Distributed fault detection method for large-scale industrial processes based on subspace identification

Abstract

This invention discloses a distributed fault detection method for large-scale industrial processes based on subspace identification. A fault detection unit is established on each subsystem of the large-scale system. Local residual generators are directly identified using process data from local and neighboring industrial subsystems, and the residual covariance matrix under healthy conditions is estimated. Actuator faults in neighboring industrial subsystems are decoupled, enabling the detection of faults in both local and neighboring industrial subsystems. Using this invention, a fault detection unit can be constructed within each industrial subsystem using only local and neighboring data, achieving high-performance fault detection for both local and neighboring industrial subsystems. This method boasts high detection rates, low false alarm rates, and low false negative rates, reducing computational and data transmission burdens and making it suitable for dynamic industrial processes.

Description

Technical Field

[0001] This invention belongs to the field of industrial process monitoring technology, and relates to a fault detection method for large-scale industrial processes, specifically a data-driven distributed fault detection method for large-scale industrial processes based on subspace identification. Background Technology

[0002] With the rapid development of computing and communication technologies, the scale and complexity of modern industrial processes are constantly expanding. A large-scale industrial process often consists of many interconnected subsystems. Due to this networked interconnection, the impact of a single system failure can quickly spread to the entire industrial process, causing serious consequences. Therefore, anomaly monitoring and fault diagnosis of large-scale industrial processes are of great significance for ensuring safe and stable production processes and high-quality products. Large-scale industrial processes generate massive amounts of data, and centralized monitoring methods are often unsuitable due to limitations in computing resources and communication bandwidth. Therefore, distributed monitoring schemes are often adopted to handle large-scale processes, and data-driven distributed fault detection methods have become a research hotspot in recent years.

[0003] Existing data-driven distributed fault detection technologies for industrial processes are mostly based on multivariate statistical methods, analyzing historical data of industrial processes while neglecting existing process mechanisms. They are primarily applied to static or steady-state processes in dynamic industrial systems, but not to common dynamic industrial processes. For fault detection in dynamic industrial processes, data-driven monitoring methods based on subspace identification have received widespread attention. These methods implicitly contain some mechanistic information and can directly identify residual generators from data. They are simple to operate and possess the high performance of model-based methods. However, current methods are all based on centralized frameworks, and there is a lack of high-performance, large-scale distributed monitoring solutions for industrial processes applicable to multiple operating modes. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, this invention provides a large-scale industrial process distributed fault detection method based on subspace identification. In each industrial subsystem, an industrial process fault detection unit can be constructed using only local and neighboring data, achieving high-performance fault detection of local and neighboring industrial subsystems, reducing computational and data transmission pressure, and is applicable to dynamic industrial processes.

[0005] In large-scale industrial systems, residuals are the difference between the reconstructed measurable signal and the measured signal. When the system is fault-free, residuals mainly reflect the impact of unknown inputs such as external interference and model uncertainty on the system. This invention utilizes a distributed collaborative mechanism to design a fully distributed fault detection architecture. A fault detection unit is established for each subsystem, communicating only with the input / output sensors of the local subsystem and the fault detection units of neighboring subsystems. During both offline training and online detection, only the observation information of the local subsystem is used, without requiring global information. During offline training, each subsystem's fault detection unit first collects current and historical input / output data and constructs the corresponding Hankel matrix. Then, it constructs the input / output (IO) data model of the local subsystem and uses subspace identification technology, through Singular Value Decomposition (SVD), to identify the local residual generator, decoupling completely unknown variables (local and neighboring state variables). Then, it uses historical data and the identification results from the previous step to estimate the residual covariance matrix under healthy conditions. During the online monitoring phase, the local fault detection unit receives real-time data and calculates T... 2 The statistics are compared with the threshold to obtain the local detection results.

[0006] This invention can handle complex, heterogeneous, large-scale industrial processes, and is also applicable to dynamic processes that traditional data-driven methods cannot handle. The established fully distributed framework greatly reduces the pressure of data transmission and computation, reduces offline training time, and ensures the real-time performance of online monitoring. It utilizes a subspace identification method to directly identify the local residual generator through data, simplifying operation and avoiding the complex process of pre-identifying the system's state-space model using data. Considering the interconnected characteristics of subsystems within large-scale industrial processes and the correlation between local and neighboring subsystems, it uses partial information from neighboring subsystems when identifying the local residual generator, significantly improving detection performance compared to distributed detection methods. The established local residual generator can decouple the dynamic changes of neighbors from actuator faults. Existing distributed detection methods require setting boundaries for the correlation information between local and neighboring subsystems to locate actuator faults; however, this invention can directly decouple the actuator faults of neighbors without setting such a boundary, achieving actuator fault location. Furthermore, the local residual generator can detect sensor faults in both local and neighboring systems, thus enabling simultaneous location of both actuator and sensor faults.

[0007] The technical solution provided by this invention is:

[0008] A large-scale industrial process distributed fault detection method based on subspace identification is a data-driven, fully distributed monitoring method applicable to dynamic industrial processes. It directly identifies local residual generators and estimates the residual covariance matrix under healthy conditions using process data from local and neighboring industrial subsystems. This method decouples actuator faults from neighboring industrial subsystems, enabling the detection of faults in both the local and neighboring industrial subsystems' sensors. The main steps include:

[0009] 1) Collect process data from the local subsystem and neighboring subsystems, construct an IO model for the local subsystem of a large-scale industrial process, and establish a local fault detection unit;

[0010] 2) Identify the local residual generator r for large-scale industrial processes i (k) is used to detect whether a fault has occurred in the local subsystem;

[0011] The local residual generator is used to detect local actuator faults and local sensor faults with neighboring sensors, and can decouple neighboring actuator faults.

[0012] 3) Estimate the health status residual covariance matrix of local subsystems in large-scale industrial processes and design detection statistics;

[0013] Set the appropriate detection thresholds; the local fault detection unit performs fault detection.

[0014] Through the above steps, large-scale distributed fault detection of industrial processes based on subspace identification is achieved.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0016] This invention provides a distributed fault detection method for large-scale industrial processes based on subspace identification. In each industrial subsystem, a fault detection unit can be constructed using only local and neighboring data, enabling high-performance fault detection of both local and neighboring industrial subsystems with high detection rate and low false alarm and false negative rates. This invention alleviates the computational and data transmission burden in large-scale industrial process fault detection and is applicable to dynamic industrial processes. Attached Figure Description

[0017] Figure 1 A schematic diagram of a large-scale interconnected industrial process;

[0018] In large-scale interconnected industrial processes, numerous subsystems follow a directed graph. Coupling association, where ε represents a node, and ε represents an edge.

[0019] Figure 2This is a schematic diagram of the detection system structure for a specific implementation of the distributed fault detection method proposed in this invention;

[0020] Where, x i ,u i ,y i These represent the state variables, control input signals, and output signals of subsystem i, respectively.

[0021] Figure 3 This is a flowchart illustrating the algorithm for a specific implementation of the method of the present invention.

[0022] Figure 4 This is a schematic diagram of the hot strip rolling finishing process in an embodiment of the present invention, wherein F i ,L i These represent the i-th rack and the loop, respectively.

[0023] Figure 5 This is a directed graph corresponding to the hot strip rolling finishing process in an embodiment of the present invention.

[0024] Figure 6 The result diagram is an example of the distributed fault detection provided by this invention;

[0025] Among them, J i The values ​​represent the detection statistics; (a) to (f) represent the monitoring results of detection units 1 to 6 of subsystems, respectively. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and examples, but the scope of the invention is not limited in any way.

[0027] Figure 1 This is a schematic diagram of a large-scale interconnected industrial process, in which numerous subsystems are coupled and connected according to a directed graph. Figure 2 This is a schematic diagram of the detection system structure for a specific implementation of the distributed fault detection method proposed in this invention. A fault detection unit is established on each subsystem. During offline training and online monitoring, each local fault detection unit only needs to receive local input and output signals from local sensors and output signals from neighboring fault detection units to perform distributed training and detection, which greatly reduces the computational resources and data transmission pressure in large-scale systems.

[0028] This invention provides a distributed fault detection method for large-scale interconnected industrial processes based on subspace identification. It is a data-driven, fully distributed monitoring method applicable to dynamic industrial processes. It directly identifies local residual generators using process data from local and neighboring industrial subsystems and estimates the residual covariance matrix under healthy conditions. This decouples actuator faults from neighboring industrial subsystems, enabling the detection of faults in both the local and neighboring industrial subsystems' sensors. The specific algorithm for implementing this invention is as follows:

[0029]

[0030] Figure 3 The illustration shows the specific implementation process of the method of the present invention, which mainly includes the following steps:

[0031] 1) Collect data, construct IO models of local subsystems in large-scale industrial processes, and establish local fault detection units; including:

[0032] 11) First, we introduce the definition of the correlation matrix: For the variable θ, the corresponding dimension of the Hankel data matrix is ​​defined as:

[0033]

[0034] Where k is the time parameter, and N and s are dimension parameters, which can take any positive integer values, and N is the Hankel matrix Θ. k，s The number of columns (horizontal dimension), s is Θ k，s The number of rows (vertical dimension).

[0035] When variable θ takes the local input u i Local output y i and neighbor output When this is done, the corresponding Hankel matrix is ​​obtained:

[0036]

[0037] 12) The corresponding lower triangular convolution matrix Ξ[CA s The general definition of B is expressed as:

[0038]

[0039] Where C, A, and B are matrices of corresponding dimensions.

[0040] like Figure 1 and 2 As shown, a large-scale system consists of multiple subsystems. This invention establishes a fault detection unit in each subsystem. The i-th subsystem is... Figure 1In the subsystem i, when establishing a fault detection unit for the i-th subsystem, the i-th subsystem is the local subsystem, and i can take any positive integer not exceeding the total number of subsystems.

[0041] Taking the i-th subsystem as an example, firstly, the local input signal for the time interval [k-2s-1, k+N-1] is collected. Local output signal (obtained from local sensors) and neighbor output signals (Obtained from the neighbor's detection unit, by the set of neighbor subsystems) (composed of the stacked output signals of all neighbors), l i For control input u i The dimension of m, taking values ​​that are any positive integers; i To output y i The dimension of the matrix is ​​taken as any positive integer; thus, the Hankel data matrix at the current time is constructed. And the historical Hankel matrix

[0042] In the finishing process of hot strip rolling, the input signals include: work roll speed control signal. Loop torque control signal Roll gap control torque Right now The output signal includes: loop angle θ i Export strip tension σ i Work roll speed v i Loop torque T i Strip export thickness h i ,Right now

[0043] 13) Construct the I / O model of the subsystem;

[0044] The following analysis examines the dynamic model of the i-th subsystem in a large-scale industrial process, and subsequently constructs the IO model. Assume the large-scale industrial process comprises M observable linear time-invariant subsystems, each closely connected to a subset of neighboring subsystems, exhibiting complex state coupling relationships in their dynamics. Let the i-th subsystem have p... i There are *i* neighboring subsystems, and the state space of the *i*-th subsystem is expressed as:

[0045] (Equation 3)

[0046] Among them, u i E represents the control input signal of the state-space model. i =[A i1 … A ij … AiM ], t i (k)=[x1 T (k) … x i T (k) … x M T (k)] T , All contain p i Each sub-block corresponds to p i A neighbor subsystem, This represents the set of neighboring subsystems. For the state, input, output, and measurement noise signal of local subsystem i, A i B i C i E i For an unknown, time-invariant system matrix, A ij Let represent the coupled system matrix of the neighboring system j acting on the dynamics of the local subsystem i.

[0047] As can be seen from Equation 3, the input signal of the local subsystem i includes the known control input u. i and unknown input signal t i (Constituted by a stack of neighboring state variables), using the output equation of the neighboring system, the unknown input signal t i It can be decomposed into two parts, one part being the stacked output signal of the known neighboring systems. (Structure and t) i The signal consists of two parts: one part is consistent with the other part is a completely unknown signal. The decomposed unknown input t i It can be represented as:

[0048]

[0049] The unknown parameter matrices Q1 and Q2 are two block diagonal matrices, and [Q1 Q2] is nonsingular. This reduces the dimensionality of the local unknown input, allowing some of it to be represented by known signals.

[0050] Substituting (Equation 4) into (Equation 3) yields another expression for the state-space model of the local subsystem i:

[0051]

[0052] Where E i1 =E i Q1, E i2 =E i Q2, The signal is a stack of measurement noise from all neighboring subsystems, and The structure is consistent.

[0053] From Equation 5, the corresponding I / O model for subsystem i can be derived, expressed as:

[0054]

[0055] Where the convolution matrix Observation matrix State variable data matrix The output signals y of local subsystem i are respectively i Stacked output signals of neighboring subsystems Stacked measurement noise of neighboring subsystems The input signal of the local subsystem i is completely unknown. The control input u of the local subsystem i i and the measurement noise signal ω of the local subsystem i i The corresponding Hankel data matrix.

[0056] 2) Identify and construct the residual generator r of the local subsystem of a large-scale industrial process. i (k) is used to detect whether a fault has occurred in the local subsystem;

[0057] Construct the current data matrix and historical data matrix

[0058]

[0059] The current data matrix includes the control input Hankel data matrix of the local subsystem, the output signal Hankel data matrix, and the output signal Hankel data matrix of the neighboring subsystem.

[0060] SVD decomposition is performed on the product of the current data matrix and the historical data matrix to identify the local residual generator:

[0061]

[0062] Where the singular value matrix ∑2≈0, the corresponding singular vector matrix is ​​U2, denoted as U2.

[0063] In equation (8), [U1 U2] is a matrix composed of left singular vectors. It is a matrix composed of right singular vectors, a diagonal matrix. The diagonal elements are singular values.

[0064] The expression for the local residual generator is:

[0065]

[0066] Where, r i The residual signal generated by the fault detection unit corresponding to local subsystem i. For the unknown local system i, the state variable x i and completely unknown input signals Decoupling. in Together they constitute the singular vector U2 in (Equation 8), that is Data collected by online monitoring These are the Hankel matrices. The first column contains the data for the time interval [ks, k].

[0067] The following analysis examines the effectiveness of the local residual generator constructed using Equation 9 in locating actuator and sensor failures that may occur in large-scale industrial processes.

[0068] Under fault conditions, the state-space model update of the local system i shown in Equation 5 is as follows:

[0069]

[0070] Where f ai f si This indicates actuator and sensor failures occurring in local subsystem i. It is caused by a stack of sensor failures from neighboring subsystems.

[0071] The IO model corresponding to (Equation 10) is:

[0072]

[0073] in, They have the same structure, each consisting of completely unknown state variables of the neighbor system in the time interval [ks, k]. Output noise and sensor failure constitute.

[0074] Based on (Equation 11), the expression for the local residual generator under fault conditions is:

[0075]

[0076] As shown in Equation 12, the local residual generator is affected by the actuator failure of its neighbor. Decoupling, the residual signal r of local subsystem i i (k) Only affected by local actuator failure Sensor malfunctions in the local area and neighboring areas Due to the influence of [the fault], the distributed fault detection method proposed in this invention can locate actuator faults and sensor faults.

[0077] 3) Estimate the health status residual covariance matrix of the local system in a large-scale industrial process and design detection statistics;

[0078] Based on process data, identify according to (Equation 8) Then, by combining historical data without failures, the residual covariance matrix of the health status can be obtained. Represented as:

[0079]

[0080] Then construct T for detection 2 Statistic J i :

[0081]

[0082] Due to the local residual vector For completely unknown input signals Since the dimension is , the corresponding detection threshold is set as follows:

[0083]

[0084] Where, χ 2 (β) represents the chi-square distribution with β degrees of freedom, and α represents the given significance level.

[0085] Based on this, the detection method of the local fault detection unit is set as follows: z

[0086]

[0087] That is: during the online monitoring process, each fault detection unit calculates the detection statistic J in real time. i When J i Exceeding threshold J ith When the fault detection unit alarms, it indicates that a fault has occurred; J i Below the threshold J ith If the condition is met, it is considered that no fault has occurred.

[0088] Through the above steps, large-scale distributed fault detection of industrial processes based on subspace identification is achieved.

[0089] This invention specifically addresses the typical large-scale industrial process of hot strip rolling, which features a complex multi-stage and multi-system structure, high output, high product quality requirements, and cumbersome operation. Among these, the finishing rolling process is one of the most critical, and ensuring its proper operation plays a decisive role in the overall efficiency and product quality of the rolling process. This invention specifically focuses on the finishing rolling process within hot strip rolling. The entire finishing rolling process comprises 7 stands (F1–F7) and 6 loopers (L1–L6). Each stand and looper works in conjunction to adjust the strip tension, thickness, and speed, making it a complex, multi-variable, coupled, interconnected large-scale process. Figure 4 As shown, the entire finishing rolling process can be decomposed into 6 subsystems. By analyzing the mechanism model of the finishing rolling process, the correlation and coupling relationships between the subsystems are derived, which conform to the following... Figure 5 The directed graph shown. This is achieved by recording and collecting data on the strip exit tension σ, work roll speed v, strip exit thickness h, and work roll speed control input u for each stand. v Roll gap control input u s The loop angle θ, the applied torque T, and the loop torque adjustment control input u for each loop are also considered. T These process data can be used to establish fault detection units in each subsystem, identify local residual generators, and perform distributed fault detection.

[0090] Fault Setup: At the 290th sampling point, a fault is introduced in the speed sensor of the F4 work roller on the frame. The fault detection unit of each subsystem works in parallel, and the monitoring results are as follows: Figure 6 As shown, both the faulty system and its neighboring systems successfully detected the sensor fault with a fault detection rate approaching 100%.

[0091] It should be noted that the purpose of disclosing the embodiments is to help further understand the present invention. However, those skilled in the art will understand that various substitutions and modifications are possible without departing from the scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the embodiments, and the scope of protection of the present invention is defined by the scope of the claims.

Claims

1. A distributed fault detection method for large scale industrial processes based on subspace identification, characterized in that, The large-scale system comprises multiple subsystems, with a fault detection unit established on each subsystem. Local residual generators are directly identified using process data from local and neighboring industrial subsystems, and the residual covariance matrix under healthy conditions is estimated. Actuator faults in neighboring industrial subsystems are decoupled, enabling the detection of faults in both the local and neighboring industrial subsystems' sensors. The system includes the following steps: 1) Collect process data from the local subsystem and neighboring subsystems, construct the input-output model of the local subsystem of a large-scale industrial process, and establish a local fault detection unit; Local input Local output and Neighbor output Hankel data matrix of corresponding dimension, denoted as: (Equation 1) Among them, variables Take local input Local output Output with neighbor The corresponding Hankel matrix is ​​obtained: , k is the time parameter, and N and s are dimension parameters, which can take any positive integer value, where N is the Hankel matrix. The number of columns is the horizontal dimension, s is The number of rows is the vertical dimension; Lower triangular convolution matrix Defined as: (Equation 2) in, For a matrix of the corresponding dimension; Let the i-th subsystem have There are *i* neighboring subsystems, and the state space of the *i*-th subsystem is expressed as: (Equation 3) in, This represents the control input signal of the state-space model; All include Each sub-block corresponds to A neighbor subsystem, Represents the set of neighboring subsystems; For the state, input, output, and measurement noise signals of the local subsystem i, The system matrix is ​​unknown and is stationary. This represents the coupling system matrix of the neighboring system j acting on the dynamics of the local subsystem i; i and j are the indices. It is the output signal of the local subsystem i dimensionality It is the state variable of the local subsystem i dimensionality; It is the input signal of subsystem i dimensionality; The input signals of local subsystem i include known control inputs. and unknown input signals The unknown input signal is formed by stacking the state variables of the neighbors; the unknown input signal is decomposed into two parts, one part being the stacked output signal of the known neighbor system. The other part consists of completely unknown signals. The decomposed unknown input Represented as: (Equation 4) Where the unknown parameter matrix It is a 2-block diagonal matrix and Non-singular; Substituting (Equation 4) into (Equation 3) yields another expression for the state-space model of the local subsystem i: (Equation 5) in, , The signal is a stack of measurement noise from all neighboring subsystems; The input-output model of subsystem i can be represented as follows: (Equation 6) Wherein, the convolution matrix Observation matrix State variable data matrix ; , The output signals of local subsystem i are respectively Stacked output signals of neighboring subsystems Stacked measurement noise of neighboring subsystems The input signal is completely unknown to the local subsystem i. Control input of local subsystem i and the measured noise signal of local subsystem i The corresponding Hankel data matrix; 2) Identification of local residual generators for large-scale industrial processes. It is used to detect whether a fault has occurred in the local subsystem; The local residual generator is used to detect local actuator faults and sensor faults in both local and neighboring systems, and can decouple neighboring actuator faults; including: Construct the current data matrix and historical data matrix : (Equation 7) The current data matrix includes the control input Hankel data matrix, the output signal Hankel data matrix of the local subsystem, and the output signal Hankel data matrix of the neighboring subsystem. Singular value decomposition is performed on the product of the current data matrix and the historical data matrix to identify the local residual generator, as shown below: (Equation 8) Among them, singular value matrix The corresponding singular vector matrix is ,remember ; It is a singular value matrix, containing a number of singular values ​​much larger than... The singular values ​​contained therein; for The corresponding singular vector matrix; The local residual generator is represented as: (Equation 9) in, The residual signal generated by the fault detection unit corresponding to local subsystem i. For unknown local system i, state variables and completely unknown input signals Decoupling; ,in Together they constitute the singular vectors in (Equation 8) ,Right now Data collected through online monitoring These are the Hankel matrices. The first column contains data for the time interval [ks,k]. The local residual generator can locate actuator and sensor faults in large-scale industrial processes. 3) Estimate the health state residual covariance matrix of local subsystems in large-scale industrial processes, design detection statistics; set corresponding detection thresholds; local fault detection units perform fault detection, including: Obtain the residual covariance matrix of health status , represented as: (Equation 13) in, ; Then construct a detection method Statistic : (Equation 14) Local residual vector The input signal is completely unknown. The dimension of is such that the corresponding detection threshold is set as follows: (Equation 15) in, For significance level of The chi-square distribution, For a given significance level; It is the dimension of the state variables of the neighboring subsystem j. It is the dimension of the output signal of the neighboring subsystem j; The detection method for the local fault detection unit is as follows: during online monitoring, each fault detection unit calculates the detection statistics in real time. ,when Exceeding the threshold When the fault detection unit alarms, a fault has occurred; when Below the threshold When this happens, no fault occurs; Through the above steps, large-scale distributed fault detection of industrial processes based on subspace identification is achieved.

2. The large-scale industrial process distributed fault detection method based on subspace identification as described in claim 1, characterized in that, The local residual generator constructed by (Equation 9) enables the location of actuator and sensor faults in large-scale industrial processes, specifically: In the fault case, the state-space model of local system i is represented as: (Equation 10) in, , This indicates actuator and sensor failures occurring in local subsystem i. It was caused by a stack of sensor failures from neighboring subsystems; The input-output model corresponding to (Equation 10) is expressed as follows: (Equation 11) in, , , and They have the same structure, each consisting of completely unknown state variables of the neighbor system in the time interval [ks,k]. Output noise and sensor failure constitute; Based on (Equation 11), the expression for the local residual generator under fault conditions is: (Equation 12) Equation 12 indicates that the local residual generator is affected by the actuator failure of its neighbor. Decoupling, the residual signal of local subsystem i Only affected by local actuator failure Sensor malfunctions in the local area and neighboring areas The impact.

3. The large-scale industrial process distributed fault detection method based on subspace identification as described in claim 1, characterized in that, Input signals include: work roll speed control signal Loop torque control signal Roll gap control torque ,Right now .

4. The large-scale industrial process distributed fault detection method based on subspace identification as described in claim 3, characterized in that, Output signals include: looper angle Export strip tension Working roll speed Loop torque Export thickness of strip steel ,Right now .

Images