A converged multi-system plant monitoring and safety system and method

Abstract

The application discloses a kind of factory monitoring and safety system and method of fusing multiple systems, it is related to industrial data processing technical field, pre-collection subsystem information set, historical monitoring data set and historical abnormal label set, based on subsystem information set and historical monitoring data set, correlation matrix calculation model is constructed, from historical monitoring data set, using correlation matrix calculation model extracts out historical correlation matrix, and from historical correlation matrix, extract abnormal early warning feature vector, based on abnormal early warning feature vector and historical abnormal label set, construct and train abnormal early warning judgment model, using the time series of historical correlation matrix, construct and train matrix prediction model;So that under the condition of large-scale data, system can quickly respond and provide accurate early warning prediction, improve the intelligent level and safety management ability of factory monitoring system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial data processing, and particularly relates to a factory monitoring and safety system and method fusing multiple systems. BACKGROUND

[0002] In modern industrial production, a factory is often composed of multiple complex subsystems, including production processes, material transportation, energy management, environmental control, safety and monitoring, etc. These subsystems have complex interdependent relationships with each other, and any abnormality in any one subsystem can have a significant impact on the overall production process. Therefore, intelligent safety monitoring and early warning prediction of these subsystems have become an important task to ensure the safe and efficient operation of the factory. However, existing technologies often face problems such as data silos, independent system operation, and poor information sharing, which limit the effectiveness of overall monitoring and risk management. Traditional monitoring systems usually focus only on a single subsystem or device, and lack the ability to identify complex relationships between systems, making it difficult to provide early warning of potential failures.

[0003] Patent application CN118134093A discloses a smart factory quality control system based on machine learning, including an adaptive module, a multi-modal analysis module, a simulation module, and a feedback and adjustment module. The adaptive module is used to adjust and optimize the changes in production data in real time through a machine learning model, and to adapt to new data patterns. The multi-modal analysis module integrates and identifies features of different types of data, and fuses multi-modal data to identify data similarity and differences. However, this technical solution does not solve the problem of data fusion of multiple subsystems.

[0004] Therefore, the present application proposes a factory monitoring and safety system and method fusing multiple systems. SUMMARY

[0005] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application proposes a factory monitoring and safety system and method fusing multiple systems, so that the system can quickly respond and provide accurate early warning and prediction under large-scale data conditions, improving the intelligent level and safety management capability of the factory monitoring system.

[0006] To achieve the above-mentioned purpose, a factory monitoring and safety method fusing multiple systems is proposed, comprising the following steps:

[0007] Step 1: Pre-collecting a subsystem information set, a historical monitoring data set, and a historical abnormality label set;

[0008] Step 2: Based on the subsystem information set and the historical monitoring data set, a correlation matrix calculation model is constructed;

[0009] Step three: extract the historical correlation matrix from the historical correlation matrix calculation model, and extract the abnormal early warning feature vector from the historical correlation matrix;

[0010] Step four: based on the abnormal early warning feature vector and the historical abnormal label set, build and train an abnormal early warning judgment model;

[0011] Step five: use the time series of the historical correlation matrix to build and train a matrix prediction model;

[0012] Step six: in the actual monitoring process, collect real-time monitoring data, obtain real-time correlation matrix based on real-time monitoring data and correlation matrix calculation model, obtain future correlation matrix based on real-time correlation matrix and matrix prediction model, and obtain early warning judgment for the future based on future correlation matrix and abnormal early warning judgment model;

[0013] The collection method of the subsystem information set is:

[0014] Obtain the basic information of each factory device in each subsystem in the factory from the technical specification of the equipment and sensors, as the subsystem information set;

[0015] The collection method of the historical monitoring data set is:

[0016] Extract or directly export the historical monitoring data generated by each subsystem and each device in the historical operation process from the factory's monitoring and data acquisition system or directly from the factory's background database;

[0017] The collection method of the historical abnormal label set is:

[0018] Extract or collect from the factory's event management or reporting system, get abnormal label data each time an abnormal event occurs, and form a historical abnormal label set;

[0019] The method for building a correlation matrix calculation model based on the subsystem information set and the historical monitoring data set is:

[0020] For each subsystem included in the subsystem information set:

[0021] According to the measurement data of each device of the subsystem in the historical monitoring data set, construct a device correlation matrix for each subsystem;

[0022] Extract the principal feature vector of the device correlation matrix of each subsystem, and construct a subsystem correlation matrix between each subsystem based on the principal feature vector of each subsystem;

[0023] The method for constructing a device correlation matrix for each subsystem is:

[0024] Setting a time window size T and a sliding step s;

[0025] Mark each unit time in the historical monitoring data set as t;

[0026] Mark the number of each subsystem as i;

[0027] Mark the number of each device in the i-th subsystem as ij;

[0028] For each unit time t, extract the parameter set composed of the measurement data of the ij-th device in each subsystem from unit time t-T to unit time t to form a data window W(Sij, t); wherein Sij represents a parameter time sequence composed of the measurement data of the ij-th device in the i-th subsystem arranged in chronological order;

[0029] Using a correlation coefficient analysis method to calculate the correlation between devices in the i-th subsystem to generate a device correlation matrix;

[0030] The way to extract the principal eigenvector of the device correlation matrix of each subsystem is:

[0031] Obtain the principal eigenvector by singular value decomposition of the device correlation matrix;

[0032] The way to obtain the principal eigenvector using singular value decomposition is:

[0033] Mark the device correlation matrix as A;

[0034] Using singular value decomposition, the matrix A is converted to the form of A = U∑V T ;

[0035] Wherein, U and V are both orthogonal matrices, the column vector of U is called the left singular vector, and each column vector of V is the right singular vector; Σ is a singular value matrix, which is a diagonal matrix containing singular values arranged in descending order;

[0036] Select the largest singular value from the singular value matrix, and take the left singular vector corresponding to the largest singular value as the principal eigenvector;

[0037] The way to construct the subsystem correlation matrix between each subsystem based on the principal eigenvector of each subsystem is:

[0038] Mark any two subsystems as i1 and i2;

[0039] Using the Pearson correlation coefficient formula, the subsystem correlation degree between subsystem i1 and subsystem i2 is calculated from the principal eigenvector of subsystem i1 and the principal eigenvector of subsystem i2.

[0040] According to the subsystem correlation degree between each subsystem, a subsystem correlation matrix is formed;

[0041] The manner of extracting the historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model is that:

[0042] For each time window in the historical monitoring process, the subsystem correlation matrix between each subsystem in each time window is calculated by the calculation manner of the subsystem correlation matrix, as the historical correlation matrix;

[0043] The manner of extracting the abnormal early warning feature vector from the historical correlation matrix is that:

[0044] The principal eigenvector of the historical correlation matrix is extracted as the abnormal early warning feature vector by using the singular value decomposition method on the historical correlation matrix of each time window;

[0045] The manner of constructing and training the abnormal early warning judgment model based on the abnormal early warning feature vector and the historical abnormal label set is that:

[0046] Taking any classification model as the abnormal early warning judgment model;

[0047] The abnormal early warning judgment model takes the abnormal early warning feature vector as input, and takes the predicted value of the abnormal type number corresponding to the time window of the abnormal early warning feature vector as output;

[0048] The abnormal early warning judgment model takes the cross-entropy error between the predicted value of the abnormal type number and the actual value of the abnormal type number corresponding to the time window as the first error function;

[0049] The abnormal early warning judgment model is trained using the abnormal early warning feature vector and the actual value of the abnormal type number in all time windows in the historical monitoring process until the first error function converges;

[0050] The manner of constructing and training the matrix prediction model using the time sequence of the historical correlation matrix is that:

[0051] The matrix prediction model is set to an LSTM network model;

[0052] The matrix prediction model takes the sequence of the subsystem correlation matrix in the past time window as input;

[0053] The output layer of the matrix prediction model outputs the predicted matrix of the subsystem correlation matrix in the time window after L time steps in the future as output; L is a preset prediction time step;

[0054] The real subsystem correlation matrix after L time steps corresponding to each group of input pairs is taken as a prediction target of a prediction matrix, and the matrix cosine similarity between the real subsystem correlation matrix and the prediction matrix is taken as a second error function;

[0055] The matrix prediction model is trained until the second error function converges.

[0056] The application provides a factory monitoring and safety system fusing multiple systems, which comprises a historical data collection module, a correlation matrix calculation module, a warning model training module, a matrix prediction module and a safety warning module.

[0057] The historical data collection module collects a subsystem information set, a historical monitoring data set and a historical abnormal label set in advance, and sends the subsystem information set and the historical monitoring data set to the correlation matrix calculation module and sends the historical abnormal label set to the warning model training module.

[0058] The correlation matrix calculation module constructs a correlation matrix calculation model based on the subsystem information set and the historical monitoring data set, and sends the correlation matrix calculation model to the warning model training module.

[0059] The warning model training module extracts a historical correlation matrix from the historical monitoring data set by using the correlation matrix calculation model, extracts an abnormal warning feature vector from the historical correlation matrix, constructs and trains an abnormal warning judgment model based on the abnormal warning feature vector and the historical abnormal label set, sends the historical correlation matrix to the matrix prediction module and sends the abnormal warning judgment model to the safety warning module.

[0060] The matrix prediction module constructs and trains a matrix prediction model by using a time sequence of the historical correlation matrix, and sends the matrix prediction model to the safety warning module.

[0061] The safety warning module collects real-time monitoring data in an actual monitoring process, obtains a real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model, obtains a future correlation matrix based on the real-time correlation matrix and the matrix prediction model, and obtains a future warning judgment based on the future correlation matrix and the abnormal warning judgment model.

[0062] Compared with the prior art, the application has the advantages that:

[0063] The application pre-collects a subsystem information set, a historical monitoring data set and a historical abnormal label set, constructs a correlation matrix calculation model based on the subsystem information set and the historical monitoring data set, extracts a historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model, extracts an abnormal early warning feature vector from the historical correlation matrix, constructs and trains an abnormal early warning judgment model based on the abnormal early warning feature vector and the historical abnormal label set, constructs and trains a matrix prediction model using a time sequence of the historical correlation matrix, and finally collects real-time monitoring data in an actual monitoring process, obtains a real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model, obtains a future correlation matrix based on the real-time correlation matrix and the matrix prediction model, and obtains a future early warning judgment based on the future correlation matrix and the abnormal early warning judgment model; the scheme realizes high efficiency of calculation in the correlation analysis process. The main feature vector of the equipment and the abnormal correlation degree between the subsystems are extracted through a dimension reduction technology such as singular value decomposition, the data calculation amount is greatly reduced, the operation speed is improved, the integrity of the main information is maintained, the system can quickly respond and provide accurate early warning prediction under the condition of large-scale data, and the intelligent level and the safety management ability of the factory monitoring system are improved. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 A flowchart of a factory monitoring and safety method fusing multiple systems in embodiment 1 of the application;

[0065] Figure 2 A topological relationship example diagram of each device and production line in a production system in embodiment 1 of the application;

[0066] Figure 3 A module connection relationship diagram of a factory monitoring and safety system fusing multiple systems in embodiment 2 of the application. DETAILED DESCRIPTION

[0067] The technical solutions of the application will be described clearly and completely in combination with embodiments. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0068] Embodiment 1

[0069] As shown in Figure 1 A factory monitoring and safety method fusing multiple systems, comprising the following steps:

[0070] Step 1: pre-collect a subsystem information set, a historical monitoring data set and a historical abnormal label set;

[0071] Step two: based on the subsystem information set and the historical monitoring data set, a correlation matrix calculation model is constructed;

[0072] Step three: from the historical monitoring data set, the correlation matrix calculation model is used to extract the historical correlation matrix, and the abnormal early warning feature vector is extracted from the historical correlation matrix;

[0073] Step four: based on the abnormal early warning feature vector and the historical abnormal label set, an abnormal early warning judgment model is constructed and trained;

[0074] Step five: using the time series of the historical correlation matrix, a matrix prediction model is constructed and trained;

[0075] Step six: in the actual monitoring process, real-time monitoring data is collected, real-time correlation matrix is obtained based on real-time monitoring data and correlation matrix calculation model, future correlation matrix is obtained based on real-time correlation matrix and matrix prediction model, and future warning judgment is obtained based on future correlation matrix and abnormal early warning judgment model;

[0076] The collection method of the subsystem information set is:

[0077] The basic information of each factory device in each subsystem in the factory is obtained from the technical specification of the device and the sensor, which is the subsystem information set;

[0078] Specifically, the subsystems are different according to different products produced by the factory. For example, the subsystems include but are not limited to production process subsystems, including but not limited to machining equipment such as CNC machine tools, punching machines, milling machines, lathes, assembly equipment such as automated assembly lines and robot welding equipment, molding equipment such as injection molding machines and die casting machines, material and product conveying subsystems, including but not limited to conveying belt systems, automatic guided vehicles and lifting equipment; warehouse subsystems, including but not limited to stackers, automatic shelves, bar code scanners and RFID systems; and energy management subsystems, environmental control subsystems, safety monitoring subsystems, quality control subsystems and communication subsystems, etc.

[0079] For example, an automatic assembly line, a data collector 1, a central controller, a CNC machining center, a data collector 2, an injection molding machine and other devices or production lines are included in a certain production system, and the topological relationship of each device and production line is as shown in Figure 2

[0080] ​The subsystem information set includes, but is not limited to, device information of each device in each subsystem, sensor information (if the subsystem contains sensors), network topology between devices, and operating parameters of each device. Specifically, the device information includes, but is not limited to, device type, model and function description of each subsystem, and the sensor information includes, but is not limited to, sensor list used in each subsystem, sensor type (such as temperature sensor, pressure sensor), measurement accuracy and unit, etc.

[0081] The collection of the historical monitoring data set is as follows:

[0082] The historical monitoring data of each subsystem and each device in the subsystem generated in the historical operation process are extracted from the monitoring and data acquisition system of the factory or directly exported from the background database of the factory.

[0083] Specifically, the historical monitoring data includes, but is not limited to, accurate time stamp, measurement data of each subsystem in the monitoring period, and parameter record of all manual operation events of each device by the operator.

[0084] The collection of the historical abnormal label set is as follows:

[0085] The abnormal label data is extracted from the operator event log or abnormal report or collected from the event management or report system of the factory, and the abnormal label data is obtained each time an abnormal event occurs to form the historical abnormal label set.

[0086] Specifically, the abnormal label data includes event type, event reason, event time, affected subsystem, and severity of the impact on the subsystem each time an abnormal event occurs. Specifically, the severity can be set as no harm, low risk, medium risk, high risk, and shutdown for maintenance, and can be further quantified as 1, 2, 3, 4, 5, etc. digital labels for subsequent model use.

[0087] Further, the method for constructing the correlation matrix calculation model based on the subsystem information set and the historical monitoring data set is as follows:

[0088] For each subsystem included in the subsystem information set:

[0089] A device correlation matrix is constructed for each subsystem according to the measurement data of each device of the subsystem in the historical monitoring data set.

[0090] The principal component vector of the device correlation matrix of each subsystem is extracted, and a subsystem correlation matrix between each subsystem is constructed based on the principal component vector of each subsystem.

[0091] Specifically, the way of constructing the device correlation matrix for each subsystem is:

[0092] Set a time window size T and a sliding step s; preferably, the time window size T can be set to 5 minutes, that is, analyze the correlation between each subsystem in every 5 minutes; the sliding step s can be set to 1 minute, that is, the time window changes every 1 minute;

[0093] Mark each unit time in the historical monitoring data set as t; the unit time can generally be set to 1 minute;

[0094] Mark the number of each subsystem as i;

[0095] Mark the number of each device in the i-th subsystem as ij;

[0096] For each unit time t, extract the parameter set composed of the measurement data of the ij-th device in each subsystem from the unit time t-T to the unit time t to form a data window W(Sij, t); wherein Sij represents a parameter time series composed of the measurement data of the ij-th device in the i-th subsystem arranged in time order;

[0097] Use a correlation coefficient analysis method to calculate the correlation between devices in the i-th subsystem to generate a device correlation matrix;

[0098] Specifically, the correlation coefficient analysis method includes but is not limited to Pearson, Spearman or mutual information;

[0099] For example, the embodiment provided herein calculates the device correlation matrix using the Pearson correlation coefficient method:

[0100] Mark the number of other devices in the i-th subsystem except the ij-th device as ik;

[0101] Mark the correlation degree between the ij-th device and the jk-th device as Rijk;

[0102] The calculation formula of Riijk is:

[0103]

[0104] Wherein sijt' is the t'-th measurement data in the time series of Sij, that is, the measurement data of the ij-th device after the t'-th unit time in the data window corresponding to the time series of Sij;

[0105] is the average of all measurement data in the time series of Sij;

[0106] Similarly, sikt' is the t'th measurement data in the time series of Sik, i.e. the measurement data of the i k device after the t'th unit time in the data window corresponding to the time series of Sik;

[0107] is the average of all measurement data in the time series of Sik;

[0108] The correlation degree between each device in the i'th subsystem is constructed into a device correlation matrix;

[0109] The value corresponding to the j'th row and k'th column in the device correlation matrix is Rijk;

[0110] Further, the way of extracting the principal eigenvector of the device correlation matrix of each subsystem is:

[0111] The principal eigenvector is obtained by singular value decomposition of the device correlation matrix using the singular value decomposition method;

[0112] As an example, the way of obtaining the principal eigenvector using the singular value decomposition method can be:

[0113] The device correlation matrix is denoted as A;

[0114] The matrix A is converted into A = U∑V using the singular value decomposition method; T ;

[0115] Wherein, U and V are both orthogonal matrices, the column vector of U is called left singular vector, and each column vector of V is right singular vector; Σ is a singular value matrix, which is a diagonal matrix containing singular values arranged in descending order;

[0116] It should be noted that in the matrix A, the left singular vector corresponds to the characteristic mode in the original data space. These vectors define the direction of different correlation modes; while the right singular vector corresponds to the characteristic mode of the column space, which is usually related to the column of the original data;

[0117] The largest singular value is selected from the singular value matrix, and the left singular vector corresponding to the largest singular value is taken as the principal eigenvector;

[0118] It should be noted that the principal eigenvector represents a main correlation mode between devices, which indicates the potential correlation relationship between devices in the subsystem, and each element in the principal eigenvector represents the relative contribution or weight of each device, while the positive or negative sign of each element indicates the consistency or symmetry of the change direction of each device in the mode; the absolute value of the element represents the strength contribution of the device to the main correlation mode.

[0119] Further, the way of constructing the subsystem correlation matrix between each subsystem based on the principal eigenvector of each subsystem is:

[0120] Labeling any two subsystems as i1 and i2 respectively;

[0121] Using the Pearson correlation coefficient formula, the subsystem correlation degree between subsystem i1 and subsystem i2 is calculated based on the principal eigenvector of subsystem i1 and the principal eigenvector of subsystem i2;

[0122] It can be understood that the calculation method of the subsystem correlation degree is consistent with the calculation method of the correlation degree between the devices described above, which will not be described here again. The difference is that the input vector of the correlation degree between the devices is the time series of the measurement data of each device, while the input vector of the subsystem correlation degree is the principal eigenvector of each subsystem;

[0123] The subsystem correlation matrix is constructed according to the subsystem correlation degree between each subsystem;

[0124] Further, the way of extracting the historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model is:

[0125] For each time window in the historical monitoring process, the subsystem correlation matrix between each subsystem in each time window is calculated by the calculation method of the subsystem correlation matrix, which is used as the historical correlation matrix;

[0126] Then the way of extracting the abnormal early warning eigenvector from the historical correlation matrix is:

[0127] The principal eigenvector of the historical correlation matrix is extracted as the abnormal early warning eigenvector by using the singular value decomposition method for each time window of the historical correlation matrix;

[0128] It can be understood that each element in the abnormal early warning eigenvector represents the contribution weight of a subsystem to the entire system anomaly, thereby effectively representing the main correlation pattern between each subsystem. The correlation pattern may be an abnormal state;

[0129] Further, the way of constructing and training the abnormal early warning judgment model based on the abnormal early warning eigenvector and the historical abnormal label set is:

[0130] Using any classification model as the abnormal early warning judgment model; the classification model includes but is not limited to DNN model, SVM model, etc.;

[0131] The abnormal early warning judgment model takes the abnormal early warning eigenvector as the input and takes the predicted value of the abnormal type number of the time window corresponding to the abnormal early warning eigenvector as the output;

[0132] Specifically, the predicted value of the abnormal type number can contain information of the number of event type and the number of severity, for example, the number of short circuit event type is 0, and the short circuit event type can be divided into 5 abnormal type numbers, which are 1, 2, 3, 4 and 5, respectively, representing non-harmful short circuit, low-risk short circuit, medium-risk short circuit, high-risk short circuit and short circuit for maintenance, and the abnormal type numbers of the remaining event types are similar;

[0133] The abnormal early warning judgment model takes the cross-entropy error between the predicted value of the abnormal type number and the actual value of the abnormal type number in the corresponding time window as a first error function;

[0134] The abnormal early warning judgment model is trained using the abnormal early warning feature vectors in all time windows in the historical monitoring process and the actual values of the abnormal type numbers until the first error function converges;

[0135] It can be understood that through the training of the abnormal early warning judgment model, the type and risk level of the abnormal event occurring in the factory can be determined by analyzing the real-time correlation between each subsystem and between each device in the subsystem;

[0136] Further, the time sequence of the historical correlation matrix is used to build and train a matrix prediction model in the following manner:

[0137] The matrix prediction model is set as an LSTM network model; it can be understood that the LSTM model is suitable for processing and predicting time series data, and it can capture long-time span data dependencies through memory cells, adapt to dynamic changes and correlations in time series;

[0138] The matrix prediction model takes the sequence of the subsystem correlation matrix in the past time window as input;

[0139] The output layer of the matrix prediction model outputs the predicted matrix of the subsystem correlation matrix in the time window after L time steps in the future as output; L is a preset prediction time step;

[0140] The real subsystem correlation matrix after L time steps corresponding to each group of input is taken as the prediction target of the predicted matrix, and the matrix cosine similarity between the real subsystem correlation matrix and the predicted matrix is taken as a second error function;

[0141] The matrix prediction model is trained until the second error function converges;

[0142] It can be understood that by training the matrix prediction model, the subsystem correlation matrix after L time steps in the future can be predicted through the past subsystem correlation matrix, so as to determine whether there is an abnormal state after L time steps in the future through the predicted subsystem correlation matrix;

[0143] Therefore, the real-time monitoring data includes the real-time measurement data collected by all devices in all subsystems;

[0144] And the way to obtain the real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model is to construct a real-time subsystem correlation matrix based on the real-time measurement data collected by each device in each subsystem as the real-time correlation matrix;

[0145] The way to obtain the future correlation matrix based on the real-time correlation matrix and the matrix prediction model is to combine the real-time correlation matrix and the subsystem correlation matrix in the past several time windows as the input of the matrix prediction model, and obtain the prediction matrix of L time steps output by the matrix prediction model as the future correlation matrix;

[0146] The way to obtain the future warning judgment based on the future correlation matrix and the abnormal warning judgment model is to input the future correlation matrix into the abnormal warning judgment model to obtain the prediction value of the abnormal type number output by the abnormal warning judgment model, and according to the prediction value, the type and severity of the event that may occur after L time steps in the future can be determined, so that the corresponding warning can be performed according to the pre-set warning mode according to the type and severity of the event.

[0147] Embodiment 2

[0148] As shown in Figure 3 A multi-system integrated factory monitoring and safety system includes a historical data collection module, a correlation matrix calculation module, a warning model training module, a matrix prediction module, and a safety warning module. Each module is connected by electrical means.

[0149] The historical data collection module pre-collects a subsystem information set, a historical monitoring data set, and a historical abnormal label set, and sends the subsystem information set and the historical monitoring data set to the correlation matrix calculation module, and sends the historical abnormal label set to the warning model training module.

[0150] The correlation matrix calculation module constructs a correlation matrix calculation model based on the subsystem information set and the historical monitoring data set, and sends the correlation matrix calculation model to the warning model training module.

[0151] The early warning model training module extracts a historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model, extracts an abnormal early warning feature vector from the historical correlation matrix, constructs and trains an abnormal early warning judgment model based on the abnormal early warning feature vector and the historical abnormal label set, and sends the historical correlation matrix to the matrix prediction module and the abnormal early warning judgment model to the safety early warning module;

[0152] The matrix prediction module constructs and trains a matrix prediction model using a time sequence of the historical correlation matrix, and sends the matrix prediction model to the safety early warning module.

[0153] The safety early warning module collects real-time monitoring data in an actual monitoring process, obtains a real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model, obtains a future correlation matrix based on the real-time correlation matrix and the matrix prediction model, and obtains a future early warning judgment based on the future correlation matrix and the abnormal early warning judgment model.

[0154] The method and device, apparatus of the present application can be implemented in many ways. For example, the method and device, apparatus of the present application can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is merely for illustration, and the steps of the method of the present application are not limited to the above specifically described order, unless otherwise specifically stated. In addition, in some embodiments, the present application can also be implemented as programs recorded in recording media, which include machine-readable instructions for implementing the method according to the present application. Thus, the present application also covers recording media storing programs for executing the method according to the present application.

[0155] In addition, parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of corresponding technical solutions in the prior art are not described in detail to avoid excessive repetition.

[0156] The specific embodiments described above further illustrate the objects, technical solutions, and advantages of the present application. It should be understood that the above description is merely a specific embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

[0157] The above preset parameters or preset thresholds are set by a person skilled in the art according to actual conditions or obtained through a large amount of data simulation.

[0158] The above embodiments are merely used to illustrate the technical method of the present application and are not intended to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, a person of ordinary skill in the art should understand that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application.

Claims

1. A method for plant monitoring and security that fuses multiple systems, characterized by, The method comprises the following steps: Step 1: Collecting a subsystem information set, a historical monitoring data set and a historical abnormal label set in advance; Step 2: Building a correlation matrix calculation model based on the subsystem information set and the historical monitoring data set; Step 3: Extracting a historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model, and extracting an abnormal early warning feature vector from the historical correlation matrix; Step 4: Building and training an abnormal early warning judgment model based on the abnormal early warning feature vector and the historical abnormal label set; Step 5: Building and training a matrix prediction model using a time sequence of the historical correlation matrix; Step 6: Collecting real-time monitoring data in an actual monitoring process, obtaining a real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model, obtaining a future correlation matrix based on the real-time correlation matrix and the matrix prediction model, and obtaining a future early warning judgment based on the future correlation matrix and the abnormal early warning judgment model. The way of building the correlation matrix calculation model based on the subsystem information set and the historical monitoring data set is as follows: For each subsystem contained in the subsystem information set: According to the measurement data of each device of each subsystem in the historical monitoring data set, a device correlation matrix is constructed for each subsystem; The principal eigenvector of the device correlation matrix of each subsystem is extracted, and a subsystem correlation matrix between each subsystem is constructed based on the principal eigenvector of each subsystem.

2. The method of claim 1, wherein the method is implemented in a fusion multi-system plant monitoring and security system. The collection method of the subsystem information set is as follows: The basic information of each factory device in each subsystem is obtained as the subsystem information set; The collection method of the historical monitoring data set is as follows: The historical monitoring data generated by each subsystem and each device in each subsystem in the historical running process are extracted from the monitoring and data acquisition system of the factory or exported from the background database of the factory; The collection method of the historical abnormal label set is as follows: The abnormal label data is obtained each time an abnormal event occurs to form the historical abnormal label set by extracting or collecting from the operator event log or the abnormal report or from the event management or report system of the factory.

3. The method of claim 2, wherein the method further comprises: The way of constructing a device correlation matrix for each subsystem is as follows: The size of the time window T and the sliding step s are set; Each unit time in the historical monitoring data set is marked as t; The number of each subsystem is marked as i; The number of each device in the i-th subsystem is marked as ij; For each unit time t, the parameter set composed of the measurement data of the ij-th device in each subsystem from the unit time t-T to the unit time t is extracted to form a data window W(Sij, t); wherein Sij represents a parameter time sequence composed of the measurement data of the ij-th device in the i-th subsystem arranged in time sequence; The correlation between the devices in the i-th subsystem is calculated using a correlation coefficient analysis method to generate a device correlation matrix.

4. The method of claim 3, wherein the method further comprises: The way of extracting the principal eigenvector of the device correlation matrix of each subsystem is as follows: The principal eigenvector is obtained by singular value decomposition of the device correlation matrix.

5. The method of claim 4, wherein the method further comprises: The way of obtaining the principal eigenvector using the singular value decomposition method is as follows: The device correlation matrix is marked as A; The matrix A is transformed into the form using the method of singular value decomposition; wherein, and V are both orthogonal matrices, the column vectors of U are called left singular vectors, and each column vector of V is a right singular vector; Σ is a singular value matrix, which is a diagonal matrix containing singular values arranged in descending order; Select the largest singular value from the singular value matrix and take the left singular vector corresponding to the largest singular value as the principal eigenvector.

6. The method of claim 5, wherein the method further comprises: The method for constructing the subsystem correlation matrix between subsystems based on the principal feature vectors of each subsystem is as follows: Label any two subsystems as i1 and i2 respectively; The Pearson correlation coefficient formula is used to calculate the subsystem correlation degree between subsystem i1 and subsystem i2 using the principal eigenvectors of subsystem i1 and subsystem i2. A subsystem correlation matrix is ​​constructed based on the degree of correlation between each subsystem.

7. The method of claim 6, wherein the method further comprises: The method for constructing and training the anomaly warning judgment model based on the anomaly warning feature vector and the historical anomaly label set is as follows: Using an arbitrary classification model as the anomaly warning judgment model; The abnormal warning judgment model takes the abnormal warning feature vector as input and the predicted value of the abnormal type number of the time window corresponding to the abnormal warning feature vector as output. The anomaly warning judgment model uses the cross-entropy error between the predicted value of the anomaly type number and the actual value of the anomaly type number in the corresponding time window as the first error function. The anomaly warning judgment model is trained using the actual values ​​of the anomaly warning feature vectors and anomaly type numbers within all time windows of the historical monitoring process, until the first error function converges.

8. The converged plant monitoring and security method of claim 7, wherein, The method for constructing and training a matrix prediction model using time series data from historical correlation matrices is as follows: Set the matrix prediction model to an LSTM network model; The matrix prediction model takes the sequence of subsystem correlation matrices within a past time window as input; The output layer of the matrix prediction model outputs a prediction matrix of the subsystem correlation matrix within a time window after L time steps in the future; L is the preset prediction time step. The true subsystem correlation matrix after L time steps for each input is used as the prediction target of the prediction matrix, and the matrix cosine similarity between the true subsystem correlation matrix and the prediction matrix is ​​used as the second error function. The matrix prediction model is trained until the second error function converges.

9. A converged plant monitoring and safety system for implementing the converged plant monitoring and safety method of any one of claims 1-8, wherein, It includes a historical data collection module, an association matrix calculation module, an early warning model training module, a matrix prediction module, and a security early warning module; the modules are connected electrically. The historical data collection module pre-collects subsystem information sets, historical monitoring data sets, and historical anomaly label sets, and sends the subsystem information sets and historical monitoring data sets to the correlation matrix calculation module, and the historical anomaly label sets to the early warning model training module; The correlation matrix calculation module constructs a correlation matrix calculation model based on the subsystem information set and historical monitoring data set, and sends the correlation matrix calculation model to the early warning model training module. The early warning model training module extracts the historical correlation matrix from the historical monitoring data set using the correlation matrix calculation model, extracts the abnormal early warning feature vector from the historical correlation matrix, constructs and trains the abnormal early warning judgment model based on the abnormal early warning feature vector and the historical abnormal label set, and sends the historical correlation matrix to the matrix prediction module and the abnormal early warning judgment model to the safety early warning module. The matrix prediction module uses the time series of historical correlation matrices to build and train a matrix prediction model, and then sends the matrix prediction model to the security warning module. The safety early warning module collects real-time monitoring data during actual monitoring, obtains a real-time correlation matrix based on the real-time monitoring data and the correlation matrix calculation model, obtains a future correlation matrix based on the real-time correlation matrix and the matrix prediction model, and obtains an early warning judgment for the future based on the future correlation matrix and the anomaly early warning judgment model.

Images