to the total number of common time positions.

[0078] Step S313: Traverse the symbol pairing sequence. For each pairing element read, locate the corresponding row and column intersection unit in the symbol frequency statistics container and increase the cumulative count value of that unit until all pairing elements have been processed, and obtain the original frequency statistics table.

[0079] The process of traversing the symbol pairing sequence begins with the first element and proceeds sequentially through each ordered symbol pair. For each ordered symbol pair, using the symbol from the first discrete symbol sequence as the row index and the symbol from the second discrete symbol sequence as the column index, the corresponding row-column intersection cell is located in the symbol frequency statistics container, and the currently stored count value in that cell is incremented by one unit. After all symbol pairing elements have been processed, the symbol frequency statistics container has accumulated complete frequency information, and this two-dimensional array is now the original frequency statistics table.

[0080] Step S314: Perform overall frequency reduction on the count values ​​of each unit in the original frequency statistics table, convert the cumulative count value of each unit into its proportion of the total number of paired elements, and generate a joint occurrence probability distribution table.

[0081] The overall frequency total is calculated by summing the counts of all intersecting rows and columns in the original frequency statistics table to obtain the total number of paired elements. Then, the cumulative count of each unit is used as the numerator, and the total number of paired elements is used as the denominator to calculate the proportion value of each unit. The joint occurrence probability distribution table is a two-dimensional real matrix of the original frequency statistics table after replacing the count values ​​with the corresponding proportion values. Each element in the table represents the probability of occurrence of the corresponding symbol pair.

[0082] After summing the counts of all cells in the original frequency statistics table to obtain the total number of pairings, each row and column intersection cell is traversed, its count is divided by the total number of pairings, and the resulting floating-point number is written into the corresponding position in the joint occurrence probability distribution table. After all cells have been converted, the sum of all probability values ​​in the joint occurrence probability distribution table is 1.

[0083] Step S315: Perform probability benchmark compensation on the units with zero probability in the joint occurrence probability distribution table, assign a preset small probability value to the units with zero probability, and adjust the non-zero probability units to keep the overall probability sum unchanged, thereby obtaining the modified joint probability distribution table.

[0084] Probability baseline compensation is used to avoid singularity issues in subsequent information measurement calculations that might arise from the presence of absolutely zero probability terms in the joint probability distribution. The preset microprobability is a fixed small value much smaller than the normal probability order of magnitude. Its value can be defined as the reciprocal of the total number of pairings multiplied by a coefficient less than one. After assigning this microprobability to zero-probability units, all probability values ​​are renormalized, that is, the values ​​of all non-zero-probability units are reduced proportionally, so that the sum of the probabilities of all units remains precisely one.

[0085] The joint occurrence probability distribution table is scanned, the number of units with a probability of zero is recorded, and a preset small probability value is assigned to each unit with a zero probability. The sum of the small probability values ​​assigned to all units with a probability of zero is calculated, and this sum is subtracted from the overall probability sum. The remaining overall probability share is then distributed and reduced among the original non-zero probability units according to their respective original probability values. Each original non-zero probability unit is updated one by one with the reduced probability value, and the probability values ​​of all units are then written into the modified joint probability distribution table.

[0086] Step S316: Structure the modified joint probability distribution table into a joint symbol occurrence distribution structure, which contains complete probability information of each symbol pair occurrence.

[0087] The joint symbol occurrence distribution structure is a complete encapsulation of the modified joint probability distribution table. Internally, it stores complete information such as the conditional and marginal probabilities of all possible symbol pairs, which can be directly queried and accessed for subsequent transfer entropy calculations. This structure internally maintains a row-indexed symbol list, a column-indexed symbol list, and the modified joint probability matrix, and provides an interface for querying probabilities by symbol pair.

[0088] Step S320: Sequentially set candidate delay values, shift the first discrete symbol sequence along the time axis according to each candidate delay value, re-count the frequency of common occurrence of symbol pairs between the shifted first discrete symbol sequence and the unshifted second discrete symbol sequence, and generate the joint delay distribution structure under the candidate delay.

[0089] The candidate delay values ​​are a series of integer values ​​starting from zero and increasing in steps of the sampling period. Their maximum value is set based on the maximum possible propagation delay of the production line process. Time axis shifting involves moving the entire first discrete symbol sequence in the positive direction of the time axis by the number of steps specified by the candidate delay values. Positions left blank at the beginning of the sequence and positions exceeding the time range of the second discrete symbol sequence are discarded, retaining only the portions of the two sequences within the common effective time window after shifting. The joint delay distribution structure is reconstructed after applying a specific candidate delay, using the same frequency statistics and probability compensation process as steps S311 to S316, to reconstruct the joint symbol occurrence distribution structure between the two sequences.

[0090] In this embodiment, for the first discrete symbol sequence of the clamping force sensor values ​​and the second discrete symbol sequence of the slider displacement encoder readings, candidate time delay values ​​start from zero and are sequentially selected by one step, two steps, three steps, until a preset maximum time delay is reached. For each candidate time delay value, the first discrete symbol sequence is shifted to the right by the corresponding number of steps, and the common interval where the two sequences are time-aligned after the shift is extracted as the effective analysis interval. Within this effective interval, the probability of occurrence of all symbol pairs is calculated to generate the joint time delay distribution structure corresponding to the candidate time delay value.

[0091] Step S330: For each candidate delay, compare the distribution difference between its joint delay distribution structure and the joint symbol occurrence distribution structure under the condition of no delay, extract the change in conditional information caused by time shift, and use the change in conditional information as the dependency strength measure corresponding to the candidate delay.

[0092] The comparison of distribution differences is conducted within the framework of information theory. First, the conditional entropy under the joint distribution structure with time delay and the conditional entropy under the joint symbol occurrence distribution structure without time delay are calculated separately. The difference between the two conditional entropies represents the change in conditional information caused by the time shift. The calculation of conditional entropy is based on taking the logarithm of the ratio of the joint probability distribution to the marginal probability distribution and then calculating its expectation. It reflects the degree of uncertainty regarding the value of one sequence symbol given the value of one sequence symbol. The dependency strength metric is defined as this change in conditional information; it measures the contribution of the source sequence to the reduction of uncertainty in the target sequence under a specific time delay.

[0093] For each candidate time delay, the conditional entropy of the second discrete symbol sequence under the condition of the first discrete symbol sequence after known shift is first calculated based on the joint time delay distribution structure. Then, the conditional entropy of the second discrete symbol sequence under the condition of the first discrete symbol sequence without shift is calculated based on the joint symbol occurrence distribution structure without time delay. The difference between the baseline conditional entropy without time delay and the time delay conditional entropy is the dependency strength measure under that candidate time delay. If the change in conditional information after time delay shift is greater than the change in conditional information of the baseline without time delay, it indicates that the source sequence's interpretability of the target sequence is enhanced under that time delay.

[0094] Step S340: Detect the delay point when the dependency strength metric reaches its peak from all candidate delays, and record the delay value and the peak dependency strength metric value corresponding to that delay point.

[0095] The dependency strength metric reaches its peak when a local maximum appears on the curve of the dependency strength metric versus candidate delay. This maximum may be globally unique among all candidate delays, or multiple local peaks may exist. The delay point is the candidate delay value corresponding to when the dependency strength metric reaches its peak. The peak dependency strength metric value is the dependency strength metric value calculated at this delay point.

[0096] Along the direction from the clamping force sensor reading to the slider displacement encoder reading, candidate time delay values ​​and their dependence intensity metric values ​​are plotted as a time delay dependence curve. Observe the changing trend of the dependence intensity metric along the time delay axis, find the reversal point where the dependence intensity metric changes from rising to falling, and the time delay point corresponding to this reversal point is the peak time delay point, and the dependence intensity metric value at this point is the peak dependence intensity metric value.

[0097] In one implementation, step S340 specifically includes the following steps S341 to S346: Step S341: Pair the candidate delay value sequence with the corresponding dependency strength metric sequence to obtain the delay dependency strength metric corresponding sequence, and arrange the corresponding sequence in ascending order of candidate delay value.

[0098] The candidate delay value sequence is a list of all candidate delay values ​​arranged in the order of their generation. The dependency strength metric sequence is a list of corresponding dependency strength metric values ​​arranged in the same order. The delay dependency strength metric corresponding sequence is formed by pairing elements at the same position in two sequences to form tuples, and then sorting the tuples in ascending order according to the candidate delay values.

[0099] Step S342: On the sequence corresponding to the sorted time delay dependency intensity measure, traverse each candidate time delay point, compare its dependency intensity measure value with the dependency intensity measure value of the adjacent time delay points, find the local peak point that is larger than the dependency intensity measure value of the adjacent points, and mark it as the candidate peak point.

[0100] The process iterates from the second candidate delay point to the penultimate candidate delay point. For the current candidate delay point, its dependency strength metric is extracted, along with the dependency strength metrics of the preceding and following candidate delay points. If the dependency strength metric of the current point is greater than both the preceding and following candidate delay points, then the current point is marked as a candidate peak point.

[0101] Step S343: Among the candidate peak points, further evaluate the degree of distinction between the dependency strength metric value of each candidate peak point and the average level of the entire dependency strength metric sequence, filter out candidate peak points with insufficient distinction, and retain significant candidate peak points.

[0102] The degree of discrimination is measured by calculating the difference between the dependency strength metric of the candidate peak point and the median or mean of the entire dependency strength metric sequence. If the ratio of this difference to the standard deviation of the entire dependency strength metric sequence is less than a preset significant multiple, the degree of discrimination is deemed insufficient, and the candidate peak point is filtered out; otherwise, the candidate peak point is retained and marked as a significant candidate peak point.

[0103] Step S344: If there are multiple significant candidate peak points, compare the magnitude of the dependency strength metric of each significant candidate peak point, and select the significant candidate peak point with the largest dependency strength metric as the first peak point.

[0104] All remaining significant candidate peaks are sorted in descending order of their dependence strength metric, with the first one being designated as the first peak. If only one significant candidate peak exists, it is directly designated as the first peak.

[0105] Step S345: Extract the candidate delay value corresponding to the first peak point as the representative delay value, and extract the dependency strength metric value corresponding to the first peak point as the peak dependency strength metric value.

[0106] The representative delay value is the optimal time delay corresponding to the strongest information transmission observed from the directed relation, and the peak dependency strength metric is the quantitative expression of the strongest information transmission.

[0107] Step S346: Associate and store the representative time delay value and the peak dependence strength metric value, and output them as the causal strength and effect time delay information of the target measurement point variable pair.

[0108] By using associated storage, the strength and latency values ​​of the directed relationship can be directly retrieved when constructing causal edges later.

[0109] Step S350: Determine the direction of interaction between the measurement point variable pairs based on the time delay value, take the peak dependence strength metric value as the directed causal strength value from the cause measurement point variable to the effect measurement point variable, and save the time delay value as the optimal action time delay for the measurement point variable pair.

[0110] The direction of action is determined as follows: The peak dependency strength metric of the first discrete symbol sequence on the second discrete symbol sequence at a representative time delay is compared with the peak dependency strength metric of the second discrete symbol sequence on the first discrete symbol sequence at a representative time delay. The direction corresponding to the larger value, from the source sequence measurement point variable to the target sequence measurement point variable, is determined as the information causal driving direction. The result measurement point variable is the measurement point variable from which the driving direction originates, and the effect measurement point variable is the measurement point variable to which the driving direction points. The directed causal strength value is assigned the peak dependency strength metric value in that direction, and the optimal action delay is the corresponding representative time delay value.

[0111] Step S360: Traverse all measurement point variable pairs on the production line in the manner described above, summarize the directed causal strength values ​​and optimal action delays for each measurement point variable pair, and obtain the set of directed causal strengths.

[0112] All the directions of positive causal relationships between all measured point variables, the values ​​of directional causal strength, and the optimal effect delay are collected into a common list. This list is the directional causal strength set. Each record in the set consists of four fields: cause measured point variable identifier, effect measured point variable identifier, directional causal strength value, and optimal effect delay.

[0113] Step S400: Extract measurement point variable pairs from the directed causal strength set whose strength values ​​satisfy the preset causal significance condition and whose direction is clear from cause to effect. For each extracted measurement point variable pair, instantiate measurement point nodes and device nodes of the devices to which the measurement point variables belong. Establish causal relationship edges from the cause measurement point node to the effect measurement point node. Based on the transfer entropy time delay analysis, assign measurement point-level action delay attributes to each causal relationship edge to form a set of causal relationship edges at the measurement point node level.

[0114] The preset causal significance condition is a criterion for filtering causal strength values, used to eliminate pseudo-causal relationships with extremely small strength values, possibly generated by the co-occurrence of random symbol sequences. A clear direction, pointing from cause to effect, means that the positive directed causal strength value of the measured variable pair is significantly greater than the negative directed causal strength value, and the difference exceeds the preset direction difference tolerance, thus indicating that the causal driving direction is unambiguous. Instantiating a measured node involves creating a vertex object representing a specific measured variable in the data structure. This object contains attributes such as the measured variable identifier and the device identifier to which the measured variable belongs. Instantiating a device node involves creating a vertex object representing a device in the data structure. This object contains attributes such as the device identifier and device type. A causal relationship edge is a directed edge from the cause measured node to the effect measured node. Besides recording the references to the two endpoints, this edge also carries the measured-level effect delay attribute.

[0115] In this embodiment, the directed causal strength set is traversed, and the directed causal strength value of each record is compared with the strength threshold in the preset causal significance condition. The strength threshold is set based on the statistical distribution of all directed causal strength values, usually by adding a certain multiple of the median absolute deviation. For records whose directed causal strength values ​​exceed the strength threshold, the difference between their forward and reverse directed causal strength values ​​is further examined, and a directional dominance measure is calculated. Records whose directional dominance measure exceeds a preset directional difference tolerance are identified as causal relationship pairs with clear directions. For each causal relationship pair with clear directions, a cause-measurement node object and an effect-measurement node object are created in the data container, and a causal relationship edge object is created from the cause-measurement node to the effect-measurement node. At the same time, the optimal action delay value extracted from the directed causal strength set is written into the measurement-level action delay attribute field of this edge. All causal relationship edges created in the above manner are aggregated to obtain the causal relationship edge set at the measurement node level.

[0116] In one implementation, step S400 specifically includes the following steps S410 to S460: Step S410: For each directed causal strength value in the directed causal strength set, compare it with the preset causal significance judgment benchmark, filter out measurement point variable pairs whose strength values ​​are lower than the preset causal significance judgment benchmark, and retain causal relationship candidate pairs that meet the significance requirements.

[0117] The pre-defined causal significance criterion is an adaptive threshold derived from the numerical distribution of directed causal intensity. Its construction process involves: calculating the median of all directed causal intensity values ​​in the set; calculating the median of the absolute difference between each value and the median; multiplying this median by a pre-defined factor and adding it to the median; and using this median as the significance criterion. Candidate causal pairs are those pairs of measurement point variables whose directed causal intensity values ​​meet the significance requirement.

[0118] Step S420: Verify the directional attributes of the candidate causal relationship pairs, confirm that the directional causal strength value of the cause-effect pair is higher than the reverse directional causal strength value and exceeds the preset directional difference tolerance, retain the pair of measurement variables with a clear direction from cause to effect, and obtain the confirmed causal relationship pair set.

[0119] Directional attribute verification involves extracting the directional causal strength values ​​for each candidate causal relationship pair in both the forward and reverse directions to determine the superiority or inferiority of the direction. A preset directional difference tolerance is a threshold value designed to ensure the statistical reliability of the directional judgment.

[0120] In one implementation, step S420 specifically includes the following steps S421 to S425: Step S421: For each candidate causal relationship pair, extract the positive causal strength value from the first measurement point variable to the second measurement point variable, and the negative causal strength value from the second measurement point variable to the first measurement point variable from the set of directed causal strengths.

[0121] The first and second measurement point variables are the internal symbols of the two measurement point variables that constitute a candidate causal relationship pair within the data structure. The distinction between positive and negative causality is determined by the order in which they are recorded in the set of directed causal strengths.

[0122] Step S422: Calculate the difference between the positive directional causality strength value and the negative directional causality strength value, and calculate the relative proportion of this difference to the positive directional causality strength value to obtain the directional dominance measure.

[0123] The directional dominance measure is defined as the difference between the positive directional causal strength value and the negative directional causal strength value, divided by the positive directional causal strength value plus a very small regularization constant. The value ranges from zero to one, and the larger the value, the more significant the positive dominance.

[0124] Step S423: Compare the directional advantage metric with the preset directional advantage criterion. If the directional advantage metric exceeds the preset directional advantage criterion, then determine that the direction of the causal relationship candidate pair meets the directional definiteness requirement.

[0125] The preset directional dominance determination benchmark is an empirical threshold, determined based on the distribution of directional dominance metrics from a large number of known causal relationships. When the directional dominance metric exceeds this benchmark, the difference between forward and reverse directions becomes substantially significant.

[0126] Step S424: For candidate causal relationships that meet the requirement of directional clarification, based on the relationship between the magnitudes of the positive directional causal strength values ​​and the negative directional causal strength values, the direction from the side with the larger strength value to the side with the smaller strength value is determined as the causal pointing direction.

[0127] If the positive directional causal strength is greater than the negative directional causal strength, then the causal direction is from the first measurement point variable to the second measurement point variable; otherwise, it is from the second measurement point variable to the first measurement point variable.

[0128] Step S425: Encapsulate the positive directional causal strength value, negative directional causal strength value, and causal direction of the candidate causal relationship pairs with a definite causal direction into a direction confirmation record, and summarize all the measurement point variable pairs in the direction confirmation record to form a set of confirmed causal relationship pairs.

[0129] The directional weighting record is a structure containing forward intensity values, reverse intensity values, causal pointing direction, and optimal action delay. The set of confirmed causal relationships is the set of all measurement point variable pairs corresponding to all directional weighting records.

[0130] Step S430: To confirm each causal relationship pair in the set of causal relationship pairs, create a causal point node instance representing the causal point variable and an effect point node instance representing the effect point variable in the data structure, and establish a pointing relationship record for the causal point node instance and the effect point node instance.

[0131] A cause measurement point node instance is a data object containing attributes such as measurement point variable identifier, physical meaning description of the measurement point variable, and identification of the associated equipment. Result measurement point node instances have the same structure. The pointer relationship record associates the object references of cause and result measurement point node instances with directed edges.

[0132] Step S440: Obtain the equipment identifier of each measurement point variable from the production line equipment configuration description, associate the cause measurement point node instance with the first equipment node instance to which the cause measurement point variable belongs, and associate the result measurement point node instance with the second equipment node instance to which the result measurement point variable belongs, thus completing the attribution link from the measurement point layer to the equipment layer.

[0133] A production line equipment configuration description is a structured document that describes the physical composition of a production line. It is typically stored in Extensible Markup Language (EXPLAIN) or a similar format, recording the identifier of each piece of equipment, its type, a list of measurement point variables installed on the equipment, and the mechanical connections between the equipment. The first and second equipment node instances are the vertex objects in the knowledge graph representing the equipment to which the cause and effect measurement point variables belong, respectively.

[0134] Step S450: Read and confirm the optimal action delay recorded during the transfer entropy analysis of the causal relationship pair, and directly attach the optimal action delay as the measurement point level action delay attribute of the causal relationship edge to the causal relationship edge pointing from the cause measurement point node to the effect measurement point node, and mark the measurement point identifier source corresponding to the delay value of each causal relationship edge, indicating that the delay value belongs to a specific measurement point layer causal pair.

[0135] The delay attribute at the measurement point level is delay information directly bound to the causal relationship edges at the measurement point layer. This information originates from the representative delay value determined during the transfer entropy delay scanning phase. The measurement point identifier source refers to the ordered pair of cause and effect measurement point variable identifiers to which this delay value belongs. Marking this source is to preserve the traceability of the delay contribution of each measurement point layer when subsequently aggregated to the device level.

[0136] In one implementation, step S450 specifically includes the following steps S451 to S455: Step S451: For the confirmed causal relationship pair, retrieve the original record of the best action delay stored in the directed causal strength set, obtain the best action delay value corresponding to the direction of the cause measurement point variable pointing to the effect measurement point variable, and save the delay value in the form of measurement point-level delay record, clearly indicating the source of the measurement point variable to which the delay value belongs.

[0137] Each record in the directed causal strength set uses a tuple of cause and effect measurement point variable identifiers as a joint key, through which the optimal action delay value can be directly retrieved. A measurement point-level delay record is a structure containing the delay value, cause and effect measurement point variable identifiers.

[0138] Step S452: If it is confirmed that there are multiple measurement point hierarchical causal relationship edges in the causal relationship pair corresponding to the same device pair, then extract the optimal action delay value and the measurement point variable source identifier of each measurement point hierarchical causal relationship edge to obtain the set of delay values ​​with measurement point source labels within the same device pair.

[0139] A single device pair is an unordered pair consisting of a first device node and a second device node. When multiple causal measurement point variables installed on the first device have measurement point-level causal relationship edges with multiple effect measurement point variables installed on the second device, these edges all belong to the same device pair. The set of delay values ​​with measurement point source annotations is a list formed by aggregating the measurement point-level delay records of these edges.

[0140] Step S453: Classify and group the delay values ​​in the delay value set according to the measurement point source identifier. Calculate the distribution center measure within the same measurement point variable source group for delay values ​​to obtain the characteristic delay representative quantity of each group. Encapsulate the characteristic delay representative quantity of each group together with the corresponding measurement point identifier into a delay distribution attribute record for the device node pair, retaining the delay difference information of each measurement point source instead of merging it into a single mean.

[0141] The within-group distribution center measure can be calculated using the arithmetic mean or median of the within-group time delay values ​​as the representative quantity of the characteristic time delay. The time delay distribution attribute record is a key-value mapping structure, where the key is the source identifier of the measurement point variable, and the value is the representative quantity of the characteristic time delay corresponding to that source group.

[0142] Step S454: Record the time delay distribution attributes and write them into the description structure of the semantic relationship edge of the corresponding device node, so that the semantic relationship edge carries multi-source time delay distribution information that retains the granularity difference of the measurement point.

[0143] The descriptive structure of a semantic relation edge is an attribute container attached to the edge object, supporting multi-key-value storage. Multi-source latency distribution information is the complete write result of the latency distribution attribute record.

[0144] Step S455: For a device pair containing only a single measurement point level causal relationship edge, the measurement point level delay record of that single edge is directly written as its delay distribution attribute record.

[0145] Step S460: Summarize all established causal relationship edges and their measurement point level effect delay attributes and measurement point identification sources, and group them according to the index of the cause measurement point node and the effect measurement point node to generate a set of causal relationship edges at the measurement point node level.

[0146] Step S500: Obtain a static structural constraint model pre-constructed based on the fixed mechanical transmission topology and physical adjacency relationship between production line equipment. Perform equipment-level semantic reduction on the causal relationship edge set at the measurement point node level through the static structural constraint model. Aggregate and deduce the causal relationship edges at the measurement point node level belonging to the same equipment pair into semantic relationship edges between equipment nodes. Assign relationship type labels to each semantic relationship edge according to the mechanical connection type of the equipment at both ends of the edge. Generate a production line knowledge graph based on the set of semantic relationship edges of equipment nodes and assigned relationship type labels.

[0147] The static structural constraint model is a priori knowledge structure established based on the physical design and engineering installation data of the production line before performing any data-driven causal analysis. Its core function is to accurately describe which equipment constitutes the production line and how these devices are connected in a fixed manner that does not change with the production process. The information sources for building this model include the overall mechanical design drawing of the production line, the equipment installation foundation layout drawing, the transmission system assembly drawing, the hydraulic and pneumatic circuit diagram, and the equipment procurement list and on-site installation records. The model contains two related components: a set of production line equipment nodes and a description of the mechanical connection types between the equipment. The set of production line equipment nodes is a summary of vertices obtained by digitally abstracting each piece of equipment on the production line that has an independent mechanical body and a clear process function. Each piece of equipment is assigned a globally unique equipment identifier in the model, along with attribute fields such as equipment type, equipment tag number, work section, and rated technical parameters. The granularity of equipment type classification is sufficient to distinguish different functional machines such as stamping presses, feeding robots, finished product conveyors, hydraulic pump stations, and mold cooling devices. The description of mechanical connection types between equipment defines the properties and categories of each physical connection path between devices. It is strictly divided into two main categories: transmission relationship type identifiers and physical adjacency relationship type identifiers. Transmission relationship type identifiers are specifically used to mark connections where power or motion is directly transmitted from one device to another. These connections use clearly defined transmission elements as the medium, such as gear pairs, pulley sets, chains and sprockets, universal couplings, or hydraulic lines. This indicates a forced motion coupling between the two devices during operation; changes in the operating speed, torque, or displacement of one device will forcibly cause changes in the corresponding physical quantities of the other device through the transmission chain. Physical adjacency relationship type identifiers are specifically used to mark connections that do not intend to transmit power, but exist due to spatial proximity, shared base mounting, or rigid fixation via brackets or fasteners. Examples include two devices sharing the same mounting base plate, one device's frame being bolted to the outer shell of another device, or two devices having only a small gap in space without direct contact. This indicates that there is only a spatial constraint and structural support relationship between the two devices, without continuous motion or forced power transmission.

[0148] In this model, each device connection record is organized using ordered pairs of source and target device identifiers as indexes, and includes a structured connection type field. The connection type field employs a two-layer encoding mechanism: the first layer is an enumerated category code, with values ​​limited to two categories: transmission connection and physical adjacency connection, used to distinguish the essential attributes of the connection at the highest level; the second layer is a fine-grained code, further subdividing the transmission connection category into gear transmission, belt transmission, chain transmission, hydraulic transmission, etc., and the physical adjacency connection category into fixed adjacency, sliding adjacency, and adjacent non-contact, etc. This hierarchical encoding design allows the static structural constraint model to provide both a binary judgment of whether devices are associated and a detailed semantic description of the association nature. Therefore, when subsequently semantically reducing the causal relationships discovered through data-driven discovery to the device level, the model can provide authoritative constraint information from the mechanical design level.

[0149] In one implementation, step S500 specifically includes the following steps S510 to S560: Step S510: Read the static structural constraint model and extract the set of production line equipment nodes and the mechanical connection type description between the equipment. The mechanical connection type description includes the transmission relationship type identifier and the physical adjacency relationship type identifier.

[0150] The process of reading the static structural constraint model involves parsing and traversing its internal storage structure. This model is stored in the production line design database or engineering configuration file as a graph data structure or a relational table structure. During reading, the equipment node registry is first loaded, and each equipment node record in the registry is traversed to extract the equipment identifier, equipment type, and equipment tag number, assembling them into a production line equipment node set. Next, the equipment connection relationship table is loaded, and each connection record is traversed, extracting the source equipment identifier, target equipment identifier, and connection type fields. The connection type field is then parsed record by record: the first-level enumerated category code and the second-level fine-class code are separated from the connection type field, classifying them as either a transmission relationship type identifier or a physical adjacency relationship type identifier, while fully preserving the transmission method or adjacency method information in the fine-class code.

[0151] In one embodiment, the production line consists of five pieces of equipment: a closed-loop double-point punch press, a multi-joint feeding robot, a roller conveyor belt, a hydraulic pump station, and a mold cooling device. In the static structural constraint model, the production line equipment node set includes the equipment node records for each of these five pieces of equipment. Regarding the connection relationships between the equipment, the model records that the hydraulic pump station and the closed-loop double-point punch press are connected by hydraulic transmission (detailed code: hydraulic transmission); the closed-loop double-point punch press and the multi-joint feeding robot are connected by gear transmission (detailed code: gear transmission); the multi-joint feeding robot and the roller conveyor belt are physically adjacent (detailed code: adjacent but not contacting); and the mold cooling device and the closed-loop double-point punch press are physically adjacent (detailed code: fixed adjacency). These connection records completely cover the entire fixed mechanical transmission topology and physical adjacency relationships of the production line.

[0152] Step S520: Traverse each causal relationship edge in the set of causal relationship edges at the measurement point node level, and look up the device node to which the edge belongs through the measurement point nodes at both ends of the edge. Map the causal relationship edges at the measurement point node level to the device node pair level to generate edge mapping records for device node pairs.

[0153] Each edge in the causal relationship edge set stores the identifiers of both the cause and effect measurement point nodes. The reverse lookup operation is as follows: after obtaining the cause measurement point node identifier, the device identifier installed on that measurement point variable is retrieved from the measurement point registration information table. Then, the device node corresponding to that device identifier is retrieved from the production line device node set. The same method is used to obtain the device node to which the effect measurement point node belongs. Thus, a causal relationship edge at the measurement point level is transformed into an association record between a pair of device nodes. This record is the edge mapping record for the device node pair, which internally retains the directed causal strength value and the measurement point-level effect delay attribute of the original measurement point level edge for use in subsequent aggregation stages.

[0154] In this embodiment, it is assumed that the causal relationship edge set at the measuring point node level includes a causal relationship edge pointing from the clamping force sensor value to the hydraulic system oil supply pressure value. The clamping force sensor is installed on the closed-type double-point press, and the hydraulic system oil supply pressure value sensor is installed on the hydraulic pump station. Through reverse lookup, the cause measuring point node corresponds to the closed-type double-point press equipment node, and the effect measuring point node corresponds to the hydraulic pump station equipment node. The edge mapping record of the generated equipment node pair records the source equipment as the closed-type double-point press, the target equipment as the hydraulic pump station, and the direction as from the closed-type double-point press to the hydraulic pump station.

[0155] Step S530: Aggregate all the causal relationship edges of the measurement point node level belonging to the same device node pair, extract the directed causal strength value and measurement point level action delay attribute of each edge and its measurement point identification source, and fuse them to obtain the comprehensive causal strength description and delay distribution attribute record between the device node pairs.

[0156] Grouping edges belonging to the same device node means grouping edges whose source and target device identifiers are identical. Within this group, all causal relationship edges at the measurement point level are collected, each carrying its own causal strength and latency information. The aggregation process merges the multi-source causal information carried by these edges into a unified representation at the device level. This aims to summarize the overall strength from a statistical perspective while fully preserving the latency differences between the sources of each measurement point, avoiding the loss of important measurement point-level information during aggregation.

[0157] In this embodiment, the closed-loop two-point press is equipped with two measuring point variables: a clamping force sensor and a slider displacement encoder. The hydraulic pump station is equipped with one measuring point variable: the hydraulic system supply pressure value. Therefore, there are two causal relationship edges at the measuring point level from the closed-loop two-point press to the hydraulic pump station, one from the clamping force sensor value to the hydraulic system supply pressure value, and the other from the slider displacement encoder reading to the hydraulic system supply pressure value. These two edges are aggregated within the same device node pair group.

[0158] As one implementation method, step S530 specifically includes the following steps S531 to S535: Step S531: For a specified pair of device nodes, collect all causal relationship edges at the level of the measurement point node belonging to the pair of device nodes to obtain a list of causal relationship edges within the pair of devices. Each entry in the list of causal relationship edges within the pair of devices contains the directed causal strength value of a causal relationship edge, the measurement point level action delay attribute, and its measurement point identifier source.

[0159] The criteria for determining a dependency relationship are: the device to which the cause measurement point node belongs in a causal relationship edge is equal to the source device of that device node pair, and the device to which the effect measurement point node belongs is equal to the target device of that device node pair. All edges that satisfy this condition are added to the list of causal relationship edges within the device pair. Each entry in the list fully retains the original directed causal strength value, the measurement point-level action delay attribute expressed in sampling periods or time units, and the measurement point identifier source composed of the cause measurement point variable identifier and the effect measurement point variable identifier.

[0160] Step S532: Extract the directed causal strength values ​​one by one from the list of causal relationship edges within the device pair, calculate the central tendency representative quantity of all directed causal strength values, and use the central tendency representative quantity as the comprehensive causal strength expression between device node pairs.

[0161] The central tendency measure is a statistical descriptive measure that summarizes multiple directed causal strength values. When there are multiple edges in the list of causal relationships within a device, the central tendency measure is taken as the arithmetic mean of their numerical distributions to reflect the overall level of causal driving strength between device nodes. When the list contains only a single edge, the central tendency measure is simply the directed causal strength value of that edge itself. The comprehensive causal strength representation will serve as one of the important attributes of semantic relationship edges at the device level.

[0162] Step S533: Extract the measurement point-level action delay attribute and its measurement point identification source one by one from the causal relationship edge list of the device. Classify and group the delay attributes according to the measurement point identification source. Calculate the distribution center measure within the same measurement point identification source group for the delay values ​​to obtain the characteristic delay representative quantity of each group. Encapsulate the characteristic delay representative quantity of each group and the corresponding measurement point identification source together as the delay distribution attribute record of the device node pair.

[0163] The source of measurement point identification is distinguished by the combination key formed by concatenating the causal measurement point variable identifier and the effect measurement point variable identifier. Classification and grouping divides all entries in the list into non-overlapping groups based on their measurement point identification sources. Each group corresponds to a specific measurement point-level causal pair. For the time delay values ​​within each group, the distribution center measure is calculated. If there is only one time delay record in the group, the center measure is that record itself. If there are multiple time delay records in the group, their arithmetic mean is taken as the characteristic time delay representative quantity. The characteristic time delay representative quantities of all groups and the measurement point identification sources are combined to form a time delay distribution attribute record. This record completely preserves the multi-source details of time delay transmitted from the measurement point layer to the equipment layer.

[0164] Step S534: Identify whether there are conflicting edges in the directional attribute of the causal relationship edge list within the device pair. If there are edges pointing from the first device node to the second device node and edges pointing from the second device node to the first device node, then converge the causal strength values ​​of the two directions respectively, and select the direction with the dominant convergence strength as the semantic relationship direction of the device node pair.

[0165] A conflicting edge refers to a situation where two directed causal strength records in opposite directions exist simultaneously within the same device node pair. For such cases, a comprehensive causal strength representation is calculated for all edges pointing from the first device to the second device, and a comprehensive causal strength representation can also be calculated for edges pointing from the second device to the first device. The values ​​of the comprehensive causal strength representations in the two directions are compared, and the direction with the larger value is considered to have a dominant causal driving effect at the device level. This direction is then determined as the direction in which the semantic relationship edge between the device node pairs points.

[0166] Step S535: Combine the comprehensive causal strength description, delay distribution attribute record and semantic relationship direction into an aggregated causal attribute set for the device node pair.

[0167] The aggregated causal attribute set is a data container that stores three pieces of information: causal strength, latency distribution, and causal direction. It is passed to the knowledge graph construction stage along with the semantic relationship edges of the device node pair.

[0168] Step S540: Compare the comprehensive causal strength description with the preset strength filtering standard, remove the device node pairs whose comprehensive causal strength description does not meet the preset strength filtering standard, and retain the device node pairs that meet the strength requirements as semantic relation edge carrying objects.

[0169] The purpose of the preset strength filtering standard is to perform a further saliency screening of causal discovery results at the device semantic level. This eliminates weak connections that still lack sufficient strength support after aggregation at the causal point level, ensuring that the semantic relationship edges in the final knowledge graph have statistical reliability and physical interpretability. The standard is set as follows: collect the comprehensive causal strength representation values ​​of all device node pairs, sort all values ​​in ascending order, calculate the difference between the upper and lower quartiles as the interquartile range, and add a preset multiple of the interquartile range to the upper quartile as the strength filtering threshold. If the comprehensive causal strength representation of a device node pair is lower than this threshold, the device node pair is filtered out; otherwise, it is retained as a semantic relationship edge carrier and allowed to enter the formal creation process of semantic relationship edges.

[0170] Step S550: In the static structural constraint model, find the mechanical connection type description between the device nodes at both ends of the semantic relation edge carrying object, and designate the found transmission relation type identifier or physical adjacency relation type identifier as the relation type label of the semantic relation edge.

[0171] The device connection relationships stored in the static structural constraint model are physical facts verified by engineering. Applying them as structural constraints to the causal relationship edges discovered through data-driven methods can assign precise technological semantics to these semantic relationship edges. During a search, the source device identifier and target device identifier of the object carried by the semantic relationship edge are used as search keys to perform an exact match in the connection relationship table of the static structural constraint model. If a record is matched, the corresponding transmission relationship type identifier or physical adjacency relationship type identifier is extracted from the record and designated as the relationship type label for the semantic relationship edge. If a match is not found because the static structural constraint model does not record a direct connection between the device pair, the relationship type label is marked as an undefined type.

[0172] As one implementation method, step S550 specifically includes the following steps S551 to S556: Step S551: Obtain the source device node identifier and target device node identifier of the semantic relation edge carrying object, and form a device node identifier pair.

[0173] The source device node identifier is a unique identifier of the originating device specified by the semantic relationship direction in the aggregated causal attribute set, and the target device node identifier is a unique identifier of the pointing end device specified by the semantic relationship direction. The two constitute an ordered pair of device node identifiers.

[0174] Step S552: Match and retrieve the device node identifiers in the topology connection table of the static structural constraint model to locate the topology record describing the connection method between the source device node and the target device node.

[0175] The topology connection table is an index structure in the static structural constraint model specifically for storing connection information between devices. The matching retrieval process is as follows: the source device node identifier is used as the row index key and the target device node identifier is used as the column index key to search the topology connection table; if a record is located, it is returned directly; if the first search fails, the roles of the source device node identifier and the target device node identifier are swapped and the search is repeated to cover cases where the connection relationships may be stored in any order in the table.

[0176] Step S553: ​​Parse the connection type field from the topology record. The connection type field contains an enumerated category code used to distinguish between transit connections and physical adjacency connections.

[0177] The connection type field is a structured field with a fixed bit allocation. Its high-order bits store the enumeration type category code, and its low-order bits store the sub-category code. During parsing, the high-order enumeration type category code value is extracted through bitmasking operations. This code value has a predefined mapping relationship with the transmission connection category and the physical adjacency connection category in the model metadata.

[0178] Step S554: If the connection type field indicates a transmission connection, extract the transmission connection sub-class code. The transmission connection sub-class code is used to distinguish between gear transmission, belt transmission, or chain transmission. The transmission connection sub-class code is then bound to the transmission relationship type identifier as a relationship type label.

[0179] After the enumerated category code is parsed into the major category of transmission connection, the minor category code of the transmission connection is extracted from the lower part of the connection type field. The mapping relationship between the value set of the minor category code of the transmission connection and the specific transmission method is predefined in the model semantic dictionary. The transmission relationship type identifier is used as a prefix and the minor category code of the transmission connection is used as a suffix to combine into a complete relationship type label string, which is assigned to the label attribute of the semantic relationship edge.

[0180] Step S555: If the connection type field indicates physical adjacency connection, extract the physical adjacency connection subclass code. The physical adjacency connection subclass code is used to distinguish between fixed adjacency, sliding adjacency, or adjacent non-contact mode, and bind the physical adjacency connection subclass code with the physical adjacency relationship type identifier as a relationship type label.

[0181] Similarly to step S554, after the enumerated category code is parsed into the physical adjacency connection category, the physical adjacency connection sub-category code is extracted from the low part of the connection type field, its meaning is parsed according to the preset dictionary, and the physical adjacency relationship type identifier and the physical adjacency connection sub-category code are combined into a relationship type label.

[0182] Step S556: Write the relation type label into the label attribute of the semantic relation edge to complete the type marking of the semantic relation edge.

[0183] The label attribute of a semantic relation edge is a text field. After the concatenated relation type label string is directly written into this field, the semantic relation edge carries both data-driven causal strength information and structure-driven connection type semantics. The combination of the two makes the knowledge graph's expression of production line process relationships more complete.

[0184] Step S560: Using production line equipment nodes as vertices and semantic relationship edges with assigned relation type labels as connecting edges, construct a graph structure data containing a set of equipment nodes and a set of semantic relationship edges, and output the graph structure data as a production line knowledge graph.

[0185] The production line equipment nodes directly utilize the set of production line equipment nodes extracted from the static structural constraint model in step S510. Each node retains attributes such as equipment identifier, equipment type, and equipment tag number. Each edge in the semantic relationship edge set carries the source equipment node index, target equipment node index, relationship type label, comprehensive causal strength description, and time delay distribution attribute record. The final constructed graph structure data is organized in the form of a node list and an edge list. Each entry in the node list corresponds to the complete attribute record of a piece of equipment, and each entry in the edge list corresponds to the causal semantic relationship between a pair of equipment and its type label. This graph structure data constitutes the production line knowledge graph, which integrates the prior constraints of the production line's mechanical structure with the causal driving relationships discovered from time series data through symbolic aggregation and transfer entropy analysis. It can provide knowledge support that is both data-driven and physically interpretable for production line fault tracing, process fluctuation propagation path analysis, and equipment health management.

[0186] This invention also provides a server, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the production line knowledge graph generation method based on the Industrial Internet provided in this invention.

[0187] Please see details. Figure 4 This is a schematic diagram of the structure of a server provided in an embodiment of the present invention. Figure 4As shown, the server 1000 may include a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the server 1000 may also include a user interface 1003 and at least one communication bus 1002. The communication bus 1002 is used to implement communication between these components. The user interface 1003 may include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the processor 1001. Figure 4 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0188] exist Figure 4 In the server 1000 shown, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface; and the processor 1001 can be used to call the device control application stored in the memory 1005 to implement the methods provided in the above embodiments.

[0189] It should be understood that the server 1000 described in this embodiment of the invention can execute the foregoing text. Figure 2 The implementation principle and beneficial effects of the method for generating production line knowledge graphs based on the Industrial Internet described in the corresponding embodiments will not be elaborated here.

Claims

1. A method for generating a production line knowledge graph based on an industrial internet, characterized in that, The method includes: Obtain a continuous time series data set consisting of multiple measurement point variables from the same data source on the production line; For each of the measured point variables, an adaptive segmented symbolic aggregation based on trend inflection points is performed on the continuous time series data set. By extracting key locations in the time series where the trend direction changes, adaptive segmentation is performed, and each segment is mapped to a preset symbol set to obtain the corresponding discrete symbol sequence. For any two measurement point variables, perform multi-delay transfer entropy calculation to obtain the directed causal intensity value that characterizes the direction and driving intensity of information flow between the two measurement point variables. Traverse all measurement point variable pairs and summarize all directed causal intensity values ​​to obtain the directed causal intensity set. Extract measurement point variable pairs from the directed causal strength set whose strength values ​​satisfy the preset causal significance condition and whose direction is clear from cause to effect. For each extracted measurement point variable pair, instantiate measurement point nodes and device nodes of the devices to which the measurement point variables belong. Establish causal relationship edges from the cause measurement point node to the effect measurement point node. Based on the transmission entropy time delay analysis, assign measurement point-level action delay attributes to each causal relationship edge to form a set of causal relationship edges at the measurement point node level. A static structural constraint model is pre-constructed based on the fixed mechanical transmission topology and physical adjacency relationships between production line equipment. The causal relationship edge set at the measurement point node level is semantically reduced at the equipment level using the static structural constraint model. Causal relationship edges at the measurement point node level belonging to the same equipment pair are aggregated and deduced into semantic relationship edges between equipment nodes. Relationship type labels are assigned to each semantic relationship edge according to the mechanical connection type of the equipment at both ends of the edge. The production line knowledge graph is generated based on the equipment nodes and the set of semantic relationship edges with assigned relationship type labels.

2. The method according to claim 1, characterized in that, The process involves performing adaptive segmented symbolic aggregation based on trend inflection points on the continuous time series data set for each measured point variable. This is achieved by extracting key locations in the time series where the trend direction changes, adaptively segmenting the data, and mapping each segment to a preset symbol set to obtain the corresponding discrete symbol sequence. This includes: Perform sequential adjacent point slope trend analysis on the continuous time series data set of the measured point variables, generate local direction vector markers between each adjacent point, and detect the data sampling positions where the local direction vector markers are flipped. Record the data sampling positions as the initial selection turning points to form the initial selection turning point sequence. For each initial inflection point in the initial inflection point sequence, a neighborhood box with a fixed time span centered on that point is defined. The directional attributes of the initial inflection points contained in each neighborhood box are extracted and the concentrated expression of the directional attributes within the box is analyzed. The center points of the neighborhood boxes whose concentrated expression of directional attributes meets the preset stable aggregation conditions are retained, and the remaining scattered points are removed to generate an aggregation-strengthening inflection point sequence. Calculate the time interval between adjacent inflection points in the clustered strengthening inflection point sequence, compare the time interval with the preset shortest duration constraint of trend segments, and for short-term trend segments with time intervals lower than the shortest duration constraint, merge and resolve the inflection points at both ends, and reorganize the resolved inflection points into a key inflection point sequence. Using each key inflection point in the key inflection point sequence as a dividing boundary, the continuous time series data set is truncated with boundary alignment to obtain multiple non-overlapping subsequence segments arranged in chronological order. The central tendency measure and dispersion measure of the internal numerical distribution of each subsequence segment are extracted to obtain the statistical description pair of the subsequence segment. The statistical description pair is spatially compared with the symbol reference representation of each prototype symbol in the preset symbol dictionary. The offset distance of the statistical description pair relative to each symbol reference representation is calculated. The symbol character corresponding to the prototype symbol with the smallest offset distance is selected as the replacement symbol of the subsequence segment. The replacement symbols of all subsequence segments are concatenated in chronological order to generate the discrete symbol sequence of the measurement point variable. The discrete symbol sequences generated by all measurement point variables in the production line are summarized, and cataloged and arranged using the measurement point variable identifier as an index to obtain a discrete symbol sequence concatenation table.

3. The method according to claim 2, characterized in that, The process involves defining a fixed-time neighborhood bounding box centered on each initially selected turning point in the initial turning point sequence, extracting the directional attributes of the initially selected turning points contained within each neighborhood bounding box, analyzing the concentrated expression of the directional attributes within the box, retaining the center points of neighborhood bounding boxes whose concentrated expression of directional attributes meets preset stable clustering conditions, removing the remaining scattered points, and generating a clustering-enhancing turning point sequence, including: For the initial selection of turning point sequence, a sliding processing window is established in the time dimension and the window length is set. The sliding processing window is moved along the time axis in single step. After each movement, the initial selection of turning point subset contained in the window and the directional attribute distribution of each subset member are extracted. For each window position, the distribution of directional attributes within the initially selected inflection point subset is statistically analyzed, and the consistency representation of directional attributes is calculated. The consistency representation reflects the degree of aggregation of the directional directions of the inflection points within the window. The consistency representation quantity is compared with a preset consistency threshold. The window position where the consistency representation quantity exceeds the preset consistency threshold is marked as an effective aggregation window. The initial inflection point corresponding to the center time of the effective aggregation window is determined as the candidate aggregation center point. Overlap analysis is performed on adjacent effective clustering windows on the time axis. If the coverage areas of adjacent windows overlap and the time interval between their respective candidate clustering centers is less than the preset deduplication interval, only the candidate clustering centers with larger consistency characterization values ​​are retained, and the remaining candidate clustering centers are filtered out to obtain a set of clustering centers without redundancy. Using each cluster center point in the set of non-redundant cluster center points as a reference, the directional attributes of all initially selected turning points within the corresponding effective cluster window are referenced back. Based on the time boundary of the initially selected turning points with consistent directional attributes, the precise representative time coordinates of the cluster center point are repositioned to obtain the time-corrected cluster center point. All time-corrected cluster centers are arranged in chronological order and grouped into the cluster strengthening inflection point sequence.

4. The method according to claim 2, characterized in that, The step involves spatially comparing the statistical description pair with the symbolic reference representations of each prototype symbol in a preset symbol dictionary, calculating the offset distance of the statistical description pair relative to each symbolic reference representation, selecting the symbol character corresponding to the prototype symbol with the smallest offset distance as the replacement symbol for that subsequence segment, and concatenating the replacement symbols of all subsequence segments in chronological order to generate the discrete symbol sequence of the measured point variable, including: The central tendency measure and the dispersion measure in the statistical description pair are mapped across measurement points in a unified dimension. Based on the global distribution benchmark of all measurement point variable sequences, the central tendency measure and the dispersion measure are aligned in position and scaled in magnitude to obtain the standardized description coordinates of the subsequence segment. Read the preset symbol normalization coordinates of each prototype symbol from the preset symbol dictionary, calculate the vector end distance between the normalization description coordinates and the normalization coordinates of each symbol, and obtain the initial separation degree sorting of the subsequence segment and each prototype symbol. Identify the pre-defined substitute symbol in the time sequence of the current subsequence segment to form the pre-order symbol background, and extract the set of legal subsequent symbols allowed to appear under the pre-order symbol background according to the preset symbol conversion protocol table. By applying a transformation constraint to the initial separation degree ranking using the set of legal subsequent symbols, the separation degree values ​​corresponding to prototype symbols that do not belong to the set of legal subsequent symbols are expanded to a preset penalty value, resulting in a modified separation degree ranking that takes into account context constraints. In the corrected separation sorting, the prototype symbol with the smallest separation is selected, and the symbol character bound to the prototype symbol is extracted as the final replacement symbol of the subsequence segment. The final substitution symbol is appended to the end of the generated symbol sequence. After all subsequence segments have been processed, the complete symbol sequence is output as the discrete symbol sequence of the measurement point variable.

5. The method according to claim 1, characterized in that, For any two measured variable discrete symbol sequences, multi-delay propagation entropy calculation is performed to obtain directed causal strength values ​​characterizing the direction and driving intensity of information flow between the two measured variable pairs. All measured variable pairs are traversed, and all directed causal strength values ​​are summarized to obtain a set of directed causal strength values, including: Read the first discrete symbol sequence and the second discrete symbol sequence of the target measurement point variable pair from the discrete symbol sequence concatenation list, and count the frequency of co-occurrence of each symbol pair in the first discrete symbol sequence and the second discrete symbol sequence without applying time delay, and generate a joint symbol occurrence distribution structure; Candidate delay values ​​are set sequentially. The first discrete symbol sequence is shifted along the time axis according to each candidate delay value. The frequency of co-occurrence of symbol pairs is recounted between the shifted first discrete symbol sequence and the unshifted second discrete symbol sequence to generate the joint delay distribution structure under the candidate delay. For each candidate delay, compare the distribution difference between its joint delay distribution structure and the joint symbol occurrence distribution structure under the condition of no delay, extract the change in conditional information caused by time shift, and use the change in conditional information as the dependency strength measure corresponding to the candidate delay. Detect the latency point where the dependency strength metric reaches its peak from all candidate latency points, and record the latency value and peak dependency strength metric value corresponding to that latency point. The interaction direction between the measurement point variable pairs is determined based on the time delay value. The peak dependence strength metric is used as the directional causal strength value from the cause measurement point variable to the effect measurement point variable. At the same time, the time delay value is saved as the optimal interaction time delay for the measurement point variable pair. By iterating through all measurement point variable pairs on the production line, summarizing the directed causal strength values ​​and optimal action delays for each measurement point variable pair, the set of directed causal strengths is obtained.

6. The method according to claim 5, characterized in that, The step of calculating the frequency of co-occurrence of each symbol pair in the first and second discrete symbol sequences under conditions without applying time delay, and generating a joint symbol occurrence distribution structure, includes: Obtain the row and column dimensions of the preset symbol frequency statistics container, where the rows correspond to all possible symbols of the first discrete symbol sequence and the columns correspond to all possible symbols of the second discrete symbol sequence. Initialize the count value of each row and column intersection unit in the symbol frequency statistics container. The first discrete symbol sequence and the second discrete symbol sequence are paired one by one according to the same time position to generate a symbol pairing sequence. Each pairing element in the symbol pairing sequence consists of the first discrete symbol sequence symbol and the second discrete symbol sequence symbol at the same time position. Traverse the symbol pairing sequence. For each pairing element read, locate the corresponding row and column intersection unit in the symbol frequency statistics container and increment the cumulative count value of that unit until all pairing elements have been processed to obtain the original frequency statistics table. The total frequency of each unit in the original frequency statistics table is calculated, and the cumulative count of each unit is converted into its proportion of the total number of paired elements to generate a joint occurrence probability distribution table. Probability baseline compensation is performed on the units with a probability of zero in the joint occurrence probability distribution table. A preset small probability value is assigned to the units with a probability of zero, and the non-zero probability units are adjusted to keep the overall probability sum unchanged, so as to obtain the modified joint probability distribution table. The modified joint probability distribution table is structured into the joint symbol occurrence distribution structure, which contains complete probability information for the occurrence of each symbol pair.

7. The method according to claim 5, characterized in that, The step of detecting the delay point where the dependency strength metric reaches its peak from all candidate delays, and recording the delay value and the peak dependency strength metric value corresponding to that delay point, includes: The candidate delay value sequence is paired with the corresponding dependency strength metric sequence to obtain the delay-dependency strength metric correspondence sequence, and the corresponding sequence is arranged in ascending order of candidate delay value; On the sorted time delay-dependency strength metric correspondence sequence, traverse each candidate time delay point, compare its dependency strength metric value with the dependency strength metric value of the adjacent time delay points, find the local peak point that is larger than the dependency strength metric value of the adjacent neighbor, and mark it as the candidate peak point. Among the candidate peak points, the degree of distinction between the dependency strength metric value of each candidate peak point and the average level of the entire dependency strength metric sequence is further evaluated. Candidate peak points with insufficient distinction are filtered out, and significant candidate peak points are retained. If there are multiple significant candidate peaks, compare the magnitude of the dependency strength metric of each significant candidate peak and select the significant candidate peak with the largest dependency strength metric as the first peak. Extract the candidate delay value corresponding to the first peak point as the representative delay value, and extract the dependency strength metric value corresponding to the first peak point as the peak dependency strength metric value. The representative time delay value and the peak dependence strength metric value are associated and stored, and output as the causal strength and effect time delay information of the target measurement point variable pair.

8. The method according to claim 1, characterized in that, The process involves extracting measurement point variable pairs from the directed causal strength set whose strength values ​​satisfy a preset causal significance condition and whose direction is clearly from cause to effect. For each extracted measurement point variable pair, a measurement point node and the device node of the device to which the measurement point variable belongs are instantiated. A causal relationship edge is established from the cause measurement point node to the effect measurement point node. Based on the propagation entropy time delay analysis, a measurement point-level action delay attribute is assigned to each causal relationship edge, forming a set of causal relationship edges at the measurement point node level, including: For each directed causal strength value in the set of directed causal strengths, compare it with a preset causal significance judgment benchmark, filter out measurement point variable pairs whose strength values ​​are lower than the preset causal significance judgment benchmark, and retain causal relationship candidate pairs that meet the significance requirements; The directional attributes of the candidate causal relationship pairs are verified. If the directional causal strength value of the cause-effect pair is higher than the directional causal strength value of the reverse pair and exceeds the preset directional difference tolerance, the pair of measurement point variables with a clear direction from cause to effect is retained to obtain the set of confirmed causal relationship pairs. For each confirmed causal relationship pair in the confirmed causal relationship pair set, create a cause measurement point node instance representing the cause measurement point variable and an effect measurement point node instance representing the effect measurement point variable in the data structure, and establish a pointing relationship record for the cause measurement point node instance and the effect measurement point node instance. Obtain the equipment identifier of each measurement point variable from the production line equipment configuration description, associate the cause measurement point node instance with the first equipment node instance to which the cause measurement point variable belongs, and associate the result measurement point node instance with the second equipment node instance to which the result measurement point variable belongs, thus completing the attribution link from the measurement point layer to the equipment layer. Read the optimal action delay recorded during the transfer entropy analysis of the confirmed causal relationship pair, and directly attach the optimal action delay as the measurement point level action delay attribute of the causal relationship edge to the causal relationship edge pointing from the cause measurement point node to the effect measurement point node. Mark the measurement point identifier source corresponding to the delay value of each causal relationship edge to indicate that the delay value belongs to a specific measurement point layer causal pair. All established causal relationship edges, their measurement point-level effect delay attributes, and measurement point identifier sources are summarized and grouped according to the indexes of cause measurement point nodes and effect measurement point nodes to generate the causal relationship edge set at the measurement point node level.

9. The method according to claim 8, characterized in that, The directional attributes of the candidate causal relationship pairs are verified. If the directional causal strength value of the cause-effect pair is higher than the reverse directional causal strength value and exceeds a preset directional difference tolerance, then the pair of measure point variables with a clear directional direction from cause to effect is retained, resulting in a confirmed causal relationship pair set, including: For each candidate causal relationship, the positive causal strength value pointing from the first measurement point variable to the second measurement point variable and the negative causal strength value pointing from the second measurement point variable to the first measurement point variable are extracted from the set of directed causal strengths. Calculate the difference between the positive directional causality strength value and the negative directional causality strength value, and calculate the relative proportion of the difference to the positive directional causality strength value to obtain a directional dominance measure. The directional advantage metric is compared with a preset directional advantage criterion. If the directional advantage metric exceeds the preset directional advantage criterion, the direction of the causal candidate pair is determined to meet the directional definiteness requirement. For candidate causal relationships that meet the requirement of directional clarity, the direction from which the stronger causal force points to the weaker causal force is determined as the causal direction based on the relationship between the magnitudes of the positive and negative causal strength values. The positive directional causal strength value, negative directional causal strength value, and causal direction of the candidate causal relationship with a definite causal direction are encapsulated into a direction confirmation record, and all measurement point variable pairs in the direction confirmation record are summarized to form the confirmed causal relationship pair set.

10. A server, characterized in that, include: processor; And a memory, wherein the memory stores computer-readable code that, when executed by the processor, causes the processor to perform the method as described in any one of claims 1 to 9.

Images