Fault determination method and device and electronic equipment

Abstract

The invention discloses a fault determination method and device and electronic equipment. The method comprises the following steps: collecting original state data of a target system; performing phase-space reconstruction on the original state data to obtain a joint phase-space vector; reconstructing a target vector from the joint phase-space vector by using a variational auto-encoder; determining a fault monitoring index corresponding to the joint phase space vector and the target vector; and when the fault monitoring index is greater than a preset threshold value, determining that the target system has a fault. According to the method and the device, the technical problem that the fault monitoring technology adopted by the related technology is insensitive to early-stage and slowly-changing faults due to the fact that the complex non-linear relation and the dynamic characteristic which are generally existing in the industrial process are difficult to describe is solved.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and more specifically, to a method, apparatus, and electronic device for fault determination. Background Technology

[0002] Currently, data-driven fault monitoring methods have become a research hotspot. Among them, multivariate statistical process monitoring methods, such as principal component analysis (PCA) and partial least squares (PLS), are widely used because they do not require precise mechanistic models.

[0003] However, the fault monitoring techniques used in these technologies are difficult to characterize the complex nonlinear relationships that are common in industrial processes. Furthermore, they treat time-series data as independent samples, completely ignoring the dynamic characteristics of the system, which leads to insensitivity to early and slowly changing faults.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a method, apparatus, and electronic device for fault determination, which at least solves the technical problem that fault monitoring technologies used in related technologies are unable to characterize the complex nonlinear relationships and dynamic characteristics that are common in industrial processes, resulting in insensitivity to early and slowly changing faults.

[0006] According to one aspect of the embodiments of this application, a method for fault determination is provided, comprising: collecting raw state data of a target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions; performing phase space reconstruction on the raw state data to obtain a joint phase space vector, wherein the joint phase space vector is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables; reconstructing a target vector from the joint phase space vector using a variational autoencoder, wherein the target vector is used to characterize the normal state of the target system under operating conditions; determining a fault monitoring index corresponding to the joint phase space vector and the target vector, wherein the fault monitoring index is used to quantify the degree to which the state of the target system deviates from the normal state; and determining that the target system has failed when the fault monitoring index is greater than a preset threshold.

[0007] In some embodiments of this application, phase space reconstruction is performed on the original state data to obtain a joint phase space vector, including: determining reconstruction parameters, wherein the reconstruction parameters include a time delay parameter for determining the distance between two adjacent historical data points in the original state data, and an embedding dimension parameter for determining the dimension of the phase space vector of each variable; and converting the original state data into a joint phase space vector according to the reconstruction parameters.

[0008] In some embodiments of this application, converting the original state data into a joint phase space vector based on reconstruction parameters includes: at each time point, extracting multiple embedding vectors corresponding to multiple variables from the time series data corresponding to multiple variables, wherein the dimension of the embedding vector is determined by the embedding dimension parameter, and the distance between two adjacent elements in the embedding vector is determined by the time delay parameter; concatenating the embedding vectors corresponding to multiple variables to obtain an initial joint phase space vector, wherein the initial joint phase space vector is composed of the values ​​of multiple variables at the time point and the values ​​of multiple historical time points before the time point; and determining the joint phase space vector based on the initial joint phase space vectors corresponding to all time points within the time window.

[0009] In some embodiments of this application, the variational autoencoder is trained in the following manner: obtaining a training dataset, wherein the training dataset includes a historical joint phase space vector corresponding to historical state data, wherein each element in the historical joint phase space vector is used to characterize the historical state of the target system at each historical moment; processing the training dataset with an initial variational autoencoder to obtain a first reconstructed vector; determining the loss between the historical joint phase space vector and the first reconstructed vector; and adjusting the parameters of the initial variational autoencoder according to the loss to obtain the variational autoencoder.

[0010] In some embodiments of this application, determining the loss between the historical joint phase space vector and the first reconstructed vector includes: determining a first loss between the historical joint phase space vector and the first reconstructed vector, wherein the first loss is used to quantify the degree of difference between the historical joint phase space vector and the first reconstructed vector; determining a second loss to quantify the difference between the posterior distribution generated by the initial variational autoencoder and a preset prior distribution, wherein the posterior distribution includes the probability distribution of latent variables inferred from the training dataset, and the prior distribution includes a normal distribution; and determining the loss based on the first loss and the second loss.

[0011] In some embodiments of this application, a variational autoencoder is used to reconstruct a target vector from a joint phase space vector, including: using the encoder of the variational autoencoder to perform a nonlinear transformation on the joint phase space vector to obtain probability distribution parameters of the latent space, wherein the probability distribution parameters include a mean vector and a log-variance vector, and the latent space is used to encode the normal state of the target system; using the variational autoencoder to sample from the latent space according to the probability distribution parameters to obtain latent variables, wherein the latent variables are used to characterize the features of the normal state; and using the decoder of the variational autoencoder to reconstruct the latent variables to obtain the target vector.

[0012] In some embodiments of this application, the preset threshold is determined in the following way: a variational autoencoder is used to reconstruct each historical joint phase space vector in the training dataset to obtain a second reconstructed vector; the squared prediction error between each second reconstructed vector and the corresponding historical joint phase space vector is determined; and the preset threshold is determined based on the squared prediction errors corresponding to all historical joint phase space vectors in the training dataset.

[0013] According to another aspect of the embodiments of this application, a fault determination apparatus is also provided, comprising: a data acquisition module for acquiring raw state data of a target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions; a reconstruction module for performing phase space reconstruction on the raw state data to obtain a joint phase space vector, wherein the joint phase space vector is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables; a parsing module for reconstructing a target vector from the joint phase space vector using a variational autoencoder, wherein the target vector is used to characterize the normal state of the target system under operating conditions; a determination module for determining fault monitoring indicators corresponding to the joint phase space vector and the target vector, wherein the fault monitoring indicators are used to quantify the degree to which the state of the target system deviates from the normal state; and a judgment module for determining that the target system has failed when the fault monitoring indicators are greater than a preset threshold.

[0014] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the above-described fault determination method.

[0015] According to another aspect of the embodiments of this application, a non-volatile storage medium is also provided, the non-volatile storage medium including a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned fault determination method by running the computer program.

[0016] According to another aspect of the embodiments of this application, a computer program product is also provided, including computer instructions that, when executed by a processor, implement the above-described fault determination method.

[0017] In this embodiment, by collecting multivariate time series data of the target system under operating conditions, reconstructing the phase space to obtain a joint phase space vector, and using a variational autoencoder to learn a low-dimensional representation of the normal state, the goal of accurately characterizing the dynamic characteristics and nonlinear relationships of the system is achieved. This realizes the technical effect of high-sensitivity monitoring and early warning of early and slowly changing faults, thereby solving the technical problem that the fault monitoring technology used in related technologies is difficult to characterize the complex nonlinear relationships and dynamic characteristics that are common in industrial processes, resulting in insensitivity to early and slowly changing faults. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0019] Figure 1 This is a hardware structure block diagram of a computer terminal for a fault determination method according to an embodiment of this application;

[0020] Figure 2 This is a flowchart of a fault determination method according to an embodiment of this application;

[0021] Figure 3 This is an overall flowchart of a fault determination method according to an embodiment of this application;

[0022] Figure 4 This is a flowchart of a multivariable phase space reconstruction method for fault determination according to an embodiment of this application;

[0023] Figure 5 This is a flowchart of a variational autoencoder training method for fault determination according to an embodiment of this application;

[0024] Figure 6 This is a statistical distribution diagram of SPE (Solution Processing) of a fault determination method according to an embodiment of this application;

[0025] Figure 7 This is a schematic diagram illustrating the determination of embedded dimension parameters in a fault determination method according to an embodiment of this application.

[0026] Figure 8 This is a schematic diagram illustrating the determination of time delay parameters in a fault determination method according to an embodiment of this application;

[0027] Figure 9 This is a schematic diagram of a fault determination device according to an embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] To better understand the embodiments of this application, the technical terms involved in the embodiments of this application are explained below:

[0031] Phase space reconstruction of multivariate time series: In this embodiment, nonlinear time series analysis based on Takens' embedding theorem maps historical observation data of multiple process variables to a high-dimensional state space, constructing state vectors to reveal the dynamic characteristics of the system and the interactions between variables. Phase space reconstruction technology is used to convert the collected raw state data into a joint phase space vector containing rich dynamic information and coupling relationships, providing a data foundation for subsequent fault monitoring models. It should be noted that each reconstructed data point represents the complete "state" of the system at a certain moment, rather than an isolated measurement value.

[0032] Joint phase space reconstruction: In one embodiment of this application, multivariate time series phase space reconstruction is performed simultaneously on multiple variables to form a unified state vector that can simultaneously describe the dynamic characteristics of each variable and their mutual coupling relationships. The joint phase space vector serves as the input data for the variational autoencoder. It not only preserves the temporal dynamics of each variable but also incorporates the correlation between multiple variables, thereby improving the comprehensiveness and sensitivity of fault detection.

[0033] Normal-State Manifold Learning Based on Variational Autoencoders: In this embodiment, a deep generative model, the variational autoencoder, is used to perform nonlinear dimensionality reduction and probabilistic modeling on data under normal operating conditions, learning a low-dimensional, probabilistic normal-state manifold (which is a concentrated representation of all normal operating states in the latent space). The variational autoencoder learns the probability distribution of normal states from the joint phase space vectors, and the generated target vector represents the manifold of normal system operation, serving as a reference benchmark for subsequent fault monitoring.

[0034] Squared Prediction Error Monitoring Statistic (SPE): In this embodiment, SPE measures the difference between online monitoring data reconstructed by a variational autoencoder and the original data, and is used to monitor the occurrence of faults. SPE is used as a real-time calculated fault monitoring indicator. By comparing its value with a preset threshold, it can determine whether the target system deviates from its normal state, thus achieving early warning of faults. For example, it can specifically be the sum of squares of the differences between newly acquired online data reconstructed in phase space and data reconstructed by a VAE model. This error directly reflects the degree to which the current system state deviates from the "normal state manifold" learned by the VAE.

[0035] Dynamic and Static Feature Fusion Monitoring: In this embodiment, dynamic features extracted by phase space reconstruction and static features learned by variational autoencoders are fused and applied during the monitoring process. The fusion of dynamic and static features makes the SPE index more comprehensive, capable of simultaneously capturing anomalies in both dynamic changes and static distributions of the system, thereby improving the accuracy and robustness of fault monitoring. For example, it can simultaneously utilize two types of information from the process: first, the system dynamic evolution features extracted through phase space reconstruction (the patterns of change over time); and second, the system static distribution features extracted through VAE (the probability distribution structure of normal data), thus achieving more comprehensive and sensitive monitoring.

[0036] Fault monitoring / fault detection: A technology that identifies and diagnoses whether a system deviates from its normal operating mode by analyzing process data, and locates faults or abnormal states. In this application embodiment, fault monitoring is the ultimate goal of the entire technical solution. By determining that the SPE index exceeds a preset threshold, automatic detection and early warning of faults are achieved.

[0037] Manifold: A mathematical concept referring to a topological space that is locally homeomorphic to Euclidean space. In the embodiments of this application, the normal state manifold refers to a low-dimensional, continuous region in the latent space of the variational autoencoder that represents the normal operating state. This manifold is the probabilistic set of all normal state vectors in the latent space.

[0038] Latent Space: In a variational autoencoder, the latent space is the low-dimensional space to which the encoder maps the input data. It captures the key features and latent structure of the input data. In this embodiment, the latent space is the result of the variational autoencoder performing nonlinear feature compression on the joint phase space vector, and is used to extract hidden representations describing the dynamic nature of the system from the data.

[0039] Variational Autoencoder (VAE): In this embodiment of the application, a generative deep learning model consists of an encoder and a decoder. The encoder maps the input data to a probabilistic latent space, and the decoder reconstructs the data from the latent space.

[0040] In the fields of industrial automation and intelligent manufacturing technology, especially for fault monitoring and health management of complex industrial systems, traditional multivariate statistical process monitoring methods (such as PCA and PLS) have long held a dominant position. However, the inherent limitations of these methods are becoming increasingly apparent, specifically:

[0041] (1) Limitation of linear assumption: PCA and other methods are essentially linear transformation techniques, which are difficult to effectively characterize the complex nonlinear relationships that are common in industrial processes, resulting in a decline in the monitoring performance of nonlinear processes.

[0042] (2) Ignoring dynamic characteristics: Traditional methods usually treat the data at each sampling time as independent and identically distributed static samples, completely ignoring the important temporal correlation and dynamic evolution laws contained in industrial process data, which makes them insensitive to early and slowly changing faults.

[0043] (3) Limited feature extraction capability: For high-dimensional, non-Gaussian distributed data, linear methods such as PCA have insufficient feature extraction capability and cannot learn the complex probability distribution structure (i.e., "manifold") at the bottom layer of the data, resulting in low accuracy and robustness of fault detection.

[0044] Although some studies have attempted to introduce deep learning models (such as standard autoencoders, AE) or consider temporal models (such as recurrent neural networks, RNN), they still have shortcomings. Standard AE lacks probabilistic interpretation and does not clearly characterize the boundaries of normal states; although RNN can handle temporal sequences, its "black box" characteristics are too strong, and it faces challenges in capturing long-term dependencies and training stability.

[0045] To address the aforementioned technical problems, this application provides corresponding solutions, which are detailed below.

[0046] The fault determination method provided in this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal for implementing a fault determination method is shown. Figure 1As shown, the computer terminal 10 may include one or more processors (shown as 102a, 102b, ..., 102n in the figure) (the processor may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission module 106 for communication functions connected via wired and / or wireless networks. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0047] It should be noted that the aforementioned one or more processors and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10. As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0048] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the fault determination method in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned fault determination method. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0049] The transmission module 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission module 106 may be a radio frequency (RF) module, used for wireless communication with the Internet.

[0050] The display can be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10.

[0051] It should be noted here that, in some optional embodiments, the above... Figure 1 The computer terminal shown may include hardware elements (including circuitry), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware and software elements. It should be noted that... Figure 1 This is only one instance of a specific particular instance, and is intended to illustrate the types of components that may exist in the aforementioned computer terminal.

[0052] In the above operating environment, this application provides a method embodiment for fault determination. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0053] Figure 2 This is a flowchart of a fault determination method according to an embodiment of this application, such as... Figure 2 As shown, the method includes the following steps:

[0054] Step S202: Collect the raw state data of the target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions.

[0055] In step S202 above, the target system refers to the industrial process or equipment that needs to be monitored for faults in this application. For example, in a chemical production scenario, the target system can be a high-temperature and high-pressure reactor and its related control system, which is monitored by multiple sensors. Each sensor measures a specific variable, such as temperature, pressure, flow rate, etc.

[0056] Raw state data refers to the raw sensor readings or operational records of the target system under normal or specific operating conditions. For example, it can be multivariate time series data. This data is collected directly from the industrial site without any preprocessing and contains real-time state information of the system during operation, such as time series data of reactor internal temperature and curves of cooling water flow rate over time.

[0057] Operating conditions refer to the environment or settings parameters during the operation of an industrial process or equipment, which can affect the system state. Operating conditions include, but are not limited to, temperature setpoint, pressure level, feed rate, etc. The raw state data collected is the dynamic response under these conditions.

[0058] In some embodiments of this application, a sensor network in the industrial field can be used to continuously record real-time readings of each process variable. For example, temperature sensors can be used to record the internal temperature of a reactor, and flow meters can be used to record the flow rate of cooling water. The outputs of these sensors are typically analog signals, which need to be converted into digital signals by a data acquisition card. Furthermore, historical data can be extracted from the log files of the process control system. This historical data records the specific values ​​of process variables at different points in time.

[0059] After collecting the raw state data, preprocessing operations such as standardization, noise reduction, and missing value handling can be performed on the data to form a standardized multivariate time series dataset X. raw Let the number of process variables be k and the number of sampling points be N, then X raw It is an N×k matrix.

[0060] Step S204: The original state data is reconstructed in phase space to obtain a joint phase space vector, which is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables.

[0061] In step S204 above, phase space reconstruction, based on Takens embedding theory, is the process of converting one-dimensional time series data into a multi-dimensional state vector. The aim is to restore the original dynamic structure of the system, capturing the nonlinear dynamic relationships and interactions between variables within the system. The joint phase space vector integrates the time series data of all monitored variables to form a high-dimensional vector that comprehensively describes the dynamic state of the system, providing a comprehensive state representation for subsequent fault monitoring.

[0062] Temporal dynamic characteristics refer to the system state characteristics that change over time, such as trends, periodicity, and time delay effects. Coupling relationships refer to the degree and manner in which variables within the system influence each other. In this embodiment, by combining phase space vectors, these characteristics and relationships can be captured simultaneously.

[0063] To address the challenge of traditional fault detection methods in capturing complex dynamic characteristics and inter-variable coupling relationships in nonlinear dynamic systems, the original state data can be reconstructed into a joint phase space vector using the following method: Determine the reconstruction parameters, including a time delay parameter for determining the distance between two adjacent historical data points in the original state data, and an embedding dimension parameter for determining the dimension of the phase space vector for each variable; then, based on the reconstruction parameters, convert the original state data into a joint phase space vector.

[0064] It should be noted that in phase space reconstruction, the time delay parameter (τ) determines the interval between adjacent data points extracted from the time series, and the embedding dimension parameter (m) refers to the dimension of the vector generated in phase space reconstruction, that is, how many historical points are used to represent the current state.

[0065] In some embodiments of this application, the mutual information method can be used to determine the time delay τ. Specifically, for the time series data of each variable, the mutual information (i.e., the information correlation between the current value and the value τ steps ago) is calculated. The choice of τ should make the mutual information decrease significantly for the first time, which means that the data before τ steps no longer provides additional information, thereby avoiding data redundancy.

[0066] For the embedding dimension, the False Nearest Neighbor Method can be used to determine it. This method observes how points that were originally considered close to each other in the phase space (false neighbors) gradually separate as m increases. When the number of false neighbors decreases to a certain level (e.g., less than 5% of the total number of data points), m is considered appropriate.

[0067] After determining the reconstruction parameters, for each variable x(t) of the target system, a phase space vector X(t) = [x(t), x(t-τ), ..., x(t-(m-1)τ)] is constructed using a determined time delay τ and embedding dimension m. Then, X(t) of all variables is concatenated to form a joint phase space vector Y(t).

[0068] To facilitate understanding of the process of determining the above reconstruction parameters, the following explanation is based on some specific embodiments: Two key parameters are selected for phase space reconstruction, namely time delay τ and embedding dimension m.

[0069] Determining the time delay τ: For each variable, the mutual information method can be used for automatic calculation. Specifically, the mutual information between the time series x(t) and its sequence x(t+τ) after a delay of τ steps is calculated, and the τ corresponding to the first mutual information minimum point is the optimal time delay. For multivariate systems, the optimal τ for each variable can be calculated separately, and then the average or maximum value can be taken as the global τ; alternatively, a unified τ can be determined starting from the key variables.

[0070] Determining the embedding dimension m: This can be automatically calculated using the spurious nearest neighbor method. The basic idea is that when the embedding dimension is insufficient, some neighboring points in the phase space (spurious neighbors) are due to projection onto a lower-dimensional space. As the embedding dimension m increases, the proportion of spurious neighbors decreases. When the proportion of spurious neighbors decreases below a certain threshold (e.g., 5%) or tends to stabilize, the corresponding m is the appropriate embedding dimension.

[0071] Ultimately, a set of reconstruction parameters (m,τ) can be obtained.

[0072] Figure 7 This is a schematic diagram illustrating the determination of embedding dimension parameters in a fault determination method according to an embodiment of this application. Figure 8 This is a schematic diagram illustrating the determination of time delay parameters in a fault determination method according to an embodiment of this application, as shown below. Figure 7 and Figure 8 As shown, the values ​​corresponding to the dashed lines are the determined reconstruction parameter values. It should be noted that... Figure 7 In this context, the FNN ratio refers to the False Nearest Neighbor Ratio, which is the proportion of false nearest neighbors to the total number of data points.

[0073] In some embodiments of this application, the original state data can be converted into a joint phase space vector based on the reconstruction parameters through the following steps: at each time point, multiple embedding vectors corresponding to multiple variables are extracted from the time series data corresponding to multiple variables, wherein the dimension of the embedding vector is determined by the embedding dimension parameter, and the distance between two adjacent elements in the embedding vector is determined by the time delay parameter; the embedding vectors corresponding to multiple variables are concatenated to obtain an initial joint phase space vector, wherein the initial joint phase space vector is composed of the values ​​of multiple variables at the time point and the values ​​of multiple historical time points before the time point; the joint phase space vector is determined based on the initial joint phase space vectors corresponding to all time points within the time window.

[0074] An embedding vector is a vector composed of a segment of historical data selected based on the embedding dimension m and time delay τ in the time series of a variable. It is used to represent the local dynamic characteristics near the current state.

[0075] Specifically, for any time point t (where the value of t ranges from (m-1)... From τ+1 to N, a joint phase space vector Y(t) can be constructed, which is the initial joint phase space vector. Y(t) is composed of the values ​​of all k variables at the current time t and their respective historical times. Its mathematical expression is as follows:

[0076]

[0077] in, Let represent the value of the i-th variable at time t, and let the dimension of this vector be D = k. m.

[0078] By performing the above operation on each valid time point t using a sliding time window, an N × k original data matrix can be transformed into an M × D phase space state matrix Y (i.e., a joint phase space vector), where M = N - (m-1). τ is the number of valid state vectors.

[0079] Figure 4 This is a flowchart of a multivariable phase space reconstruction method for fault determination according to an embodiment of this application, as shown below. Figure 4 As shown:

[0080] Step S401, Input: The input is the preprocessed multivariate time series. .

[0081] Step S402, Parameter module: The module marked "parameter (m, τ)" indicates that the optimal reconstruction parameters determined in the above process are used.

[0082] Step S403, Reconstruction Process: The time series of each variable is fed into a "DelayStack", and the embedding vector of the variable is generated according to the parameters m and τ. .

[0083] Step S404, joint splicing: The embedding vectors of all k variables are spliced ​​together to form the final joint phase space vector Y(t).

[0084] Step S405, Output: The output is the complete phase space state matrix, Y, with dimensions M × (k m).

[0085] Through the above steps, data from a single point in time can be transformed into "state" data containing historical information, enabling the model to perceive the dynamic evolution of the system, rather than just a static snapshot. Furthermore, by constructing a joint phase space vector, the interrelationships between multiple variables are naturally embedded into the state representation, providing a more comprehensive description of the overall behavior of the system. Phase space reconstruction is the cornerstone of analyzing nonlinear dynamic systems, providing the most suitable data foundation for subsequent nonlinear models such as VAEs to learn complex normal state manifolds.

[0086] Step S206: A variational autoencoder is used to reconstruct the target vector from the joint phase space vector, wherein the target vector is used to characterize the normal state of the target system under operating conditions.

[0087] In step S206 above, the variational autoencoder (VAE) learns and memorizes the dynamic characteristics of the target system under normal operating conditions. Its target vector can represent the normal state manifold. In the VAE, the decoder generates an output similar to the input data (here, the joint phase space vector) based on the latent variable z generated by the encoder; this is the target vector. The reconstructed target vector is compared with the original joint phase space vector to define the reconstruction error. The target vector is the reconstructed output generated by the decoder after the VAE training that best matches the joint phase space vector under normal conditions; it can represent the state pattern of the system during normal operation.

[0088] In some embodiments of this application, the VAE may include encoder and decoder modules of a multilayer neural network. For example, the encoder may be a deep neural network for extracting latent variables z from the joint phase space vector; the decoder may also be a deep neural network for transforming the latent variables z back to an output similar to the joint phase space vector.

[0089] To effectively identify and reconstruct the normal state of the target system, the variational autoencoder is trained as follows: A training dataset is obtained, which includes historical joint phase space vectors corresponding to historical state data. Each element in the historical joint phase space vector represents the historical state of the target system at each historical moment. The training dataset is processed using an initial variational autoencoder to obtain a first reconstructed vector. The loss between the historical joint phase space vector and the first reconstructed vector is determined. The parameters of the initial variational autoencoder are adjusted based on the loss to obtain the variational autoencoder.

[0090] Specifically, (1) In order to ensure that the model learns the normal operating mode of the system rather than the fault mode, historical state data under normal operating conditions can be extracted from historical data, and phase space can be reconstructed to obtain a series of historical joint phase space vectors. For example, multivariate time series can be extracted from the fault-free operation data of the past few months to construct joint phase space vectors.

[0091] (2) The historical joint phase space vector is fed into the encoder of the VAE to obtain a low-dimensional latent variable representation z; then z is fed into the decoder to generate the first reconstructed vector. For example, if the dimension of the historical joint phase space vector is 20, it can be converted into a 6-dimensional z by the VAE and then reconstructed into a 20-dimensional first reconstructed vector.

[0092] (3) Calculate the reconstruction error (e.g., MSE) and KL divergence. The reconstruction error measures the difference between the first reconstruction vector and the original historical joint phase space vector, while the KL divergence ensures that the distribution of the latent variable z is close to the preset prior distribution (e.g., standard normal distribution).

[0093] In some embodiments of this application, the loss between the historical joint phase space vector and the first reconstructed vector can be determined in the following manner: determining a first loss between the historical joint phase space vector and the first reconstructed vector, wherein the first loss is used to quantify the degree of difference between the historical joint phase space vector and the first reconstructed vector; determining a second loss to quantify the difference between the posterior distribution generated by the initial variational autoencoder and a preset prior distribution, wherein the posterior distribution includes the probability distribution of latent variables inferred from the training dataset, and the prior distribution includes a normal distribution; and determining the loss based on the first loss and the second loss.

[0094] Specifically, the first loss quantifies the degree of difference between the historical joint phase space vector and the first reconstructed vector. For example, the mean squared error (MSE) can be used, which can measure the distance between the reconstructed data output by the decoder and the actual normal data. The second loss is used to quantify the difference between the posterior distribution generated by the initial variational autoencoder and the preset prior distribution (usually a normal distribution). The second loss can be achieved, for example, by calculating the KL divergence, to ensure that the latent space distribution is regular and to avoid the model overfitting the training data.

[0095] It should be noted that the posterior distribution is the probability distribution of the latent variable z output by the encoder part of the variational autoencoder. It reflects the model's specific understanding of the input data (historical joint phase space vector). The posterior distribution can be obtained, for example, by the mean μ and log-variance log generated by the encoder. var The parameters describe the prior distribution, which is the distribution of the latent variable z that is pre-defined before training the model. For example, it can be simplified to the standard normal distribution N(0, I), where I is the identity matrix.

[0096] To facilitate understanding of the loss determination process described above, the following explanation will be provided in conjunction with some specific embodiments.

[0097] 1) Reconstruction Loss:

[0098] To measure the difference between the decoder's reconstructed output Y'(t) and the original input Y(t), for continuous data, the mean squared error (MSE) can be used: L recon = ||Y(t) - Y'(t)||².

[0099] Its physical significance lies in ensuring that the VAE can accurately recover the normal system dynamic state from the latent representation.

[0100] 2) KL Divergence (Kullback-Leibler Divergence):

[0101] Measuring the distribution of latent variables generated by the encoder The difference between (z|Y) and the prior distribution (such as the standard normal distribution p(z) = N(0, I)) is calculated using the formula: L KL = D KL ( (z|Y) || p(z)).

[0102] This means regularizing the latent space to make its distribution regular and continuous, avoiding overfitting, and ensuring that each region of the latent space corresponds to a meaningful output.

[0103] 3) Total loss function: L total = L recon + β·L KL , where β is a hyperparameter used to balance reconstruction accuracy and latent space regularity.

[0104] (4) Use the backpropagation algorithm and gradient descent method to update the encoder and decoder parameters of the VAE according to the gradient of the loss function. For example, adjust the network weights through the Adam optimizer until the loss function converges or reaches the predetermined training rounds.

[0105] In some embodiments of this application, the phase space state matrix Y can be randomly divided into a training set and a validation set, and the network weights of the encoder and decoder can be initialized. For a batch of training data Y... batch :

[0106] μ and log are obtained through the encoder. var Use the reparameterization technique to sample latent variables: z = μ + ε exp(0.5 log var ), where ε ~ N(0, I), this step allows the gradient to propagate back through random nodes; input z into the decoder to obtain the reconstructed output Y' batch .

[0107] Subsequently, the total loss L for this batch was calculated. total And update the parameters of the encoder and decoder using the backpropagation algorithm. And θ, repeat the above process until the model's loss on the validation set converges.

[0108] Figure 5 This is a flowchart of a variational autoencoder training method for fault determination according to an embodiment of this application, as shown below. Figure 5 As shown, it includes:

[0109] Step S501, Input: The input is the joint phase space vector Y(t).

[0110] Step S502, encoder: Y(t) is fed into encoder network (usually composed of fully connected layers and activation functions), outputting the parameters μ and log_var of the latent distribution.

[0111] Step S503, sampling point: There is a random node marked "ε ~ N(0, I)", which is combined with μ and log_var through the reparameterization technique to generate the latent variable z.

[0112] Step S504, Decoder: The latent variable z is fed into the decoder network to reconstruct Y'(t).

[0113] Step S505, Loss Calculation: Input the original Y(t) and the reconstructed Y'(t) into the "Loss Function" module, and calculate L. recon and L KL The Loss_total is obtained and fed back to the network for optimization.

[0114] Step S506, Output: After training is completed, the encoder and decoder are fixed for subsequent online monitoring.

[0115] Through the above steps, the joint phase space matrix Y obtained from phase space reconstruction is directly used as the input of the variational autoencoder. Phase space reconstruction solves the problem of what data to use. It transforms the original, static, and isolated sensor readings into dynamic vectors that can describe the motion state of the system. VAE training solves the problem of how to learn normal patterns. It uses the powerful capabilities of deep learning to extract the most essential, low-dimensional health state features from these dynamic vectors and build a powerful generative model.

[0116] This application combines nonlinear dynamical system theory (phase space reconstruction) with modern deep learning (variable autoencoder). Phase space reconstruction provides VAEs with the most suitable data format for analyzing dynamic systems, while VAEs provide the most powerful fitting and probabilistic modeling tools to date for complex manifolds in phase space. The combination of the two achieves a "1+1>2" effect, making the model extremely sensitive to small anomalies in dynamic processes. VAEs can learn highly complex "normal state manifolds," far exceeding the capabilities of traditional linear methods (such as PCA). Furthermore, VAEs learn the probability distribution of data, rather than simple point estimation, which gives them a more natural characterization of the boundaries of normal states, enhancing the robustness of monitoring. In addition, the latent space provides a low-dimensional, meaningful feature space that can be used for subsequent fault diagnosis and analysis.

[0117] To extract key, low-dimensional latent features from a high-dimensional joint phase space vector, a variational autoencoder (VAE) can be used to reconstruct the target vector from the joint phase space vector through the following steps: The encoder of the VAE performs a nonlinear transformation on the joint phase space vector to obtain the probability distribution parameters of the latent space, which include a mean vector and a log-variance vector. The latent space is used to encode the normal state of the target system. The VAE samples from the latent space based on the probability distribution parameters to obtain latent variables, which are used to characterize the features of the normal state. The decoder of the VAE reconstructs the latent variables to obtain the target vector.

[0118] In some embodiments of this application, the VAE model consists of three parts: an encoder, a latent space, and a decoder.

[0119] (1) Encoder: It consists of a deep neural network whose input is a joint phase space vector Y(t) (dimension D = k × m). The encoder maps the high-dimensional input data to the probability distribution parameters of the latent space through multiple nonlinear transformations. Specifically, the encoder outputs two vectors:

[0120] Mean vector μ; Logarithmic variance vector log var (Log-Variance Vector).

[0121] These two vectors together define the posterior distribution of the latent variable z. (z|Y), where These are the network parameters for the encoder.

[0122] (2) Latent Space: is a low-dimensional, continuous space (dimension d, usually d << D). The latent variable z is sampled from this space. It is a compact and meaningful representation of the original high-dimensional dynamic state and captures the essential characteristics of the normal operation of the system.

[0123] (3) Decoder: Also composed of a deep neural network, it takes the latent variable z sampled from the latent distribution as input and strives to reconstruct the output Y'(t) as similar as possible to the original input Y(t). The decoder learns the conditional distribution p. θ (Y|z), where θ is the network parameter of the decoder.

[0124] Specifically, the joint phase space vector can be input into the encoder network, which is a deep model containing multiple neural networks, typically including fully connected layers, activation functions (such as ReLU, tanh), and a final output layer. The output layer provides the probability distribution parameters (μ and log) of the latent variable z. var ); using the mean μ and logarithmic variance log of the encoder output. var By employing the reparameterization trick, the latent variable z is sampled in the latent space: z = μ + ε·exp(0.5·log var ), where ε ~ N(0, I); The sampled latent variable z is input into the decoder network. The decoder is also composed of a multi-layer neural network. Its goal is to transform z back into an output similar to the original joint phase space vector, i.e., the target vector.

[0125] Step S208: Determine the fault monitoring index corresponding to the joint phase space vector and the target vector, wherein the fault monitoring index is used to quantify the degree to which the state of the target system deviates from the normal state.

[0126] In step S208 above, the fault monitoring index is defined as the degree of difference between the calculated joint phase space vector and the target vector reconstructed by the VAE, which is used to assess the deviation between the current state and the normal state of the system.

[0127] In some embodiments of this application, the squared prediction error (SPE) can be calculated as a fault monitoring metric. SPE is determined by comparing the joint phase space vectors. and target vector The calculation is based on the difference between them, and the formula is as follows:

[0128]

[0129] The SPE index solves the problem of how to quantify the deviation between the system state and the normal state in order to achieve early fault detection. In industrial process fault monitoring, SPE can capture minute changes in the system state caused by slow changes in equipment performance, which is difficult to achieve with traditional threshold monitoring methods.

[0130] Step S210: If the fault monitoring index is greater than the preset threshold, it is determined that the target system has failed.

[0131] In step S210 above, the preset threshold is a fault judgment standard determined based on the SPE control limit, used to distinguish between normal state and potential fault state. When the fault monitoring index (such as SPE value) exceeds this threshold, the system determines that a fault has occurred.

[0132] In some embodiments of this application, the preset threshold is determined in the following way: a variational autoencoder is used to reconstruct each historical joint phase space vector in the training dataset to obtain a second reconstructed vector; the squared prediction error between each second reconstructed vector and the corresponding historical joint phase space vector is determined; and the preset threshold is determined based on the squared prediction errors corresponding to all historical joint phase space vectors in the training dataset.

[0133] Specifically, after successfully training the VAE model, which can accurately reconstruct the system's dynamic state under normal operating conditions, a statistical quantification threshold for fault determination, namely the SPE control limit, can be determined based on the trained VAE model and normal training data (i.e., historical state data under normal operating conditions). This threshold includes:

[0134] (1) Calculate the reconstruction error of the training set.

[0135] The phase space state matrix Y under all normal operating conditions train (i.e., the training dataset) is input again into the already trained VAE model. For each sample Y(i) in the training set, the following operations are performed:

[0136] The reconstructed output Y'(i) is obtained through an encoder-decoder process, and the squared prediction error (SPE) of the sample is calculated:

[0137]

[0138] Where D is the dimension of the phase space vector (D = k × m). and These are the j-th elements of the original vector and the reconstructed vector, respectively.

[0139] (2) Establish the empirical distribution of the SPE statistic.

[0140] Collect the SPE values ​​of all training samples {SPE(1), SPE(2), ..., SPE(M)}, where M is the number of training samples. These SPE values ​​constitute the empirical distribution of the reconstruction error under normal operating conditions, which describes the inherent and acceptable fluctuation range of the target system under "healthy conditions".

[0141] (3) Determine the upper limit of control (control limit).

[0142] Based on the obtained empirical distribution of SPE, select an appropriate statistical quantile to set the control upper limit SPE_lim, for example:

[0143]

[0144] in, Represents the empirical distribution of SPE Quantiles It is a small significance level (e.g., =0.01 or 0.05).

[0145] This means that under normal, fault-free operating conditions, there are (e.g., 99% or 95%) of the samples will have SPE values ​​falling within the control limits. Therefore, once the SPE value exceeds this threshold during online monitoring, there is a high degree of confidence that the process has become abnormal.

[0146] During application, the real-time SPE value can be... Control limits determined during the offline phase Comparison:

[0147] (1) If ≤ If so, the judgment process is in a normal state.

[0148] (2) If > If so, the judgment process has malfunctioned.

[0149] Figure 6 This is a SPE statistical distribution diagram of a fault determination method according to an embodiment of this application, such as... Figure 6 As shown, the horizontal axis represents the SPE value, and the vertical axis represents the probability density or sample frequency. The curve is an empirical distribution curve that approximates a chi-square distribution or a skewed distribution, representing the distribution of normal SPE values. The straight line perpendicular to the horizontal axis is marked as... Located at the high quantile of the distribution, in The area on the left can be marked as the normal area, and the area on the right can be marked as the fault alarm area.

[0150] In some embodiments of this application, when a fault is detected, the system immediately triggers an alarm signal (such as a message push) and records detailed information about the fault, including: the timestamp t of the fault occurrence and the exceeded SPE value. , corresponding original state data and joint phase space vector .

[0151] Through steps S202 to S210, multivariate time series data of the target system under operating conditions are collected, phase space is reconstructed to obtain a joint phase space vector, and a low-dimensional representation of the normal state is learned using a variational autoencoder. This achieves the goal of accurately characterizing the dynamic characteristics and nonlinear relationships of the system, thereby realizing the technical effect of high-sensitivity monitoring and early warning of early and slowly changing faults. This solves the technical problem that the fault monitoring technology used in related technologies is difficult to characterize the complex nonlinear relationships and dynamic characteristics that are common in industrial processes, resulting in insensitivity to early and slowly changing faults.

[0152] Figure 3 This is an overall flowchart of a fault determination method according to an embodiment of this application, such as... Figure 3 As shown, the specific architecture of this application embodiment is divided into two stages: offline modeling and online monitoring. It includes a closed-loop process of how to determine control limits in the offline stage and how to calculate SPE in real time and make fault decisions in the online stage.

[0153] In some embodiments of this application, the offline monitoring phase, specifically the offline training phase, may include the following steps:

[0154] (1) Start, that is, start the data acquisition system and prepare to obtain multivariate time series data under normal working conditions from the historical database.

[0155] (2) Data acquisition and preprocessing: Obtain multivariate data under normal operating conditions from historical data, such as reactor temperature, cooling water flow rate, cooling water inlet temperature, feed flow rate and reactor pressure. Then, standardize, denoise and process missing values ​​for these raw data to ensure the quality of training data.

[0156] (3) Multivariable phase space reconstruction, including: according to the phase space reconstruction theory, determine the appropriate embedding dimension m and time delay τ, and use these parameters to convert the preprocessed multivariable time series data into a series of joint phase space vectors. Each vector contains a set of states of multiple variables at different time points, reflecting the complete dynamic characteristics of the system at a certain moment.

[0157] (4) VAE model training, including: constructing a variational autoencoder model containing an encoder and a decoder, training only with normal data, and adjusting the model parameters to enable the encoder to learn the latent space representation of the joint phase space vector under normal operating conditions, and to enable the decoder to accurately reconstruct the joint phase space vector based on the latent variable z. The training process simultaneously minimizes the reconstruction error and KL divergence to ensure that the model learns a normal state manifold that is both accurate and regular.

[0158] (5) Calculate the SPE control limit. After the VAE model is trained, use normal training data to calculate the empirical distribution of the squared prediction error (SPE) value. Based on the high quantile point (such as 99%) of the empirical distribution, determine an SPE control limit, which is used as the fault warning threshold in the online monitoring stage.

[0159] In the online monitoring phase, the model and thresholds established in the offline phase are used to continuously assess the health status of the real-time industrial process. This is a cyclical process. In some embodiments of this application, the online monitoring phase includes the following steps:

[0160] (1) Real-time data acquisition and preprocessing.

[0161] Real-time acquisition of the latest multivariable sensor data from industrial process control systems. The real-time data is processed using preprocessing methods (such as standardization) to ensure that the data scale is consistent with that of the training phase.

[0162] (2) Real-time multivariable phase space reconstruction.

[0163] Using the reconstruction parameters (m, τ) determined in the offline phase, the phase space of the preprocessed real-time data window is reconstructed. Specifically:

[0164] We need the current time t and the previous (m-1) times. Historical data from τ time points are used to construct the current joint phase space vector. : .

[0165] (3) VAE forward calculation and real-time SPE value acquisition.

[0166] The phase space vectors constructed in real time Input it into the pre-trained VAE model.

[0167] The model performs forward computation: the encoder generates the latent variable z, and the decoder reconstructs the output vector. .

[0168] Calculate the real-time SPE value at the current moment: .

[0169] (4) Fault diagnosis.

[0170] Real-time SPE value Control limits determined during the offline phase Compare: If ≤ If the judgment process is in a normal state; if > If so, the judgment process has malfunctioned.

[0171] (5) Fault alarm and recording.

[0172] When a fault is detected, the system immediately triggers an alarm signal (such as a message push) and records detailed information about the fault. After completing the monitoring at the current moment, the system waits for the next sampling moment and then continues to collect new data for the next round of monitoring, thereby achieving 24 / 7 uninterrupted real-time fault monitoring.

[0173] It should be noted that the real-time data input is a continuous real-time data stream obtained from the data interface. The data stream, after preprocessing and phase space reconstruction, outputs... The parameters of these two modules are strictly consistent with those in the offline phase; the trained VAE model serves as the core computational unit, receiving... Output reconstructed values The SPE calculation module then calculates... The comparator (decision module) is used to... Read from the storage unit The comparison is performed; the comparator outputs a binary signal: "0" represents normal, and "1" represents fault. The fault signal triggers an alarm.

[0174] This application's embodiments achieve a complete engineering closed loop through a clear two-stage design of offline training and online monitoring. Knowledge solidification (offline stage) solidifies complex knowledge of normal system behavior into VAE model parameters and... Within this simple threshold, efficient execution (online phase) utilizes solidified knowledge, requiring only simple matrix operations and neural network forward propagation to achieve rapid, real-time fault diagnosis. It boasts high computational efficiency, meeting the real-time requirements of industrial sites. The entire solution only requires normal data, not fault data, enabling its application to new equipment and processes with unknown or difficult-to-obtain fault modes. It possesses strong practicality and universality.

[0175] To facilitate understanding of the various embodiments of the above-described fault determination method, a specific embodiment will be used as an example for explanation below. This embodiment takes the H2 production line of a chemical plant as the application object. This production line had previously experienced a production trip due to reactor temperature runaway, resulting in significant economic losses.

[0176] The H2 production line is a typical continuous chemical process, with the core equipment being a high-temperature, high-pressure reactor. This process is characterized by strong nonlinearity, large time delay, and multivariate coupling. Historically, a slow decline in cooling system efficiency has led to a gradual increase in reactor temperature, ultimately triggering the safety interlock system and causing an unplanned shutdown. Traditional DCS alarm systems only alert when the temperature significantly exceeds limits, which is too late. The embodiments in this application aim to achieve early warning of faults through dynamic monitoring of multiple key variables.

[0177] (1) Data acquisition and preprocessing.

[0178] Monitoring Variable Selection: Five key variables directly related to the jump-out event were selected: Reactor internal temperature (core parameter); Cooling water flow rate; Cooling water inlet temperature; Feed flow rate; Reactor pressure.

[0179] Data source: Sampled data from two consecutive months of normal operating conditions were extracted from the factory's historical data, with a sampling interval of 10 seconds, totaling approximately 250,000 sample points. Data processing included: outlier removal: using the 3σ criterion to eliminate obvious outliers; missing value imputation: using linear interpolation; and standardization: normalizing each variable to a mean of 0 and a standard deviation of 1.

[0180] (2) Multivariable phase space reconstruction.

[0181] Parameters determined:

[0182] The time delay τ of each variable is determined using the mutual information method, and the average optimal τ = 6 (corresponding to 60 seconds) is calculated. Figure 8 As shown, when τ = 0: x(t), x(t) and x(t+0), x(t+0) are the same point, they are completely correlated, have the maximum mutual information value, but the information is completely redundant.

[0183] As τ begins to increase: x(t+τ) and x(t+τ) begin to differ from x(t) and x(t), and the mutual information value decreases. Hopefully, this will stop when the statistical dependence between them (i.e., mutual information) just begins to become less strong.

[0184] The first local minimum point: This point means that if the delay is smaller (τ-1), the correlation between the two points is still too strong and the redundancy is high; if the delay is larger (τ+1), although the independence is stronger, the two points may have lost the dynamic connection that should exist in the dynamic system and become like random noise. This point is where the balance is achieved between "reducing redundancy" and "preserving dynamic correlation".

[0185] The spurious nearest neighbor method is used to determine the embedding dimension m. When m=4, the proportion of spurious nearest neighbors drops to below 5% and tends to converge.

[0186] The final reconstruction parameters are determined as follows: (m=4, τ=6).

[0187] Phase space vector construction: For each time point t, construct a joint phase space vector:

[0188]

[0189] The dimensions are: 5 variables × 4 embedding dimensions = 20 dimensions.

[0190] (3) VAE model training.

[0191] Network architecture design:

[0192] Encoder: Input layer 20 nodes → Hidden layer 64 nodes (ReLU) → Hidden layer 32 nodes (ReLU) → Output layer (mean vector and variance vector each 6 nodes);

[0193] Latent space dimension: 6 dimensions (much smaller than the 20 dimensions of the input);

[0194] Decoder: 6 nodes in latent space → 32 nodes in hidden layer (ReLU) → 64 nodes in hidden layer (ReLU) → 20 nodes in output layer (linear activation).

[0195] Training parameters: Batch size: 256; Training epochs: 200; Optimizer: Adam, learning rate: 0.001; KL divergence weight β: 0.5.

[0196] Training results: The reconstruction error MSE of the model on the validation set is stable at 0.015, indicating that the model has fully learned the dynamic pattern of normal operating conditions.

[0197] (4) Calculation of SPE control limits.

[0198] All normal training data are used to calculate the SPE value of each sample through the trained VAE model. The 99th percentile of the SPE value is 2.85. Therefore, the upper limit of control is set: SPE_lim = 2.85. This means that under normal working conditions, the SPE value of 99% of the samples is lower than this threshold.

[0199] (5) Online monitoring and early warning effect.

[0200] Test data: Data from the 15 days prior to the actual shutdown of the H2 production line was used as the test set. During this period, the cooling system had already begun to show efficiency degradation. The monitoring results are shown in Table 1 below:

[0201] Table 1: Comparison of monitoring results.

[0202]

[0203] The table above compares in detail the response of the temperature alarm mechanism of the traditional DCS (Distributed Control System) and the fault early warning method based on SPE (Square Prediction Error) proposed in this application at several key time points before the H2 production line trip, as well as the relationship with the actual process state. It shows that the SPE early warning method of this application can detect fault signals significantly earlier than the temperature alarm mechanism of the traditional DCS, especially in the early stage of the fault. Even if the direct measured values ​​such as temperature have not exceeded the limit, the abnormal changes in the system state in the phase space can be accurately captured.

[0204] (6) Verification of technical effects.

[0205] Early warning capability: Traditional DCS systems have difficulty detecting anomalies in advance, while the embodiment of this application shows abnormal signs two days before the trip, providing an early warning time of more than 40 hours.

[0206] Dynamic sensitivity: The SPE index is extremely sensitive to the slow decline in cooling system efficiency and can capture dynamic characteristic changes that cannot be detected by traditional methods.

[0207] Avoiding economic losses: If a warning is received in a timely manner, there is sufficient time to arrange subsequent emergency plans, thus avoiding huge economic losses caused by improper handling.

[0208] It is evident that the technical solution of this application possesses extremely strong early warning capabilities, enabling it to detect subtle abnormal dynamic characteristics even under normal operating conditions of traditional monitoring systems. Furthermore, it exhibits high sensitivity to slowly evolving faults, making it particularly suitable for monitoring slowly developing fault modes such as equipment performance degradation, catalyst deactivation, and scaling. Moreover, it enables multi-variable collaborative analysis, detecting changes in the dynamic relationships between variables, rather than simply identifying instances where a single variable exceeds limits. In addition, it demonstrates strong engineering practicality; once the offline model training is complete, the computational burden of online monitoring is light, making it suitable for real-time operation in industrial settings.

[0209] This application's embodiments introduce phase space reconstruction theory to convert multivariate time series into state-space data that reveals the system's inherent dynamic characteristics, thereby achieving an accurate description of the process's dynamic behavior. Furthermore, by employing an advanced deep learning model, the Variational Autoencoder (VAE), the complex probability distribution of normal process data is learned to accurately construct a "normal state manifold." By combining these two techniques and defining the squared prediction error (SPE) as a monitoring statistic, highly sensitive and reliable monitoring of early-stage faults in nonlinear industrial processes is achieved.

[0210] Compared with related technologies, this solution has the following significant advantages:

[0211] (1) Strong dynamic perception capability: Due to the use of phase space reconstruction, it can keenly capture abnormal changes in the dynamic characteristics of the system, and its detection capability for early and slow-changing faults far exceeds that of static methods.

[0212] (2) Excellent nonlinear processing capability: As a powerful nonlinear generative model, VAE can accurately learn the normal operating mode of complex nonlinear processes, overcoming the limitations of linear methods such as PCA.

[0213] (3) Unsupervised and self-learning: This method only requires normal historical data to complete model training, without the need to obtain various fault data in advance. It is highly practical and easy to implement.

[0214] (4) High early warning sensitivity: Combining dynamic and static features, the monitoring statistics (SPE) are extremely sensitive to small abnormal pattern changes, enabling early warning of faults.

[0215] To achieve fault monitoring based on dynamic characteristic analysis, the core concept of this application lies in the combination of "dynamic feature extraction" and "normal manifold learning". Around this core concept, there are feasible alternatives in many aspects of the technical solution, and these alternatives should be regarded as reasonable extensions of the scope of protection of this application.

[0216] For example, in dynamic feature extraction, this step is to convert the original time series into features that can reflect the dynamics of the system. Phase space reconstruction is the preferred solution to achieve this goal, but it is not the only solution.

[0217] Alternative Solution 1: Feature extraction based on recurrent neural networks.

[0218] That is, the embodiments of this application can also use recurrent neural networks (RNNs), especially long short-term memory networks (LSTM) or gated recurrent units (GRUs), as dynamic feature extractors. Specifically, the original multivariate time series is directly input into an RNN encoder, and the last hidden state or the sum of all hidden states of the RNN can be used as the latent vector representing the dynamic features of the system. This scheme uses the memory function of the RNN to replace the state space construction function of phase space reconstruction. Both aim to capture temporal dynamic dependencies and are different technical means to achieve the same goal.

[0219] Alternative Solution 2: Feature extraction based on temporal convolutional networks.

[0220] The embodiments of this application can also use a Temporal Convolutional Network (TCN) with causal dilated convolutions to extract dynamic features. TCN captures dependency patterns of long-term sequences through multiple layers of convolutional kernels, and its output can be used as a representation of dynamic features. TCN is another network structure for capturing long-term dependencies, and in this scenario, it can be used interchangeably with phase space reconstruction or RNN.

[0221] Alternative Solution 3: Extraction based on time-series feature engineering.

[0222] This application embodiment can also manually extract statistical features (such as mean, variance, and slope), time-domain features (such as zero-crossing rate and autocorrelation coefficient), or frequency-domain features (such as the dominant frequency energy of Fourier transform and wavelet transform) from the original time series, and combine these features into a feature vector to describe the system state. This is a feature engineering method based on prior knowledge, and its purpose is consistent with phase space reconstruction, that is, to construct a feature representation that better reveals the system state.

[0223] Figure 9 This is a structural diagram of a fault determination device according to an embodiment of this application, such as... Figure 9 As shown, the device includes:

[0224] The acquisition module 902 is used to acquire the raw state data of the target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions;

[0225] The reconstruction module 904 is used to reconstruct the phase space of the original state data to obtain a joint phase space vector, wherein the joint phase space vector is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables.

[0226] The parsing module 906 is used to reconstruct the target vector from the joint phase space vector using a variational autoencoder, wherein the target vector is used to characterize the normal state of the target system under operating conditions;

[0227] The determination module 908 is used to determine the fault monitoring index corresponding to the joint phase space vector and the target vector, wherein the fault monitoring index is used to quantify the degree to which the state of the target system deviates from the normal state;

[0228] The judgment module 910 is used to determine that the target system has failed when the fault monitoring index is greater than a preset threshold.

[0229] It should be noted that, Figure 9 The fault determination device shown is used to perform Figure 2 The method for determining the fault shown, therefore Figure 2 The relevant explanations in the fault determination method also apply to Figure 9 The device used to determine the fault is shown, and will not be described in detail here.

[0230] This application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the steps of the fault determination method implemented in various embodiments of this application.

[0231] This application also provides a non-volatile storage medium including a stored computer program, wherein the device containing the non-volatile storage medium executes the steps of the fault determination method in various embodiments of this application by running the computer program.

[0232] This application also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the fault determination method in various embodiments of this application.

[0233] This application also provides a computer program that, when executed by a processor, implements the steps of the fault determination method in various embodiments of this application.

[0234] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0235] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0236] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0237] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0238] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0239] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0240] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for fault determination, characterized in that, include: Collect raw state data of the target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions; The original state data is reconstructed in phase space to obtain a joint phase space vector, wherein the joint phase space vector is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables; A target vector is reconstructed from the joint phase space vector using a variational autoencoder, wherein the target vector is used to characterize the normal state of the target system under the operating conditions; Determine the fault monitoring index corresponding to the joint phase space vector and the target vector, wherein the fault monitoring index is used to quantify the degree to which the state of the target system deviates from the normal state; If the fault monitoring index exceeds a preset threshold, the target system is determined to have malfunctioned.

2. The method according to claim 1, characterized in that, The original state data is reconstructed into a phase space to obtain a joint phase space vector, including: Determine the reconstruction parameters, wherein the reconstruction parameters include a time delay parameter for determining the distance between two adjacent historical data points in the original state data, and an embedding dimension parameter for determining the dimension of the phase space vector of each variable; The original state data is converted into the joint phase space vector based on the reconstruction parameters.

3. The method according to claim 2, characterized in that, Converting the original state data into the joint phase space vector based on the reconstruction parameters includes: At each time point, embedding vectors corresponding to the multiple variables are extracted from the time series data corresponding to the multiple variables respectively, wherein the dimension of the embedding vector is determined by the embedding dimension parameter, and the distance between two adjacent elements in the embedding vector is determined by the time delay parameter; The embedding vectors corresponding to the multiple variables are concatenated to obtain an initial joint phase space vector, wherein the initial joint phase space vector is composed of the values ​​of the multiple variables at the time point and the values ​​of multiple historical time points before the time point; The joint phase space vector is determined based on the initial joint phase space vectors corresponding to all time points within the time window.

4. The method according to claim 1, characterized in that, The variational autoencoder is trained in the following manner: Obtain a training dataset, wherein the training dataset includes a historical joint phase space vector corresponding to historical state data, wherein each element in the historical joint phase space vector is used to characterize the historical state of the target system at each historical moment; The training dataset is processed using an initial variational autoencoder to obtain the first reconstructed vector; Determine the loss between the historical joint phase space vector and the first reconstructed vector; The parameters of the initial variational autoencoder are adjusted based on the loss to obtain the variational autoencoder.

5. The method according to claim 4, characterized in that, Determining the loss between the historical joint phase space vector and the first reconstructed vector includes: A first loss is determined between the historical joint phase space vector and the first reconstructed vector, wherein the first loss is used to quantify the degree of difference between the historical joint phase space vector and the first reconstructed vector; A second loss is determined to quantify the difference between the posterior distribution generated by the initial variational autoencoder and a preset prior distribution, wherein the posterior distribution includes the probability distribution of latent variables inferred from the training dataset, and the prior distribution includes a normal distribution; The loss is determined based on the first loss and the second loss.

6. The method according to claim 1, characterized in that, The target vector is reconstructed from the joint phase space vector using a variational autoencoder, including: The encoder of the variational autoencoder performs a nonlinear transformation on the joint phase space vector to obtain the probability distribution parameters of the latent space, wherein the probability distribution parameters include a mean vector and a log-variance vector, and the latent space is used to encode the normal state of the target system. The variational autoencoder is used to sample from the latent space according to the probability distribution parameters to obtain latent variables, wherein the latent variables are used to characterize the features of the normal state; The latent variables are reconstructed using the decoder of the variational autoencoder to obtain the target vector.

7. The method according to claim 4, characterized in that, The preset threshold is determined in the following way: The variational autoencoder is used to reconstruct each historical joint phase space vector in the training dataset to obtain a second reconstructed vector. Determine the squared prediction error between each second reconstruction vector and its corresponding historical joint phase space vector; The preset threshold is determined based on the squared prediction errors corresponding to all historical joint phase space vectors in the training dataset.

8. A fault determination device, characterized in that, include: The acquisition module is used to acquire the raw state data of the target system, wherein the raw state data includes time series data corresponding to multiple variables of the target system under operating conditions; The reconstruction module is used to reconstruct the phase space of the original state data to obtain a joint phase space vector, wherein the joint phase space vector is used to characterize the temporal dynamic characteristics of each variable and the coupling relationship between multiple variables; The parsing module is used to reconstruct the target vector from the joint phase space vector using a variational autoencoder, wherein the target vector is used to characterize the normal state of the target system under the operating conditions; A determination module is used to determine fault monitoring indicators corresponding to the joint phase space vector and the target vector, wherein the fault monitoring indicators are used to quantify the degree to which the state of the target system deviates from the normal state; The judgment module is used to determine that the target system has failed when the fault monitoring index is greater than a preset threshold.

9. An electronic device, characterized in that, include: A memory and a processor, the memory being used to store program instructions; the processor being connected to the memory and used to execute the fault determination method according to any one of claims 1 to 7.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored computer program, wherein the device containing the non-volatile storage medium executes the fault determination method according to any one of claims 1 to 7 by running the computer program.

11. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the fault determination method according to any one of claims 1 to 7.

Images