Valve stiction detection for closed-loop controllers

Abstract

A deep convolutional neural network (CNN) can be used to identify nonlinear control valve events in a chemical process control loop. Identifying the nonlinear control valve events can be based on predetermined features of red, green, and blue (RGB) images obtained by merging respective matrices assembled from continuous wavelet transform coefficients derived from chemical plant operation controller signals. Once a nonlinear control valve event is identified in the control loop, controller settings can be modified to minimize the nonlinear control valve event and / or a root cause investigation can be performed to address the problem.

Description

[Technical Field]

[0001] The present disclosure relates to identifying nonlinear control valve events in multivariate time series data. Such techniques can be particularly useful for identifying the state of a controller in a control valve loop of a chemical process and notifying an operator to adjust the controller settings to address the nonlinear control valve events. [Background technology]

[0002] Control loop performance monitoring can provide safe and consistent operation of a chemical plant. However, there are situations where equipment performance issues can result in less than optimal performance of a chemical plant. For example, stiction is an equipment problem that causes resistance to proper valve movement, resulting in a delay between the controller output and the valve stem position. When a control valve experiences stiction, its stem may not move when the controller output changes, and this nonlinearity causes sustained oscillations in the control loop.

[0003] Accurately identifying problems such as valve stiction in closed-loop controllers is a difficult task. Often, chemical plants contain hundreds, if not thousands, of control loops with controllers. Because the valves controlled by each controller are susceptible to stiction, the possibility of having to address this condition certainly arises. As a result, there is a need in the art to monitor and evaluate control loops to help improve the operation of chemical plants. Summary of the Invention

[0004] The present disclosure is directed to improving chemical plant operation by monitoring and evaluating control loops for nonlinear control valve events. Specifically, the present disclosure provides methods and systems for identifying nonlinear control valve events in a chemical process control loop. Identifying nonlinear control valve events is based on predetermined features of RGB images detected by a deep convolutional neural network (CNN) trained to identify nonlinear control valve events. For various embodiments, such nonlinear control valve events may include valve stiction, and identifying such nonlinear control valve events may help improve chemical plant operation.

[0005] As provided herein, a method of the present disclosure includes receiving a set point (SP) signal, a process variable (PV) signal, and a controller output (CO) signal, which constitute multivariate time series data, from a control loop for a control valve of a chemical process, where the CO signal is received from a controller that controls the position of the control valve in the chemical process. The multivariate time series data from each of the SP signal, the PV signal, and the CO signal are transformed using a continuous wavelet transform (CWT) to provide respective CWT coefficients for each of the SP signal, the PV signal, and the CO signal. The respective CWT coefficients are assembled into respective matrices, and the respective matrices are merged to provide a red, green, and blue (RGB) image. In various embodiments, the RGB image is input to a deep convolutional neural network (CNN) trained to identify nonlinear control valve events in the control loop of the chemical process based on predetermined features of the RGB image. In response to the CNN identifying the predetermined features, the controller settings can be adjusted to minimize or eliminate the nonlinear control valve event (e.g., a stiction valve event). Additionally, identifying a control valve problem allows the root cause of a non-linear control valve event (e.g., a stiction problem) to be further investigated so that it can be addressed (e.g., corrected).

[0006] For various embodiments, the multivariate time series data is acquired at a predetermined frequency over a predetermined time interval. For example, the nonlinear control valve event may be a vibration event. In certain embodiments, the nonlinear control valve event is the result of a valve controlled by a controller being in a state of stiction. In additional embodiments, various wavelets may be used when transforming the multivariate time series data using the CWT. For example, the Morlet wavelet is used when transforming the multivariate time series data using the CWT. For various embodiments, transforming the at least one pairwise comparison of the SP signal, the PV signal, and the CO signal of the multivariate time series data includes transforming the multivariate time series data without preprocessing the multivariate time series data.

[0007] For various embodiments, the CNN identifies predetermined features based on changes in frequency information from a time-series multivariate signal. For example, predetermined features in an RGB image are associated with dynamic frequency spectral activations that are identified by the CNN when a nonlinear control valve event is identified.

[0008] For various embodiments, the CNN is pre-trained to identify nonlinear control valve events in the RGB images, and the method further includes training the CNN via transfer learning using the plurality of RGB images of at least one pairwise comparison of SP, PV, and CO signals of the multivariate time series data of the controller from the previously identified nonlinear control valve events of the controller in the plurality of RGB images. In some embodiments, the CNN may be trained using RGB images based on signals, such as SP, PV, and CO signals, associated with different controllers (e.g., the training of the CNN need not be controller-specific).

[0009] For various embodiments, the controller is a flow controller in a control loop of a chemical process. Alternatively, the controller may be a level controller, a temperature controller, or a pressure controller in a control loop in a chemical process.

[0010]

[0006] Embodiments of the present disclosure also include a system including a detector, a CNN, and a controller. The detector is configured to receive SP, PV, and CO signals constituting multivariate time series data from a control loop for a control valve of a chemical process. The CNN is trained with a plurality of RGB images of a comparison of at least one pair of SP, PV, and CO signals of the multivariate time series data of the controller with previously identified nonlinear control valve events of the controller in the plurality of RGB images. The controller is coupled to the detector and the CNN, and is configured to: convert at least one pair of SP, PV, and CO signals of the multivariate time series data into an RGB image using a CWT; input the RGB image to the CNN; identify nonlinear control valve events of the controller using the CNN; and adjust settings of the controller to minimize or eliminate the nonlinear control valve events of the controller identified by the CNN. For adjustment purposes, the CNN can allow an operator to recognize nonlinear control valve events of the controller, at which point one or more of further investigations into the root cause of the nonlinear control valve events (e.g., stiction problems) can be initiated and / or adjustments can be made by the operator to the controller settings to minimize or eliminate the nonlinear control valve events of the controller identified by the CNN. For various embodiments, the controller is configured to convert the multivariate time series data into an RGB image without preprocessing the multivariate time series data.

[0011] For various embodiments, the nonlinear control valve event is the result of the valve controlled by the controller being in a state of stiction.

[0012] The above summary of the present disclosure is not intended to describe each disclosed embodiment or every implementation of the present disclosure. More particularly, the present specification exemplifies exemplary embodiments. In several places throughout the application, guidance is provided through lists of examples, which examples can be used in various combinations. In each instance, the recited list serves only as a representative group and should not be interpreted as an exclusive list. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 1 illustrates an embodiment of a control loop for a chemical process for an embodiment of the present disclosure. [Figure 2A] Two examples of controllers of RGB images and feature maps by CNN networks are shown in the presence of stiction (Figures 2A and 2B) and absence of stiction (Figures 2C and 2D). [Figure 2B] Two examples of controllers of RGB images and feature maps by CNN networks are shown in the presence of stiction (Figures 2A and 2B) and absence of stiction (Figures 2C and 2D). [Figure 2C] Two examples of controllers of RGB images and feature maps by CNN networks are shown in the presence of stiction (Figures 2A and 2B) and absence of stiction (Figures 2C and 2D). [Figure 2D] Two examples of controllers of RGB images and feature maps by CNN networks are shown in the presence of stiction (Figures 2A and 2B) and absence of stiction (Figures 2C and 2D). DETAILED DESCRIPTION OF THE INVENTION

[0014] The present disclosure is directed to improving chemical plant operation by monitoring and evaluating control loops for nonlinear control valve events. Specifically, the present disclosure provides methods and systems for identifying nonlinear control valve events in a chemical process control loop. The nonlinear control valve events are based on predetermined features of RGB images that are identified by a deep convolutional neural network (CNN) trained to identify nonlinear control valve events. For various embodiments, such nonlinear control valve events may include valve stiction, and identifying and addressing such nonlinear control valve events (e.g., changing valve settings and / or replacing failed equipment) can help improve chemical plant operation.

[0015] By way of background, CNNs are a type of deep neural network that can be used to identify patterns in images. For example, CNNs can be used to assign importance (learnable weights and biases) to various aspects / objects in an image and distinguish one from another. Such images can include so-called three-channel representations (e.g., RGB images) with red, green, and blue components, and CNNs use neural networks to process the red, green, and blue components of many signals simultaneously. The "deep" aspect of a convolutional neural network refers to the use of multiple layers in a CNN, which extract successively higher-level features from the raw input. In the case of image analysis, examples of such features in RGB images include edges or color as low-level features, and CNNs can be trained through a combination of convolution and pooling, among other steps, to identify high-level features such as shape and gradient orientation, which enable a more complete interpretation of the RGB image.

[0016] The RGB images provided herein are derived from multiple non-stationary and / or time-series signals using the continuous wavelet transform (CWT). The CWT is useful for converting these one-dimensional time-series signals into coefficients that fully account for both the scale and location of the signal while being resistant to signal noise. Each CWT is a scaled and shifted version of a wavelet function, the sum of which is taken over time to provide what are called wavelet coefficients, which represent the time, frequency, and characteristics of the signal. The wavelet coefficients can then be provided in a scalogram, which provides high-resolution detail of relative changes in both time and frequency. Such detail allows for both pattern recognition and / or feature extraction of the signal.

[0017] As used herein, the singular forms "a," "an," and "the" include both singular and plural referents unless the context clearly dictates otherwise. Furthermore, the word "may" is used throughout this application in its permissive (i.e., could, can) sense rather than its obligatory (i.e., must) sense. The term "comprises" and its derivatives mean "including, but not limited to." The term "coupled," unless otherwise specified, means directly or indirectly connected, and can include wireless connections.

[0018] As will be understood, elements shown in the various embodiments herein may be added, interchanged, and / or eliminated to provide additional embodiments of the present disclosure. Additionally, it will be understood that the proportions and relative scales of the elements provided in the figures are intended to illustrate certain embodiments of the present invention and should not be construed in a limiting sense.

[0019] As previously mentioned, the present disclosure is directed to improving chemical plant operation by monitoring and evaluating control loops for nonlinear control valve events. Chemical plants in industries such as oil refineries, petrochemical industries, polymer industries, pulp and paper industries, power plants, and pharmaceutical industries have many control loops that help maintain critical process variables at their respective set points. Control loops are components of a chemical plant control system, each including physical components (e.g., control elements such as valves and sensors) and control functions (e.g., controllers) that work together to regulate each of the process variables to meet the desired set point.

[0020] FIG. 1 provides a diagram of a control loop 100 for a chemical process 102, which includes elements for measuring and controlling a process variable at a desired setpoint. The process variable may include, for example, flow rate, level, pressure, or temperature, among other process variables that need to be controlled in a chemical process. In this example, the control loop 100 includes a controller 104, a flow control valve 106, and a flow transmitter 108, both of which are integrated into a process fluid flow line 110. The flow control valve 106, under the control of the controller 104, can change its valve stem position to regulate fluid flow through the process fluid flow line 110. Examples of the controller 104 include, among other things, a programmable proportional-integral-derivative (PID) controller, as known in the art. The controller 104 provides a controller output (CO) signal 112 to the flow control valve 106, which allows for any number of valve positions (e.g., 0-100% open).

[0021] During operation, the CO signal 112 is received from the controller 104 to control the position of a control valve in a chemical process. For example, the CO signal 112 received from the controller 104 sets the position of a valve stem based on a setpoint (SP) signal 114. For various embodiments, the SP signal 114 can be received by the controller 104 manually or cascaded from another source. Fluid flow through a process fluid flow line 110 in this example is measured by a flow transmitter 108, which provides a process variable (PV) signal 116 to the controller 104. Because chemical processes are typically continuous processes, the CO signal 112, the SP signal 114, and the PV signal 116 can each provide multivariate time series data from a control loop 100 for a control valve (e.g., a flow control valve 106) of the chemical process.

[0022] When a control loop (e.g., control loop 100) is properly implemented, it operates to automatically adjust the value of PV to equal the desired value of SP. However, poor control loop performance (e.g., nonlinear control valve events such as oscillations within the control loop) can result in detrimental problems in a chemical plant that disrupt normal plant operation. Such detrimental problems can include reduced product quality, increased wear on plant equipment, increased both energy and material consumption, and manual manipulation of the control loop. Nonlinear control valve events such as oscillations in a control loop can result from multiple causes, including improper control tuning, multi-loop interactions, sensor failures, external vibration disturbances, and control valve issues. Control valve issues can include, for example, stiction, backlash, deadband, deadzone, hysteresis, and saturation, which can cause the controller to excessively move the valve stem when attempting to regulate the control loop.

[0023] Of the valve problems mentioned above, stiction accounts for a large proportion of control valve problems in chemical plants. Stiction occurs when static friction within the valve resists proper valve movement, resulting in a delay between the CO signal and the movement of the valve stem position. When a control valve experiences stiction, its stem may not move even though the CO signal continues to change. As a result, the relationship between the CO signal and the valve stem position becomes nonlinear. This situation causes oscillations in the control loop, which affects the control and performance of the chemical process.

[0024] The present disclosure is directed to improving chemical plant operation by monitoring and evaluating control loops for nonlinear control valve events using the techniques described herein. Specifically, the present disclosure provides methods and systems for identifying nonlinear control valve events in a chemical process control loop. The nonlinear control valve events are based on predetermined features of RGB images that are identified by a CNN trained to identify nonlinear control valve events. For various embodiments, such nonlinear control valve events may include valve stiction, and identifying and addressing such nonlinear control valve events may help improve chemical plant operation.

[0025] As provided herein, a method of the present disclosure includes receiving a set point (SP) signal, a process variable (PV) signal, and a controller output (CO) signal that constitute multivariate time series data from a control loop for a control valve of a chemical process, where the CO signal is received from a controller that controls the position of a control valve in the chemical process, as seen and described in Figure 1. The multivariate time series data from each of the SP signal, the PV signal, and the CO signal are transformed using a CWT to provide respective CWT coefficients for each of the SP signal, the PV signal, and the CO signal.

[0026] The CWT of each of the above signals to provide the resulting CWT coefficients is calculated using Equation #1:

[0027]

number

[0028] The CWT generally decomposes signals using wavelets that are highly localized in time. The CWT can provide higher resolution for discrete transforms, thus offering the ability to glean more information from a signal than typical frequency transforms such as the Fourier transform (or any other spectral technique) or the discrete wavelet transform. The CWT allows for the use of a range of wavelets with scales that span the scales of interest in the signal, such that small-scale signal components correlate well with smaller-scale wavelets and therefore appear with higher energy at smaller scales in the transform. Similarly, large-scale signal components correlate well with larger-scale wavelets and therefore appear with higher energy at larger scales in the transform. Thus, components at different scales can be separated and extracted in the wavelet transform domain. Furthermore, the use of a continuous range of wavelets in scale and time location allows for higher-resolution transforms than are possible with discrete techniques.

[0029] As provided herein, the CWT coefficients of a signal may be used to create a scalogram, which provides a visualization of the signal's energy in time and frequency and allows for differentiation between signal fluctuation types, including the nonlinear control valve events discussed herein. For various embodiments, a variety of suitable wavelet functions may be used in conjunction with the present disclosure. Preferably, the present disclosure uses the Morlet wavelet, which is a complex wave within a scaled Gaussian envelope.

[0030] For various embodiments, each CWT coefficient is assembled into a respective matrix, which is then merged to provide a red, green, and blue (RGB) image. For example, the matrix values ​​can be converted to an RGB color matrix to generate an RGB image. For example, to achieve this, a Gramian angle field can be used, where the matrix is ​​first converted to a GAF image and the coordinates are mapped using a polar coordinate system so that the GAF represents the temporal correlation between each time point in the time series data. The GAF can be combined into a larger image whose values ​​are converted to an RGB color matrix. The RGB image provides a three-channel representation of each signal that can be analyzed by a CNN. As described herein, a CNN is a deep learning algorithm that can take an input image and distinguish and assign importance to various aspects of the image through learnable weights and biases. In various embodiments, the RGB image is input to a CNN trained to identify nonlinear control valve events in a chemical process control loop based on predetermined features of the RGB image. In response to the CNN identifying the predetermined features, controller settings are adjusted to eliminate the nonlinear control valve events. As described herein, operations can be notified of controller anomalies and adjustments can then be made to the controller settings to minimize and / or eliminate nonlinear control valve events. Additionally, identifying control valve problems allows for investigation of the root cause of the nonlinear control valve events (e.g., stiction problems) for subsequent correction.

[0031] In various embodiments, CNNs involve the use of convolution, a linear operation involving the multiplication of a matrix of weights, called a filter or kernel, by a matrix of an RGB image. The kernel is systematically applied to each overlapping portion or filter-sized patch of input data (channels of the RGB image of the signal) from left to right and top to bottom. In this way, convolutional layers within CNNs systematically apply learned kernels to the input image to create feature maps that summarize the presence of those features in the input. Convolutional layers can perform dot products of convolutional kernels with the layer's input matrix to generate intermediate results. Kernels can be transposed, multiplied individually by all corresponding values, and added together. Collectively, the intermediate results form a result (e.g., a feature map) that can serve as input to the next layer. Because there are multiple kernels for each window in the input, the final result can provide a three-channel representation.

[0032] CNNs also include pooling in convolutional layers, which helps downsample features within the layer. Two common pooling methods are average pooling and max pooling, which summarize the average presence and most activated presence of a feature, respectively, with max pooling being preferred. The result of using a pooling layer after a convolutional layer (often repeated multiple times) is to create a downsampled or pooled feature map, which summarizes the features detected in the input. Fully connected layers can also be added to learn nonlinear control valve events as provided herein. Examples of CNNs useful in the present disclosure include LeNet, AlexNet, VGGNet, GoogLeNet, ResNet, and ZFNet, among others.

[0033] For various embodiments, the predetermined features of the RGB images identified by the CNN as representing nonlinear control valve events can be used to identify the control valve generating the nonlinear control valve event. For various embodiments, the CWT transform of the SP, PV, and OP signals generates unique RGB images, as shown in Figures 2A (RGB image showing valve stiction) and 2C (RGB image showing non-valve stiction) for stiction and non-stiction control loops, respectively. The magnitude of the CWT coefficients results in the variability of the image color and pattern intensity in Figures 2A and 2C, respectively, and the difference is shown at 220 in Figures 2A and 2C. The CWT transforms seen in Figures 2A and 2C are then processed within the CNN through trained filters and layers to provide the feature maps illustrated in Figures 2B and 2D. The final classification score associated with stiction or non-stiction is calculated using a softmax activation function following the final fully connected layer in the neural network. The final decision is made based on the highest classification score or highest likelihood that the observation belongs to a particular class. The CNN identifies discrepancies in image intensity and overlap to determine whether the control loop has valve stiction characteristics, as shown in the CNN feature maps in Figures 2B (information detected by the CNN network to identify stiction) and 2D (information detected by the CNN network to identify non-stiction) for control loops with stiction and non-stiction, respectively. Differences in intensity are identified by the CNN (e.g., at 230), and these areas are regions where the CWT information from the three signals overlap or diverge. The intensity of these activations is used to determine whether the control loop exhibits stiction or non-stiction. Valves exhibiting stiction, where the PV does not track the output well, exhibit different intensities and combined color profiles than when stiction is not present, as shown in Figures 2B and 2D.

[0034] Although specific embodiments have been described above, these embodiments are not intended to limit the scope of the present disclosure, even if only a single embodiment is described with respect to a particular feature. The example features provided in this disclosure are intended to be illustrative rather than limiting, unless otherwise stated. The above description is intended to cover such alternatives, modifications, and equivalents as would be apparent to one skilled in the art having the benefit of this disclosure.

[0035] The scope of the present disclosure includes any feature or combination of features (either explicit or implicit) disclosed herein, or any generalization thereof, whether or not it alleviates any or all of the problems addressed herein. Various advantages of the present disclosure have been described herein, but embodiments may provide some, all, or none of such advantages, or may provide other advantages.

[0036] In the Detailed Description, some features are grouped together in a single embodiment for the purpose of streamlining the disclosure. This method of disclosure is not to be interpreted as reflecting an intention that the disclosed embodiments of the present disclosure require more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive subject matter lies in less than all features of a single disclosed embodiment. Accordingly, the following claims are incorporated into the Detailed Description herein, with each claim standing on its own as a separate embodiment.

Claims

1. 1. A method comprising: receiving a set point (SP) signal, a process variable (PV) signal, and a controller output (CO) signal comprising multivariate time series data from a control loop for a control valve of a chemical process; receiving the CO signal from a controller that controls the position of the control valve; transforming the multivariate time series data from each of the SP signals, the PV signals, and the CO signals using a continuous wavelet transform (CWT) to provide respective CWT coefficients for each of the SP signals, the PV signals, and the CO signals; assembling the respective CWT coefficients into respective matrices; merging the respective matrices to provide a red, green, and blue (RGB) image; inputting the RGB image into a deep convolutional neural network (CNN) trained to identify nonlinear control valve events in the control loop of the chemical process based on predetermined features of the RGB image; and adjusting settings of the controller to minimize the nonlinear control valve event in response to the CNN identifying the predetermined feature.

2. The method of claim 1 , wherein the multivariate time series data is obtained at a predetermined frequency over a predetermined time interval.

3. The method of claim 2 , wherein the nonlinear control valve event is a vibration event.

4. The method according to any one of claims 1 to 3, wherein Morlet wavelets are used when transforming the multivariate time series data using the CWT.

5. The method of any one of claims 1 to 4, wherein the CNN identifies the predetermined features based on changes in frequency information from a time-series multivariate signal.

6. 6. The method of claim 5, wherein the predetermined features of the RGB image are associated with dynamic frequency spectral activations identified by the CNN when the nonlinear control valve event is identified.

7. the CNN is pre-trained to identify the nonlinear control valve events in the RGB images; 7. The method of claim 1, further comprising training the CNN via transfer learning using the plurality of RGB images of at least one pairwise comparison of the SP signal, the PV signal, and the CO signal of the multivariate time series data of the controller from previously identified nonlinear control valve events of the controller in a plurality of RGB images.

8. The method of any one of claims 1 to 7, wherein the non-linear control valve event is a result of the controller being in a state of stiction.

9. The method of any one of claims 1 to 8, wherein the controller is a flow controller in the control loop of the chemical process.

10. The method of any one of claims 1 to 8, wherein the controller can be a level controller, a flow controller, a pressure controller, or a temperature controller in the control loop of the chemical process.

11. 11. The method of claim 1, wherein transforming the comparison of at least one pair of the SP signals, the PV signals, and the CO signals of the multivariate time series data comprises transforming the multivariate time series data without preprocessing the multivariate time series data.

12. 1. A system comprising: A detector comprising: a detector configured to receive a set point (SP) signal, a process variable (PV) signal, and a controller output (CO) signal, the set point (SP), a process variable (PV), and a controller output (CO) signal comprising multivariate time series data from a control loop for a control valve of a chemical process; a deep convolutional neural network (CNN) trained on the plurality of RGB images of at least one pairwise comparison of the SP signal, the PV signal, and the CO signal of the multivariate time series data of the controller from previously identified nonlinear control valve events of the controller in the plurality of RGB images; a controller coupled to the detector and the CNN, converting the at least one pair of the SP signal, the PV signal, and the CO signal of the multivariate time series data into an RGB image using a continuous wavelet transform; Input the RGB image to the CNN, using the CNN to identify nonlinear control valve events of the controller; a controller configured to adjust settings of the controller to eliminate the nonlinear control valve events of the controller identified by the CNN.

13. The system of claim 12 , wherein the controller is selected from a flow controller and a level controller.

14. 14. The system of claim 12 or 13, wherein the non-linear control valve event is a result of the controller being in a state of stiction.

15. The system of any one of claims 12 to 14, wherein the controller is configured to convert the multivariate time series data into the RGB image without pre-processing the multivariate time series data.

Images