Analysis method, computer-readable recording medium, and information processing apparatus

Abstract

The present application provides an analysis method, a computer-readable recording medium, and an information processing apparatus. The information processing apparatus acquires an inference result when a precondition is provided to a causal model having a plurality of variables related to an operation of a plant. Furthermore, the information processing apparatus determines, from the plurality of variables, a relevant variable that depends on the precondition, based on the inference result. Subsequently, the information processing apparatus displays, for the relevant variable, information related to a state of the relevant variable obtained by the inference result and a statistical amount of plant data that matches the relevant variable among plant data generated in the plant.

Description

Technical Field

[0001] This invention relates to analytical methods, computer-readable recording media, and information processing apparatus. Background Technology

[0002] In various plants using oil, petrochemical, chemical, and natural gas, operational control is implemented using process data. The relationships between process data and various physical phenomena are intricate, and the data is complex and multidimensional, presented in a dispersed environment with 4Ms (Machine, Method, Man, Material). By analyzing this complex multidimensional data, the factors causing anomalies are identified, causal relationships between plant components and between processes are generated, and information is provided to operators.

[0003] Existing technical documents

[0004] Patent Document 1: Japanese Patent Publication No. 2013-41448

[0005] Patent Document 2: Japanese Patent Publication No. 2013-218725

[0006] Patent Document 3: Japanese Patent Publication No. 2018-128855

[0007] Patent Document 4: Japanese Patent Publication No. 2020-9080

[0008] However, relying solely on the display of causal relationships makes it difficult for operators to take immediate and actionable responses. For example, while experienced operators can immediately determine the appropriate action based on the display of causal relationships, inexperienced operators are prone to getting overwhelmed by the information and becoming confused. Summary of the Invention

[0009] One object of the present invention is to provide an analytical method, a computer-readable recording medium, and an information processing apparatus that can assist operators in making rapid decisions.

[0010] According to one analytical method, the computer performs the following processing: obtaining inference results when providing preconditions for a causal model with multiple variables related to the operation of a plant; determining relevant variables from the multiple variables that depend on the preconditions based on the inference results; and displaying, for the relevant variables, information related to the state of the relevant variables obtained through the inference results and statistics of plant data that conform to the relevant variables in plant data generated in the plant.

[0011] A computer-readable recording medium in one manner records an analysis program that causes a computer to perform the following processes: obtaining inference results when providing preconditions for a causal model having multiple variables related to the operation of a plant; determining relevant variables from the multiple variables that depend on the preconditions based on the inference results; and displaying, for the relevant variables, information related to the state of the relevant variables obtained through the inference results and statistics of plant data that conform to the relevant variables in plant data generated in the plant.

[0012] An information processing apparatus according to one approach includes: obtaining an inference result when providing preconditions for a causal model having multiple variables related to the operation of a plant; a determination unit that determines relevant variables dependent on the preconditions from the multiple variables based on the inference result; and a display unit that displays, for the relevant variables, information related to the state of the relevant variables obtained through the inference result and statistics of plant data that conform to the relevant variables in plant data generated in the plant.

[0013] One implementation method can assist operators in making quick decisions. Attached Figure Description

[0014] Figure 1 This is a diagram illustrating the system structure of Implementation Method 1.

[0015] Figure 2 This is a functional block diagram illustrating the functional structure of the information processing device in Implementation Method 1.

[0016] Figure 3 This is an example of the process data collection.

[0017] Figure 4 This is a diagram illustrating the preprocessed data.

[0018] Figure 5 This is a diagram illustrating an example of clustering results based on probabilistic latent semantic analysis.

[0019] Figure 6 This is a diagram illustrating the decision examples of causal relationships.

[0020] Figure 7 This is a diagram illustrating Example 1 of the generation of the learning dataset used for causal model learning.

[0021] Figure 8 This is a diagram illustrating Example 2 of the generation of the learning dataset used for causal model learning.

[0022] Figure 9 This is a diagram illustrating an example of a Bayesian network after it has completed learning.

[0023] Figure 10This is a diagram illustrating an example of how Bayesian network-based inference results can be visualized.

[0024] Figure 11 This is a diagram illustrating an example of how QMM-related information is derived from inferences based on Bayesian networks.

[0025] Figure 12 This is a flowchart illustrating the processing flow of Implementation Method 1.

[0026] Figure 13 This is a functional block diagram showing the functional structure of the information processing device 10 in Embodiment 2.

[0027] Figure 14 This is a diagram illustrating the process of Implementation Method 2.

[0028] Figure 15 This is a diagram illustrating an application example of cause and effect.

[0029] Figure 16 This is a diagram illustrating an example of hardware structure. Detailed Implementation

[0030] Hereinafter, embodiments of the analytical method, computer-readable recording medium, and information processing apparatus disclosed in this application will be described in detail with reference to the accompanying drawings. However, the present invention is not limited to the described embodiments. Furthermore, the same reference numerals are used to denote the same elements, and repeated descriptions are appropriately omitted; various embodiments can be appropriately combined without contradiction.

[0031] (Overall structure)

[0032] Figure 1 This is a diagram illustrating the system structure of Implementation Method 1. For example... Figure 1 As shown, the system includes a factory 1, a historical database 12, and an information processing device 10. Furthermore, the factory 1 and the historical database 12 are communicatively connected via a dedicated line or similar means, regardless of whether the connection is wired or wireless. Similarly, the historical database 12 and the information processing device 10 can be communicatively connected via a network N, such as the Internet or a dedicated line, regardless of whether the connection is wired or wireless.

[0033] Plant 1, comprising multiple pieces of equipment or machines and a control system 11, is an example of various plants using oil, petrochemical, chemical, or natural gas resources. The control system 11 controls the operation of the various pieces of equipment installed within Plant 1. Plant 1 is equipped with a distributed control system (DCS). The control system 11 acquires process data such as measured values ​​(PV), setpoint values ​​(SV), and manipulated variables (MV) from control devices (not shown) installed on the controlled equipment and from corresponding (not shown) operating devices.

[0034] Here, the field device is a field device such as an operator that has the function of measuring the operating status of the set equipment (e.g., pressure, temperature, flow rate, etc.) and controlling the operation of the set equipment according to the input control signal (e.g., an actuator, etc.). The field device outputs the operating status of the set equipment as process data to the control system 11 sequentially. In addition, the process data also includes information about the type of output measured value (e.g., pressure, temperature, flow rate, etc.). Furthermore, the process data includes information such as a label name assigned to identify the field device itself. In addition, the measured values ​​output as process data may not only be measured values ​​measured by the field device, but may also include calculated values ​​calculated from the measured values. The calculation from the measured value to the calculated value can be performed by the field device or by an external device (not shown) connected to the field device.

[0035] The historical database 12 is a device that stores long-term historical data by saving process data acquired by the control system 11 in chronological order. Examples include various types of memory such as ROM (Read Only Memory), RAM (Random Access Memory), flash memory, and storage devices such as HDD (Hard Disk Drive). The stored process data is output to the information processing device 10 via, for example, a dedicated communication network N constructed within the factory 1. Furthermore, the number of control systems 11 and historical databases 12 connected to the information processing device 10 is not limited to... Figure 1 The number shown can be composed of multiple devices. In addition, the historical database 12 can be built into the control system 11, becoming a structural element in constructing a distributed control system or other control systems.

[0036] The information processing device 10 uses the process data stored in the historical database 12 and the parent-child relationships of the structural elements constituting the factory 1 to generate a causal model. Moreover, the information processing device 10 is an example of a computer device that takes the state of the factory 1 as a premise and inputs it into a causal model such as a Bayesian network to generate operator-operable information and output it.

[0037] (Reference technology displayed by the operator)

[0038] To accurately analyze factors related to PQCDS (Productivity, Quality, Cost, Delivery, Safety) in the factory, the following steps are taken: Based on improving data quality by decomposing the data according to certain regularities and general terms and according to each similar operating state, factor analysis is performed according to each decomposed operating state and based on various machine learning models.

[0039] Typically, dimensionality compression and clustering are known techniques for state decomposition. For example, in anomaly detection and diagnosis of equipment in a factory, dimensionality compression is used for feature extraction, followed by clustering to categorize sensor data into several classes based on operating modes. Furthermore, techniques are known to improve anomaly detection sensitivity and diagnostic accuracy by modeling each class separately. These techniques, by representing multidimensional data with low-dimensional models, decompose complex states into simpler models, thus offering the advantage of easily understood or interpreted phenomena. Examples of dimensionality compression methods used here include principal component analysis, independent component analysis, nonnegative matrix factorization, latent structure projection, and positive quasi-correlation analysis. Examples of clustering methods include the EM algorithm and k-means for time trajectory segmentation and mixed distributions.

[0040] Furthermore, causal analysis based on machine learning models typically tabulates the correlation between the objective (outcome) and the explanatory variables (causes) using correlation coefficients and contribution rates. It is known that graphical models can represent the probability distributions between explanatory variables using undirected and directed graphs. For example, directed graphs, with their directional flow from "causes" to "outcomes," are easily understood by humans, allowing users to intuitively grasp the direct and indirect factors that cause the impact and to refocus attention on previously overlooked factors.

[0041] Bayesian networks are known as graphical models that utilize directed graphs to represent causal relationships between variables. Because Bayesian networks preserve the quantitative relationships between variables with conditional probabilities, by providing evidence states (evidence) to nodes of interest, the probability distribution of the states of other nodes at that time and their probabilities can be inferred. For example, Bayesian networks are used to analyze causal relationships in factory equipment alarms and operator actions, changes in process operation status, and causal relationships in equipment, parts, and deterioration conditions.

[0042] (Improvements in the reference technology)

[0043] Regarding the aforementioned state decomposition techniques, dimensionality compression typically involves mapping new components (axes) to a lower-dimensional space while preserving as much useful information as possible. This is a method of extracting features. However, these new components themselves may not have physical meaning, and their interpretation is often difficult. For example, in anomaly detection, explaining the causes of anomalies in a feature space lacking physical meaning is challenging. In situations where explanations of these causes are crucial, insufficient justification can lead to false detections.

[0044] On the other hand, typical clustering methods do not sparsify the data, but rather group data based on similarity while maintaining the original data structure. For example, k-means, a hard clustering method, which judges similarity based on certain "distance scales," can sometimes struggle to group data appropriately if the data becomes large-scale and multi-dimensional, such as process data. This difficulty is also manifested in terms of the so-called "curse of dimensionality."

[0045] Furthermore, given the complex relationships between process data and various physical phenomena, classification as simply "0%" or "100%" as in hard clustering is often unsuitable. Therefore, to avoid the "curse of dimensionality" that can occur when processing process data, and to represent the degree of similarity with probability values, a soft clustering method is used, which judges similarity based on "probabilistic frequency of occurrence (co-occurrence rate under latent meaning conditions)" instead of distance. A common soft clustering method is Probabilistic Latent Semantic Analysis (PLSA).

[0046] Furthermore, regarding causal analysis, Bayesian networks, which demonstrate how causal relationships between variables can be represented using directed graphs, are algorithms for processing discrete variables. Therefore, when applied to process data, processing the discrete numerical data obtained from sensors at specified periods results in a massive number of nodes and states. This leads to computational explosion and complex networks. Consequently, while Bayesian networks typically categorize variables like "Unstable" and "Increase" based on their numerical meaning (abstract representation), making it easier to grasp the overall qualitative trend, they struggle to analyze the specific numerical values ​​leading to the reaction process.

[0047] Furthermore, as a method for highlighting the results of causal analysis, methods such as emphasizing or listing high-probability paths obtained through Bayesian network learning for equipment, parts, and deterioration conditions can help users easily understand the causal relationships of the causes. However, in purposes such as quality stabilization in chemical processes, it is not only necessary to grasp the causal relationships of the causes, but also, importantly, to provide information that, like a Quality Management Matrix (QMM) that operators typically refer to during operation (equivalent to a manufacturing formula), allows operators to easily compare and discuss with existing benchmarks and immediately reflect the results from the operator's perspective and what needs to be done.

[0048] Here, the information processing device 10 of Embodiment 1 utilizes probabilistic latent semantic analysis and Bayesian networks to extract factors affecting production management indicators of quality from complex operational data containing environmental changes such as the four production elements of products in the factory through machine learning. Then, the information processing device 10 assists operators in making rapid decisions by converting the machine learning results into a form easily viewed and understood by the operator and providing prompts.

[0049] (Explanation of terminology)

[0050] Furthermore, the four production elements used in Implementation Method 1 are equipment (Machine), process and steps (Method), operator (Man), and raw materials (Material). Probabilistic Latent Semantic Analysis (PLSA) is a type of soft clustering technique that judges similarity based on the probabilistic frequency of occurrence, and can represent the degree of cluster affiliation in terms of probability. Moreover, PLSA can cluster rows and columns simultaneously. This PLSA is also known as Probabilistic Latent Semantic Analysis.

[0051] Furthermore, Bayesian networks are an example of probabilistic and causal models that use directed graphs to visualize qualitative dependencies between multiple probabilistic variables and represent quantitative relationships between variables using conditional probabilities. Production management indicators encompass concepts such as productivity, quality, cost, delivery time, and safety. A quality management sheet is analogous to a manufacturing recipe; to ensure product quality, it records information such as which management points must be controlled within which baseline range (specific numerical range), serving as crucial information for operators during operation.

[0052] (Functional Structure)

[0053] Next, it indicates that it has Figure 1 The diagram shows the functional structure of each device in the system. Furthermore, since the control system 11 and historical database 12 have the same structure as the control system and historical database commonly used in the control and management of plant 1, detailed descriptions are omitted. Here, we will describe the information processing device 10, which has functions different from the monitoring and management devices commonly used in the control and management of plant 1.

[0054] Figure 2 This is a functional block diagram illustrating the functional structure of the information processing device 10 in Embodiment 1. For example... Figure 2 As shown, the information processing device 10 includes a communication unit 100, a storage unit 101, and a control unit 110. Furthermore, the functional units of the information processing device 10 are not limited to those shown in the figure; they may also include other functional units such as a display unit implemented by a display or the like.

[0055] The communication unit 100 is a processing unit that controls communication with other devices, and is implemented, for example, by a communication interface. For example, the communication unit 100 controls communication with the historical database 12, receives process data from the historical database 12, or sends the results executed by the control unit 110 (described later) to a terminal used by the administrator.

[0056] Storage unit 101 is a processing unit that stores various data and various programs executed by control unit 110, and is implemented, for example, by memory and hard disk. Storage unit 101 stores data obtained during the execution of various processes by control unit 110 and processing results obtained by executing various processes, as well as various data generated during the processing executed by information processing device 10.

[0057] The control unit 110 is a processing unit responsible for the overall information processing device 10, and is implemented by, for example, a processor. The control unit 110 includes a process data collection unit 111, a clustering unit 112, a causal relationship candidate decision unit 113, a causal model construction unit 114, an analysis unit 115, and a display unit 116.

[0058] The process data collection unit 111 is a processing unit that collects process data in chronological order. Specifically, the process data collection unit 111 requests the output of process data records from the historical database 12 at predetermined time intervals when the information processing device 10 begins analysis and processing, and obtains the process data output according to the request. In addition, the process data collection unit 111 stores the collected process data in the storage unit 101 or outputs it to the clustering unit 112.

[0059] Figure 3 This is an example of data collected during the process. For example... Figure 3 As shown, the process data includes "Time, TagA1, TagA2, TagA3, TagB1, ...". Here, "Time" refers to the time when the process data was collected. "TagA1, TagA2, TagA3, TagB1, etc." represent information about the process data, such as measured values, setpoints, and operating quantities obtained from Plant 1. Figure 3 In the example, it means that at time "t1", as process data "TagA1, TagA2, TagA3, TagB1", "15, 110, 1.8, 70" were collected.

[0060] The clustering unit 112 is a processing unit that outputs the results of clustering time-related elements and label-related elements using probabilistic latent semantic analysis according to their respective probabilities to the causal model construction unit 114. Specifically, as preprocessing, the clustering unit 112 performs segmentation of the desired analysis object period, handling missing values ​​and outliers in the original data. In addition, the clustering unit 112 can perform the calculation of derived variables such as differential and integral values, moving averages, etc., as needed.

[0061] Furthermore, since probabilistic latent semantic analysis processes discrete variables (categorical variables), the clustering unit 112 performs discretization processing on numerical data, i.e., process data, converting numerical data such as "1.2" into categorical values ​​such as "1.0-2.0". As discretization processing, methods such as equal-frequency partitioning, equal-quantity partitioning, and chi-square binning can be used. Moreover, when there are variables of interest, such as those equivalent to the objective variable, clustering suitable for the characteristics of those variables can be performed by applying weights to those variables.

[0062] Figure 4 This is a graph illustrating the preprocessed data. For example... Figure 4 As shown, the clustering unit 112 performs clustering by... Figure 3 The process data shown is discretized to generate... Figure 4The preprocessed data is shown below. For example, the clustering unit 112 converts the process data "time = t1, TagA1 = 15, TagA2 = 110, TagA3 = 1.8..." into "time = t1, TagA1 = 10 - 20, TagA2 = 100 - 150, TagA3 = 1.5 - 2.5...".

[0063] Subsequently, the clustering unit 112 uses the preprocessed dataset to simultaneously cluster the time elements and label elements of the process data through probabilistic latent semantic analysis, and calculates the probability (P) of each cluster. Here, the number of clusters can be determined based on the operator's domain knowledge, or it can be determined using an indicator such as Akaike's Information Criterion (AIC) used to evaluate the goodness of statistical models.

[0064] Furthermore, multiple clustering operations can be performed in stages here. For example, based on the clustering results of the obtained time elements (equivalent to the results of decomposition according to each operation state), the clustering unit 112 decomposes the data in the time direction and then uses probabilistic latent semantic analysis to cluster the data of each decomposed data again. This enables the extraction of highly relevant labels in the same operation state (cluster) and the staged subdivision of operation states.

[0065] Figure 5 This is a diagram illustrating an example of clustering results using probabilistic latent semantic analysis. Figure 5 This represents an example with a cluster size of 3. For example... Figure 5 As shown, the clustering unit 112 can obtain clustering results for row directions during similar extraction operations by performing probabilistic latent semantic analysis on the preprocessed data (see reference). Figure 5 Similarly, we can obtain the vertical clustering results of the relevant labels (see (a)). Figure 5 (b)

[0066] For example, Figure 5 The clustering results shown in (a) represent the probability that each process data point, determined by time, belongs to each cluster (Cluster1, Cluster2, Cluster3). More specifically, it indicates that the probability of process data at time t1 belonging to Cluster1 is 40%, to Cluster2 is 30%, and to Cluster3 is 30%. Here, Cluster1, etc., represent the state of factory 1, such as stable operation (normal state) and abnormal operation (abnormal state).

[0067] also, Figure 5The clustering results shown in (b) represent the probability that the tag of each process data belongs to each cluster (Cluster1, Cluster2, Cluster3). More specifically, it means that the probability of TagA1 belonging to Cluster1 is 30%, the probability of belonging to Cluster2 is 30%, and the probability of belonging to Cluster3 is 40%. Here, Cluster1, etc., represent the state of Factory 1, such as stable operation and abnormal operation. Furthermore, Figure 5 When the clustering results shown in (b) are used for the processing described later, it is preferable to attach time-order elements such as the average and dispersion values ​​of the obtained time to each tag.

[0068] The causal relationship candidate determination unit 113 is a processing unit that defines parent-child causal relationships as candidates based on plant structure information such as P&ID (Piping and Instrumentation Diagram), control loops, and monitoring screen definitions, taking into account the correlation between tags of field devices and other field devices, and outputs this information to the causal model construction unit 114. Additionally, P&ID is a graphical representation of plant structure information such as the location of piping and field devices within the plant.

[0069] Figure 6 This is a diagram illustrating a decision example of causal relationship candidates. The causal relationship candidate decision unit 113 considers the relevance based on the operator's domain knowledge, such as the positional relationship between upstream and downstream of piping, and defines the relevance between the tags of field devices and other field devices as causal parent-child relationship candidates, and outputs them to the causal model construction unit 114.

[0070] For example Figure 6 As shown, tags such as "TagA1, TagA2" are obtained from device A, "TagB1, TagB2" from device B, "TagC1, TagC2" from device C, and "TagD1, TagD2" from device D. In this case, if the causal relationship candidate determination unit 113 determines, based on predefined piping information, that devices B and C are downstream of device A, and device D is downstream of both devices B and C, then device A is determined as the parent candidate, devices B and C as child candidates, and device D as a grandchild candidate. Furthermore, the causal relationship candidate determination unit 113... Figure 6 As shown in (a), numerical data representing the parent-child relationship is generated. For example, "-" indicates that the candidate is not defined as a parent-child relationship, i.e., it is not included in the causal exploration scope during learning. Furthermore, "1" indicates upstream and "0" indicates downstream. Figure 6The example illustrates a causal relationship candidate via piping connection, but this is merely an example and not a limitation. For instance, causal relationship candidates can be determined based on various information such as equipment level, installation location, and installation site. Furthermore, equipment that becomes a causal relationship candidate does not necessarily have multiple tags; equipment with only one tag can be set as the object of causal relationship determination.

[0071] The causal model construction unit 114 is a processing unit that uses records of process data collected by the process data collection unit 111, classification results of the clustering unit 112, and information on parent-child relationship candidates from the causal relationship candidate decision unit 113 to construct causal models between various variables (tags) and environmental factors (e.g., changes in external temperature), clusters, and objectives (e.g., quality) within the plant 1 through a Bayesian network.

[0072] For example, the causal model construction unit 114 creates a learning dataset for use in learning a Bayesian network-based causal model based on preprocessed data and clustering results based on the probability of belonging to each cluster. Here, the clustering results based on the probability of belonging to each cluster can be learned by reflecting the frequency of data occurrence. This is because Bayesian networks are statistical probabilistic models that represent the relationships between variables with conditional probabilities, so this method is feasible. However, since computation time is prioritized here, this method is not necessarily used when data is intentionally assigned to the cluster with the highest probability as "0 or 1" (a hard clustering application of soft clustering results).

[0073] Figure 7 This is a diagram illustrating Example 1 of the generation of the learning dataset used for causal model learning. (See Example 1.) Figure 7 As shown, the causal model construction unit 114 links the preprocessed data and clustering results according to time, and replicates the combined data according to their respective probabilities. For example, for data at time t1, since the probability of belonging to Cluster1 is "40%", the causal model construction unit 114 generates four data points at time t1 representing belonging to Cluster1: "Cluster1=1, Cluster2=0, Cluster3=0". Similarly, for data at time t1, since the probability of belonging to Cluster2 is "30%", the causal model construction unit 114 generates three data points at time t1 representing belonging to Cluster2: "Cluster1=0, Cluster2=1, Cluster3=0". Furthermore, for data at time t1, since the probability of belonging to Cluster3 is "30%", the causal model construction unit 114 generates three data points at time t1 representing belonging to Cluster3: "Cluster1=0, Cluster2=0, Cluster3=1".

[0074] Figure 8 This is a diagram illustrating Example 2 of the generation of the learning dataset used for causal model learning. (See example 2.) Figure 8 As shown, the causal model construction unit 114 links the preprocessed data and clustering results according to time, and discretizes the combined data into clusters according to their probabilities. For example, for data at time t1, since the probability of belonging to Cluster 1 is the highest, the causal model construction unit 114 generates data at time t1 representing "Cluster1=1, Cluster2=0, Cluster3=0" belonging to Cluster 1. Similarly, for data at time t2, since the probability of belonging to Cluster 2 is the highest, the causal model construction unit 114 generates data at time t2 representing "Cluster1=0, Cluster2=1, Cluster3=0" belonging to Cluster 2. Furthermore, for data at time t3, since the probability of belonging to Cluster 3 is the highest, the causal model construction unit 114 generates data at time t3 representing "Cluster1=0, Cluster2=0, Cluster3=1" belonging to Cluster 3.

[0075] like Figure 7 and Figure 8 The causal model construction unit 114 can expand the learning data based on probability for each data point. Here, the causal model construction unit 114 adds information about the "quality" of the desired factory 1 to each data point. As an example, the "quality" is set to "1" in a stable state and "0" in an abnormal state. The information about the "quality" can be obtained along with the process data, or it can be set by managers, etc.

[0076] Next, the causal model construction unit 114 performs structure learning of an example of a causal model, namely a Bayesian network, based on the aforementioned learning dataset and the information on parent-child relationship candidates for causality generated by the causal relationship candidate decision unit 113. Here, in the parent-child relationship candidates for causality, nodes with high probabilistic dependencies are represented by a directed graph, and each node is maintained as quantitative information using a Conditional Probability Table (CPT). In addition, as information beneficial to the operator, the causal model construction unit 114 highlights nodes that correspond to controllable labels within each node.

[0077] Figure 9 This is a diagram illustrating an example of a Bayesian network after learning. Causal model construction section 114 will... Figure 7 or Figure 8 The dataset shown is for learning purposes. Figure 6The causal relationship shown in (a) is used as learning data to perform structure learning (training) of the Bayesian network and generate Figure 9 The Bayesian network shown is generated as follows. The generated Bayesian network contains nodes corresponding to the objective ("quality"), nodes corresponding to the probabilistic latent semantic analysis results ("Cluster1, Cluster2, Cluster3"), and nodes corresponding to the explanatory variables, i.e., the discretized sensor values ​​(Tags). Additionally, the nodes corresponding to each Tag contain variables calculated from the sensor values, such as differential and integral values.

[0078] Here, each node corresponding to each descriptive variable, i.e., each tag, contains a conditional probability table. For example... Figure 9 The "TagC2" shown contains a probability table indicating that the probability of the state "40-50" is 20%, the probability of the state "50-60" is 70%, and the probability of the state "60-70" is 10%. Furthermore, algorithms for learning the structure of Bayesian networks can employ well-known methods. Figure 9 In the diagram, nodes that can be set and changed by the operator, which are equivalent to controllable labels, are indicated with thick boxes.

[0079] return Figure 2 The analysis unit 115 is a processing unit that extracts elements with high probability (influence) and their state values, paths with high influence (probability), etc., based on the causal model (Bayesian network) constructed by the causal model construction unit 114 and using analysis results such as the post-hoc probability inferred from the desired scenario that meets various preconditions. Furthermore, the analysis unit 115 is a processing unit that converts the analysis results into a form equivalent to a QMM.

[0080] In the fully learned Bayesian network obtained by the causal model construction unit 114, the analysis unit 115 infers the desired scenario by providing evidence states (evidence) to each node, thereby calculating the posterior probability distribution of each node. Here, by extracting elements with high posterior probability values, the analysis unit 115 can determine the nodes (equivalent to management points in a QMM) and their state values ​​(equivalent to management baselines in a QMM) and their probability values ​​that have a significant impact on the scenario. Furthermore, by using the objective variable as a base point and tracing back to the parent node with a high posterior probability value, the analysis unit 115 can determine the propagation path with a significant impact on the scenario. Additionally, the analysis unit 115 emphasizes the directed graph, thereby visually capturing the path with the highest probability. Moreover, to make it easier for the operator to understand, the analysis unit 115 can also simulate the path and state value corresponding to the path with the highest probability on the Bayesian network on the P&ID.

[0081] Figure 10This diagram illustrates an example of how Bayesian network-based inference results can be visualized. Here, the operator specifies "quality is unstable when TagA3 is low" as a prerequisite. Figure 10 As shown, the analysis unit 115 sets the probability value of the lowest state (0.5-1.5) in the conditional probability table of node "TagA3" to "1" according to the preconditions, and sets the others to "0". Furthermore, the analysis unit 115 sets the probability value of "state" equivalent to "unstable" in the conditional probability table of node "quality" to "1", and sets the probability value equivalent to "stable" to "0". After setting the evidence in this way, the analysis unit 115 executes the Bayesian network and obtains the inference result.

[0082] As a result, the analysis unit 115 determines the conditional dependencies of each node by updating the probability values ​​of each variable (node) that satisfies the preconditions. For example, the post-hoc probability distribution of node "Cluster1" is updated to "State 1 (belongs to), probability value (0.7)", "State 2 (not belonging to), probability value (0.3)", and the post-hoc probability distribution of node "Cluster2" is updated to "State 1 (belongs to), probability value (0.8)", "State 2 (not belonging to), probability value (0.2)". Furthermore, the post-hoc probability distribution of node "TagD3" is updated to "State (130-140), probability value (0.2)", "State (140-150), probability value (0.5)", and "State (150-160), probability value (0.3)".

[0083] Then, the analysis unit 115 selects the node with the highest probability value from the upstream direction (the higher-level direction of the Bayesian network) of the objective variable, i.e., the node "quality". This allows it to determine the node associated with the precondition "quality is unstable when TagA3 is low". For example, the analysis unit 115 determines the node quality, node "Cluster2", node "TagD3", node "TagB3", and node "TagA1".

[0084] Subsequently, in order to make it easy for the operator to compare and discuss the inference results under the desired premise with existing benchmarks, the analysis unit 115 generates information equivalent to QMM in a form that can be immediately reflected in the operation from the operator's perspective. Figure 11 This is a diagram illustrating an example of how QMM-related information can be derived using inference from Bayesian networks. For example... Figure 11 As shown, the analysis unit 115 is aimed at... Figure 10 For nodes with high influence from the preconditions determined in the middle, generate Figure 11 The information shown in (a) is equivalent to that of the QMM and Figure 11 The comparison information shown in (b) is displayed.

[0085] Figure 11 The information shown in (a) that corresponds to the QMM includes "management point, management baseline, probability value, and compliance rate". Here, "management point" represents the information related to... Figure 10 The nodes identified in the process are those with high relevance to the preconditions. "Management baseline" indicates the state where the probability value is highest based on the above inferences; "probability value" is its probability value. "Compliance level" is an example of degree information, representing the proportion of the management baseline value contained in the collected data from the entire process.

[0086] Figure 11 The comparison information shown in (b) includes "the existing QMM management baseline, the overall average value of the data, the highest frequency value of the data, the highest value of the data, the lowest value of the data, and the standard deviation of the data". Here, "the existing QMM management baseline" is a pre-set baseline value. "The overall average value, the highest frequency value of the data, the highest value of the data, the lowest value of the data, and the standard deviation of the data" are statistical measures of the corresponding data in the collected full-process data.

[0087] Using the example above, the analysis unit 115, for TagA1, which is determined to be a prerequisite condition with a high degree of influence, uses it as "management benchmark, probability value, compliance rate" to determine or calculate and display "20-23℃, 74%, 88%", etc., and uses it as "existing QMM management benchmark, overall average value of data, highest frequency value of data, maximum value of data, minimum value of data, standard deviation of data" to determine or calculate and display "0-0℃, 0℃, 0℃, 0℃, 0℃, 0℃, 0℃". Furthermore, although simplified notation is used here, the numerical values ​​are substituted into 0.

[0088] In this way, the analysis unit 115 can define "compliance rate," which quantitatively represents the probability (frequency) of actually complying with management standards during the target period. Furthermore, for the extracted nodes (management points) and their status values ​​(management standards), the analysis unit 115 outputs basic statistics such as the overall average, highest frequency, maximum, minimum, and standard deviation of the data as comparison information with the overall trend of the analyzed data. In addition, if there is an existing QMM referenced in Plant 1, the analysis unit 115 displays its content as comparison information with previous data.

[0089] return Figure 2The display unit 116 is a processing unit that displays and outputs various information. Specifically, the display unit 116 displays the learned Bayesian network. Furthermore, based on the aforementioned scenarios (various preconditions and assumptions), the display unit 116 provides visual prompts to users such as process managers and operators in the factory, highlighting the nodes with significant impact, their state and probability values, the most probable path, and information equivalent to a QMM (Quality of Model) based on the inference results. This allows the unit to determine whether the user finds the results credible, i.e., based on the mechanism of process changes and known insights, whether the user accepts the results and explanations, and whether they are appropriate.

[0090] (Processing flow)

[0091] Figure 12 This is a flowchart illustrating the processing flow of Implementation Method 1. For example... Figure 12 As shown, after the start of analysis and processing is indicated by users including managers and operators, the process data collection unit 111 obtains process data from the historical database 12 (S101).

[0092] Next, the clustering unit 112 performs preprocessing on the collected process data, including discretization, missing values, and outliers (S102). Then, it uses probabilistic latent semantic analysis to simultaneously cluster the temporal and label elements of the process data (S103). For example, the process data may sometimes contain missing tags. In this case, the clustering unit 112 performs clustering based on a set mean and pre-specified values.

[0093] Furthermore, the causal relationship candidate determination unit 113 determines the causal relationship candidates based on the parent-child relationship of the devices such as the “Tag” contained in the output process data (S104).

[0094] Next, the causal model construction unit 114 generates a training dataset for a Bayesian network-based causal model based on the preprocessed data obtained in S102 and the clustering results based on the probability of belonging to each cluster obtained in S103 (S105). Subsequently, the causal model construction unit 114 performs Bayesian network structure learning based on the training dataset obtained in S105 and the parent-child relationship candidate information of causality obtained in S104 (S106).

[0095] Furthermore, in the learned Bayesian network obtained in S106, the analysis unit 115 provides evidence states to each desired node as the scenario to be known and performs inference (S107). In addition, the analysis unit 115 uses the inference results under the desired preconditions to generate... Figure 11 The information shown is equivalent to QMM (S108). As a result, the display unit 116 can perform the display output of inference results and information equivalent to QMM.

[0096] Here, it is determined whether the user accepts the result (Explanation) (S109). Here, if the information processing device 10 receives input indicating acceptance from the user (S109: Yes), the series of analyses ends. On the other hand, if the information processing device 10 receives input indicating rejection from the user (S109: No), it returns to S103, appropriately changes the analysis object and clustering conditions, changes the parent-child relationship alternative assumptions in S104, and performs the analysis again.

[0097] (Effect)

[0098] As described above, the information processing device 10 utilizes probabilistic latent semantic analysis and Bayesian networks to extract factors affecting production management indicators such as quality from complex operational data containing environmental changes such as the four production elements of products in factory 1 through machine learning. Furthermore, by converting the machine learning results into a form easily reviewed and understood by the operator and providing prompts, the information processing device 10 can assist operators in making rapid decisions during operations.

[0099] Furthermore, the information processing device 10 simplifies the situation by classifying multidimensional process data, which is intricately intertwined with the influence of various physical phenomena and environmental changes, into similar operational states and related labels by avoiding the so-called curse of dimensionality, and analyzes the complex factors of the situation, thereby improving the interpretability of the results.

[0100] Furthermore, by applying soft clustering results based on the probability of association to model learning, the information processing device 10 can improve the accuracy of causal analysis, even in process data with complex relationships to various physical phenomena. Moreover, by importing information based on clustering results, physical correlations between labels, known environmental changes, and the operator's domain insights and experience into the model, the information processing device 10 can form an analysis rooted in the reaction process, enabling the construction of models with high reliability and acceptability.

[0101] Furthermore, based on the inference results in the desired scenario that meets various preconditions and assumptions, the information processing device 10 visualizes the nodes with significant impact, propagation paths, and controllable tags, thereby effectively identifying elements with high control effectiveness. In addition, the information processing device 10 provides prompts in a QMM-like format from the operator's perspective, allowing the operator to compare with existing conditions, quickly grasp the current situation, discover new challenges, and utilize the results as new operating conditions.

[0102] However, trend and correlation analyses in process data are often comprehensive analyses. If interpretation of the results is also included, it requires a significant amount of time. Furthermore, in general machine learning models such as deep learning, considering that more explanatory variables (features) lead to poorer interpretability, it can also increase learning time and reduce generalizability due to overfitting.

[0103] Here, Embodiment 2 uses the results of Embodiment 1 to illustrate the information processing apparatus 10 that achieves improved accuracy of subsequent various analysis and machine learning models. Figure 13 This is a functional block diagram representing the functional structure of the information processing device 10 in Embodiment 2. Here, we will explain the functions that are different from those in Embodiment 1, namely the trend analysis unit 117 and the prediction unit 118.

[0104] Trend Analysis Unit 117 is a processing unit that uses the analysis results obtained from Analysis Unit 115 to perform trend analysis and correlation analysis. Prediction Unit 118 is a processing unit that uses the analysis results obtained from Analysis Unit 115 to generate a machine learning model and uses the generated machine learning model to predict the state of Factory 1 and the values ​​of each tag, etc.

[0105] Figure 14 This is a diagram illustrating the process of Implementation Method 2. (For example...) Figure 14 As shown, the analysis unit 115 performs sensitivity analysis on the target variable when providing evidence to various explanatory variables by executing the processing described in Embodiment 1. That is, the analysis unit 115 can extract variables (Tags) that have a greater influence on the target variable by calculating the post-hoc probability value, the difference between the pre-hoc probability and the post-hoc probability, etc. Here, an example is shown in which "TagD1, Cluster2, TagA1" are extracted as important tags.

[0106] Furthermore, the trend analysis unit 117, referring to the analysis results, uses the metadata of the analysis, i.e., the process data, to perform trend analysis and correlation analysis based on the important tags. To illustrate with the above example, the trend analysis unit 117 uses the process data, which are respectively equivalent to the important tags "TagD1, Cluster2, TagA1", to calculate the displacement of the time sequence of each important tag, the correlation of each important tag, etc.

[0107] As a result, since important tags for the target can be extracted in advance, and the analysis can be advanced based on these key points, local depth mining can be carried out as needed, thus improving the efficiency of the analysis.

[0108] The prediction unit 118 uses the key labels based on the analysis results as feature quantities for general machine learning models such as deep learning, and performs model learning. Using the example above, the prediction unit 118 obtains the process data and current quality for each of the key labels "TagD1, Cluster2, TagA1". That is, the prediction unit 118 generates "process data for TagD1, quality", etc. Furthermore, the prediction unit 118 generates a quality prediction model by performing machine learning, setting "process data for TagD1" as the explanatory variable and "quality" as the objective variable in the data "process data for TagD1, quality". Subsequently, if the prediction unit 118 obtains the latest process data, it inputs this latest process data into the quality prediction model to obtain the quality prediction result for plant 1, and displays the output to operators, etc.

[0109] In this way, the prediction unit 118 can omit features that have no or minimal impact on the target variable in advance. As a result, important features (labels and clusters, etc.) that have a significant impact on the target variable can be extracted in advance. By using these features to construct the prediction model, improved analysis efficiency can be expected.

[0110] The above describes the embodiments of the present invention. In addition to the embodiments described above, the present invention can also be implemented in various other ways.

[0111] (causation)

[0112] For example, Figure 6 The causal relationship shown is just one example; other elements can be added, and levels can be increased or decreased. Figure 15 This is a diagram illustrating an application example of cause and effect. For example... Figure 15 As shown, for example, as a grandchild alternative, information about the temperature of device E, "Tag M," can also be added to the causal relationship (parent-child relationship). As another example, external temperature, etc., can also be added as... Figure 6 The parent candidates for all devices shown are listed. By adding new elements, the dimensionality of the objects to be learned by the Bayesian network can be increased, thereby improving the accuracy of the Bayesian network. Furthermore, beyond temperature, causal relationship candidates based on factors such as changes in the environment (external temperature, human intervention, equipment maintenance, etc.) and operator experience (such as higher quality at night) can be added.

[0113] (numerical values, etc.)

[0114] The types of process data, number of tags, number of clusters, thresholds, and data quantity used in the above implementation method are just examples and can be arbitrarily changed. Furthermore, "quality" has been used as an example of an objective, but it is not limited to this. For example, more specific objectives can be set, such as the types of faults within factory 1 and the status of device X in factory 1, and human factors such as operator error can also be set.

[0115] Furthermore, Bayesian networks are an example of causal models, and various graphical causal models and probabilities can also be employed. In addition, each node (tag) in a causal model such as a Bayesian network corresponds to multiple variables related to the operation of factory 1. Furthermore, the variables with the highest probability values ​​determined based on the inference results correspond to relevant variables dependent on preconditions. Moreover, the learning and inference of Bayesian networks can be performed periodically over a certain period, or even after a day's operations using batch processing. Furthermore, deep learning is also an example of machine learning, and various algorithms such as neural networks, deep learning algorithms, and support vector machines can be employed.

[0116] (system)

[0117] The information in the above text and accompanying drawings, including processing steps, control steps, specific names, various data, and parameters, can be arbitrarily changed, unless specifically stated otherwise. Furthermore, Figure 6 The devices shown are examples of structural equipment. Furthermore, Figure 11 The display format is only one example, and can be arbitrarily changed to a drop-down format, etc., and the selection of comparison information can also be arbitrarily changed. In addition, the information processing device 10 can also directly obtain factory data from the factory 1.

[0118] Furthermore, the structural elements of each device illustrated are functional conceptual elements, and their physical structure may not necessarily be as shown in the diagram. That is, the specific methods of distributing and integrating the devices are not limited to those illustrated. In other words, they can be distributed or integrated, functionally or physically, in any unit, depending on various loads and usage conditions.

[0119] Furthermore, each processing function performed by each device may be implemented, in whole or in part, by a CPU and a program that is analyzed and executed by the CPU, or may be implemented as hardware based on wiring logic.

[0120] (hardware)

[0121] Next, an example of the hardware structure of the information processing device 10 will be described. Figure 16 This is a diagram illustrating an example of hardware structure. For example... Figure 16 As shown, the information processing device 10 includes a communication device 10a, an HDD (Hard Disk Drive) 10b, a memory 10c, and a processor 10d. Furthermore, Figure 16The various parts shown are interconnected via buses, etc.

[0122] Communication device 10a, such as a network interface card, communicates with other servers. HDD 10b is the storage implementation. Figure 2 The program and database shown have the functions.

[0123] Processor 10d reads data from HDD 10b, etc., for execution and... Figure 2 The same processing program for each of the processing units shown is expanded in memory 10c, and then executed. Figure 2 The process operates as described above. For example, this process performs the same functions as each processing unit in the information processing device 10. Specifically, the processor 10d reads programs from the HDD 10b, etc., that have the same functions as the process data collection unit 111, clustering unit 112, causal relationship candidate determination unit 113, causal model construction unit 114, analysis unit 115, and display unit 116. Moreover, the process executed by the processor 10d performs the same processing as the process data collection unit 111, clustering unit 112, causal relationship candidate determination unit 113, causal model construction unit 114, analysis unit 115, and display unit 116.

[0124] Thus, the information processing device 10 operates as an information processing device that performs analysis methods by reading and executing the program. Furthermore, the information processing device 10 can also read the aforementioned program from a recording medium by a media reading device, and by executing the read program, achieve the same function as in the embodiments described above. Additionally, the programs in other embodiments are not limited to being executed by the information processing device 10. For example, the present invention can also be applied when other computers or servers execute programs, or when programs are executed in conjunction with other computers.

[0125] The program can be distributed via networks such as the Internet. Furthermore, the program can be recorded on computer-readable recording media such as hard disks, floppy disks (FD), CD-ROMs, MO (Magneto-Optical disk), and DVDs (Digital Versatile Discs), and can be read and executed by a computer from the recording media.

[0126] Explanation of reference numerals in the attached figures

[0127] 10. Information processing device

[0128] 100 Ministry of Communications

[0129] 101 Storage Department

[0130] 110 Control Department

[0131] 111 Process Data Collection Department

[0132] 112 Clustering Department

[0133] 113 Causal Relationship Alternative Decision-Making Department

[0134] 114 Causal Model Construction Department

[0135] 115 Analysis Department

[0136] 116 Display Section

Claims

1. An analytical method, characterized in that, The computer performs the following processing: To obtain inference results when given preconditions for a causal model with multiple variables related to plant operations. Based on the inference results, relevant variables dependent on the preconditions are determined from the plurality of variables. For the relevant variables, the management benchmarks related to the state of the relevant variables as operational targets, obtained through the inference results, and the statistics of factory data that conform to the relevant variables in the factory data generated in the factory are displayed in a comparable form. As information relating to the state of the relevant variables, the displayed processing shows the conditions and probability values ​​obtained from the inference results, as well as information on the degree to which the conditions are quantitatively represented in the operation of the plant.

2. The analytical method according to claim 1, characterized in that, The computer performs the following processing: Collect multiple process data, including the multiple variables, output from the plant. Perform clustering to classify the multiple process data according to the operating status of the factory. The structural learning of the causal model is performed using the learning data derived from the process data and the clustering results.

3. The analytical method according to claim 2, characterized in that, In the process of performing structural learning of the causal model, Based on the correlation between the structural equipment constituting the factory, the parent-child relationship of the structural equipment is determined. The process data, the clustering results, and the parent-child relationships are used in the learning data to perform structural learning of the causal model.

4. The analytical method according to claim 3, characterized in that, In the process of performing structure learning for the causal model, the structure learning of the Bayesian network is performed using the learning data and the objective variable representing the state of the factory. In the aforementioned processing, the variables designated as the objective and the preconditions of the factory's state are input into the learned Bayesian network for inference, and the inference result is obtained. In the determined process, within each cluster to which each node in the Bayesian network belongs, the node with the highest probability value obtained through the inference is determined as the relevant variable. In the display process, for the relevant variables, the conditions, probability values, degree information, and statistics obtained through the inference results are displayed in a comparable manner.

5. A computer-readable recording medium recording an analysis program, characterized in that, The analysis program causes the computer to perform the following processes: To obtain inference results when given preconditions for a causal model with multiple variables related to plant operations. Based on the inference results, relevant variables dependent on the preconditions are determined from the plurality of variables. For the relevant variables, the management benchmarks related to the state of the relevant variables as operational targets, obtained through the inference results, and the statistics of factory data that conform to the relevant variables in the factory data generated in the factory are displayed in a comparable form. As information relating to the state of the relevant variables, the displayed processing shows the conditions and probability values ​​obtained from the inference results, as well as information on the degree to which the conditions are quantitatively represented in the operation of the plant.

6. An information processing device, characterized in that... include: The acquisition department obtains inferences when providing preconditions for a causal model with multiple variables related to the operation of the plant. The determination unit determines, based on the inference results, relevant variables dependent on the preconditions from the plurality of variables; as well as The display unit, for the aforementioned relevant variables, displays in a comparable form the management benchmarks related to the state of the relevant variables as operational targets, obtained through the inference results, as well as statistics of factory data that conform to the relevant variables in the factory data generated in the factory. As information relating to the state of the relevant variables, the display unit shows the conditions and probability values ​​obtained from the inference results, as well as information on the degree to which the conditions are complied with in the operation of the factory.

Images