Anomaly detection and diagnosis in factory automation systems using preconditioned time-delay neural networks with loss function adaptation

Abstract

A computer-implemented preprocessed time-delay autoencoder-based anomaly detection method is provided for detecting an abnormal condition of a machine disposed in a factory automation (FA) system or a manufacturing production line. The method includes the steps of acquiring a source signal from the machine via an interface, performing a data preprocessing process for the acquired source signal by normalizing a value range of the acquired source signal and filtering undesirable features from the acquired source signal, performing a time-delayed data correction process for the preprocessed source signal based on a time-delay window to generate preprocessed time-delay data, and submitting the preprocessed time-delay test data to a preprocessed time-delay autoencoder (Prep-TDAE) neural network, where the preprocessed time-delay test data is collected online while the machine is operated, and the Prep-TDAE neural network is pre-trained by using the preprocessed time-delay training data, and the method further includes the steps of detecting whether an abnormal condition is encountered for the machine by calculating an anomaly score for the preprocessed time-delay test data, and if an abnormal condition is detected, determining an anomaly occurrence time, duration, and severity for each abnormal condition of the machine.

Description

[Technical field]

[0001] The present invention relates generally to anomaly detection and diagnosis using machine learning techniques, and more particularly to anomaly detection and diagnosis using preprocessed time-delay neural networks with loss function adaptation. [Background technology]

[0002] Downtime in a factory automation (FA) system refers to the period during which the system stops its production. Machines or equipment in an FA system may stop working due to abnormality, failure, or intrusion. Therefore, it is essential to accurately detect anomalies in an FA system. Once an abnormality is detected in the system, the cause of the abnormality needs to be identified in order to quickly restore the FA system to a normal state. Anomaly diagnosis is to find the exact source that caused the abnormality. Therefore, anomaly diagnosis is also critical in an FA system. As a result, both anomaly detection and anomaly diagnosis are needed to improve the production efficiency in an FA system.

[0003] Machine learning methods have been applied for anomaly detection in manufacturing systems. Machine learning methods consist of supervised learning and unsupervised learning. Supervised learning is not suitable for FA systems due to the data complexity and difficulty of labeling data in FA systems. Therefore, unsupervised learning methods are typically used for anomaly detection in FA systems.

[0004] Traditional machine learning algorithms for anomaly detection, such as isolation forest (IF), one-class support vector machine (OC-SVM), and local outlier factor (LOF), have relatively low computational complexity and can detect anomalies when the data does not have a complex structure, which is not the case in factory automation systems where various data structures exist.

[0005] Modern machine learning methods, especially autoencoder neural networks, have been shown to be applicable for anomaly detection in both process and discrete manufacturing. Inheriting from a large number of neural network designs, autoencoders can learn the complex structure of data and achieve relatively better anomaly detection performance compared to other types of learning methods. However, FA systems require accurate anomaly detection. Due to the complexity of FA data, traditional autoencoders still cannot meet the requirements of FA systems.

[0006] To improve the production efficiency of FA systems, autoencoders must address the following issues: 1) loss function (objective function) adaptation, 2) time information discovery, and 3) data cleaning, i.e., data preprocessing. Traditionally, in machine learning neural networks such as autoencoders, only one loss function is used. This approach is effective for applications with a single type of data, such as image processing. On the other hand, FA systems have various data types, and a loss function that works for one type of data may not work well for other types of data. As a result, different loss functions should be used for different types of data. Thus, loss function adaptation is needed. In FA systems, examining historical data can reveal critical information in anomaly detection and diagnosis. For example, an anomaly may be caused by a series of operational errors. To identify such anomalies, corresponding data samples are needed. Noise in the raw data may cause the autoencoder to issue false alarms or miss real anomalies. As a result, how to build such complex autoencoder neural networks becomes a problem to be addressed.

[0007] Although abnormality diagnosis is critical in FA system, there is not much work on abnormality diagnosis in FA system. When an abnormality is detected in FA system, what is the source or cause of the abnormality? Abnormality diagnosis refers to the task of identifying the exact source that caused the abnormality. In FA system, the source can be a machine, a hardware part, or a software program. To avoid downtime, the exact source must be identified to restore the system to normal state quickly.

[0008] To improve the performance of FA systems, both anomaly detection and anomaly diagnosis mechanisms are necessary. Due to the data complexity in FA systems, machine learning methods require loss function adaptation. In addition, time information must be considered to improve the accuracy of anomaly detection and diagnosis. Furthermore, both anomaly detection and anomaly diagnosis methods rely on the accuracy of training on training data. The source data (raw data) may contain noise, and it may be difficult to learn the data characteristics. Therefore, to achieve high-performance anomaly detection and diagnosis in FA systems, it becomes desirable for machine learning methods to adaptively select loss functions, explore time-data relationships, and preprocess source data. Summary of the Invention

[0009] Some embodiments of the present invention are based on the recognition that supervised learning is not suitable for FA system anomaly detection due to data complexity and the difficulty of labeling or classifying the data.

[0010] Therefore, some embodiments of the present invention provide autoencoder-based unsupervised anomaly detection in FA systems via unsupervised learning using unlabeled training data.

[0011] Some embodiments of the present invention are based on the recognition that FA systems contain multiple types of data, such as analog data and binary data, and different types of data have different characteristics and different requirements for the loss function, and need to be trained with different loss functions. Furthermore, data in FA systems can be correlated, and therefore all types of data must be trained together. As a result, loss function adaptation is key to achieving accurate anomaly detection performance.

[0012] Therefore, some embodiments of the present invention adaptively select a loss function in the training process. The choice of loss function depends on the data type.

[0013] Some embodiments of the present invention are based on the recognition that data history can provide important information for FA system anomaly detection, for example, that an anomaly may be caused by a series of operation errors. Therefore, taking into account the time factor can detect an anomaly caused by a series of operation errors.

[0014] Therefore, some embodiments of the present invention provide that the pre-processed data is available to a time-delay autoencoder neural network, which examines the time information of the data signal. Time-delayed modifications of the data signal can be applied directly to the pre-processed data signal.

[0015] Some embodiments of the present invention are based on the recognition that FA system data contains various noises, for example, some machines may emit strong radio noise, which may seriously affect data quality and anomaly detection accuracy. Therefore, data cleaning, especially training data cleaning to reduce various noises, is crucial to improve anomaly detection accuracy.

[0016] Therefore, some embodiments of the present invention pre-process the raw data by applying data filters to clean the data before feeding it to a neural network.

[0017] Some embodiments of the present invention are also based on the recognition that a computer-implemented preprocessed time-delay autoencoder based anomaly detection method may be provided for detecting abnormal conditions in machines located in a factory automation (FA) system or a manufacturing production line. The computer-implemented preprocessed time-delay autoencoder based anomaly detection method may include the steps of acquiring source signals from a machine via an interface; performing a data preprocessing process for the acquired source signals by normalizing a value range of the acquired source signals and filtering undesirable features from the acquired source signals; performing a time-delayed data correction process for the preprocessed source signals based on a time delay window to generate preprocessed time-delay data; and submitting the preprocessed time-delay test data to a preprocessed time-delay autoencoder (Prep-TDAE) neural network, where the preprocessed time-delay test data is collected online while the machine is operated, and the Prep-TDAE neural network is pre-trained by using the preprocessed time-delay training data, and the method may further include the steps of detecting whether an abnormal condition is encountered for the machine by calculating an anomaly score for the preprocessed time-delay test data; and if an abnormal condition is detected, determining an anomaly occurrence time, duration, and severity for each abnormal condition of the machine.

[0018] Also provided is a pre-processed time-delay autoencoder based anomaly detection system for detecting an abnormal condition of a machine disposed in a factory automation (FA) system or manufacturing production line. The system includes an interface configured to acquire source signals from the machine, a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, a pre-processed time-delay autoencoder (Pre-TDAE) neural network, an anomaly diagnosis algorithm, and a computer-implemented statistical-based anomaly diagnosis method, and a processor, the processor performing the data pre-processing process algorithm for the acquired source signals by normalizing a value range of the acquired source signals and filtering undesirable features from the acquired source signals, and filtering the pre-processed source signals based on a time delay window to generate pre-processed time-delay data. The processor is configured to perform the steps of performing a time-delayed data correction process algorithm for the source signal and submitting the preprocessed time-delay test data to a Prep-TDAE neural network, where the preprocessed time-delay test data is collected online while the machine is operated, and the Prep-TDAE neural network is pre-trained by using the preprocessed time-delay training data, and the processor is further configured to perform the steps of detecting whether an abnormal condition has been encountered for the machine by calculating an anomaly score for the preprocessed time-delay test data, and if an abnormal condition is detected, determining an anomaly occurrence time, duration, and severity for each abnormal condition of the machine.

[0019] Some embodiments of the present invention are based on the recognition that once an anomaly is detected in an FA system, an anomaly diagnosis is required to identify the exact source that caused the anomaly and therefore quickly restore the system to a normal state.

[0020] Further, a computer-implemented statistical-based anomaly diagnosis method may be provided for diagnosing anomaly sources in machines disposed in a factory automation (FA) system or manufacturing production line. The method may include the steps of acquiring source signals from the machines via an interface, performing a data pre-processing process for the acquired source signals based on a data pre-processing process algorithm, performing a time-delayed data correction process algorithm for the acquired source signals based on a time delay window to generate pre-processed time-delayed data, normalizing the pre-processed time-delayed data to have zero mean and unit variance, calculating a statistical reduced covariance matrix such that the covariance matrix is ​​not irreversible, calculating a whitening matrix for a whitening transformation, performing a whitening transformation of the signals for mean and variance calculation, calculating the mean and variance of the transformed signals for anomaly score calculation, calculating an anomaly score for each data signal for anomaly data signal diagnosis, diagnosing the anomaly data signal, and identifying a machine having an anomaly among the machines using the anomaly diagnosis algorithm.

[0021] Some embodiments may also provide a statistics-based anomaly diagnosis system for diagnosing anomaly sources in a machine disposed in a factory automation (FA) system or a manufacturing production line. The statistics-based anomaly diagnosis system may include an interface configured to acquire a source signal from the machine, a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, an anomaly diagnosis algorithm, and a computer-implemented statistics-based anomaly diagnosis method, and a processor configured to execute the computer-executable program. The processor performs the steps of performing a data pre-processing process algorithm for the acquired source signals, performing a time delayed data correction process algorithm for the acquired source signals based on a time delay window to generate pre-processed time delayed data, normalizing the pre-processed time delayed data to have zero mean and unit variance, calculating a statistical reduced covariance matrix such that the covariance matrix is ​​not irreversible, calculating a whitening matrix for a whitening transformation, performing a whitening transformation of the signals for mean and variance calculation, calculating the mean and variance of the transformed signals for anomaly score calculation, calculating an anomaly score for each data signal for anomaly data signal diagnosis, diagnosing the anomaly data signal, and identifying machines having anomalies from among the machines using the anomaly diagnosis algorithm.

[0022] According to some embodiments of the present invention, a computer-implemented residual error-based anomaly diagnosis method may be provided for diagnosing anomaly sources of machines arranged in a factory automation (FA) system or manufacturing production line. The computer-implemented residual error diagnosis method may include: acquiring a source signal from a machine through an interface; performing a data pre-processing process for the acquired source signal; performing a time-delayed data correction process algorithm for the acquired source signal based on a time delay window to generate pre-processed time-delayed data; and submitting the pre-processed time-delayed data to a pre-processed time-delay autoencoder (Pre-TDAE) neural network to generate a reconstructed signal, where the Pre-TDAE neural network is pre-trained by training data; and the method may further include: calculating a residual error based on the reconstructed signal; calculating the mean and variance of the residual error; calculating an anomaly score for the reconstructed signal; diagnosing an anomaly signal corresponding to the source signal of the machine based on the anomaly score; and identifying a machine having an anomaly among the machines using an anomaly diagnosis algorithm.

[0023] Also provided is a residual error-based anomaly diagnosis system for diagnosing anomaly source in a machine disposed in a factory automation (FA) system or a manufacturing production line. The residual-based anomaly diagnosis may include an interface configured to acquire a source signal from the machine, a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, a pre-processed time-delay autoencoder (Pre-TDAE) neural network, an anomaly diagnosis algorithm, and a computer-implemented residual-based anomaly diagnosis method, and a processor configured to execute the computer-executable program. The processor is configured to perform the steps of: performing a data pre-processing process algorithm for the acquired source signal; performing a time-delayed data correction process algorithm for the acquired source signal based on a time delay window to generate pre-processed time-delayed data; and submitting the pre-processed time-delayed data to a Pre-TDAE neural network to generate a reconstructed signal, where the Pre-TDAE neural network is pre-trained with the training data; and the processor is further configured to perform the steps of: calculating a residual error based on the reconstructed signal; calculating a mean and variance of the residual error; calculating an anomaly score for each data signal; diagnosing an anomaly signal corresponding to the machine's source signal based on the anomaly score; and identifying a machine having an anomaly from among the machines using the anomaly diagnosis algorithm.

[0024] Therefore, some embodiments of the present invention provide that the residual error of the autoencoder neural network can indicate an anomaly score for each individual data signal, and that performing pre-processing and time-delayed correction can improve the anomaly diagnosis accuracy.

[0025] The presently disclosed embodiments are further described with reference to the accompanying drawings, in which the drawings are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the presently disclosed embodiments. [Brief description of the drawings]

[0026] [Figure 1] FIG. 2 is a flowchart showing an anomaly detection and anomaly diagnosis in a factory automation system according to an embodiment of the present invention. [Figure 2A] FIG. 1 illustrates a multi-loss function autoencoder neural network structure in which the loss function is composed of multiple sub-functions adapted depending on the data type of the input data signal, according to an embodiment of the present invention. [Figure 2B] FIG. 1 illustrates an example of multiple loss function adaptation for an autoencoder, where a cross-entropy function is selected for binary data and a squared error function is selected for analog data, according to an embodiment of the present invention. [Diagram 3] FIG. 2 illustrates a flowchart of time-delayed data modification in which source data is modified by a time delay window of size 2 according to an embodiment of the present invention. [Figure 4] FIG. 2 illustrates a flow chart of data pre-processing in which input data is split into analog data and binary data according to an embodiment of the present invention. [Diagram 5] FIG. 2 illustrates a flowchart of a method for statistically-based anomaly diagnosis in accordance with an embodiment of the present invention. [Figure 6] FIG. 2 shows a flowchart of a residual error-based anomaly diagnosis method according to an embodiment of the present invention. [Figure 7] FIG. 1 illustrates a block diagram of an anomaly detection system in accordance with an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] Various embodiments of the present invention are described below with reference to the drawings. It should be noted that the drawings are not drawn to scale, and elements of similar structure or function are represented by the same reference numerals throughout the drawings. Furthermore, the drawings are merely intended to facilitate the description of specific embodiments of the present invention. The drawings are not intended as an exhaustive description of the present invention or as limitations on the scope of the present invention. In addition, aspects described in connection with a particular embodiment of the present invention are not necessarily limited to that embodiment, and may be practiced in any other embodiment of the present invention.

[0028] Anomaly detection and anomaly diagnosis are two different processes that are critical for FA systems to avoid downtime and improve FA system efficiency. Anomaly detection is to determine whether the FA system is operating with anomalies or failures. If so, what is the number of anomalies, what time the anomalies occurred, and how long did the anomalies last? Once anomalies are detected, they need to be resolved quickly so that the FA system can be restored to normal. FA systems can be very complex systems. To resolve anomalies, the cause of the anomalies must be known. Anomaly diagnosis is the process for identifying the exact source that caused the anomaly. FIG. 1 shows a schematic diagram illustrating a flowchart of factory system anomaly detection and diagnosis according to some embodiments of the present invention. There is an anomaly detection block 100 and an anomaly diagnosis block 500. The anomaly detection block 100 is composed of an anomaly detection training block 101 and an anomaly detection block 102. Similarly, the anomaly diagnosis block 500 is composed of an anomaly diagnosis training block 501 and an anomaly diagnosis block 502. Anomaly detection using preprocessed time delay autoencoder with loss function adaptation

[0029] Anomaly detection is very important for FA system maintenance and FA system efficiency improvement. FIG. 1 shows a flowchart of FA system anomaly detection 100 according to some embodiments of the present invention. There are two stages for anomaly detection: anomaly detection training stage 101 and anomaly detection stage 102. It is assumed that an optimization algorithm such as a momentum algorithm is performed to train an anomaly detection algorithm such as a neural network. Next, during the anomaly detection training stage 101, data is collected (111) from a normally operating FA system 001. The normal data is used as input data for training (112) the anomaly detection algorithm. Once the anomaly detection algorithm is fully trained, the model is saved (113). For example, if an autoencoder neural network is used as the learning model, the multiplicative weights, biases, and hyperparameters are saved (114). During the anomaly detection stage, test data is collected (115) from a FA system 002 that may be operating abnormally. The collected data is used as input data to a fully trained anomaly detection model 116. The anomaly detection algorithm calculates an overall anomaly score for the test data (117). The corresponding anomaly scores of the test data are calculated to identify whether an anomaly has been encountered. The anomaly scores are determined according to a specific anomaly detection algorithm, such as an autoencoder neural network. If an anomaly is detected, its corresponding anomaly occurrence time, duration, and severity 118 may also be determined for anomaly diagnosis and FA system maintenance. Once an anomaly is detected, the detected anomaly is diagnosed by applying an anomaly diagnosis algorithm to identify the exact source that caused the anomaly. The process of anomaly diagnosis 500 shown in FIG. 5 and FIG. 6 is particularly important for quickly restoring the system to a normal state.

[0030] Machine learning methods such as neural networks can be employed as algorithms to detect anomalies in FA systems. In general, unsupervised learning methods are used for anomaly detection in FA systems because it is impractical to enumerate and label all types of anomalies in FA systems.

[0031] Autoencoder neural network is an unsupervised learning method. Traditional autoencoders work well for homogeneous data with simple structure. Due to the data complexity and stringent requirements of FA systems, traditional autoencoders face the challenge of detecting anomalies in FA systems. The issues of loss function adaptation, temporal information consideration, and data preprocessing must be addressed.

[0032]

number

[0033] It is important to note that while FIG. 2A shows a four-layer asymmetric autoencoder neural network structure, all embodiments of the present invention are equally valid for various layers of both asymmetric and symmetric autoencoder neural networks.

[0034] 2B shows an example of loss function adaptation using binary data 220 and analog data 221, because binary data and analog data are two common data types in FA systems. To exploit the prior information on data types, a binary cross-entropy loss function 222 is applied to the binary data signal 220, and a squared error loss function 223 is applied to the analog data signal 221. The final loss function is the combination of the two loss functions 222 and 223. By adopting a more suitable loss function, the output layer can learn the input data signal more accurately after the encoding and decoding processes of the autoencoder.

[0035] Combining historical data can provide important information for FA system anomaly detection. For example, an anomaly may be caused by a series of operational errors. Therefore, considering the time factor can detect anomalies caused by a series of operational errors. For this purpose, an embodiment of the present invention provides a time-delayed data correction process 300 shown in FIG. 3. An autoencoder neural network that receives the time-delay corrected data as input is called a time-delay autoencoder (TDAE) 301, which can be applied for anomaly detection in FA systems. As a result, the TDAE is an autoencoder in which source data (input data) 302 is corrected based on a time-delay window w 303 that specifies the length of the history to be considered. It is assumed that the source data 302 is collected continuously in time.

number

[0036] The time-delayed data modification affects the input data on the spatial and temporal domains. The spatial dimension of the modified data, i.e., the number of input signals, is increased by a factor of w, and the number of samples of each signal is reduced by w-1. As the number of input signals increases, the TDAE301 has a much larger and more complex network structure compared to conventional autoencoders. For example, if w=2 and 10 data signals are collected, the time-delayed data modification reduces the number of samples by 10. * produces a signal with w=20 samples. Meanwhile, the time-delayed data modification reduces the number of data samples in the time domain. For example, if w=2 and the source data signal contains 1000 samples, the time-delayed data modification reduces the number of samples to 1000-(w-1)=999 samples.

[0037]

number

[0038] TDAE is one of the autoencoder structures to leverage neural networks for learning temporal information. It learns the temporal relationship between data signals over time, which can reveal significant information in FA systems. Because the temporal information is explored by TDAE, improved anomaly detection performance can be obtained.

[0039] Furthermore, to detect anomalies in FA systems using an autoencoder neural network, the quality of the input data is crucial because the neural network learns the input data characteristics intentionally. Noisy or uncleaned data may result in false or missing detections. Some embodiments of the present invention provide a method for pre-processing source data before feeding the data to an autoencoder neural network. In order to increase the signal difference between normal and abnormal data and increase the accuracy of anomaly detection, the data signal is pre-processed by applying a digital filter. The digital filter is selected to filter out undesirable features of the data signal. For example, the data signal may be filtered in a pre-set frequency band. As a result, the magnitude of the constituent frequencies of each data signal that are outside the designed frequency band is reduced. There are different types of digital filters, and a Chebyshev filter is used as an example to illustrate the filtering procedure as shown in FIG. 4.

[0040]

number

[0041] The pre-processed data 409 can be fed directly to an autoencoder neural network or can undergo a time-delayed modification as shown in FIG. 4. If the pre-processed data 409 is fed directly to an autoencoder, the resulting autoencoder is called a pre-processed autoencoder (Prep-AE). However, as mentioned above, time information is important in FA systems. Therefore, in FIG. 4, the pre-processed data 409 undergoes a time-delayed modification 300, which outputs pre-processed time-delayed data 410. An autoencoder neural network that receives the pre-processed time-delayed data as input data is called a pre-processed time-delay autoencoder (Prep-TDAE) 411. Prep-TDAE outputs the result 412 as

number

[0042] During the anomaly detection phase, the training data and the test data are first preprocessed and then undergo time-delayed correction. The preprocessed time-delayed training data is used to train the Prep-TDAE. The preprocessed time-delayed test data is input to the trained Prep-TDAE for anomaly detection. If the test error exceeds a predefined threshold, the anomaly is detected. The anomaly occurrence time, the anomaly duration, and the anomaly severity may also be determined. The anomaly occurrence time corresponds to a time index of the test data at which the test error exceeds a threshold. For example, if the 1st time index corresponds to 1:00 PM and the data is sampled every 100 milliseconds, the 10th time index indicates the time 1 second after 1:00 PM. The anomaly duration indicates the length of time the anomaly exists. For example, if the data samples are collected every 10 milliseconds and the anomaly lasts for five consecutive time indexes, the anomaly duration is 50 milliseconds. The anomaly severity may be determined based on the magnitude of the test error. The greater the magnitude of the test error, the more severe the anomaly.

[0043] Note that the anomaly index detected by Prep-TDAE is the TDAE input index, and the corresponding time index can be calculated as time index=TDAE input index+time delayed window size. Abnormality diagnosis in FA systems and manufacturing production lines

[0044] Anomaly diagnosis is identifying the exact source that caused the anomaly. FA systems are composed of various hardware and software components. Some anomalies may be caused by a single component, while other anomalies may be caused by a combination of multiple components. Some embodiments of the present invention provide a statistics-based anomaly diagnosis method and a residual error-based anomaly diagnosis method. Anomaly diagnosis using statistics-based methods

[0045]

number

[0046] In the anomaly diagnosis stage 501, similar to diagnostic training, test data 520 is preprocessed, time delayed and rectified. The preprocessed time delayed test data 521 is first standard normalized (511). Next, the stored whitening matrix is ​​used to whiten transform (522) the normalized test data. After that, the stored mean and variance in the last step of anomaly training are used to standard normalize (523) the resulting sequence after whitening transformation.

[0047]

number

[0048] It can be observed that the anomaly score of an analog sensor is calculated by adding the anomaly scores of the pre-processing filter outputs, which indicates that if an anomalous signal is detected among the pre-processing filter outputs, the corresponding sensor will be identified as anomalous.

[0049] It can also be seen that a final anomaly score is calculated for each data signal. For a data signal, if its anomaly score is above a predefined threshold, the data signal is identified as an anomalous signal 525. For each anomalous signal, the sensor that collected the signal is identified 526. The anomalous sensor corresponds to an FA system component that is operating abnormally. Anomaly diagnosis using residual error-based methods

[0050]

number

[0051]

number

[0052]

number

[0053]

number

[0054] It can be observed that if the Prep-TDAE residual errors are independent, the occurrence of anomalies will induce large reconstruction errors of the corresponding sensors. A threshold can be pre-set to determine the abnormal sensors. For a data signal, if its anomaly score exceeds the pre-defined threshold, the data signal is identified as an abnormal signal 644. For each abnormal signal, the sensor that collected the signal is identified (645). The abnormal sensors correspond to FA system components that are operating abnormally.

[0055] FIG. 7 illustrates a block diagram of an anomaly detection system according to an embodiment of the present invention. The diagram illustrates a block diagram of an apparatus 700 for controlling a system including multiple signal sources causing multiple events, according to some embodiments. One example of a system is a manufacturing production line. The apparatus 700 includes a processor 720 configured to execute stored instructions and a memory 740 that stores instructions executable by the processor. The processor 720 may be a single-core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 740 may include random access memory (RAM), read only memory (ROM), flash memory, or any other suitable memory system. The processor 720 is connected to one or more input and output devices through a bus 706.

[0056] The instructions implement a method for detecting and / or diagnosing anomalies in a plurality of events of a system. The apparatus 700 is configured to detect and diagnose anomalies using a stored program 730. The stored program includes a data pre-processing function 731, a time-delayed data correction function 732, an anomaly detection algorithm 733, and an anomaly diagnosis algorithm 734. The anomaly detection algorithm 733 may be a Prep-TDAE, and the anomaly diagnosis algorithm 734 may be a statistical-based method or a residual error-based method. The anomaly detection algorithm 733 and the anomaly diagnosis algorithm 734 are trained to detect and diagnose control conditions of the system. For example, the Prep-TDAE neural network 733 may be trained offline by using operational data and used to detect anomalies online. Similarly, the anomaly diagnosis algorithm may be trained offline and used to diagnose anomalies online. Examples of operational data include signals from signal sources collected during operation of the system, e.g., events of the system. Examples of training data include signals from signal sources collected over a period of time. The period of time may be a time interval before operation / production begins and / or during operation of the system.

[0057] The above-described embodiments of the present invention may be implemented in any of many ways. For example, the embodiments may be implemented using hardware, software, or a combination thereof. When implemented in software, the software code may be executed on any suitable processor or collection of processors, whether the processors are provided in a single computer or distributed among multiple computers. Such a processor may be implemented as an integrated circuit having one or more processors within an integrated circuit component. The processor may be implemented using circuitry in any suitable format.

[0058] Also, embodiments of the invention may be embodied as a method, examples of which are provided. The operations performed as part of the method may be ordered in any suitable manner. Thus, while an example embodiment may show operations as sequential, embodiments may be constructed in which operations are performed in an order different from that illustrated, including even including performing some operations simultaneously.

[0059] The use of ordinal terms such as "first," "second," etc. in the claims to modify claim elements does not, by itself, imply any priority, precedence, or ordering of one claim element over another claim element, or the temporal order in which method operations are performed, but is merely used as a label to distinguish one claim element having a certain name from another element having the same name (except for the use of the ordinal terminology).

[0060] Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention.

[0061] Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.

Claims

1. 1. A computer-implemented preprocessed time-delay autoencoder based anomaly detection method for detecting an abnormal condition in a machine located in a factory automation (FA) system or manufacturing production line, comprising: acquiring a source signal from the machine via an interface; performing a data pre-processing process for the acquired source signal by normalizing a value range of the acquired source signal and filtering undesirable features from the acquired source signal; performing a time delayed data correction process for the preprocessed source signal based on a time delay window to generate preprocessed time delayed data; and submitting preprocessed time-delay test data to a preprocessed time-delayed autoencoder (Prep-TDAE) neural network, the preprocessed time-delay test data being collected online while the machine is being operated, the Prep-TDAE neural network being pre-trained by using the preprocessed time-delay training data, the method further comprising: detecting whether an abnormal condition has been encountered with respect to the machine by calculating an anomaly score of the pre-processed time delay test data; and if said abnormal condition is detected, determining an abnormality time, duration, and severity for said abnormal condition for each of said machines.

2. The method of claim 1 , wherein the data pre-processing process splits the normalized source signal into a normalized analog signal and a normalized binary signal.

3. The method of claim 1 , wherein in the determining step, the abnormal condition is detected by calculating the anomaly score of a normalized version of the signal.

4. 2. The method of claim 1 , wherein an input layer of the Prep-TEAE neural network is configured to accept binary portions of the preprocessed time delayed data via binary acceptors of the input layer and to accept analog portions of the preprocessed time delayed data via analog acceptors of the input layer.

5. 2. The method of claim 1, wherein in a pre-processing step, the acquired signal is split into a binary signal and an analog signal, the analog signal is processed by a band-pass filter and then by a high-pass filter, and the binary signal is processed by a unit impulse filter.

6. The method of claim 1 , wherein the Prep-TDAE neural network performs loss function adaptation by applying different loss functions to different data types.

7. The method of claim 6 , wherein the binary signals are trained with a cross-entropy loss function and the analog signals are trained with a squared error loss function.

8. The method of claim 1 , wherein the anomaly occurrence time corresponds to a time index in the test data at which an anomaly score exceeds a predefined threshold.

9. The method of claim 1 , wherein the anomaly duration is the length of time that the anomaly score of test data persistently exceeds a predefined threshold across successive time indexes.

10. The method of claim 1 , wherein the anomaly severity is determined based on a magnitude of the anomaly score.

11. 1. A computer-implemented, statistics-based anomaly diagnosis method for diagnosing anomaly sources in a machine located in a factory automation (FA) system or manufacturing production line, comprising: acquiring a source signal from the machine via an interface; performing a data pre-processing process for the acquired source signal based on a data pre-processing algorithm; performing a time delayed data correction process algorithm for the acquired source signals based on a time delay window to generate pre-processed time delayed data; normalizing the preprocessed time delay data to have zero mean and unit variance; calculating a statistically reduced covariance matrix such that said covariance matrix is ​​not irreversible; computing a whitening matrix for a whitening transformation; performing said whitening transformation of said signal for the calculation of said mean and said variance; calculating the mean and the variance of the transformed signal for anomaly score calculation; calculating an anomaly score for each data signal for an anomaly data signal diagnosis; diagnosing the abnormal data signal; and identifying, from among said machines, a machine having said anomaly using said anomaly diagnosis algorithm.

12. 【Number 1】

13. 1. A computer-implemented, residual-based anomaly diagnosis method for diagnosing anomaly sources in a machine located in a factory automation (FA) system or manufacturing production line, comprising: acquiring a source signal from the machine via an interface; performing a data pre-processing process for the acquired source signals; performing a time delayed data correction process algorithm for the acquired source signals based on a time delay window to generate pre-processed time delayed data; and submitting the preprocessed time-delayed data to a preprocessed time-delay autoencoder (Pre-TDAE) neural network to generate a reconstructed signal, the Prep-TDAE neural network being pre-trained with training data, the method further comprising: calculating a residual error based on the reconstructed signal; calculating the mean and variance of the residual error; calculating an anomaly score for the reconstructed signal; diagnosing an abnormality signal corresponding to the source signal of the machine based on the anomaly score; and identifying from among said machines those machines having an anomaly using an anomaly diagnosis algorithm.

14. 【Fig. 2】

15. 【Number 3】

16. 【FIG. 4】

17. 1. A preprocessed time-delay autoencoder based anomaly detection system for detecting abnormal conditions in a machine located in a factory automation (FA) system or manufacturing production line, comprising: an interface configured to obtain a source signal from the machine; a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, a pre-processed time-delay autoencoder (Pre-TDAE) neural network, an anomaly diagnosis algorithm, and the computer-implemented statistical-based anomaly diagnosis method of claim 1; and a processor, the processor comprising: performing the data pre-processing algorithm for the acquired source signals by normalizing a value range of the acquired source signals and filtering undesirable features from the acquired source signals; performing a time delayed data correction process algorithm on the preprocessed source signal based on a time delay window to generate preprocessed time delayed data; and submitting pre-processed time delay test data to the Prep-TDAE neural network, the pre-processed time delay test data being collected online while the machine is in operation, the Prep-TDAE neural network being pre-trained by using the pre-processed time delay training data, the processor further comprising: detecting whether an abnormal condition has been encountered with respect to the machine by calculating an anomaly score of the pre-processed time delay test data; and if the abnormal condition is detected, determining an abnormality time, duration, and severity for the abnormal condition for each of the machines.

18. 20. The system of claim 17, wherein the data pre-processing process splits the normalized source signal into a normalized analog signal and a normalized binary signal.

19. 1. A statistics-based anomaly diagnosis system for diagnosing anomaly sources in a factory automation (FA) system or a machine located in a manufacturing production line, comprising: an interface configured to obtain a source signal from the machine; a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, an anomaly diagnosis algorithm, and the computer-implemented statistical-based anomaly diagnosis method of claim 11; and a processor configured to execute the computer-executable program, the processor comprising: performing said data pre-processing algorithm on said acquired source signals; performing a time delayed data correction process algorithm for the acquired source signal based on a time delay window to generate pre-processed time delayed data; normalizing the preprocessed time delay data to have zero mean and unit variance; calculating a statistically reduced covariance matrix such that said covariance matrix is ​​not irreversible; computing a whitening matrix for a whitening transformation; performing said whitening transformation of said signal for the calculation of said mean and said variance; calculating the mean and the variance of the transformed signal for anomaly score calculation; calculating an anomaly score for each data signal for an anomaly data signal diagnosis; diagnosing the abnormal data signal; and identifying, from among said machines, a machine having said anomaly using said anomaly diagnosis algorithm.

20. 【Number 5】

21. 1. A residual-based anomaly diagnosis system for diagnosing anomaly sources in a factory automation (FA) system or a machine located in a manufacturing production line, comprising: an interface configured to obtain a source signal from the machine; a memory for storing a computer-executable program including a data pre-processing algorithm, a time-delayed data correction process algorithm, a pre-processed time-delay autoencoder (Pre-TDAE) neural network, an anomaly diagnosis algorithm, and the computer-implemented residual-based anomaly diagnosis method of claim 13; and a processor configured to execute the computer-executable program, the processor comprising: performing said data pre-processing algorithm on said acquired source signals; performing a time delayed data correction process algorithm for the acquired source signal based on a time delay window to generate pre-processed time delayed data; and submitting the pre-processed time delay data to the Pre-TDAE neural network to generate a reconstructed signal, the Prep-TDAE neural network being pre-trained with training data, the processor further comprising: calculating a residual error based on the reconstructed signal; calculating the mean and variance of the residual error; calculating an anomaly score for each data signal; diagnosing the abnormal signal corresponding to the source signal of the machine based on the anomaly score; and identifying a machine having an anomaly from among the machines using the anomaly diagnosis algorithm.

22. 【Number 6】

23. 【Number 7】

24. 【Number 8】

Images