Endpoint detection in manufacturing processes using near-infrared spectroscopy and machine learning techniques
By using a support vector machine (SVM) classification model and generating an iterative model based on multivariate spectral data, the accuracy and robustness issues of steady-state detection in the manufacturing process in existing technologies are solved, and efficient steady-state identification is achieved.
Patent Information
- Application Number
- CN202311307168.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2017-05-04
- Filing Date
- 2018-05-04
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2038-05-04
AI Technical Summary
Existing univariate and principal component analysis techniques are difficult to accurately and robustly detect steady states during manufacturing processes, especially in the case of multivariate data.
A support vector machine (SVM) classification model is used to generate an iterative model based on multivariate spectral data. The model is trained on the spectral data to determine whether the manufacturing process has reached a steady state.
It improves the accuracy and robustness of steady-state detection in the manufacturing process, enabling real-time or near-real-time identification of process transition points.
Smart Images

Figure CN117290780B_ABST
Abstract
Description
[0001] This application is a divisional application of the application filed on May 4, 2018, with application number 201810419832.5, entitled "Endpoint detection of manufacturing process using near-infrared spectroscopy and machine learning technology". Technical Field
[0002] This disclosure generally relates to endpoint detection during the manufacturing process. Background Technology
[0003] As part of implementing a manufacturing process (e.g., continuous manufacturing, batch manufacturing), process analysis technology (PAT) systems can be used to generate real-time or near real-time data (e.g., spectral data) for monitoring and controlling the manufacturing process. Continuous manufacturing processes allow raw materials to be fed into the system and finished products (e.g., pharmaceutical products) to be discharged from the system in a continuous manner. In other words, in continuous manufacturing, the individual steps of the manufacturing process are transformed into a single, integrated manufacturing process (e.g., rather than a series of discrete steps as in batch manufacturing). Summary of the Invention
[0004] According to some possible implementations, the device may include one or more processors to: receive training spectral data associated with a manufacturing process transitioning from an unstable state to a stable state; generate multiple iterations of a support vector machine (SVM) classification model based on the training spectral data; determine multiple predicted transition times associated with the manufacturing process based on the multiple iterations of the SVM classification model, wherein a predicted transition time among the multiple predicted transition times may identify the time during which the corresponding iteration of the SVM classification model predicts the manufacturing process transitioning from an unstable state to a stable state during the manufacturing process; and generate a final SVM classification model associated with determining whether the manufacturing process has reached a stable state based on the multiple predicted transition times.
[0005] According to some possible implementations, a non-transitory computer-readable medium may store one or more instructions that, when executed by one or more processors, cause the one or more processors to: receive training spectral data associated with a first execution of a manufacturing process transitioning from an unstable state to a stable state; iteratively generate an SVM classification model based on the training spectral data associated with determining whether another execution of the manufacturing process has transitioned from an unstable state to a stable state; receive additional spectral data associated with a second execution of the manufacturing process; and determine, based on the SVM classification model and the additional spectral data, whether the second execution of the manufacturing process has transitioned from an unstable state to a stable state.
[0006] According to some possible implementations, a method may include: receiving first spectral data associated with a first execution of a manufacturing process transitioning from an unstable state to a stable state by a device; generating multiple iterations of an SVM classification model by the device and based on the first spectral data; determining multiple predicted transition times associated with the first execution of the manufacturing process by the device and based on the multiple iterations of the SVM classification model; generating a final SVM classification model associated with determining whether another execution of the manufacturing process has reached a stable state by the device and based on the multiple predicted transition times; receiving second spectral data associated with a second execution of the manufacturing process by the device; and determining whether the second execution of the manufacturing process has reached a stable state by the device based on the final SVM classification model and the second spectral data. Attached Figure Description
[0007] Figures 1A-1C This is a diagram providing an overview of the exemplary implementation described herein;
[0008] Figure 2 A diagram of an example environment in which the systems and / or methods described herein can be implemented;
[0009] Figure 3 yes Figure 2 A diagram of example components of one or more devices;
[0010] Figure 4 This is a flowchart of an example process for generating an SVM classification model to detect when a manufacturing process has reached a steady state;
[0011] Figure 5A and Figure 5B This is an example graphical representation related to determining the transition time of a manufacturing process based on the transition time predicted iteratively by an SVM classification model associated with the manufacturing process;
[0012] Figure 6 This is a flowchart of an example process used to determine whether a manufacturing process has reached a steady state based on spectral data and an SVM classification model.
[0013] Figure 7A and Figure 7B This is an example graphical representation illustrating a simplified decision boundary associated with an SVM classification model; and
[0014] Figure 8 Based on Figure 7A and Figure 7B A graphical representation of example decision values where the decision boundary is determined. Detailed Implementation
[0015] The following detailed description of the exemplary embodiments is with reference to the accompanying drawings. The same reference numerals in different drawings may identify the same or similar elements.
[0016] Manufacturing processes (e.g., continuous or batch manufacturing processes for manufacturing pharmaceutical products) may involve one or more state transitions, such as a transition from an unstable state (e.g., a state in which the properties of a material and / or compound change over time) to a stable state (e.g., a state in which the properties of a material and / or compound remain substantially constant over time). For example, a mixing process included in a manufacturing process for manufacturing a pharmaceutical product may involve a transition of the spectral properties of a compound from an unstable state (e.g., at the start of the mixing process) to a stable state (e.g., indicating that the mixing process is complete).
[0017] Therefore, to improve efficiency and / or optimize manufacturing processes, the manufacturing process should be monitored to determine (e.g., in real-time or near real-time) when it has reached a steady state. A possible technique for detecting the state of a manufacturing process is a univariate model that detects the state based on a single variable associated with the manufacturing process (such as total spectral intensity). Another possible technique is to use principal component analysis (PCA) to identify a set of variables (i.e., principal components) for detecting the state of the manufacturing process and to detect when the process has reached a steady state based on monitoring this set of variables. However, in some cases, the data measured during the manufacturing process can be multivariate data (e.g., NIR spectra including data associated with hundreds of variables). Therefore, because univariate or PCA techniques focus on a relatively small number of variables, these techniques may lead to inaccurate state detection and / or may not be robust enough to ensure accurate state detection.
[0018] The embodiments described herein provide support vector machine (SVM) classification models capable of generating support vector machine (SVM) classification models for determining whether a manufacturing process (e.g., continuous manufacturing process, batch manufacturing process, etc.) has reached a steady state, and detection devices that use SVM classification models and multivariate spectral data associated with the manufacturing process to determine whether the manufacturing process has reached a steady state. In some embodiments, the SVM classification model may consider multiple variables (e.g., 80 variables, 120 variables, 150 variables, etc.), thereby increasing the accuracy and / or robustness of the SVM classification model (e.g., compared to the techniques described above).
[0019] Figures 1A-1C This is a diagram providing an overview of the exemplary implementation 100 described herein. (See diagram for example.) Figure 1AAs shown, and indicated by reference numeral 102, the detection device can receive training spectral data associated with the manufacturing process. Based on iterations of support vector machine (SVM) classification models that may be generated to correlate with whether the manufacturing process has reached a steady state, the training spectral data (sometimes referred to as first spectral data) can include spectral data associated with the manufacturing process. For example, the training spectral data can include spectra measured by a spectrometer during the execution of the manufacturing process (e.g., multivariate time-series data such as NIR spectra). The execution (performance) of the manufacturing process during which the training spectral data is collected can be referred to as the first execution of the manufacturing process.
[0020] As shown, the training spectral data can include spectral data measured at the beginning of the manufacturing process (time t0) and at a time when the manufacturing process is known to be in a non-steady state (time t). us0 (Measured spectral data), when the manufacturing process is known to be in a steady state (time t) ss0 ) Spectral data measured at the end of the manufacturing process (time t) e The measured spectral data and the time t0 and time t2 when the state of the manufacturing process is unknown. e Spectral data measured between.
[0021] As shown by reference numerals 104 to 116 in the attached figures, the detection device can generate iterations of the SVM classification model based on training spectral data. For example, as shown by reference numeral 104, in order to generate the initial iteration (iteration 0) of the SVM classification model, the detection device can create an initial set of unstable state data (e.g., including data from time t0 to time t) based on the training spectral data. us0 Spectral data measured) and initial set steady-state data (e.g., including data from time t) ss0 By time t e (Measured spectral data).
[0022] As shown by reference numeral 106 in the attached figure, the detection device can generate the initial iteration of the SVM classification model based on the initial set of unstable state data and the initial set of stable state data. As illustrated, based on providing training spectral data as input to the initial iteration of the SVM classification model, the detection device can determine the initial prediction transition time (t) associated with the initial iteration of the SVM classification model. trans0 (For example, the initial iteration of an SVM classification model predicts the time it takes for a manufacturing process to transition from an unstable state to a stable state.)
[0023] As shown in Figure 108, in order to generate the first iteration (Iteration 1) of the SVM classification model, the detection device can create a first set of non-stationary state data based on the training spectral data (e.g., including data from time t0 to time t). trans0Spectral data measured) and the first set of steady-state data (e.g., including data from time t) ss0-dt*1 By time t e (Measured spectral data). It is worth noting that the first set of non-stationary state data includes spectral data measured up to the transition time predicted by the initial iteration of the SVM classification model, while the first set of stationary state data includes the spectral data included in the initial set of stationary state data, as well as the data measured at time t. ss0 Spectral data measured at a previous time step.
[0024] As shown by reference numeral 110 in the attached figure, the detection device can generate the first iteration of the SVM classification model based on the first set of unstable state data and the first set of stable state data. As illustrated, based on providing training spectral data as input to the first iteration of the SVM classification model, the detection device can determine the transition time (t) of the first prediction associated with the first iteration of the SVM classification model. trans1 (For example, the first iteration of an SVM classification model predicts the time it takes for a manufacturing process to transition from an unstable state to a stable state.)
[0025] In some implementations, the detection device may generate n (n>1) iterations of the SVM classification model and determine the predicted transition time in such a manner until the earliest time associated with the steady-state dataset used to generate the nth iteration of the SVM classification model is a time threshold amount (e.g., a time step) from the time when the known manufacturing process is in a non-steady state (e.g., until the steady-state set includes data from t...). ss0-dt*n =t us0+dt By time t e (Up to the measured spectral data).
[0026] For example, such as Figure 1A As shown by reference numeral 112 in the attached figure, in order to generate the nth iteration (iteration n) of the SVM classification model, the detection device can create the nth set of non-stationary state data (e.g., including data from time t0 to time t) based on the training spectral data. trans(n-1) Spectral data measured) and the nth set of steady-state data (e.g., including data from time t). ss0-dt*n=tus0+dt By time t e (Measured spectral data). It is worth noting that the nth set of non-stationary state data includes spectral data measured up to the time predicted by the (n-1)th (i.e., previous) iteration of the SVM classification model, while the nth set of stationary state data includes the spectral data included in the (n-1)th set of stationary state data, as well as the data measured at time t. ss0-dt*(n-1) Spectral data measured at a previous time step.
[0027] As shown by reference numeral 114 in the attached figure, the detection device can generate the nth iteration of the SVM classification model based on the nth set of unstable state data and the nth set of stable state data. As illustrated, based on providing training spectral data as input to the nth iteration of the SVM classification model, the detection device can determine the transition time (t) of the nth prediction associated with the nth iteration of the SVM classification model. trans(n) (For example, the nth iteration of an SVM classification model predicts the time it takes for a manufacturing process to transition from an unstable state to a stable state.)
[0028] As shown by reference numeral 116 in the figures, the inspection device can determine the dominant (e.g., most predicted) transition time associated with the manufacturing process based on n transition times predicted through n iterations of the SVM classification model. As shown by reference numeral 118 in the figures, the inspection device can generate a final SVM classification model based on the transition times associated with the manufacturing process. For example, the inspection device can create a final set of unsteady-state data comprising spectral data measured before the determined dominant transition time, and a final set of steady-state data comprising training spectral data measured at or after the determined dominant transition time. Accordingly, as shown, the inspection device can generate a final SVM classification model based on the final set of unsteady-state data and the final set of steady-state data.
[0029] like Figure 1B As shown, and via reference numeral 120, the detection device can (at a later time) identify an SVM classification model (e.g., based on the final SVM classification model generated as described) to detect whether the execution of the manufacturing process has reached a steady state. For example, the detection device can identify the SVM classification model based on receiving an indication that the manufacturing process is starting or has already started.
[0030] As shown by reference numeral 122 in the attached figure, the detection device can receive spectral data associated with the manufacturing process (sometimes referred to as second spectral data or additional spectral data) during a second execution of the manufacturing process. For example, as shown, the spectrometer can receive spectral data at a given time during the execution of the manufacturing process (e.g., time t). A Spectral data is measured at [location] and can be provided to the detection equipment. The execution of the manufacturing process during which spectral data is collected as input to the SVM classification model can be referred to as the second execution of the manufacturing process.
[0031] As further shown, the detection equipment can determine the manufacturing process at time t based on spectral data and an SVM classification model. A Whether the location is in a stable state. For example, as shown by reference numeral 124 in the attached figure, the detection device can provide information at time t. AThe spectral data measured at the location is used as input to the SVM classification model. As shown in Figure 126, the inspection equipment can determine the manufacturing process at time t based on the output of the SVM classification model. A The process is not in a stable state. In some implementations, the detection device can determine whether the manufacturing process is in a stable state based on the decision boundary associated with the SVM classification model, as described below.
[0032] like Figure 1C As shown, and by reference numeral 128, the detection device can be detected at a later time (time t) during the execution of the manufacturing process. B It receives spectral data associated with the manufacturing process. For example, as shown, the spectrometer can receive spectral data at time t. B It measures spectral data and can provide the spectral data to the detection equipment.
[0033] As further shown, the detection equipment can determine the manufacturing process at time t based on spectral data and an SVM classification model. B Whether the location is in a stable state. For example, as shown by reference numeral 130 in the attached figure, the detection device can provide information at time t. B The spectral data measured at the location is used as input to the SVM classification model. As shown in Figure 132, the inspection equipment can determine the manufacturing process at time t based on the output of the SVM classification model. B It is in a stable state.
[0034] As indicated by reference numeral 134 in the accompanying drawings, in some embodiments, the detection device may (optionally) determine a quantitative measure associated with the steady state based on the determination that the manufacturing process has reached a steady state. The quantitative measure may include a measure indicating a quantitative attribute associated with the steady state, such as the concentration of a component of a compound in the steady state, the particle size in the steady state, etc. For example, the detection device may store or access a regression model (e.g., a partial least squares (PLS) regression model, a support vector regression (SVR) model) that receives spectral data based on its detection of a steady state as input and provides a quantitative measure associated with that steady state as output.
[0035] As shown by reference numeral 136 in the accompanying drawings, based on the determination that the manufacturing process has reached a steady state, the detection equipment can (e.g., provide an indication that the manufacturing process has reached a steady state to user equipment associated with monitoring the manufacturing process). As further shown, the detection equipment can also provide information associated with quantitative measurements.
[0036] In this way, the detection equipment can generate an SVM classification model to determine whether the manufacturing process has reached a steady state, and use the SVM classification model and multivariate spectral data associated with the manufacturing process to determine whether the manufacturing process has reached a steady state.
[0037] As mentioned above, Figures 1A-1C This is provided as an example only. Other examples are possible and may differ from those provided. Figures 1A-1C The example described.
[0038] Figure 2 This is a diagram of an example environment 200 that can implement the systems and / or methods described in this paper. (See diagram 200 for example environment 200.) Figure 2 As shown, environment 200 may include one or more spectrometers 210-1 to 210-X (X≥1) (collectively referred to herein as spectrometer 210, and individually as spectrometer 210), detection device 220, user equipment 230, and network 240. The devices of environment 200 may be interconnected via wired connections, wireless connections, or a combination of wired and wireless connections.
[0039] Spectrometer 210 includes a device capable of performing spectral measurements on a sample (e.g., a sample associated with a manufacturing process). For example, spectrometer 210 may include a desktop (i.e., non-handheld) spectrometer device that performs spectroscopy (e.g., vibrational spectroscopy, such as near-infrared (NIR) spectroscopy, mid-infrared spectroscopy, Raman spectroscopy, etc.). In some embodiments, spectrometer 210 may be able to provide spectral data acquired by spectrometer 210 for analysis by another device (such as detection device 220).
[0040] The detection device 220 includes one or more devices capable of detecting whether a manufacturing process has reached a steady state based on a classification model associated with the manufacturing process and spectral data associated with the manufacturing process. For example, the detection device 220 may include a server, a group of servers, a computer, a cloud computing device, etc. In some embodiments, the detection device 220 may be able to generate a classification model based on training spectral data associated with the manufacturing process. In some embodiments, the detection device 220 may receive information from and / or send information to another device in the environment 200 (such as spectrometer 210 and / or user equipment 230).
[0041] User equipment 230 includes one or more devices capable of receiving, processing, and / or providing information related to whether the manufacturing process has reached a steady state. For example, user equipment 230 may include communication and computing devices such as desktop computers, mobile phones (e.g., smartphones, cordless phones, etc.), laptop computers, tablets, handheld computers, wearable communication devices (e.g., smartwatches, smart glasses, etc.) or similar types of devices.
[0042] Network 240 includes one or more wired and / or wireless networks. For example, network 240 may include cellular networks (e.g., Long Term Evolution (LTE) networks, 3G networks, Code Division Multiple Access (CDMA) networks, etc.), Public Land Mobile Networks (PLMN), Local Area Networks (LAN), Wide Area Networks (WAN), Metropolitan Area Networks (MAN), Telephone Networks (e.g., Public Switched Telephone Network (PSTN)), Private Networks, Ad Hoc Networks, Intranets, the Internet, Fiber-based Networks, Cloud Computing Networks, etc., and / or combinations of these or other types of networks.
[0043] Provided as an example Figure 2 The number and layout of the devices and networks shown are illustrated. In reality, there may be... Figure 2 The examples shown are those compared to additional equipment and / or networks, fewer equipment and / or networks, different equipment and / or networks, or equipment and / or networks with different arrangements. Furthermore, Figure 2 The two or more devices shown can be implemented within a single device, or, Figure 2 The single device shown can be implemented as multiple, distributed devices. Alternatively or additionally, a group of devices in environment 200 (e.g., one or more devices) can perform one or more functions described as being performed by another group of devices in environment 200.
[0044] Figure 3 This is a diagram of example components of device 300. Device 300 may correspond to spectrometer 210, detection device 220, and / or user equipment 230. In some embodiments, spectrometer 210, detection device 220, and / or user equipment 230 may include one or more devices 300 and / or one or more components of device 300. Figure 3 As shown, device 300 may include bus 310, processor 320, memory 330, storage component 340, input component 350, output component 360 and communication interface 370.
[0045] Bus 310 includes components that allow communication between components of device 300. Processor 320 is implemented in hardware, firmware, or a combination of hardware and software. Processor 320 includes a central processing unit (CPU), graphics processing unit (GPU), accelerated processing unit (APU), microprocessor, microcontroller, digital signal processor, field-programmable gate array (FPGA), application-specific integrated circuit (ASIC), or another type of processing component. In some embodiments, processor 320 includes one or more processors that can be programmed to perform functions. Memory 330 includes random access memory (RAM), read-only memory (ROM), and / or another type of dynamic or static storage device (e.g., flash memory, magnetic storage, and / or optical storage) that stores information and / or instructions used by processor 320.
[0046] Storage component 340 stores information and / or software related to the operation and use of device 300. For example, storage component 340 may include hard disks (e.g., magnetic disks, optical disks, magneto-optical disks, and / or solid-state disks), compact discs (CDs), digital versatile discs (DVDs), floppy disks, cassette tapes, magnetic tapes, and / or other types of non-transitory computer-readable media and corresponding drives.
[0047] Input component 350 includes components that allow device 300 to receive information such as via user input (e.g., touchscreen display, keyboard, keypad, mouse, button, switch, and / or microphone). Alternatively or additionally, input component 350 may include sensors for sensing information (e.g., a Global Positioning System (GPS) component, accelerometer, gyroscope, and / or actuator). Output component 360 includes components that provide output information from device 300 (e.g., a display, speaker, and / or one or more light-emitting diodes (LEDs)).
[0048] Communication interface 370 includes transceiver-like components (e.g., a transceiver and / or separate receiver and transmitter) that enable device 300 to communicate with other devices, such as via a wired connection, a wireless connection, or a combination of wired and wireless connections. Communication interface 370 may allow device 300 to receive information from and / or provide information to another device. For example, communication interface 370 may include an Ethernet interface, an optical interface, a coaxial interface, an infrared interface, a radio frequency (RF) interface, a universal serial bus (USB) interface, a Wi-Fi interface, a cellular network interface, etc.
[0049] Device 300 can perform one or more of the processes described herein. Device 300 can perform these processes in response to processor 320 executing software instructions stored in non-transitory computer-readable media (such as memory 330 and / or storage component 340). Computer-readable media are defined herein as non-transitory memory devices. Memory devices include memory space within a single physical storage device or memory space distributed across multiple physical storage devices.
[0050] Software instructions may be read into memory 330 and / or storage component 340 via communication interface 370 from another computer-readable medium or from another device. When executed, the software instructions stored in memory 330 and / or storage component 340 may cause processor 320 to perform one or more processes described herein. Alternatively, hardwired circuitry may be used in place of or in combination with software instructions to perform one or more processes described herein. Therefore, the embodiments described herein are not limited to any particular combination of hardware circuitry and software.
[0051] Provided as an example Figure 3 The number and arrangement of components are shown. In practice, device 300 may include [components]. Figure 3 The components shown are those that are additional, fewer, different, or arranged differently compared to other components. Alternatively, a set of components of device 300 (e.g., one or more components) may perform one or more functions described as being performed by another set of components of device 300.
[0052] Figure 4 This is a flowchart of an example process 400 for generating a classification model used to detect when a manufacturing process reaches a steady state. In some implementations, Figure 4 One or more process frames can be executed by the detection device 220. In some embodiments, Figure 4 One or more process frames may be performed by another device or group of devices (such as spectrometer 210 and / or user equipment 230) that is separate from or includes the detection device 220.
[0053] like Figure 4 As shown, process 400 may include receiving training spectral data associated with the manufacturing process (box 410). For example, detection device 220 may receive training spectral data associated with the manufacturing process.
[0054] Based on which iterations can generate the SVM classification model, the training spectral data can include spectral data associated with the manufacturing process. For example, the training spectral data can include spectra measured by spectrometer 210 during the execution of the manufacturing process (e.g., multivariate time-series data such as NIR spectra). In some embodiments, the manufacturing process can be a continuous manufacturing process or a batch manufacturing process. In some embodiments, the detection device 220 can generate iterations of the SVM classification model based on the training spectral data, as described below.
[0055] In some implementations, the training spectral data may include historical spectra measured at different times (e.g., periodically at a series of time steps) during earlier executions of the manufacturing process. For example, the training spectral data may include spectra measured at the start time (hereinafter referred to as time t0) of an earlier execution of the manufacturing process and at the end time (hereinafter referred to as time t0) of an earlier execution of the manufacturing process. e The spectrum measured at ( ).
[0056] As another example, training spectral data could include time intervals (referred to in this paper as time t) during earlier executions of a known manufacturing process that were already in a non-stationary state. us0 The spectrum measured at ( ). In some implementations, time t us0 It can be the same as time t0 (e.g., because the manufacturing process is in a non-steady state at the beginning of the manufacturing process). Or, time t us0 This can be a time after time t0, such as the time of one time step after time t0, the time of five time steps after time t0, the time of forty time steps after time t0, etc. In some implementations, time t... us0 It could be a time after time t0, assuming that earlier executions of the manufacturing process at that time were already in a non-steady state.
[0057] As an additional example, training spectral data could include data from earlier executions of a known manufacturing process that have reached a steady state (referred to here as time t). ss0 The spectrum measured at time t. In some implementations, time t... ss0 It can be related to time t e The same amount of time (e.g., because the manufacturing process reaches a steady state at the end of the manufacturing process). Or, time t. ss0 It can be time t e Previous time, such as time t e The previous time step, time t e The previous 5 time steps, time t e The time of the previous 40 time steps, etc. In some implementations, time t ss0It can be time t e Previously, it was assumed that the earlier stages of the manufacturing process had reached a stable state.
[0058] As another example, training spectral data can include data at time t0 and time t... e The spectrum was measured at a time interval during which the state of the earlier stages of the manufacturing process was unknown.
[0059] In some implementations, the detection device 220 may receive training spectral data from one or more other devices, such as one or more spectrometers 210 that acquire training spectral data during an earlier execution of the manufacturing process or a server device that stores training spectral data measured by one or more spectrometers 210 during an earlier execution of the manufacturing process.
[0060] In some implementations, training spectral data may be associated with multiple earlier executions of the manufacturing process, where initial conditions (e.g., total weight, particle size, distribution, humidity level, etc.) vary across the executions. In this case, multiple sets of training spectral data associated with the multiple executions of the manufacturing process can be averaged and / or additionally combined to form training spectral data. In some implementations, measuring the training spectral data by varying the initial conditions results in increased accuracy and / or robustness of the SVM classification model generated based on the training spectral data (e.g., compared to an SVM classification model generated based on a manufacturing process with a single set of initial conditions).
[0061] In some embodiments, the detection device 220 can perform dimensionality reduction on the training spectral data. Dimensionality reduction can include reducing the number of variables in the multivariate training spectral data upon which an SVM classification model can be generated. In some embodiments, principal component analysis (PCA) can be used to perform dimensionality reduction, whereby principal components (i.e., subsets of the multivariates) are identified for use in generating the SVM classification model. Alternatively, or optionally, variable selection techniques can be used to perform dimensionality reduction, whereby variables among the multivariates are selected that are discriminative, for example, compounds associated with a manufacturing process. Examples of such variable selection techniques include selectivity ratio (SR) techniques, variable importance in projection (VIP) techniques, etc. In some embodiments, performing dimensionality reduction can, for example, improve the interpretability of the SVM classification model and / or improve the generation of the SVM classification model (e.g., compared to an SVM classification model generated on the entire set of training spectral data).
[0062] like Figure 4As further shown, process 400 may include creating a non-stationary state dataset and a stationary state dataset based on training spectral data (box 420). For example, detection device 220 may create a non-stationary state dataset and a stationary state dataset based on training spectral data.
[0063] To generate iterations for the SVM classification model, the non-stationary state dataset can include spectral data from the training spectral data corresponding to the times when the manufacturing process is assumed to be in a non-stationary state. For example, the initial set of non-stationary state data associated with the initial iterations for generating the SVM classification model can include data from time t0 to time t... us0 The spectral data measured at time t0. Continuing this example, another set of unsteady-state data associated with the next iteration of generating the SVM model could be included from time t0 to time t... trans0 Spectral data measured at time t, where time t trans0 This refers to the transition time predicted by the initial iteration of the SVM classification model (e.g., the time of transition from an unstable state to a stable state). Typically, the nth set of unstable state data associated with the nth iteration that generates the SVM model can include data from time t0 to time t... trans(n-1) Spectral data measured at time t, where time t trans(n-1) The transformation time is determined by the (n-1)th (i.e., the previous) iteration of the SVM classification model. Additional details regarding the iterations that generate the SVM classification model are described below.
[0064] In some implementations, the non-stationary state dataset can be updated, modified, and / or recreated to generate each iteration of the SVM model, for example, by adding additional spectral data included in the training spectral data to the non-stationary state dataset associated with the previous iteration that generated the SVM model. For example, the nth set of non-stationary state data used to generate the nth iteration of the SVM classification model may include spectral data included in the (n-1)th set of non-stationary state data used to generate the SVM classification model, as well as spectral data measured from the last time in the (n-1)th set of non-stationary state data to the transition time predicted using the (n-1)th iteration of the SVM classification model. As a specific example, the stationary state dataset used to generate the 4th iteration of the SVM classification model may include data from time t0 to time t... trans3 (For example, spectral data measured from the start time to the transition time determined by the third iteration using the model), while the steady-state dataset used to generate the fifth (i.e., the next) iteration of the SVM classification model may include data from time t0 to time t... trans4 (For example, spectral data measured from the start time to the transition time determined by the 4th iteration using the SVM classification model.)
[0065] To generate iterations for the SVM classification model, the steady-state dataset can include spectral data from the training spectral data corresponding to earlier executions of the manufacturing process that are assumed to be in a steady state. For example, the initial set of steady-state data associated with the initial iterations for generating the SVM classification model can include spectral data from time t... ss0 By time t e The spectral data measured at time t. Continuing this example, another set of steady-state data associated with the next iteration of generating the SVM model could be included from time t. ss0-dt*1 By time t e Spectral data measured at time t, where time t ss0-dt*1 It is time t ss0 The previous time step. In other words, during each iteration of generating the SVM classification model, the detection device 220 can be compared with time t. ss0 The spectral data associated with previous time steps are iteratively added to each set of steady-state data. Typically, the nth set of steady-state data associated with the nth iteration of generating the SVM model can be included from time t. ss0-dt*n By time t e Spectral data measured at time t, where time t ss0-dt*n It is time t ss0 The time of the previous n time steps.
[0066] In some implementations, the stable state dataset for each iteration of the SVM model can be updated, modified, and / or recreated, for example, by adding additional spectral data to the stable state dataset associated with the previous iteration that generated the SVM model. For example, the stable state dataset for a given iteration of generating the SVM classification model may include spectral data included in the stable state dataset of previous iterations used to generate the SVM classification model, as well as spectral data measured at a time step immediately preceding the earliest time step associated with the spectral dataset of the previous iterations used to generate the SVM classification model. As a specific example, the stable state dataset for the fourth iteration of generating the SVM classification model may include data from time t... ss0-dt*4 By time t e The measured spectral data, and the steady-state dataset used to generate the SVM classification model for the 5th (i.e., the next) iteration, can include data from time t. ss0-dt*5 By time t e Measured spectral data (i.e., t) ss0-dt*4 (Spectral data from the previous time step).
[0067] like Figure 4As further illustrated, process 400 may include an iteration (box 430) of generating an SVM classification model associated with whether the manufacturing process has reached a steady state, based on both the non-steady-state dataset and the steady-state dataset. For example, the detection device 220 may generate an iteration of an SVM classification model associated with when the manufacturing process has reached a steady state, based on both the non-steady-state dataset and the steady-state dataset.
[0068] In some implementations, the detection device 220 can generate an iterative SVM classification model based on applying SVM techniques to both the unstable and stable datasets. For example, the detection device 220 can generate an SVM classification model by mapping the unstable and stable datasets to points in space, such that the unstable and stable datasets are separated by, for example, a hyperplane.
[0069] In some implementations, detection device 220 can determine the transition time of the prediction associated with (i.e., through) the iteration of the SVM classification model based on training spectral data and iterations of the SVM classification model. For example, detection device 220 can generate iterations of the SVM classification model based on a stable-state dataset and an unstable-state dataset. Here, detection device 220 can use training spectral data (e.g., with the transition time from time t0 to time t) to determine the transition time of the prediction associated with (i.e., through) the iterations of the SVM classification model. e Each time step is associated with a mapping to the same space, and iterations of the SVM classification model are generated based on this same space. In this example, based on the location where the training spectral data items are mapped (e.g., relative to the hyperplane set), the SVM classification model can identify the transition time (t) associated with the manufacturing process, predicted by the initial iterations of the SVM classification model. trans(n) In some embodiments, the detection device 220 can determine the transition time of the prediction for each iteration of the SVM classification model. In some embodiments, the detection device 220 can store information identifying the transition time of the prediction associated with the iteration of the SVM classification model, so as to allow the detection device 220 to determine the transition time associated with the manufacturing process, as described below.
[0070] like Figure 4 As further shown, process 400 may include determining whether to generate another iteration of the SVM classification model (box 440). For example, detection device 220 may determine whether to generate another iteration of the SVM classification model.
[0071] In some implementations, the detection device 220 can determine whether to generate another iteration of the SVM classification model based on the time associated with the steady-state dataset. For example, the detection device 220 can be configured to continue iterating the generation of the SVM classification model until the earliest time associated with the steady-state dataset satisfies the time (time t) when the manufacturing process is known to be in a non-steady state. us0 The threshold time associated with the steady-state dataset is then determined. As a specific example, detection device 220 can be configured to continue iterating the generation of the SVM classification model (and determining the predicted transition time) until the earliest time associated with the steady-state dataset is reached with time t. us0 Until the difference threshold (e.g., until t) ss0-dt*n Distance t us0 A time step (t) ss0-dt*n =t us0+dt )until).
[0072] Alternatively, the detection device 220 can determine whether to generate another iteration of the SVM classification model based on an iteration threshold. For example, the detection device 220 can be configured to continue generating iterations of the SVM classification model within a threshold time period until a threshold number of iterations has been generated, etc. Here, the detection device 220 can determine whether to generate another iteration of the SVM classification model based on whether a threshold is met (e.g., whether a threshold time period has elapsed, whether a threshold number of iterations has been generated, etc.).
[0073] like Figure 4 As further shown, if another iteration of the SVM classification model is to be generated (box 440 - yes), process 400 may include creating another set of stable-state data and another set of non-stationary-state data based on the training spectral data (box 420). For example, detection device 220 may determine that another iteration of the SVM classification model is to be generated (e.g., when t ss0-dt*n >t us0+dt (When, when the iteration threshold is not met, etc.), and the detection device 220 can create another set of stable state data and another set of unstable state data based on the training spectral data.
[0074] In some embodiments, detection device 220 may create another set of non-stationary state data and another set of stationary state data in the manner described above with respect to box 420. In some embodiments, while creating the other set of stationary state data and the other set of non-stationary state data, detection device 220 may generate another iteration of the SVM classification model and determine the transition time predicted by the other iteration of the SVM classification model, as described above with respect to box 430.
[0075] As an example of the above iterative process, the detection device 220 can create an initial set of unstable state data (e.g., including data from time t0 to time t). us0 Training spectral data measured at time t) and initial set steady-state data (e.g., including data from time t) ss0 By time t e The training spectral data is measured at the specified time. In this example, detection device 220 can apply SVM technology to both the initial set of stable-state data and the initial set of unstable-state data to generate the initial iteration of the SVM classification model. Next, detection device 220 can provide the training spectral data as input to the initial iteration of the SVM classification model and determine the transition time (t) of the initial prediction associated with the initial iteration of the SVM classification model. trans0 ) as output.
[0076] Continuing with this example, detection device 220 can determine t ss0 >t us0+dt Therefore, detection device 220 can generate another iteration of the SVM classification model. Then, detection device 220 can create a first set of steady-state data (e.g., including data from time t) based on the training spectral data and the transition time of the initial prediction. ss0-dt*1 By time t e The measured training spectral data) and the first set of unsteady-state data (e.g., including data from time t0 to time t) trans0 The detection device 220 then applies SVM technology to the first set of stable-state data and the first set of unstable-state data to generate the first iteration of the SVM classification model. Next, the detection device 220 provides the training spectral data as input to the first iteration of the SVM classification model and determines the transition time (t) of the first prediction associated with the first iteration of the SVM classification model. trans1 The detection device 220 can continue to generate iterations of the SVM classification model in this manner (and determine the conversion time of the prediction) until the detection device 220 determines that no further iterations will be generated.
[0077] like Figure 4 As further shown, if another iteration of the SVM classification model will not be generated (box 440 - No), process 400 may include determining the transition time associated with the manufacturing process based on the transition time predicted by the iteration of the SVM classification model (box 450). For example, inspection device 220 may determine that another iteration of the SVM classification model will not be generated, and inspection device 220 may determine the transition time associated with the manufacturing process based on the transition time predicted by the iteration of the SVM classification model.
[0078] In some implementations, the inspection device 220 can determine the transition time associated with the manufacturing process based on the predicted transition time associated with iterations of the SVM classification model. For example, the inspection device 220 can determine n transition times predicted by n iterations of the SVM classification model in the manner described above. Here, the inspection device 220 can determine the transition time associated with the manufacturing process as the dominant transition time (e.g., the transition time with the most frequent prediction) among the transition times predicted by n iterations of the SVM classification model.
[0079] Figure 5A and Figure 5B These are example graphical representations 500 and 550 related to determining the changeover time of a manufacturing process based on iterative predictions of the changeover time using an SVM classification model associated with the manufacturing process. For Figure 5A and Figure 5B Assume that the detection device 220 has determined the n transition times corresponding to the n iterations of the SVM classification model.
[0080] exist Figure 5A In the diagram, each point represents the transformation time predicted through the iterations of the SVM classification model, relative to the time t associated with the iterations of the SVM classification model. ss0-dt*n Plot (i.e., time t) ss0-dt*n (Associated with the set of stable states associated with the iterations that generate the SVM classification model). For example... Figure 5A As shown, the dominant transition time in the set of n transition times is at time 130 (e.g., time 130 is predicted by more iterations than any other transition time). In this example, detection device 220 can determine the transition time associated with the manufacturing process as time 130.
[0081] Figure 5B Another graphical representation is shown in the image. Figure 5B The graph shows the total number of occurrences for each transition time. Again, as... Figure 5B As shown, the dominant transition time among the n transition times is at time 130. Therefore, the detection device 220 can determine the transition time associated with the manufacturing process as time 130.
[0082] As mentioned above, Figure 5A and Figure 5B Provided as an example only. Other examples are possible and may differ from those provided. Figure 5A and Figure 5B The example described.
[0083] Return to Figure 4Process 400 may include generating a final SVM classification model based on the changeover time associated with the manufacturing process (box 460). For example, inspection equipment 220 may generate a final SVM classification model based on the changeover time associated with the manufacturing process.
[0084] The final SVM classification model can include an SVM classification model generated based on the transition time associated with the manufacturing process, which is determined based on the iteration of the SVM classification model.
[0085] In some implementations, the detection device 220 can generate a final classification model based on the transition time associated with the manufacturing process. For example, as described above, the detection device 220 can determine the transition time associated with the manufacturing process. Here, the detection device 220 can create a final set of stationary-state data including training spectral data associated with the time prior to the transition time, and a final set of stationary-state data including training spectral data associated with the time at or after the transition time. In this example, the detection device 220 can apply SVM technology to both the final set of stationary-state data and the final set of stationary-state data, and can generate a final SVM classification model in a similar manner to that described above.
[0086] In some implementations, as described above, the SVM classification model may include a decision boundary (e.g., a hyperplane), which can serve as a basis for determining whether a later execution of the manufacturing process has reached a steady state. See below. Figure 6 Describe additional details about the decision boundary.
[0087] In some implementations, the inspection device 220 may store a final SVM classification model, allowing it to determine whether a later execution of the manufacturing process is in a non-stationary or stationary state, as described below. In this way, the inspection device 220 can generate an SVM classification model that receives spectral data associated with the manufacturing process as input and provides an indication of whether the manufacturing process is in a non-stationary or stationary state as output.
[0088] In some implementations, the manufacturing process may include multiple state transitions, and the detection device 220 may repeat process 400 for each stable state to determine multiple SVM classification models associated with the manufacturing process. For example, the manufacturing process may transition from a first unstable state to a first stable state, from the first stable state to a second unstable state, and from the second unstable state to a second stable state. In this example, the detection device 220 may perform process 400 (e.g., based on training spectral data measured with respect to the manufacturing process) to generate an SVM classification model associated with the transition time to the second stable state. Next, the detection device 220 may perform process 400 (e.g., based on a subset of training spectral data that does not include the training spectral data associated with the second stable state) to generate an SVM classification model associated with the transition time to the first stable state.
[0089] although Figure 4 An example block diagram of process 400 is shown; however, in some embodiments, process 400 may include... Figure 4 The boxes depicted in the process are those that are additional, fewer, different, or arranged differently compared to other boxes. Alternatively, two or more boxes in process 400 may be executed in parallel.
[0090] Figure 6 This is a flowchart of an example process 600 used to determine whether a manufacturing process has reached a steady state based on spectral data and an SVM classification model. In some implementations, Figure 6 One or more process frames can be executed by the detection device 220. In some embodiments, Figure 6 One or more process frames may be executed by another device separate from the detection device 220 or by a group of devices including the detection device 220 (such as spectrometer 210 and / or user equipment 230).
[0091] like Figure 6 As shown, process 600 may include identifying an SVM classification model for detecting whether the manufacturing process has reached a steady state (box 610). For example, detection device 220 may identify an SVM classification model for detecting whether the manufacturing process has reached a steady state.
[0092] In some implementations, the detection device 220 may identify the SVM classification model based on information stored or accessible by the detection device 220. For example, the detection device 220 may identify the SVM classification model based on the stored final SVM classification model generated by the detection device 220, as described above with respect to process 400.
[0093] In some implementations, the detection device 220 may identify an SVM classification model when it receives (e.g., from spectrometer 210, user equipment 230, based on user input, etc.) an indication that the detection device 220 wants to monitor the manufacturing process to determine when the manufacturing process has reached a steady state. For example, the detection device 220 may receive an indication from spectrometer 210 and / or user equipment 230 that a particular manufacturing process is about to start or has already started, and may (e.g., automatically) identify an SVM classification model based on the received indication, for example, when the detection device 220 is configured to automatically monitor the manufacturing process to detect when the manufacturing process has reached a steady state.
[0094] like Figure 6 As further shown, process 600 may include receiving spectral data associated with the manufacturing process (box 620). For example, detection device 220 may receive spectral data associated with the manufacturing process.
[0095] In some embodiments, the spectral data may include spectra measured by one or more spectrometers 210 during the execution of the manufacturing process. In some embodiments, the detection device 220 may receive the spectral data in real-time or near real-time during the manufacturing process. For example, the detection device 220 may receive spectral data measured by the spectrometer 210 during the execution of the manufacturing process in real-time or near real-time relative to the spectrometer 210. In some embodiments, the detection device 220 may determine whether the manufacturing process has reached a steady state based on the spectral data and an SVM classification model, as described below.
[0096] like Figure 6 As further shown, process 600 may include determining whether the manufacturing process has reached a steady state based on spectral data and an SVM classification model (box 630). For example, inspection device 220 may determine whether the manufacturing process has reached a steady state based on spectral data and an SVM classification model.
[0097] In some embodiments, the detection device 220 can determine whether the manufacturing process has reached a steady state based on a decision boundary associated with an SVM classification model. For example, based on identifying the transition times of the manufacturing process using training spectral data (as described above), the detection device 220 can generate an SVM classification model that includes a decision boundary represented by a hyperplane in the spectral space. Here, points in the spectral space within the decision boundary represent spectral conditions under which the manufacturing process is in a steady state, while points outside the decision boundary represent spectral conditions under which the manufacturing process is in a non-steady state. In some embodiments, the decision boundary can be generated after determining the transition times associated with the manufacturing process, based on applying an SVM classification model technique to the training spectral data, as described above.
[0098] Figure 7A and Figure 7B This is an example graphical representation 700 illustrating the simplified decision boundary associated with an SVM classification model. For illustrative purposes, Figure 7A and Figure 7B The decision boundary shown is depicted as being associated only with the first principal component (PC1) and the second principal component (PC2). In practice, the decision boundary can be associated with a different number of components (e.g., 80 variables, 120 variables, etc.).
[0099] exist Figure 7A and Figure 7B In the diagram, the gray dots and lines represent the training spectral data measured at the time when the manufacturing process is in a non-stationary state (i.e., from t0 to the transition time associated with the SVM classification model), while the black dots and lines represent the time when the manufacturing process is in a stationary state (i.e., from the transition time associated with the SVM classification model to time t). e Training spectral data measured at (). The light gray circles represent the last point when the manufacturing process is in a non-steady state, and the first point when the manufacturing process is in a steady state. Figure 7A It is all the training spectral data associated with the manufacturing process (e.g., from time t0 to time t). e The graphical representation of ) and Figure 7B It is by Figure 7A A close-up view of points within the space indicated by the dashed rectangles. Figure 7B In this context, the decision boundary is represented by a thick line around the points associated with the stable state (and a subset of the points associated with the unstable state). As mentioned above, Figure 7A and Figure 7B This is provided only as a simplified illustrative example. Other examples are possible and may differ from those provided. Figure 7A and Figure 7B The example described.
[0100] In some implementations, the detection device 220 can be based on a decision boundary (e.g., such as...). Figure 7B The decision boundary shown determines whether the manufacturing process has reached a steady state. For example, the detection device 220 can receive spectral data associated with the manufacturing process and can map the spectral data to points in space associated with the decision boundary. Here, if the points associated with the spectral data are located above or within the decision boundary, the detection device 220 can determine that the manufacturing process has reached a steady state. Alternatively, if the points associated with the spectral data are located outside the decision boundary, the detection device 220 can determine that the manufacturing process has not yet reached a steady state (i.e., is in an unstable state).
[0101] In some implementations, the detection device 220 can generate a confidence metric (referred to herein as a decision value) associated with determining whether the manufacturing process has reached a steady state. For example, the detection device 220 can determine the distance from the decision boundary to the point representing the spectral data (e.g., the distance to the nearest point on the decision boundary) based on a decision boundary and the point representing the spectral data. Here, points within the decision boundary can be assigned positive (or negative) decision values, while points outside the decision boundary can be assigned negative (or positive) decision values. In this example, decision values with higher absolute values (e.g., 4.0, 2.5, -2.5, -4.0, etc.) indicate a higher confidence level in determining the state of the manufacturing process than decision values with lower absolute values (e.g., 0.5, 0.2, -0.2, -0.5, etc.).
[0102] As mentioned above, Figure 7A and Figure 7B This is provided as an illustrative example only. Other examples are possible and may differ from those provided. Figure 7A and Figure 7B The example described.
[0103] Figure 8 Based on Figure 7A and Figure 7B A graphical representation of example decision values determined by the decision boundary (800). Figure 8 In this context, negative decision values correspond to spectral data measured during the manufacturing process that have points falling outside the decision boundary, while positive decision values correspond to spectral data measured during the manufacturing process that have points falling within the decision boundary. For example, from... Figure 8 As shown by the vertical line, the detection device 220 determines the first positive decision value at approximately a time step of 65.
[0104] In some implementations, the detection device 220 can determine whether the manufacturing process has reached a steady state based on a decision value threshold. For example, the detection device 220 can determine that the manufacturing process has reached a steady state when the decision value associated with a point within the decision boundary meets a threshold. Figure 8 As a specific example, if the detection device 220 is configured to determine that the manufacturing process has reached a steady state when the detection device 220 determines a positive decision value greater than or equal to 2.0, then the detection device 220 may make such a determination at approximately a time step of 75.
[0105] As another example, when the number of consecutive decision values representing spectral data associated with the number of consecutive time steps meets a threshold, the detection device 220 can determine that the manufacturing process has reached a steady state. Figure 8As a specific example, if the inspection device 220 is configured to determine that the manufacturing process has reached a steady state when the inspection device 220 determines 5 consecutive positive decision values, then the inspection device 220 can make such a determination at approximately time step 69.
[0106] As another example, when the number of consecutive decision values meets the threshold, the detection device 220 can determine that the manufacturing process has reached a steady state. Figure 8 As a specific example, if the detection device 220 is configured to determine that the manufacturing process has reached a steady state when the detection device 220 determines three consecutive positive decision values greater than or equal to 1.0, then the detection device 220 can make such a determination at approximately time step 68.
[0107] As another example, when the average or weighted average of multiple decision values (e.g., a series of consecutive positive decision values) meets a threshold, the detection device 220 can determine that the manufacturing process has reached a steady state. In some implementations, using a decision value threshold can prevent erroneous determinations that the manufacturing process has reached a steady state (e.g., since the manufacturing process can be inherently random).
[0108] As mentioned above, Figure 8 This is provided only as an illustrative example. Other examples are possible and may differ from those provided. Figure 8 The example described.
[0109] Return to Figure 6 If the manufacturing process has not yet reached a steady state (box 630 - No), process 600 may include receiving additional spectral data associated with the manufacturing process (box 620). For example, detection device 220 may determine that the manufacturing process has not yet reached a steady state (e.g., the manufacturing process is still in a non-steady state) and may wait to receive additional spectral data (e.g., collected at the next time step during the manufacturing process).
[0110] In some embodiments, it can be determined whether the manufacturing process has reached a steady state based on additional spectral data in the manner described above. In some embodiments, the detection device 220 can continue to receive spectral data and determine whether the manufacturing process has reached a steady state until the detection device 220 determines that the manufacturing process has reached a steady state.
[0111] like Figure 6As further shown, if the manufacturing process has reached a steady state (box 630 - Yes), process 600 may include determining a quantitative metric associated with the steady state (box 640). For example, detection equipment 220 may determine that the manufacturing process has reached a steady state and may determine a quantitative metric associated with the steady state.
[0112] Quantitative measures may include measures indicating quantitative properties associated with a steady state, such as the concentration of a component of a compound in its steady state, particle size in its steady state, etc. In some embodiments, a steady state detected by detection device 220 based on spectral data may correspond to a specific composition of a component compound having specific physical properties (e.g., particle size). Therefore, quantitative measures can be predicted based on spectral data associated with a steady state.
[0113] For example, in some embodiments, the detection device 220 may store or access a regression model (e.g., a PLS regression model, an SVR model) that receives spectral data based on its detection steady state as input and provides a quantitative measure associated with the steady state as output. In this example, the output of the regression model may be, for example, the concentration of each component of the compound, a specific size of the compound, etc.
[0114] In some embodiments, the detection device 220 may store or access a regression model. Alternatively or additionally, the detection device 220 may generate a regression model (at an earlier time) based on training spectral data and training quantitative data (e.g., information identifying the quantitative measure corresponding to the training spectral data). In some embodiments, the determination of the quantitative measure is optional.
[0115] like Figure 6 As further shown, process 600 may include providing an indication that the manufacturing process has reached a steady state, as well as information associated with quantitative measurements (box 650). For example, detection device 220 may provide an indication that the manufacturing process has reached a steady state, as well as information associated with quantitative measurements.
[0116] In some implementations, the detection device 220 may provide an indication that the manufacturing process has reached a steady state and / or information associated with quantitative measurements to another device, such as user device 230 (e.g., so that a user can be notified that the manufacturing process has reached a steady state and / or view information associated with quantitative measurements).
[0117] Alternatively, the detection device 220 may provide an indication that the manufacturing process has reached a stable state, so that actions are performed automatically. For example, the detection device 220 may provide this indication to the equipment associated with performing the manufacturing process, so as to stop the manufacturing process (e.g., stop the mixing process associated with the stable state), start the next step of the manufacturing process, or restart the manufacturing process (e.g., restart the mixing process with new raw materials), etc.
[0118] although Figure 6 The example block diagram of process 600 is shown, but in some implementations, process 600 may include... Figure 6 The boxes depicted in the process are those that are additional, fewer, different, or arranged differently. Alternatively or additionally, two or more boxes in process 600 may be executed in parallel.
[0119] The embodiments described herein provide a detection device capable of performing the following operations: generating an SVM classification model for determining whether a manufacturing process (e.g., a continuous manufacturing process, a batch manufacturing process, etc.) has reached a steady state, and using the SVM classification model and multivariate spectral data associated with the manufacturing process to determine whether the manufacturing process has reached a steady state. In some embodiments, the SVM classification model may consider multiple variables (e.g., 80 variables, 120 variables, 150 variables, etc.), thereby increasing the accuracy and / or robustness of the SVM classification model (e.g., compared to univariate techniques or PCA techniques).
[0120] The foregoing disclosure provides illustrations and descriptions, but is not intended to be exhaustive or to limit the embodiments to the precise forms disclosed. Modifications and variations are possible in accordance with the foregoing disclosure, or modifications and variations may be derived from practice of the embodiments.
[0121] As used herein, the term component is intended to be interpreted broadly as hardware, firmware, and / or a combination of hardware and software.
[0122] This document describes some implementation methods in conjunction with thresholds. As used herein, a threshold can refer to a value that is greater than, more than, higher than, greater than or equal to, less than, less than, lower than, less than or equal to, or equal to a threshold.
[0123] It is evident that the systems and / or methods described herein can be implemented in various forms of hardware, firmware, or a combination of hardware and software. The actual dedicated control hardware or software code used to implement these systems and / or methods is not limited to the implementation method. Therefore, the operation and behavior of the systems and / or methods are described herein without reference to specific software code—it should be understood that software and hardware can be designed to implement the systems and / or methods based on the description herein.
[0124] Although specific combinations of features are listed in the claims and / or disclosed in the specification, these combinations are not intended to limit the disclosure of possible embodiments. In fact, many of these features can be combined in ways not specifically listed in the claims and / or disclosed in the specification. While each dependent claim listed below may directly depend on only one claim, the disclosure of possible embodiments includes combinations of each dependent claim with each other claim in the claim set.
[0125] Unless explicitly stated otherwise, no element, action, or instruction herein should be construed as critical or necessary. Furthermore, as used herein, the articles “a” and “an” are intended to include one or more items and are interchangeable with “one or more.” Additionally, as used herein, the term “set” is intended to include one or more items (e.g., related items, unrelated items, combinations of related and unrelated items, etc.) and is interchangeable with “one or more.” The term “an” or similar language is used when the intention is to indicate only one item. Moreover, as used herein, the terms “have,” “possess,” etc., are intended to be open-ended terms. Furthermore, unless explicitly stated otherwise, the phrase “based on” is intended to mean “at least partially based on.”
[0126] This application also relates to the following aspects:
[0127] 1) An apparatus comprising:
[0128] One or more processors, for:
[0129] Receive training spectral data associated with the manufacturing process from an unstable state to a stable state;
[0130] Multiple iterations of generating a support vector machine (SVM) classification model based on the trained spectral data;
[0131] Based on multiple iterations of the SVM classification model, the conversion times of multiple predictions associated with the manufacturing process are determined.
[0132] The predicted transition time among the multiple predicted transition times is identified as the time during which the SVM classification model predicts the transition from an unstable state to a stable state during the manufacturing process, based on the corresponding iteration of the model; and
[0133] Based on the multiple predicted transition times, a final SVM classification model is generated and associated with determining whether the manufacturing process has reached a steady state.
[0134] 2) The device according to claim 1), wherein the one or more processors are further configured to:
[0135] Receive additional spectral data associated with the manufacturing process; and
[0136] Based on the final SVM classification model and the additional spectral data, it was determined that the manufacturing process had not yet reached a stable state.
[0137] 3) The device according to claim 1), wherein the one or more processors are further configured to:
[0138] Receive additional spectral data associated with the manufacturing process; and
[0139] Based on the final SVM classification model and the additional spectral data, it was determined that the manufacturing process had reached a stable state.
[0140] 4) The device according to 3), wherein the one or more processors are further configured to:
[0141] Provides an indication that the manufacturing process has reached a stable state.
[0142] 5) The device according to claim 3), wherein the one or more processors are further configured to:
[0143] Based on the determination that the manufacturing process has reached a steady state, a quantitative measure associated with the steady state is determined; and
[0144] Provide information associated with the quantitative measure.
[0145] 6) The apparatus according to 1), wherein, during multiple iterations of generating the SVM classification model, the one or more processors are configured to:
[0146] Based on the training spectral data, create an unstable state dataset and a stable state dataset; and
[0147] Based on the unstable state dataset and the stable state dataset, multiple iterations of the SVM classification model are generated.
[0148] 7) The apparatus according to 6), wherein, in determining the conversion time of the plurality of predictions, the one or more processors are configured to:
[0149] Based on the training spectral data and the iteration of the SVM classification model, the prediction transition time associated with the iteration of the SVM classification model is determined.
[0150] The predicted transition time is one of the plurality of predicted transition times.
[0151] 8) The device according to claim 1), wherein the one or more processors are further configured to:
[0152] Dimensionality reduction is performed based on the training spectral data before generating the SVM classification model through multiple iterations.
[0153] 9) A non-transitory computer-readable medium storing instructions, the instructions comprising:
[0154] One or more instructions, when executed by one or more processors, cause the one or more processors to:
[0155] Receive training spectral data associated with the first execution of the manufacturing process transitioning from an unstable state to a stable state;
[0156] Based on the training spectral data, a support vector machine (SVM) classification model is iteratively generated to determine whether another execution of the manufacturing process has transitioned from an unstable state to a stable state;
[0157] Receive additional spectral data associated with a second execution of the manufacturing process; and
[0158] Based on the SVM classification model and the additional spectral data, it is determined whether the second execution of the manufacturing process has transitioned from an unstable state to a stable state.
[0159] 10) The non-transitory computer-readable medium according to 9), wherein the one or more instructions that cause the one or more processors to determine whether the second execution of the manufacturing process has transitioned from an unstable state to a stable state cause the one or more processors to:
[0160] It has been determined that the second execution of the manufacturing process has transitioned from an unstable state to a stable state; and
[0161] Provides an indication that the second execution of the manufacturing process has transitioned from an unstable state to a stable state.
[0162] 11) The non-transitory computer-readable medium according to 10), wherein, when executed by the one or more processors, the one or more instructions further cause the one or more processors to:
[0163] Based on the determination that the second execution of the manufacturing process has transitioned from an unstable state to a stable state, a quantitative measure associated with the second execution of the manufacturing process is determined; and
[0164] Provide information associated with the quantitative measure.
[0165] 12) The non-transitory computer-readable medium according to 11), wherein, when executed by the one or more processors, the one or more instructions further cause the one or more processors to:
[0166] Identify the regression model associated with the quantitative measure; and
[0167] The one or more instructions that cause the one or more processors to determine the quantitative metric cause the one or more processors to:
[0168] The quantitative measure is determined based on the regression model.
[0169] 13) The non-transitory computer-readable medium according to 9), wherein the one or more instructions that cause the one or more processors to iteratively generate the SVM classification model cause the one or more processors to:
[0170] A first set of unstable state data and a first set of stable state data are created based on the training spectral data;
[0171] The first iteration of the SVM classification model is generated based on the first set of unstable state data and the first set of stable state data;
[0172] A second set of unstable-state data and a second set of stable-state data are created based on the training spectral data; and
[0173] The second iteration of the SVM classification model is generated based on the second set of unstable state data and the second set of stable state data.
[0174] 14) The non-transitory computer-readable medium according to 13), wherein, when executed by the one or more processors, the one or more instructions further cause the one or more processors to:
[0175] The first iteration of the SVM classification model determines the prediction transition time associated with the first iteration of the SVM classification model; and
[0176] The one or more instructions that cause the one or more processors to create the second set of unstable state data cause the one or more processors to:
[0177] A second set of non-stationary state data is created based on the predicted transition times associated with the first iteration of the SVM classification model.
[0178] 15) A method comprising:
[0179] The device receives first spectral data associated with the first execution of the manufacturing process, which transitions from an unstable state to a stable state.
[0180] Multiple iterations of the support vector machine (SVM) classification model generated by the device based on the first spectral data;
[0181] The conversion time of multiple predictions associated with the first execution of the manufacturing process is determined by the device and based on multiple iterations of the SVM classification model;
[0182] The device generates a final SVM classification model based on the multiple predicted transition times, which is associated with determining whether another execution of the manufacturing process has reached a steady state.
[0183] The device receives second spectral data associated with a second execution of the manufacturing process; and
[0184] The device determines whether the second execution of the manufacturing process has reached a stable state based on the final SVM classification model and the second spectral data.
[0185] 16) According to the method of 15), wherein determining whether the second execution of the manufacturing process has reached a steady state includes:
[0186] It was determined that the second execution of the manufacturing process had not yet reached a stable state.
[0187] 17) According to the method of 15), wherein determining whether the second execution of the manufacturing process has reached a steady state includes:
[0188] It was determined that the second execution of the manufacturing process had reached a stable state.
[0189] 18) The method according to 17) further includes:
[0190] Provides an indication that the manufacturing process has reached a stable state.
[0191] The instructions are provided such that actions associated with the manufacturing process are executed automatically.
[0192] 19) The method according to 17) further includes:
[0193] A quantitative measure associated with the steady state is determined based on a regression model related to the manufacturing process; and
[0194] Provide information associated with the quantitative measure.
[0195] 20). The method according to 15) further includes:
[0196] Before generating the SVM classification model, dimensionality reduction associated with the first spectral data is performed.
Claims
1. A computer-implemented method, comprising: The device identifies a support vector machine (SVM) classification model, which is based on a final set of unstable state data and a final set of stable state data. The final set of unstable state data includes spectral data measured before the dominant transition time of a plurality of transition times in the manufacturing process, and the final set of stable state data includes training spectral data measured at or after one or more of the dominant transition times of the plurality of transition times in the manufacturing process. The device receives measured multivariate spectral data from one or more spectrometers. Using the device, based on the multivariate spectral data received from the one or more spectrometers, and using the SVM classification model based on the final set of unstable state data and the final set of stable state data, it is determined that the manufacturing process has not yet reached a stable state where the mixed processes included in the manufacturing process are completed; After determining that the manufacturing process has not reached the steady state of the completion of the mixing process, the device receives additional spectral data from the one or more spectrometers until the device determines that the manufacturing process has reached the steady state of the completion of the mixing process. as well as After determining that the manufacturing process has reached the stable state where the mixing process is complete, the device triggers an action including restarting the mixing process on new raw materials.
2. The method of claim 1, wherein determining that the manufacturing process has not reached the steady state at which the mixing process is completed comprises: The multivariate spectral data is used as input to the SVM classification model; as well as Based on the output of the SVM classification model, it is determined that the manufacturing process has not reached the stable state at which the hybrid process is completed.
3. The method according to claim 1, further comprising: Based on determining that the manufacturing process has reached a stable state indicating the completion of the mixing process, a quantitative measure is determined indicating the particle size or concentration of a component of the compound in the stable state; and Provide information associated with the quantitative measure.
4. The method of claim 3, wherein the quantitative measurement includes the concentration of the component of the compound in the steady state.
5. The method of claim 3, wherein determining the quantitative measure comprises: Based on the determination that the manufacturing process has reached the steady state of completion of the mixing process, the multivariate spectral data is provided as input to the regression model, and The quantitative measure is received as output from the regression model.
6. The method of claim 1, wherein, The multivariate spectral data includes near-infrared (NIR) spectral data, which includes data associated with more than one hundred variables.
7. A computer device, comprising: One or more memory units; as well as One or more processors coupled to the one or more memories, the one or more processors being configured to: The classification model is identified based on a first set of data measured before the dominant transition time of a plurality of transition times in the manufacturing process and a second set of data measured at or after the dominant transition time of the plurality of transition times in the manufacturing process. Receive spectral data measured during the execution of the manufacturing process from one or more spectrometers; Based on the spectral data and using the classification model, it is determined that the manufacturing process is in a stable state, in which the mixing process in the manufacturing process is completed; as well as Based on the manufacturing process being in a stable state after the mixing process is completed, one or more of the following actions are triggered: stopping the mixing process, starting the next step in the manufacturing process, or restarting the mixing process on new raw materials.
8. The device according to claim 7, To determine that the manufacturing process is in the stable state after the mixing process is completed, the one or more processors are configured to: It is determined that the manufacturing process has reached the stable state at which the mixing process is completed, and The one or more processors mentioned therein are further configured to: Based on determining that the manufacturing process has reached a stable state indicating the completion of the mixing process, a quantitative measure associated with the stable state is determined, the quantitative measure indicating the particle size at the stable state; and Provide information associated with the quantitative measure.
9. The apparatus of claim 7, wherein, The spectral data includes multivariate spectral data.
10. A non-transitory computer-readable medium storing an instruction set, the instruction set comprising: One or more instructions, which, when executed by one or more processors of the device, cause the device to: Receive multivariate spectral data measured during the execution of the manufacturing process from one or more spectrometers; Based on the multivariate spectral data and using a support vector machine (SVM) classification model based on the final set of unstable state data and the final set of stable state data, it is determined that the manufacturing process has not yet reached a stable state where the mixed processes included in the manufacturing process are completed. After determining that the manufacturing process has not reached the steady state of the completion of the mixing process, additional spectral data are received from the one or more spectrometers until it is determined that the manufacturing process has reached the steady state of the completion of the mixing process. as well as After determining that the manufacturing process has reached the stable state where the mixing process is complete, an action is initiated that includes restarting the mixing process on new raw materials. The final set of unsteady-state data includes spectral data measured before the dominant transition time of a plurality of transition times in the manufacturing process, and the final set of steady-state data includes training spectral data measured at or after one or more of the dominant transition times of the plurality of transition times in the manufacturing process.
11. The non-transitory computer readable medium of claim 10, wherein, The multivariate spectral data includes near-infrared (NIR) spectral data, which includes data associated with more than one hundred variables.
Citation Information
Patent Citations
Adaptive knowledge-based reasoning in autonomic computing systems
US20090327172A1
