Ferric sulfate knowledge graph construction method based on multi-modal heterogeneous data

Abstract

The application relates to the technical field of knowledge graphs, in particular to a ferrous sulfate knowledge graph construction method based on multi-modal heterogeneous data, which comprises the following steps: acquiring multi-modal heterogeneous data of a ferrous sulfate production process; taking mechanical operation rhythms obtained by analyzing multi-axis high-frequency vibration signal data as a main time reference to form a common time reference; aligning and correlating tomographic imaging data and acoustic emission data based on the common time reference to obtain instantiated process paradigm entities in a knowledge graph; analyzing time series of the process paradigm entities according to predefined ontology relationship types, establishing a directed relationship with a state transition probability attribute, and forming a probabilistic state machine data structure that can be traversed by a graph query language in the knowledge graph; and defining a filling structured attribute for the process paradigm entities in the knowledge graph. The application realizes the construction of a ferrous sulfate knowledge graph by fusing multi-modal data and state analysis.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically a method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data. Background Technology

[0002] Ferric sulfate, as an important inorganic chemical product, has wide applications in water treatment, metallurgical additives, and pigment manufacturing. Its industrial production process typically involves multiple continuous or intermittent process steps, with complex and variable equipment operating conditions and material changes. To ensure product quality and production safety, various types of industrial monitoring and data acquisition systems are commonly deployed on production sites to record various process data reflecting process conditions and equipment status in real time.

[0003] In the ferric sulfate production process, these process data come from diverse sources and have varying structures, often containing records with different sampling frequencies, time bases, format specifications, and acquisition conditions, forming a typical multimodal heterogeneous dataset. Traditional production monitoring systems mostly adopt a distributed storage and independent analysis approach, making it difficult to achieve comprehensive analysis, information retrieval, and intelligent reasoning across data sources.

[0004] Knowledge graphs, as structured semantic models capable of representing entities, attributes, and their relationships, have been widely applied in information retrieval, semantic querying, intelligent reasoning, and decision support. Combining multimodal heterogeneous data from the ferric sulfate production process with an ontology model and mapping it to entity nodes and relational connections in a knowledge graph helps achieve semantic organization, traceable management, and association analysis of process data. However, industrial process data generally exhibits significant temporal characteristics and state change patterns. How to effectively integrate this temporal information and probabilistic features in knowledge graph construction while maintaining accurate correlation with the original collected data remains a pressing technological challenge.

[0005] To address this, a method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data. By integrating multimodal data and state analysis, a ferric sulfate knowledge graph is constructed. The method involves acquiring multimodal heterogeneous data of the ferric sulfate production process; using the mechanical operating rhythm obtained from the analysis of multi-axis high-frequency vibration signal data as the master time reference to form a common time reference; aligning and associating tomographic imaging data and acoustic emission data based on the common time reference to obtain instantiated process paradigm entities in the knowledge graph; analyzing the time series of the process paradigm entities according to predefined ontology relation types to establish directed relations with state transition probability attributes, forming a probabilistic state machine data structure for graph query language traversal in the knowledge graph; and defining and filling structured attributes for the process paradigm entities in the knowledge graph.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data includes:

[0009] Acquire multimodal heterogeneous data of the ferric sulfate production process, including electrical resistivity tomography data, acoustic emission signal data, and multiaxial high-frequency vibration signal data;

[0010] The mechanical running rhythm obtained by parsing multi-axis high-frequency vibration signal data is used as the main time reference. The main time reference is periodically calibrated using timestamps to form a common time reference. Based on the common time reference, the resistive tomography data and acoustic emission signal data are aligned and correlated to obtain the instantiated process paradigm entity in the knowledge graph.

[0011] Based on predefined ontology relation types, the time series of process paradigm entities are analyzed, directed relations with state transition probability attributes are established between process paradigm entities, and a probabilistic state machine data structure for graph query language traversal is formed in the knowledge graph; for process paradigm entities in the knowledge graph, structured attributes are defined and populated to complete the construction of the knowledge graph. The data structure of the structured attributes includes type identifiers and parameter fields for storing probability distribution models.

[0012] Preferably, the step of acquiring multimodal heterogeneous data of the ferric sulfate production process includes: acquiring resistivity tomography data, acoustic emission signal data, and multi-axis high-frequency vibration signal data from sensors installed on the production equipment; during data acquisition, attaching a unified format timestamp, sensor identifier, and process stage identifier to each data record, and correcting the timestamp using a time delay estimation algorithm; storing the multimodal heterogeneous data in the order of timestamps as structured records with modal tags, and retaining metadata about the data source and acquisition conditions in the records; and establishing an index table for each modal data in the storage structure.

[0013] Preferably, the process of forming a common time reference includes: analyzing the acquired multi-axis high-frequency vibration signals to generate time reference information describing the mechanical operating cycle length, phase position, and operating stability, and storing the time reference information as a common time reference in the metadata record; based on the common time reference, dividing the electrical resistance tomography data and acoustic emission signal data into window sets with a unified time index; calculating the cross-correlation coefficient of the electrical resistance tomography data and acoustic emission signal data within the corresponding time window, and confirming that the windows of the electrical resistance tomography data and acoustic emission signal data are aligned when the correlation coefficient exceeds a preset threshold; converting the aligned electrical resistance tomography data and acoustic emission data into structured data objects, with each object having corresponding time reference information and data source identifier attached.

[0014] Preferably, the process of obtaining instantiated process paradigm entities in the knowledge graph includes: establishing a cross-modal association index for structured data objects under a unified time reference; adding credibility weights based on operational stability information to form fusion feature records arranged in chronological order; using the set of fusion feature records as input; applying a density-based noise application space clustering algorithm for processing; automatically identifying clusters representing different process states; based on the density reachability relationship of data points in the feature space; aggregating core data points that meet the minimum density requirement and reachable boundary points to form clusters; identifying discrete points that do not belong to any cluster as process noise; defining each identified cluster as a specific process paradigm entity; creating a unique entity node for the process paradigm entity in the knowledge graph; and using the original fusion feature records as instance data of the entity node.

[0015] Preferably, the step of analyzing the time series of process paradigm entities based on predefined ontology relation types includes:

[0016] The fused feature records are arranged in chronological order to form a process paradigm entity sequence with timestamp indexes. According to the process categories and attributes defined in the ontology, the entities in adjacent time slices are compared to identify state change events between entities. The state transition conditions are determined by combining the occurrence frequency and duration of the state transition events, and a unique identifier is generated for each state transition condition. The state transition events are bound to the corresponding condition identifiers to form state transition records with semantic annotations, which are stored in the intermediate analysis result set.

[0017] Preferably, the step of forming a probabilistic state machine data structure for graph query language traversal in the knowledge graph includes:

[0018] A set of state nodes is constructed based on the intermediate analysis result set. Each state node corresponds to a process paradigm entity and carries the entity's type information and timestamp range. The transition probabilities between state nodes are calculated, and the probability values ​​are written into the graph database as attributes of directed relation edges. Context labels and event history summaries related to the nodes are added to the attribute table of the state nodes. The state nodes and directed relation edges are organized in the knowledge graph according to the structure of a probabilistic state machine, so that traversal operations based on graph query language can perform inference and prediction according to the state transition probability path.

[0019] Preferably, the step of defining structured attributes for each process paradigm entity in the knowledge graph includes:

[0020] A property template is specified for the process paradigm entity. The property template includes a set of fields for describing the entity's physical characteristics, process context information, and statistical model characteristics. The data structure of the fields is fixed as a type identifier of a parameterized probability distribution model and the corresponding parameter value. A probability distribution model type identifier field is set in the property template to indicate the distribution type used by the entity in statistical modeling. A set of parameter fields is set in the property template to store the key numerical parameters of the distribution model, and unit information and value range constraints are added to each field. A metadata field is set in the property template to store the generation time, data source identifier, and references to relationships with other entities.

[0021] Preferably, the step of filling in the structured attributes includes: fitting the parameters of the probability distribution model using statistical analysis based on the fused feature samples corresponding to the process paradigm entity within a historical time period; filling the parameter values ​​obtained from the fitting into the parameter fields of the attribute template, and synchronously updating the value of the probability distribution model type identifier field; storing the parameter fields and type identifier fields together in the node attribute table of the graph database, which can be directly called by the graph query language and can be used for inference calculations; after the attribute filling is completed, generating a verification record for the entity, including the filling timestamp, a summary of the parameter calculation method, and the data integrity check result.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. By extracting mechanical operating rhythm parameters from multi-axis high-frequency vibration signal data and using them as the core reference for time synchronization and feature fusion, precise alignment and contextual association between resistive tomography data and acoustic emission data are achieved, effectively improving the consistency and reliability of multimodal data fusion, thereby significantly improving the accuracy of production process status identification.

[0024] 2. Instantiate the fusion features into process paradigm entities, and establish directed relations with state transition probability attributes based on time series analysis results, so that the knowledge graph has probabilistic state machine characteristics, which can directly support state prediction, path deduction and abnormal trend analysis based on graph query language, and improve the real-time performance and intelligence level of process monitoring and prediction.

[0025] 3. Define and populate the probability distribution model type and parameter fields for each process paradigm entity node in the knowledge graph, so that probability calculation and state determination can be completed directly within the graph database, reducing the latency of external calculation calls. At the same time, it supports multi-condition semantic retrieval combined with node context tags, improving the speed of production monitoring and decision response. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the process for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data, as provided in an embodiment of the present invention.

[0027] Figure 2 A flowchart illustrating the process of forming a common time reference provided in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram illustrating the process of obtaining the instantiation paradigm entity in the knowledge graph, as provided in an embodiment of the present invention. Detailed Implementation

[0029] 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.

[0030] Please see Figures 1 to 3 This invention provides a method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data. The technical solution is as follows:

[0031] A method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data, the specific process of which is as follows: Figure 1 As shown, it includes:

[0032] Acquire multimodal heterogeneous data of the ferric sulfate production process, including electrical resistivity tomography data, acoustic emission signal data, and multiaxial high-frequency vibration signal data;

[0033] The mechanical running rhythm obtained by parsing multi-axis high-frequency vibration signal data is used as the main time reference. The main time reference is periodically calibrated using timestamps to form a common time reference. Based on the common time reference, the resistive tomography data and acoustic emission signal data are aligned and correlated to obtain the instantiated process paradigm entity in the knowledge graph.

[0034] Based on predefined ontology relation types, the time series of process paradigm entities are analyzed, directed relations with state transition probability attributes are established between process paradigm entities, and a probabilistic state machine data structure for graph query language traversal is formed in the knowledge graph; for process paradigm entities in the knowledge graph, structured attributes are defined and populated to complete the construction of the knowledge graph. The data structure of the structured attributes includes type identifiers and parameter fields for storing probability distribution models.

[0035] Furthermore, the step of acquiring multimodal heterogeneous data in the ferric sulfate production process includes: acquiring resistive tomography data, acoustic emission signal data, and multi-axis high-frequency vibration signal data from sensors installed on the production equipment; during data acquisition, attaching a unified format timestamp, sensor identifier, and process stage identifier to each data record, and correcting the timestamp using a time delay estimation algorithm; storing the multimodal heterogeneous data in a structured record with modal tags according to the timestamp order, and retaining metadata about the data source and acquisition conditions in the record; and establishing an index table for each modal data in the storage structure.

[0036] Specifically, 16 electrodes are evenly distributed around the inner wall of the reactor to form an electrical resistance tomography (ERT) system, which is used to reconstruct the conductivity distribution image of the material inside the reactor in real time, thereby generating high frame rate ERT data. Piezoelectric acoustic emission sensors are installed on the outer wall of the reactor and other key locations such as material pipelines to capture acoustic emission signal data generated by chemical reactions or equipment friction. A triaxial high-frequency vibration sensor is installed on the motor base of the agitator to monitor the running status of the agitator and output multi-axis vibration signal data.

[0037] During data acquisition, each piece of raw data from different sensors is processed and normalized. A unified format timestamp, a unique sensor identifier indicating the physical source of the data (e.g., ERT-Array-01), and the current process stage identifier (e.g., Stage-Mixing) obtained synchronously from production control are added to each data record (e.g., an ERT frame or an AE data block). In addition, to address the inherent latency issues of different data links, this embodiment incorporates a latency estimation algorithm to correct the timestamps. This method pre-determines the end-to-end latency of each modal data through offline calibration and subtracts the correction value from the raw timestamps during real-time acquisition, ensuring that the time base of all modal data can be accurately aligned to the actual occurrence time of the event.

[0038] The specific implementation of the time delay estimation algorithm is based on offline calibration. A stream of high-conductivity brine, clearly perceptible to all sensors, is instantaneously injected into the reactor through a fast valve, creating a physical synchronization pulse event. A high-precision master clock synchronized with a network time protocol server records the trigger time of this pulse event. In the data streams of various sensors such as resistivity tomography, acoustic emission, and vibration, the response time to this pulse event is accurately identified using a peak detection algorithm. The inherent end-to-end delay of each data link is calculated by subtracting the trigger time of the master clock from the response time of each data link. The calculated delays are used as fixed correction values ​​and are automatically deducted from the original timestamps corresponding to each mode in subsequent real-time acquisition, aligning the time base of all data to the actual occurrence time of the event.

[0039] After the above processing, all modal data are uniformly formatted into structured records with modal tags (such as ERT, AE). These records are written into a time series database according to the corrected timestamp order. Each record not only contains the core data payload (such as the conductivity matrix of ERT or the waveform array of AE), but also retains key information such as data source and acquisition conditions (such as sampling rate and gain) in independent metadata fields to ensure data integrity and traceability. For example, an AE data record will explicitly include its timestamp, sensor ID, process stage, waveform data, and related acquisition metadata.

[0040] In addition, this embodiment establishes indexes for key fields in the data storage structure, creates a primary index for the timestamp field to enable fast queries based on any time range, and creates auxiliary indexes for fields such as modal labels and sensor identifiers. These indexing mechanisms ensure that data from different sources and of different types can be efficiently filtered, aggregated, and associated, providing high-performance data access capabilities for subsequent multimodal data fusion and knowledge graph construction.

[0041] By combining high-precision timestamps with latency estimation algorithms, the precise temporal alignment of heterogeneous data is ensured, laying a solid foundation for subsequent causal analysis and feature fusion. In addition, standardized structured records and metadata retention guarantee the standardization and traceability of data, while an efficient indexing mechanism supports rapid retrieval of massive amounts of data, providing key performance guarantees for achieving near real-time process analysis and intelligent decision-making.

[0042] Furthermore, the process of forming a common time reference includes: analyzing the acquired multi-axis high-frequency vibration signals to generate time reference information describing the mechanical operating cycle length, phase position, and operating stability, and storing the time reference information as a common time reference in the metadata record; based on the common time reference, dividing the electrical resistance tomography data and acoustic emission signal data into window sets with unified time indexes; calculating the cross-correlation coefficient of the electrical resistance tomography data and acoustic emission signal data within the corresponding time windows, and confirming the alignment of the electrical resistance tomography data and acoustic emission signal data windows when the correlation coefficient exceeds a preset threshold; converting the aligned electrical resistance tomography data and acoustic emission data into structured data objects, with each object appended with corresponding time reference information and data source identifiers, specifically as follows: Figure 2 As shown.

[0043] Specifically, the first step is to analyze the acquired multi-axis high-frequency vibration signals to generate a global common time reference. To generate this reference, the acquired multi-axis high-frequency vibration signals are analyzed. A short-time Fourier transform is applied to the vertical vibration signals, and the frequency with the highest energy within each time window is identified as the main frequency of the agitator; its reciprocal is the cycle length. A Hilbert transform is applied to the narrow-band filtered signal of the main frequency to extract its instantaneous phase, thus obtaining the phase position of each sampling point. The operational stability index is defined as the coefficient of variation of the most recent one hundred consecutive cycle lengths; the smaller the coefficient, the more stable the mechanical operation. These analyzed indices collectively constitute the time base information and are stored as core metadata. The analysis module performs real-time spectral analysis on the vibration signals to calculate the agitator's rotational speed and determine the cycle length, for example, 500 milliseconds. Based on the cycle, continuous time is divided into discrete, unique mechanical operating cycles (e.g., identified by a cycle ID), and the phase position and operational stability index are calculated for each cycle. This time base information, containing cycle length, phase, and stability, is stored as core metadata.

[0044] Furthermore, using the common time reference, the continuous electrical resistance tomography (ERT) data stream and acoustic emission signal (AE) data are synchronously sliced. Each mechanical operation cycle (e.g., cycle ID_101) is treated as an independent window, and all ERT data frames and acoustic emission signal data signal sample points whose timestamps fall within this window are aggregated to form a one-to-one correspondence of ERT_Window_101 and AE_Window_101. In this way, the heterogeneous data streams that were originally acquired independently at different rates are effectively organized into a series of discrete data window sets with a unified time index (i.e., cycle ID), completing the mapping from the continuous time domain to discrete event cycles.

[0045] To verify the inherent correlation between different modal data within the same period window, alignment confirmation is performed by calculating the cross-correlation coefficient of the signals. For example, the image sequence within the ERT window is converted into a one-dimensional time series of average conductivity changes, and the normalized cross-correlation coefficient between it and the corresponding AE signal energy sequence within the window is calculated. When the calculated coefficient value exceeds a preset threshold, it is confirmed that the two data windows are strongly correlated, indicating that they jointly capture a significant event in the production process, thus completing the final alignment confirmation. To address the problem of poor adaptability of fixed thresholds under different operating conditions, the preset threshold further adopts an adaptive adjustment mechanism dynamically correlated with the operating stability index. This mechanism uses a base threshold as a benchmark and compares the operating stability index of the current period with the average stability index over a historical period. When the current operating state is detected to be more stable than the historical average level, the system automatically raises the judgment threshold; conversely, when the current operating state is detected to be more volatile, the judgment threshold is lowered accordingly. This dynamic adjustment ensures that the requirements for data alignment are more stringent when the equipment is running stably, while it can more flexibly capture effective transient correlation events during process fluctuations.

[0046] The preset threshold was determined through the following experimental method: two sets of sample data were prepared. One set consisted of positive sample data window pairs that were confirmed by manual verification and were strongly correlated under the same event. The other set consisted of negative sample data window pairs that were confirmed to be uncorrelated. The cross-correlation coefficient values ​​of all positive and negative sample pairs were calculated. Using different coefficient values ​​as thresholds, the corresponding true positive rate and false positive rate were calculated, and the receiver operating characteristic curve was plotted. The coefficient value corresponding to the inflection point closest to the upper left corner (0,1) on the ROC curve was selected as the optimal preset threshold that balances recall and precision. In this embodiment, the preset threshold is 0.75.

[0047] Finally, for each pair of data windows that have been confirmed to be aligned, they will be encapsulated into a unified, structured data object as the basic information unit for knowledge graph construction. This object uses the period ID as its unique identifier and contains the original ERT and AE data, as well as complete time reference information (period length, stability), identifiers of each data source, and the cross-correlation score calculated when confirming alignment.

[0048] By establishing mechanical vibration rhythms as a common time reference, a physically meaningful anchor point is provided for heterogeneous data fusion, fundamentally solving the alignment problem of asynchronous data streams. It segments continuous data into discrete event windows synchronized with the equipment's operating cycle, and ensures the intrinsic correlation of multimodal information within each window through cross-correlation analysis, effectively improving the reliability of feature fusion. The resulting structured data object not only contains aligned data but also encapsulates rich contextual information (such as periodicity and stability), greatly simplifying the input to subsequent state recognition models and significantly improving the accuracy and interpretability of model recognition results.

[0049] Furthermore, the process of obtaining the instantiation paradigm entity in the knowledge graph includes:

[0050] Under a unified time benchmark, a cross-modal association index is established for structured data objects. A credibility weight is added based on operational stability information to form a fusion feature record arranged chronologically. The set of these fusion feature records is used as input, and a density-based noise spatial clustering algorithm is applied to process them. Clusters representing different process states are automatically identified. Based on the density reachability relationship of data points in the feature space, core data points that meet the minimum density requirement and reachable boundary points are aggregated to form clusters. Simultaneously, discrete points not belonging to any cluster are identified as process noise. Each identified cluster is defined as a specific process paradigm entity. A unique entity node is created for each process paradigm entity in the knowledge graph, and the original fusion feature records are used as instance data for these entity nodes. Specifically, as follows... Figure 3 As shown.

[0051] Specifically, the structured data objects for each time period are processed to generate standardized fusion feature records for cluster analysis. A series of quantitative features are extracted from the ERT and AE data of each object, such as the average conductivity and uniformity of ERT, and the signal energy and kurtosis of AE. These values ​​are combined into a fixed-length fusion feature vector. Using the operational stability index of that period, a confidence weight is added to this feature record. Finally, a set of fusion feature records arranged in time sequence, containing multi-dimensional feature information and confidence weights, is obtained.

[0052] Furthermore, using the set of fused feature records as input, a density-based noise spatial clustering algorithm is applied to automatically identify data clusters representing different process states. After setting two key parameters—neighborhood radius and minimum number of neighbors—the DBSCAN algorithm operates based on the density reachability relationship of data points in the feature space. It aggregates core data points that meet the minimum density requirement and their reachable boundary points to form high-density regions. Discrete points that do not belong to any cluster are automatically identified by the algorithm and marked as process noise. After clustering, the identified clusters are defined as specific process paradigm entities, semantically representing a stable and recurring operating state in the ferric sulfate production process. For example, by analyzing the feature distribution of data within a cluster and combining it with process knowledge, one cluster may be named a stable mixed state, while another may be named a violent reaction state. For these entities that have been given clear semantics, a unique entity node is created for them in the knowledge graph, for example (e: ProcessEntity{name: "Stable Mixed State"}).

[0053] The minimum neighbor count is set based on process knowledge and is defined as the minimum number of consecutive cycles that a meaningful process state should contain. For example, if the process considers any stable state to last at least 5 seconds and the machine cycle is 500 milliseconds, then the minimum neighbor count is set to 5s divided by 0.5s, which equals 10. The neighborhood radius is determined heuristically using a k-distance graph. The k-value is set to the minimum neighbor value determined in the previous step. The distance from each point in the dataset to its k-th nearest neighbor is calculated. All data points are sorted from largest to smallest k-distance values ​​and plotted on a graph. The k-distance value at the inflection point of the graph (i.e., the point where the curve slope changes most drastically) is selected as the optimal neighborhood radius eps.

[0054] To compensate for the shortcomings of the master time reference method in potentially ignoring important asynchronous events during modeling, this invention further includes a secondary noise screening step. This step does not directly discard discrete data points identified as noise, but rather aggregates them and examines their original signal characteristics (especially the energy or peak amplitude of the acoustic emission signal). If multiple noise points appear consecutively within a short period, or if the signal energy of a single noise point exceeds a preset anomaly threshold, then that point or set of points is not considered meaningless noise, but is determined as an asynchronous critical event. For each determined asynchronous critical event, a special event type node (e.g., Asynchronous_Event) is created in the knowledge graph and associated with the original data instance at the time of occurrence. The preset anomaly threshold is objectively determined using a data-driven method. The signal energy values ​​corresponding to all data points successfully clustered into stable state clusters (i.e., non-noise points) are extracted to form a normal energy sample set; then, the mean and standard deviation of this sample set are calculated; finally, the anomaly threshold is determined as the mean plus three times the standard deviation.

[0055] Finally, the original fusion feature records are associated with the newly created process paradigm entity nodes to complete the instantiation process. For each fusion feature record classified into a certain cluster, a relationship (e.g., IS_INSTANCE_OF) is created in the knowledge graph from the record to its corresponding entity node. Through this concrete, timestamped data point, it is explicitly linked to the abstract, semantic process state concept, realizing the transformation from raw data to structured knowledge and filling the knowledge graph with rich and verifiable instance data.

[0056] Furthermore, to address the issue that density-based clustering algorithms are sensitive to input parameters and struggle to adapt to multi-density data clusters in a single operation, the application of a density-based noise-based spatial clustering algorithm further includes: before clustering, using the OPTICS algorithm to process and fuse the feature record set, generating an accessibility distance map that arranges the reachability distances of all data points in ascending order; defining a baseline line connecting the first and last data points on the accessibility distance map, and calculating the perpendicular distance from each data point to the baseline line; identifying the unique inflection point in the current map by determining the maximum distance value among all perpendicular distances; and inputting the reachability distance value corresponding to the inflection point as an adaptively determined neighborhood radius parameter into the subsequent density-based clustering algorithm, thereby achieving parameter self-optimization in the clustering process. This parameter self-optimization process solves the problem of manually setting clustering parameters, realizes automatic parameter optimization, and improves the accuracy of condition identification and the level of automated analysis.

[0057] DBSCAN clustering enables the unsupervised discovery of objectively existing and meaningful stable states (process paradigm entities) in the production process, and effectively identifies transition periods or abnormal noise, achieving data-driven operational condition definitions. These states are created as entity nodes in a knowledge graph and associated through instances, ensuring that each abstract state is supported by concrete and traceable data. This not only greatly enhances the reliability and interpretability of knowledge but also lays a solid foundation for subsequent construction of probabilistic models based on state transitions and for advanced semantic queries.

[0058] Furthermore, the step of analyzing the time series of process paradigm entities based on predefined ontology relation types includes:

[0059] The fused feature records are arranged in chronological order to form a process paradigm entity sequence with timestamp indexes. According to the process categories and attributes defined in the ontology, the entities in adjacent time slices are compared to identify state change events between entities. The state transition conditions are determined by combining the occurrence frequency and duration of the state transition events, and a unique identifier is generated for each state transition condition. The state transition events are bound to the corresponding condition identifiers to form state transition records with semantic annotations, which are stored in the intermediate analysis result set.

[0060] First, the fusion features associated with the process paradigm entities are recorded and strictly sorted according to their timestamps or period IDs, thus constructing a discrete sequence of process paradigm entities. This sequence with timestamp indexes intuitively reflects the evolution of the production process over time, for example, ... -> "Stable Mixed State" -> "Stable Mixed State" -> "Violent Reaction State" -> ... This sequence forms the basis for all subsequent time series analyses, transforming continuous process data into a clear and ordered state evolution path.

[0061] Furthermore, the entity sequence is traversed, and feature comparisons are performed on entities in adjacent time slices based on a predefined ontology. When a change in entity category is detected between two consecutive time points, such as a change from a stable mixed state to a violently reactive state, a state change event is identified. For each such event, the time point of occurrence, source state, and target state are recorded in detail, transforming the static description of the state into the capture of dynamic changes in the process.

[0062] Furthermore, statistical analysis is performed on all historical state change events of the same type to determine the state transition conditions that occurred, including calculating the frequency of such events and the average duration of the source state before each event. For example, analysis revealed that the transition from a stable mixed state to a violently reactive state typically occurs after the state has lasted for about 3 minutes. Based on these quantitative statistical results, specific transition conditions with physical or technological significance are defined, and a unique identifier, such as TC-SM-VR-01, is generated for each condition.

[0063] Finally, specific state change events are bound to abstract state transition conditions to form a state transition record with rich semantic annotations. This record not only contains detailed information about the event (time, source / target state) but also links to the condition identifiers it follows and can be appended with semantic annotations provided by process experts, such as "normal startup: enter reaction after mixing meets criteria." These structured records are uniformly stored in an intermediate analysis result set, providing a direct, reliable, and semantically rich data source for the next step of building a probabilistic state machine.

[0064] By analyzing entity sequences, the system accurately captures transition events between states, connecting static knowledge points into dynamic logical chains. Transition conditions are determined through statistical frequency and duration, transforming state changes from random events into predictable behaviors that follow specific patterns. The resulting semantically annotated transition records not only provide quantitative data for subsequent probabilistic model construction but, more importantly, endow the knowledge graph with causal reasoning and process prediction capabilities, enabling it to understand and predict the direction of production processes.

[0065] Furthermore, the step of forming a probabilistic state machine data structure for graph query language traversal in the knowledge graph includes:

[0066] A set of state nodes is constructed based on the intermediate analysis result set. Each state node corresponds to a process paradigm entity and carries the entity's type information and timestamp range. The transition probabilities between state nodes are calculated, and the probability values ​​are written into the graph database as attributes of directed relation edges. Context labels and event history summaries related to the nodes are added to the attribute table of the state nodes. The state nodes and directed relation edges are organized in the knowledge graph according to the structure of a probabilistic state machine, so that traversal operations based on graph query language can perform inference and prediction according to the state transition probability path.

[0067] First, based on the generated intermediate analysis result set, state nodes are constructed and enriched in the graph database. The MERGE operation of the graph database is used to create or match existing state nodes with the unique name of the process paradigm entity (such as a stable mixed state) as the key, ensuring that each process state corresponds to only one unique node in the graph. In addition, detailed attributes are filled in for each node: in addition to the entity type information, the earliest and latest timestamps of its appearance in historical data are calculated and recorded by scanning all instance data of the state, forming a clear timestamp range. This range is of great value for understanding the life cycle, activity level, and whether the state may have become a historical state, providing data support for subsequent process evolution analysis.

[0068] Next, the transition probabilities between state nodes are calculated, and weighted directed edges are created accordingly. This process follows the maximum likelihood estimation method. Specifically, the intermediate analysis result set is grouped by source state; for each group, the number of events that occur when transitioning from that source state to each different target state is counted; the total number of transitions and transition probabilities from that source state are calculated. For example, P(stable mixed state -> violently reactive state) = 0.977 is calculated. Subsequently, a directed edge of type TRANSFERS_TO is created in the knowledge graph, pointing from the source node to the target node, and the calculated probability value of 0.977 and the original count value of 125 are stored as the core attributes of this edge.

[0069] To enhance node interpretability and support more complex semantic queries, context labels and event history summaries are added to each state node. The context labels can be sourced from two sources: one part is automatically generated using preset rules, for example, if the average AE energy in a state's instance data exceeds a certain threshold, a "High_Energy" label is automatically added; the other part is manually annotated through a user interface provided to process experts, such as adding a "Key_Control_Phase" label. Simultaneously, by querying the TRANSFERS_TO relationships of all outgoing directions of the node, sorting them in descending probability, a text-formatted event history summary is automatically generated, such as "major transition to violent reaction states (97.7%)", and this is used as a node attribute. This makes each node a self-explanatory, information-rich knowledge unit.

[0070] Through the above steps, state nodes and directed edges with probability weights together constitute a complete probabilistic state machine in the knowledge graph, which can be directly traversed by the standard graph query language. For example, users can execute graph queries to calculate and return the most likely evolutionary path after N steps from any starting point by recursively multiplying the probability attribute values ​​of each edge on the path. In addition, this structure also supports anomaly detection (identifying transitions that actually occur but have extremely low probability) and process optimization (finding paths that can avoid low-yield or high-risk states), transforming the knowledge graph from a static data repository into a dynamic, computable production process prediction and decision support engine.

[0071] To address the issue of concept drift caused by factors such as equipment aging or changes in raw material batches—that is, the gradual invalidation of state transition probabilities over time—the process of calculating the transition probabilities between state nodes further includes: employing a time-sliding window mechanism to obtain the state transition frequencies of the old window representing the old process and the new window representing the current process; constructing a contingency table with the old and new windows as rows and all possible target transition states as columns; performing a chi-square homogeneity test based on the contingency table, and quantitatively determining whether there is a statistically significant difference in the transition probability distributions of the two windows by calculating the chi-square statistic and comparing it with the critical value at a preset significance level (α = 0.05 in this embodiment); when the test result is significant, a concept drift event is determined to have occurred, and the state transition probability values ​​in the knowledge graph are recalculated and updated using the data from the new window to maintain the timeliness and predictive accuracy of the model. By quantifying concept drift, the model can adapt to changes in the production line, ensuring that it always reflects the latest operating conditions, significantly improving the accuracy and timeliness of predictions.

[0072] By assigning precise probability weights to transition paths between states using the frequency of historical events, the dynamic behavior of the process is quantified. This makes it possible to directly predict paths, detect anomalies, and optimize processes using standard graph query languages, greatly enhancing the decision support value and intelligence level of knowledge graphs.

[0073] Furthermore, the step of defining structured attributes for each process paradigm entity in the knowledge graph includes:

[0074] A property template is specified for the process paradigm entity. The property template includes a set of fields for describing the entity's physical characteristics, process context information, and statistical model characteristics. The data structure of the fields is fixed as a type identifier of a parameterized probability distribution model and the corresponding parameter value. A probability distribution model type identifier field is set in the property template to indicate the distribution type used by the entity in statistical modeling. A set of parameter fields is set in the property template to store the key numerical parameters of the distribution model, and unit information and value range constraints are added to each field. A metadata field is set in the property template to store the generation time, data source identifier, and references to relationships with other entities.

[0075] Specifically, a unique, standardized attribute template is first assigned to each type of process paradigm entity (such as a stable hybrid state) in the knowledge graph. This template is designed as a structured schema definition, with the aim of elevating the entity description from simple text labels to a rich, computable dataset containing physical features, process context, and statistical model features. This templated approach ensures the uniformity of data structure for all entities of the same type, providing a foundation for subsequent automated processing and batch analysis.

[0076] In the core part of the attribute template, namely the statistical model feature field set, this embodiment innovatively fixes the data structure of the fields as a parameterized probability distribution model. The template explicitly sets a probability distribution model type identifier field, whose value is constrained to the name of a standard statistical distribution (such as Normal or Gamma), indicating the specific distribution type used for statistical modeling of entity features. This allows subsequent calculation modules to dynamically call the corresponding probability density function for calculation based on this identifier, ensuring the executability of the model.

[0077] Furthermore, the template includes a set of parameter fields corresponding to the distribution type, used to store the key numerical parameters of the distribution model. For example, when the type is Normal, the parameter set must include two fields: mean and standard deviation. To ensure the physical meaning and validity of the data, this embodiment adds explicit unit information and legal value range constraints to each parameter field. These constraints not only serve as the basis for validation during data population but also provide domain experts with an intuitive and understandable reference, preventing meaningless or erroneous parameter values ​​from being written.

[0078] Finally, to ensure the complete traceability of data sources and the relevance of knowledge, a dedicated metadata field is set in the attribute template. This field is used to store the specific timestamp of the calculation and filling of the attribute value (especially the model parameters), one or more identifiers pointing to the original fusion feature record set used to fit the model, and references to the association with other related entities (such as specific production equipment or raw material batches). This ensures that every parameter in the knowledge graph is traceable and can be effectively associated with and queried with knowledge from other fields.

[0079] By using parameterized probability distributions to describe entity characteristics, the inherent volatility and uncertainty of process states can be quantified, making it possible to conduct deeper quantitative anomaly detection and state assessment. The strict constraints on units and value ranges in the attribute templates, as well as the traceability of metadata of data sources, greatly ensure the physical meaning, quality, and reliability of knowledge. This structured and computable attribute definition makes each knowledge node a self-explanatory, verifiable, and directly computational intelligent unit, enhancing the depth and practical value of the entire knowledge graph.

[0080] Furthermore, the step of filling in the structured attributes includes:

[0081] Based on the fused feature samples corresponding to the process paradigm entity within a historical time period, statistical analysis is used to fit the parameters of the probability distribution model; the fitted parameter values ​​are filled into the parameter fields of the attribute template, and the value of the probability distribution model type identifier field is updated synchronously; the parameter fields and type identifier fields are stored together in the node attribute table of the graph database, which can be directly called by the graph query language and can be used for inference calculations; after the attribute filling is completed, a verification record is generated for the entity, including the filling timestamp, the parameter calculation method summary, and the data integrity check result.

[0082] A data sample for statistical analysis is collected by executing an explicit graph database query. This query starts with the unique ID of a stable mixed-state entity and traverses through all associated fusion feature record instances within a specified historical time period via the IS_INSTANCE_OF relation. Then, target physical features are extracted in batches from these records; in this embodiment, this is the electrical conductivity value, forming a sample set containing floating-point observations.

[0083] Next, rigorous statistical analysis and parameter fitting are performed on the collected sample set. The sample data will be fitted to a series of predefined, physically possible probability distribution models, including normal distribution, gamma distribution, and log-normal distribution. For each fit, the Akaike information criterion value is calculated. The smaller the Akaike information criterion value, the better the model is in describing the data. In this embodiment, the Akaike information criterion value of the normal distribution is the smallest. Therefore, Normal is deterministically selected as the most suitable probability distribution type to describe the conductivity fluctuation in this state. After determining the distribution type, the maximum likelihood estimation method is used to calculate the specific parameters of the normal distribution model, namely the sample mean and sample standard deviation.

[0084] After obtaining the model type and precise parameter values, they are atomically populated into the attribute template of the stable hybrid entity node and persistently stored in the graph database. An update transaction of the graph database is executed to update the value of the node's `statistical_model_features.distribution_type` field to the string "Normal", and update the mean and standard deviation values ​​in its `parameters` field to the calculated numerical values. In addition, these parameters are stored as native numeric types, rather than text, which ensures that they can be directly compared, arithmetically operated on, etc. by the graph query language in the WHERE or RETURN clauses, providing high-performance underlying support for subsequent online inference computation.

[0085] Finally, within the same transaction for attribute population, to ensure the quality and complete traceability of the entire process, a detailed verification record is generated and attached to the entity node. This record is a structured object containing a timestamp of the population operation. The record also includes a summary of the parameter calculation method, explicitly stating that the model was selected from multiple candidate distributions using the Akaike information criterion, and that the parameters were obtained based on 12,500 samples using maximum likelihood estimation, making the methodology completely transparent. Additionally, the record includes a data integrity check result. Before the transaction is committed, the populated parameter values ​​are automatically compared to the value range constraints defined in the template. Only when all parameters meet the constraints is the check result recorded as passed, and the entire transaction is successfully committed. If any parameter exceeds the range, the transaction will automatically roll back to ensure that no invalid data exists in the database. This verification record provides an indispensable and rigorous chain of evidence for future model audits, performance evaluations, and troubleshooting.

[0086] To address the issue that some process feature data have complex distribution patterns that cannot be accurately described by standard parametric probability distribution models, the process of performing statistical analysis to fit the parameters of the probability distribution model further includes: sequentially fitting a set of predefined parametric probability distribution models to the feature samples, calculating the Akaike information content criterion value for each model; selecting the model with the minimum Akaike information content criterion value as the best candidate parametric model, and performing a Kolmogorov-Smirnov test. If the p-value of the test is lower than the preset significance level (α = 0.05 in this embodiment), the model is deemed unsuitable. When the suitability is unsuitable, the non-parametric kernel density estimation method is automatically switched to for modeling. The KDE model uses a fixed Gaussian kernel function, and the optimal bandwidth parameter is determined through a cross-validation grid search method. The identifier, mean, standard deviation, and bandwidth value of the selected model are stored in the node attribute table of the knowledge graph to ensure the fidelity of the description of complex data distributions. This adaptive strategy of parametric trial and non-parametric baseline ensures that an optimal probability fitting model can be found for any complex data. It significantly improves the model's fidelity in describing process uncertainties and enhances the robustness and reliability of the entire knowledge graph.

[0087] By employing rigorous statistical methods, entity attributes are precisely quantified into computable probabilistic models, rather than simply as single numerical values. This allows for a scientific description of the inherent uncertainty of the process. Through atomic transactions, methodological documentation, and integrity verification, the knowledge generation process is ensured to be completely transparent, and the results are reliable and verifiable, greatly enhancing the quality and credibility of the knowledge graph.

[0088] This method utilizes the rhythm parameters of mechanical vibration signals as a core reference to align and correlate resistive tomography (RTG) data with acoustic emission data. This aims to improve the consistency and reliability of multimodal data fusion, supporting the improvement of the accuracy of production process state identification. Secondly, the method constructs a graph structure with state transition probability attributes from the identified process states and their temporal relationships. This structure enables the knowledge graph to support state prediction and path deduction based on graph query languages. Finally, by directly storing the state probability distribution model and parameters into the node attributes of the graph database, related probability calculations can be performed directly within the database, helping to reduce external computation latency and thus positively impacting the response speed of monitoring and decision-making.

[0089] Example 2:

[0090] A typical ferric sulfate production line comprises multiple reactors operating in parallel. Raw materials are injected through a complex piping system, and mixing and reaction are carried out by a bottom-mounted motor-driven stirring device. First, a comprehensive multimodal data acquisition network is constructed using various sensors deployed on-site. This network not only acquires core process parameters commonly seen on the production monitoring interface, such as pressure, temperature, and instantaneous and cumulative flow rates in the reactors, but also simultaneously integrates more refined proprietary data streams. These include multi-axis high-frequency vibration signals mounted on the motor base to capture the precise rhythm of mechanical operation; acoustic emission signals mounted on the outer wall of the reactor to capture chemical reactions or equipment friction sound waves; and resistivity tomography data obtained through an electrode array inside the reactor to reconstruct the material mixing state in real time. All data is assigned a uniformly formatted, high-precision timestamp, a unique sensor identifier, and a current process stage identifier during acquisition. The timestamps are then corrected using a time delay estimation algorithm, ensuring that data from diverse sources are precisely aligned in the time dimension.

[0091] Based on this, the mechanical operating cycle derived from vibration signals is used as a global common time reference. The originally asynchronous and continuous data streams of pressure, temperature, acoustic emission, and tomography are efficiently segmented into a series of synchronous analysis windows with a unified time index. To ensure the intrinsic correlation of multimodal information within a window rather than accidental overlap, the cross-correlation coefficients of different signal features within the corresponding window are calculated for verification. Only when the coefficient value exceeds a preset threshold is the data within that window confirmed as a valid, aligned event and encapsulated into a unified structured data object. Subsequently, this time-ordered set of structured data objects containing high-dimensional fusion features is used as input for processing using a density-based noise spatial clustering algorithm. This algorithm can automatically and unsupervisedly identify data clusters representing different process states, such as specific process paradigm entities defined as stable feeding state, heat preservation and stirring state, or pressurized reaction state, while identifying discrete points not belonging to any cluster as process noise or potential anomalies.

[0092] Next, each identified process paradigm entity is created as a unique entity node in the knowledge graph. By analyzing the chronological evolution sequence of these entities, dynamic changes between states can be captured, and the transition probabilities between state nodes can be statistically calculated based on historical data. For example, the probability of transitioning from the heat-preserving stirring state to the pressurized reaction state is as high as 98%. These state nodes, together with directed edges with probability weights, constitute a probabilistic state machine data structure in the knowledge graph that can be directly traversed by the standard graph query language, enabling process prediction and anomaly path detection. Finally, to enable the knowledge graph to have deeper analytical capabilities, detailed structured attributes are populated for each process paradigm entity node. By analyzing a certain physical feature (such as the temperature inside the vessel under heat-preserving stirring state) in all historical instance data of the entity, an optimal probability distribution model (such as a normal distribution) is fitted for it, and the specific parameters of the model (such as the mean and standard deviation) are calculated. These parameters, along with the model type identifier, are atomically stored in the node attribute table of the graph database, along with verification records such as generated timestamps, calculation method summaries, and data integrity check results. Through the above steps, a dynamic and computable knowledge graph of ferric sulfate production is finally constructed. It not only includes the static description of the process, but also incorporates the dynamic behavior and inherent uncertainty of the process into the graph in a precise and traceable manner, laying a solid foundation for in-depth intelligent monitoring and decision support.

[0093] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data, characterized in that, include: Acquire multimodal heterogeneous data of the ferric sulfate production process, including electrical resistivity tomography data, acoustic emission signal data, and multiaxial high-frequency vibration signal data; The mechanical running rhythm obtained by analyzing multi-axis high-frequency vibration signal data is used as the main time reference. The main time reference is periodically calibrated using timestamps to form a common time reference. Based on the common time reference, the resistive tomography data and acoustic emission signal data are aligned and correlated to obtain the instantiation process paradigm entity in the knowledge graph; The process of obtaining instantiated process paradigm entities in the knowledge graph includes: establishing a cross-modal association index for structured data objects under a unified time reference; adding credibility weights based on operational stability information to form fusion feature records arranged in chronological order; using the set of fusion feature records as input; applying a density-based noise spatial clustering algorithm to process the data; automatically identifying clusters representing different process states; based on the density reachability relationship of data points in the feature space; aggregating core data points that meet the minimum density requirement and reachable boundary points to form clusters; identifying discrete points that do not belong to any cluster as process noise; defining each identified cluster as a specific process paradigm entity; creating a unique entity node for the process paradigm entity in the knowledge graph; and using the original fusion feature records as instance data of the entity node. Based on predefined ontology relation types, the time series of process paradigm entities are analyzed, directed relations with state transition probability attributes are established between process paradigm entities, and a probabilistic state machine data structure for graph query language traversal is formed in the knowledge graph; for process paradigm entities in the knowledge graph, structured attributes are defined and populated to complete the construction of the knowledge graph. The data structure of the structured attributes includes type identifiers and parameter fields for storing probability distribution models.

2. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The steps for acquiring multimodal heterogeneous data in the ferric sulfate production process include: acquiring resistive tomography data, acoustic emission signal data, and multi-axis high-frequency vibration signal data from sensors installed on the production equipment; during data acquisition, attaching a unified format timestamp, sensor identifier, and process stage identifier to each data record, and correcting the timestamp using a time delay estimation algorithm; storing the multimodal heterogeneous data in a structured record with modal tags according to the timestamp order, and retaining metadata about the data source and acquisition conditions in the record; and establishing an index table for each modal data in the storage structure.

3. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The process of forming a common time reference includes: parsing the acquired multi-axis high-frequency vibration signal data to generate time reference information describing the mechanical operating cycle length, phase position, and operating stability, and storing the time reference information as a common time reference in the metadata record; based on the common time reference, dividing the electrical resistance tomography data and acoustic emission signal data into window sets with a unified time index; calculating the cross-correlation coefficient of the electrical resistance tomography data and acoustic emission signal data within the corresponding time window; when the correlation coefficient exceeds a preset threshold, confirming that the electrical resistance tomography data and acoustic emission signal data windows are aligned; converting the aligned electrical resistance tomography data and acoustic emission data into structured data objects, with each object having corresponding time reference information and data source identifier attached.

4. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The steps for analyzing the time series of process paradigm entities based on predefined ontology relation types include: The fused feature records are arranged in chronological order to form a process paradigm entity sequence with timestamp indexes. According to the process categories and attributes defined in the ontology, the entities in adjacent time slices are compared to identify state change events between entities. The state transition conditions are determined by combining the occurrence frequency and duration of the state transition events, and a unique identifier is generated for each state transition condition. The state transition events are bound to the corresponding condition identifiers to form state transition records with semantic annotations, which are stored in the intermediate analysis result set.

5. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The steps for forming a probabilistic state machine data structure for graph query language traversal in the knowledge graph include: A set of state nodes is constructed based on the intermediate analysis result set. Each state node corresponds to a process paradigm entity and carries the entity's type information and timestamp range. The transition probabilities between state nodes are calculated, and the probability values ​​are written into the graph database as attributes of directed relation edges. Context labels and event history summaries related to the nodes are added to the attribute table of the state nodes. The state nodes and directed relation edges are organized in the knowledge graph according to the structure of a probabilistic state machine, so that traversal operations based on graph query language can perform inference and prediction according to the state transition probability path.

6. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The steps for defining structured attributes for each process paradigm entity in the knowledge graph include: A property template is specified for the process paradigm entity. The property template includes a set of fields for describing the entity's physical characteristics, process context information, and statistical model characteristics. The data structure of the fields is fixed as a type identifier of a parameterized probability distribution model and the corresponding parameter value. A probability distribution model type identifier field is set in the property template to indicate the distribution type used by the entity in statistical modeling. A set of parameter fields is set in the property template to store the key numerical parameters of the distribution model, and unit information and value range constraints are added to each field. A metadata field is set in the property template to store the generation time, data source identifier, and references to relationships with other entities.

7. The method for constructing a ferric sulfate knowledge graph based on multimodal heterogeneous data according to claim 1, characterized in that, The steps for filling in structured attributes include: based on the fusion feature samples corresponding to the process paradigm entity within a historical time period, using statistical analysis to fit the parameters of the probability distribution model; filling the fitted parameter values ​​into the parameter fields of the attribute template, and synchronously updating the value of the probability distribution model type identifier field; storing the parameter fields and type identifier fields together in the node attribute table of the graph database, which can be directly called by the graph query language and can be used for inference calculations; after the attribute filling is completed, generating a verification record for the entity, including the filling timestamp, a summary of the parameter calculation method, and the data integrity check result.

Images