Systems and methods for anomaly detection using attention models

Abstract

An anomaly detector for controlling a system is provided. The system includes one or more tools for performing one or more tasks. The anomaly detector collects feedforward signals indicating a sequence of control inputs to a plurality of actuators and feedback signals indicating a sequence of outputs of the system resulting from the plurality of actuators operated based on the sequence of control inputs. The anomaly detector further combines input state variables extracted from the feedforward signals and output state variables extracted from the feedback signals to form a sequence of extended states of the system. An attention model further encodes the sequence of extended states to generate an encoding of each extended state of the sequence of extended states in a latent space. The anomaly detector further detects an anomaly in the current operation of the system based on the encoded sequence of extended states.

Description

[Technical Field]

[0001] The present invention relates generally to anomaly detection suitable for manufacturing automation, and more particularly to an anomaly detector and anomaly detection method using an attention model. [Background technology]

[0002] A manufacturing process is a process that transforms raw materials into a finished product using tools, machines, human labor, etc. Two types of production methods within a manufacturing process are process manufacturing and discrete manufacturing. In process manufacturing, products are generally undifferentiated, such as oil, natural gas, and salt. Discrete manufacturing produces individual items, such as automobiles, furniture, toys, and airplanes. Regardless of the type of production method, a manufacturing process can involve numerous operations that appear independent of each other but must be performed synchronously on a very strict schedule. For example, many assembly operations require two independent processes to be performed sequentially or even simultaneously with millisecond accuracy to achieve effective results.

[0003] The accuracy of the time synchronization of manufacturing operations can be affected by external and internal disturbances. External disturbances can arise from the versatility of the manufacturing process. For example, details to be assembled can be located in different positions, which can be treated as external disturbances to the assembly process. Conversely, the assembly of a detail in one position can take longer than the assembly of a detail in another position. This difference can be treated as an internal disturbance that can affect the synchronization of the various manufacturing operations.

[0004] Furthermore, controlling the performance of manufacturing processes, where fast, powerful machines can rapidly execute complex sequences of tasks, is a difficult challenge. Deviations from the intended sequence of tasks or the timing of their execution can cause anomalies, which can lead to poor quality, wasted raw materials, downtime, equipment failure, and reduced production. For this reason, great care must be taken to avoid anomalies and minimize unexpected events, and preventative measures must be built into the production line using various sensors and emergency switches.

[0005] In manufacturing process data analysis, anomaly detection is generally understood to be the identification of rare items, occurrences, or observations that deviate from the majority of the data and do not fit a well-defined notion of normal behavior. Such instances may raise suspicions that they were generated by a different mechanism or may appear inconsistent with the rest of the data set. Thus, at least in theory, anomaly detection can be used to identify unusual behavior that indicates manufacturing automation malfunctions. However, due to the complexity of modern manufacturing automation, which uses the collaborative operation of independent processes, anomaly detection based on individual processes is unreliable and / or too complex for collaborative anomaly detection.

[0006] Therefore, there is a need for a system and method for anomaly detection suitable for controlling manufacturing automation that benefits from the collaboration of independent processes. Summary of the Invention

[0007] An objective of some embodiments is to disclose an apparatus and method suitable for anomaly detection in a manufacturing process of a system. In particular, an objective of some embodiments is to provide an anomaly detector and anomaly detection method for anomaly detection using an attention model. The system may include multiple manufacturing operations in which anomalies may be detected in the presence of external and internal disturbances to the system. Examples of manufacturing operations include assembly operations performed by synchronized movements of various robotic manipulators, moving an object through a series of locations, machining a workpiece with various tools in a predetermined sequential or partially overlapping order, and combinations thereof.

[0008] Some embodiments are based on the recognition that for anomaly detection to recover quickly from external and internal disturbances in a system, the system states analyzed for anomaly detection need to be extended to both the control inputs to the system's actuators and the system states resulting from the control inputs.

[0009] Some embodiments recognize that the differing nature of these extended state variables presents challenges to the practicality of anomaly detection. For example, state variables associated with control inputs to actuators may be values ​​of physical entities such as voltage, current, tension, viscosity, resistance, valve opening, etc., while state variables of the system outputs resulting from the control inputs may be values ​​of position, orientation, sound, friction, vibration, pressure, etc.

[0010] Some embodiments are based on the recognition that in systems including modern manufacturing automation processes, the extended state may have hundreds or thousands of different state variables. Due to the number and different nature of these state variables, various anomaly detection methods, such as autoencoder reconstruction-based anomaly detection, generalize beyond the input distribution, making these anomaly detection methods suboptimal.

[0011] The objective of some embodiments is to address the issues of dimensionality and generality of various state variables of the control inputs and outputs of a system for anomaly detection using an attention model. Generally, an attention model based on an attention mechanism is an input processing technique for neural networks that allows the network to focus on specific aspects of a complex input one at a time until the entire dataset is classified. The objective of the attention mechanism is to decompose a complex task into smaller attention regions that are processed sequentially.

[0012] Attention is expressed as a function that maps a query and a "set" of key / value pairs to an output. Typically, the query, keys, values, and final output are all vectors. The output is then computed as a weighted sum of the values, with the weight assigned to each value represented by a compatibility function between the query and the set of keys.

[0013] Some embodiments disclose determining a set of exemplars to be used as keys in an attention model. Each exemplar defines an extended state of normal operation of the system for which anomaly detection is designed. Each exemplar is a vector, and each element of the vector is another vector having a value and / or values ​​of a corresponding state variable. The set of exemplars is predefined to represent various characteristic examples of extended states of normal operation, i.e., non-anomalous performance, of the system's manufacturing process.

[0014] Some embodiments disclose that a first set of values ​​is designed for a set of keys by embedding each key in the set of keys into a low-dimensional space. Thus, a first value in the first set of values ​​is an embedding of a corresponding key in the set of keys using a dimensionality lower than the dimensionality of the extended state of the system's manufacturing process. Such embedding can be performed offline, for example, during training of an attention model using a coding neural network.

[0015] Some embodiments disclose that queries in the attention model define extended states of the manufacturing process at each of a plurality of control steps of the current operation of the system. Thus, the set of queries is time-series data representing a sequence of extended states of the current operation of the system. For a first control step of the plurality of control steps, a first query of the set of queries may be compared with a first key of the set of keys. In this way, the extended states at each control step are mapped to an embedding space or latent space of the normal operation of manufacturing.

[0016] Some embodiments disclose collecting feedforward signals indicating a sequence of control inputs to a plurality of actuators of a system at a plurality of control steps and feedback signals indicating a sequence of system outputs resulting from the plurality of actuators operated based on the sequence of control inputs. The input state variables extracted from the feedforward signals and the output state variables extracted from the feedback signals are combined to form a sequence of extended states associated with the current operation of the system. The sequence of extended states is then sent to an attention model as a set of queries. The set of queries is compared with a set of keys to determine a plurality of attention weights used to weight corresponding first values ​​in a set of first values. Each key in the set of keys corresponds to a respective first value in the set of first values. A comparison metric between the set of queries and the set of keys is based on a comparison of the corresponding state variables with a statistically learned metric. Such comparisons enable anomaly detection to be employed in various manufacturing processes.

[0017] Some embodiments disclose that each of the set of queries is encoded as a corresponding weighted first value of the set of first values ​​weighted with a corresponding attention weight of the plurality of attention weights to generate an encoding of each extension state of the sequence of extension states in the latent space. Thus, the set of queries may be represented as a weighted combination of embeddings of normal operation of the manufacturing process of the system.

[0018] Some embodiments disclose that a dictionary can be designed that includes extrema of a first set of values ​​that are embedding variables that define a predetermined convex subspace of the system's normal operation in the latent space. Furthermore, an anomaly in the system's current operation may be detected when at least one coded query in the set of coded queries is outside the predetermined convex subspace of the latent space. Thus, the attention model can transform the sequence of extended states into a latent space of normal operation and perform anomaly detection in the low-dimensional latent space.

[0019] Some embodiments are based on the recognition that a set of keys and a first set of values ​​may be generated and stored by the anomaly detector during training of the anomaly detector. In this manner, the memory of the anomaly detector has key / value pairs for anomaly detection for various manufacturing processes.

[0020] Some embodiments disclose that the attention model can operate based on a multi-head attention mechanism, in which multiple independent attention mechanisms can be run in parallel, and the outputs of the multiple independent attention mechanisms are concatenated and transformed in the latent space. Through this multi-head mechanism, various manufacturing processes of a system state or a system with various components can be controlled simultaneously.

[0021] Some embodiments are based on the recognition that the results of the anomaly detection may be output to a user interface, so that if the system includes a manufacturing process in which the tools of the system are manually handled by human labor, an operator can modify the operation of the tools based on the output results, thereby making the manufacturing process more efficient.

[0022] Some embodiments are based on the recognition that the anomaly detector can generate control commands based on the results of the anomaly detection and send the control commands to actuators of the system, such that actuator signals provided to tool assemblies of the system can be modified based on the control commands.

[0023] The present disclosure is further described in the following detailed description with reference to several drawings, in which like reference numerals represent like parts throughout the several views of the drawings, shown as non-limiting examples of exemplary embodiments of the present disclosure. The drawings shown are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the presently disclosed embodiments. [Brief explanation of the drawings]

[0024] [Figure 1] FIG. 1 is a schematic diagram of a system according to some embodiments of the present disclosure. [Figure 2A] FIG. 1 illustrates an environment for an anomaly detector, according to some embodiments of the present disclosure. [Figure 2B] FIG. 1 is a block diagram of an anomaly detector according to some embodiments of the present disclosure. [Figure 2C] FIG. 1 is a block diagram of an attention model according to some embodiments of the present disclosure. [Figure 2D] FIG. 1 is a block diagram of the execution period of an attention model for multiple control steps of a manufacturing process of a system according to some embodiments of the present disclosure. [Figure 3]FIG. 10 is a block diagram for encoding an extension state of a sequence of extension states in a latent space for each control step of a plurality of control steps of a manufacturing process of a system, according to some embodiments of the present disclosure. [Figure 4] FIG. 1 is a block diagram for learning a dictionary during normal operation of a system according to some embodiments of the present disclosure. [Figure 5] FIG. 1 is a schematic diagram of a predetermined convex subspace according to some embodiments of the present disclosure. [Figure 6] FIG. 2 illustrates a flowchart illustrating the interaction between a system for detecting anomalies in current operation and an anomaly detector, according to some embodiments of the present disclosure. [Figure 7] FIG. 1 illustrates a flowchart of an exemplary method for anomaly detection in the current operation of a system. [Figure 8] FIG. 1 is a block diagram of an attention model based on a multi-head attention mechanism according to some embodiments of the present disclosure. [Figure 9] FIG. 1 is a block diagram of a use case for controlling a system using an anomaly detector, according to some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0025] In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it will be apparent to those skilled in the art that the present disclosure may be practiced without these specific details. In other instances, devices and methods are shown in block diagram form solely to avoid obscuring the disclosure. It is contemplated that various changes may be made to the function and arrangement of elements without departing from the spirit and scope of the disclosed subject matter as set forth in the appended claims.

[0026] As used in this specification and claims, the terms "for example," "for example," and "such as," as well as the verbs "comprise," "have," and "include," and each of their other verb forms, when used in conjunction with a list of one or more components or other items, should be construed as open-ended, meaning that the list should not be considered as excluding other additional components or items. The term "based on" means based at least in part on. Furthermore, it should be understood that the style and terminology used herein are for purposes of description and should not be regarded as limiting. Any headings used herein are for convenience only and have no legal or limiting effect.

[0027] Specific details are provided in the following description to provide a thorough understanding of the embodiments. However, those skilled in the art will understand that the embodiments may be practiced without these specific details. For example, systems, processes, and other elements of the disclosed subject matter may be shown as components in block diagram form so as not to obscure the embodiments in unnecessary detail. In other instances, well-known processes, structures, and techniques may be shown without unnecessary detail so as not to obscure the embodiments. Furthermore, like reference numbers and names in the various drawings indicate like elements.

[0028] In describing the disclosed embodiments, the following definitions apply throughout this disclosure.

[0029] An "anomaly detector" may refer to a device or set of devices for managing, directing, supervising, or controlling the behavior of other devices or systems. An anomaly detector may be implemented in either software or hardware and may include one or more modules. An anomaly detector that includes a feedback loop may be implemented using a microprocessor. An anomaly detector may be an embedded system.

[0030] "Electrical circuitry" refers to interconnections of wires that transmit electrical signals between components such as processors, memories, or actuators.

[0031] "Computer" refers to any device capable of accepting structured input, processing the structured input according to predetermined rules, and producing the results of the processing as output. Examples of computers include general-purpose computers, supercomputers, mainframes, superminicomputers, minicomputers, workstations, microcomputers, servers, interactive televisions, hybrid combinations of computers and interactive televisions, and special-purpose hardware for emulating computers and / or software. A computer may have a single processor or multiple processors that may or may not operate in parallel. A computer also refers to two or more computers connected to each other via a network to send or receive information from one another. An example of such a computer is a distributed computer system for processing information via computers linked by a network.

[0032] "Central Processing Unit (CPU)" or "Processor" refers to a computer or a component of a computer that reads and executes software instructions.

[0033] "Memory" or "computer-readable medium" refers to any storage device for storing data that can be accessed by a computer. Examples include magnetic hard disks, floppy disks, optical disks such as CD-ROMs or DVDs, magnetic tape, memory chips, and carrier waves used to carry computer-readable electronic data, such as those used in sending and receiving email or accessing a network, and computer memory, e.g., random access memory (RAM).

[0034] "Software" refers to predetermined rules for operating a computer. Examples of software include software, code segments, instructions, computer programs, and programmed logic. Software in intelligent systems can be self-learning.

[0035] A "module" or "unit" refers to a basic component within a computer that performs a task or part of a task, and may be implemented in either software or hardware.

[0036] "Controller," "control system," and / or "regulator" refer to a device or set of devices for managing, directing, supervising, or regulating the behavior of other devices or systems. A controller may be implemented by hardware, a processor with software-configured operations, and combinations thereof. A controller may also be an embedded system.

[0037] 1 illustrates a schematic diagram 100 of an exemplary system 102 according to an embodiment of the present disclosure. The system 102 includes a robotic arm 104 performing an assembly operation within the system. The robotic arm 104 machines a workpiece 106 with a first tool 108 and a second tool 110. The workpiece 106 is then lifted by the robotic arm 104 to form an item 112. Control inputs are provided to actuators to control the robotic arm 104, the first tool 108, and / or the second tool 110. Examples of control inputs may be, but are not limited to, voltage, current, pressure, and / or force applied to one or more joints of the robotic arm 104.

[0038] Systems 102 that perform assembly operations include a sequence of control steps to output a final product. An improper sequence of operations in the manufacturing processes performed by systems 102 or the presence of internal or external disturbances can result in the production of an inaccurate product. Therefore, anomaly detection is necessary to ensure the quality of the products produced by these manufacturing processes.

[0039] Typically, the anomaly detection process is based on features extracted from images of the manufacturing process. Such processes cannot take into account the physical interactions of variables and may not produce effective results where overlooking details could lead to improperly manufactured products. Alternatively, the anomaly detection process may be performed by a neural network, in which case data of the current operation of the system 102 may be reconstructed by the neural network, and anomalies in the manufacturing process may be detected based on the reconstructed data and the original data. In complex manufacturing processes involving thousands of state variables, such data reconstruction may be inaccurate and may lead to undesirable results.

[0040] An objective of some embodiments of the present disclosure is to propose an apparatus suitable for detecting anomalies in the manufacturing process of the system 102. Another objective of some embodiments is to propose a method suitable for detecting anomalies in the manufacturing process of the system 102. The proposed apparatus of the present disclosure may include an anomaly detector for detecting anomalies in the manufacturing process of the system 102 using an attention model. The proposed method of the present disclosure may include an anomaly detection method for detecting anomalies in the manufacturing process of the system 102 using an attention model.

[0041] Typically, in the context of neural networks, "attention" has been designed to transcribe speech. For example, when attempting to transcribe the utterance "bank of a river," the concept of a river is important for interpreting the word "bank" and distinguishing between a riverbank and a financial institution. Thus, attention mechanisms are designed to place different parts of an utterance in context with each other in a manner that can be implemented by a neural network. However, this disclosure discloses an attention model that can be adapted for anomaly detection in the manufacturing process of the system 102.

[0042] It will be understood that for purposes of anomaly detection, this disclosure considers the case of assembly operations in system 102. However, such anomaly detection may equally be used in any manufacturing process in system 102. Examples of manufacturing processes include, but are not limited to, moving an object through a series of locations or machining a workpiece with various tools in a predetermined sequential or partially overlapping order. Additionally, assembly operations may be performed by synchronized movements of various robotic manipulators.

[0043] FIG. 2A illustrates an environment 200A of an anomaly detector 202 according to an embodiment of the present disclosure. The environment 200A may further include a system 102, from which the anomaly detector 202 receives current inputs associated with the current operation of the system 102. The system 102 includes a tool assembly 204 for performing manufacturing operations in the system 102. The system 102 may further include an actuator 206 that sends multiple signals, such as S1, S2, S3, and S4, to the tool assembly 204, which includes multiple tools, such as T1, T2, and T3. For example, in the system 102, the robot arm 104, the first tool 106, and the second tool 108 may be examples of multiple tools. The robot arm 104 may use other tools from the multiple tools to shape and assemble the workpiece 106. The actuator 206 is provided with a sequence 208 of control inputs, such as voltage, current, force, viscosity, vibration, audio signals, etc., for controlling the tool assembly. Additionally, actuator 206 may be a control device such as a motor, a generator, a hydraulic device, a magnetic device, etc. The tools of tool assembly 204 may generate translational displacement, rotational displacement, vibration, friction, audio signals, etc. in response to signals provided by actuator 206.

[0044] One or more sensors 210 read the sequence 208 of control inputs and the sequence 212 of system outputs generated by the sequence 208 of control inputs. The sensors 210 may be used in an electrical circuit (not shown in FIG. 2A ). The sensors 210 may be a group of sensors such as position sensors, pressure sensors, temperature sensors, force sensors, and / or vibration sensors. The sensors 210 record all signals received from the system 102 and transmit the sensor signals to a transducer 214.

[0045] Transducers 214 are electronic devices that convert sensor signals into electrical signals that are then provided to anomaly detector 202 as current inputs 216. Thus, current inputs 216 are a combination of the sequence of control inputs 208 and the sequence of outputs 212 of system 102, based on which the current operation of the manufacturing process in the system is performed.

[0046] The anomaly detector 202 includes a processor 218 configured to control or perform the functions of all components of the anomaly detector 202. The processor 218 may be a single-core processor such as a central processing unit (CPU), a multi-core processor, a computing cluster, or any number of other configurations. The anomaly detector 202 further includes a memory 220 in which an attention model 222 is stored. The memory 220 may include non-transitory memory such as random access memory (RAM), read-only memory (ROM), flash memory, or any other suitable memory system. The anomaly detector 202 is configured to process the current input 216 to detect anomalies in the manufacturing process of the system 102 using the attention model 222. The attention model 222 may operate on an attention mechanism configured to encode the current input 216 into a latent space. Furthermore, the anomaly detector 202 is configured to generate anomaly detection results 224 and transmit the results 224 for display on a user interface 226.

[0047] The user interface 226 can be a display device such as a monitor, laptop, projector, mobile device, etc. An operator of the system may modify the actuator signals based on the displayed results 224. The anomaly detector 202 may generate control commands 228 based on the results 224 and send the control commands 228 to the actuators 206 to control the manufacturing process of the system 102. For example, the control commands 228 may be sent to actuators of the actuators 206 to stop or pause operation of components of the system 102 when the anomaly detector 202 detects an anomaly in the current input 216.

[0048] In one embodiment, the anomaly detector 202, the system 102, and the user interface 226 may communicate via a wired or wireless network (not shown in FIG. 2A ). The network may include public networks such as the Internet, telephone networks, and satellite networks, and / or local area networks (LANs) such as Ethernet, wide area networks (WANs). In addition, examples of networks may include dedicated line networks such as Internet Protocol Virtual Private Networks (IP-VPNs). In addition, the network may include wireless communication networks such as Wi-Fi, Bluetooth, etc.

[0049] 2B shows a block diagram 200B of the anomaly detector 202. As described with reference to FIG. 2A, the anomaly detector is provided with current inputs 216. The current inputs 216 include a feedforward signal 216A and a feedback signal 216B recorded by the sensor 210. The feedforward signal 216A represents a sequence 208 of control inputs provided to the actuators 206. The feedback signal 216B represents a sequence 212 of system outputs caused by a plurality of actuators operated based on the sequence 208 of control inputs.

[0050] The anomaly detector 202 further includes a first extraction unit 230, a second extraction unit 232, a combination unit 234, a first generation unit 236, and a detection unit 238. Furthermore, the processor 218 may perform the functions of the first extraction unit 230, the second extraction unit 232, the combination unit 234, the first generation unit 236, and the detection unit 238.

[0051] The first extraction unit 230 is configured to extract input state variables 240 from the feedforward signal 216A. The input state variables 240 indicate the sequence 208 of control inputs provided to the actuators 206. The input state variables 240 may be a plurality of first entities, such as values ​​of physical entities such as voltage, current, tension, viscosity, resistance, valve opening, etc. Furthermore, the input state variables 240 define a state of the system 202.

[0052] The second extraction unit 232 is configured to extract output state variables 242 from the feedback signal 216B. The output state variables 242 indicate a sequence of outputs of the system 102 in response to the sequence of control inputs 208. The output state variables 242 may be values ​​of position, orientation, sound, friction, vibration, pressure, etc.

[0053] The combination unit 234 combines the extracted input state variables 240 and output state variables 242 to generate a sequence 244 of extended states of the system 102 .

[0054] Typically, in a manufacturing process, i.e., an assembly operation in FIG. 1 , only the sequence of control inputs 208 is considered for the process of anomaly detection in the system. Such anomaly detection process can lead to inefficient data analysis at a granular level, which in turn can result in unreliable results. Because the anomaly detector 202 of the present disclosure processes a combination of input state variables 240 and output state variables 242, it can analyze a large amount of data, thereby reducing the likelihood of erroneous results in the manufacturing process of the system 102. Additionally, collecting a large amount of data directly from the sensors 210 provides a practical approach to anomaly detection, resulting in accurate and reliable results even at the granular level of the manufacturing process. Furthermore, this practical approach eliminates the need for the attention model 222 to learn the dependency of one piece of data on other pieces of data, further shortening the training time of the attention model and reducing storage space in the memory 220.

[0055] The first generating unit 236 is configured to generate a set of queries 246 based on the sequence of extended states 244. Each query in the set of queries 246 is a vector, and each element of the vector is a value and / or another vector having a value of a corresponding first entity. The first generating unit 236 may extract features of each of the plurality of first entities and generate a vector for each of the plurality of first entities. For example, the first generating unit 236 may extract various values, such as magnitude, frequency, phase, etc., from each of the plurality of first entities, such as voltage, current, pressure, etc. Then, the first generating unit 236 generates a query in the set of queries 246 for each control step of the manufacturing process of the system 102 based on the extracted various values. The manufacturing process of the system 102 may be divided into multiple control steps. Each control step may be defined as a time step of the manufacturing process. For example, the first generating unit 236 may generate a query every 30 or 40 seconds of the manufacturing process. The durations of all control steps of the manufacturing process may not be equal and may differ from each other. Furthermore, the dimensions of each first entity may differ from each other, and therefore the dimensions of the elements of the queries in the set of queries 246 may differ from each other. The generated set of queries 246 is sent to the attention model 222.

[0056] Traditional anomaly detection methods are generally based on comparing the normal operating state of the system 102 with the current operating state. Therefore, the state variables of the normal state must be compared with the state variables of the current state. Therefore, the comparison becomes problematic due to the different dimensionality of the state variables. An attention model 222 can be used to address this multiple dimensionality issue. In general, an attention model, or attention mechanism, is a neural network input processing technique that allows the network to focus on specific aspects of a complex input, one at a time, until the entire data set is classified. The purpose of an attention mechanism is to decompose a complex task into smaller attention regions that are processed sequentially.

[0057] The attention model 222 may provide a low-dimensional representation of the sequence 244 of extended states of the manufacturing process of the system 102. This adaptation may include multiple steps performed online and / or offline, i.e., before the anomaly detection process begins. The attention model 222 compares a set of queries 246 with a set of keys 248 to generate multiple attention weights. The set of queries 246 represents the sequence 244 of extended states of the current operation of the system 102, while the set of keys 248 represents the sequence of extended states of the system 102 during normal, non-anomalous operation. A set of examples of normal operation is used as the set of keys 248 in the attention model. Each example defines an extended state of normal operation of the manufacturing process of the system 102. Each example, or key, is a vector, and each element of the vector is a value and / or another vector having values ​​of a corresponding second entity. In this manner, each element of the key represents a second entity of the multiple second entities. The multiple second entities may be values ​​of physical entities such as voltage, current, pressure, friction, viscosity, etc. The set of keys 248 is predefined to represent various characteristic examples of extended states of normal behavior. Pairs of the set of keys 248 and the set of first values ​​250 are stored in the memory 220 of the anomaly detector 202. Each key in the set of keys 248 is paired with a corresponding first value in the set of first values ​​250. The set of values ​​250 is designed for the corresponding set of keys 248 by embedding each key in a low-dimensional space. Thus, the first value is an embedding of the key with a dimension lower than the dimension of the extended states of normal behavior. Such embedding can be performed offline, for example, using a coding neural network. Furthermore, the attention model 222 generates the set of coded queries 252 based on a comparison between the set of queries 246 and the set of keys 248. The coded query set 252 represents the sequence of extended states of the current behavior 244 in the context of the sequence of extended states of normal behavior. The functionality of the attention model 222 will be described in detail with reference to Figures 2C and 2D.

[0058] The detection unit 238 detects anomalies in the manufacturing process of the system 102 based on the set of coded queries 252 and the set of first values ​​250. The detection unit compares the set of coded queries 252 and the set of first values ​​250 using a predetermined convex subspace of the latent space. Details of the predetermined convex subspace are described with reference to FIG. 4 . Furthermore, the detection unit 238 may generate anomaly detection results 224. In one embodiment, the detection unit 238 may output the results 224 on the user interface 226. In another embodiment, the detection unit 238 may output a control command 230 to control the manufacturing process of the system 102 based on the anomaly detection.

[0059] 2C shows a block diagram 200C of the attention model 222. The attention model 222 includes a transformation layer 254 and an attention layer 256.

[0060] The transformation layer 254 transforms the multidimensional set of keys 248 into a lower-dimensional set of first values ​​250. In other words, the set of keys 248 is encoded in a latent space as the set of first values ​​250 in the transformation layer 254. Because the set of keys 248 is associated with the normal operation of the system 102, the encoding of the set of keys 248 into the set of first values ​​250 may be performed during training of the attention model 222. In one embodiment, the encoding of the set of keys 248 into the set of first values ​​250 may be performed by an autoencoder neural network. The autoencoder neural network may capture significant elements of the keys and encode the keys into first values ​​in an unsupervised manner based on the captured elements. Furthermore, pairs of the set of keys 248 and the set of first values ​​250 are stored in the memory 220.

[0061] The set of keys 248 may be encoded into the set of first values ​​250 during the training period of the attention model 222, while the set of queries 246 may be transformed into the set of queries 252 encoded during the execution period of the attention model 222. The execution period is defined as the execution period of the current operation of the system 102. Furthermore, each of the set of queries 246 is encoded as a weighted linear combination of the set of first values ​​250 weighted by a corresponding attention weight from the plurality of attention weights to generate an encoding of each augmented state from the sequence of augmented states 244 in the latent space. To this end, the attention layer 256 receives the set of keys 248 and the set of first values ​​250 from the memory 220. Additionally, the attention layer 256 receives the set of queries 246 associated with the current operation of the system 102 based on the sequence of control inputs 208 and the sequence of outputs of the system 102 recorded by the sensors 210. Attention layer 256 is configured to process set of queries 246, set of keys 248, and first set of values ​​250 to encode each query in set of queries 246 into a latent space. This encoding reduces the dimensionality of set of queries 246; for example, a 10-dimensional query may be converted into a 2-dimensional query in the embedding space or latent space.

[0062] 2D illustrates a block diagram 200D of the execution period of the attention model 220 for multiple control steps of the manufacturing process of the system 102, according to some embodiments of the present disclosure. The block diagram 200D includes an extension of the attention layer 256 of the attention model 222. The attention layer 256 includes an alignment layer 258, a softmax layer 260, and a context layer 262.

[0063]

number

[0064]

number

[0065]

number

[0066] FIG. 3 illustrates a block diagram 300 for encoding an extension state of the sequence of extension states 244 in a latent space for each control step of a plurality of control steps of a manufacturing process of the system 102, according to some embodiments of the present disclosure.

[0067] The system 102 includes a manufacturing process in which one manufacturing operation is performed in one control step. During the manufacturing process, a sequence of control inputs 208 is provided to the system 102 to result in a sequence of outputs 212. For example, a first operation may be performed by tool T1 shown in FIG. 2A in a first control step (t1) of the plurality of steps, a second operation may be performed by tools T2 and T3 shown in FIG. 2A in a second control step (t2) of the plurality of steps, and a third operation may be performed by tools T1 and T3 in a third control step (t3) of the plurality of steps. In the first operation, tool T1 produces a first set of outputs (O) of the sequence of outputs 212 in the first control step (t1). t1 ) to generate a first set of control inputs (I t1 ) may require a first set of control inputs (I t1 ) can be, but are not limited to, voltage, current, pressure, etc., and the first output set (O t1 Examples of the control parameters may be, but are not limited to, linear displacement, friction, etc. In the second operation, tools T2 and T3 output a second set of outputs (O) of the sequence of outputs 212 in a second control step (t2). t2 ) to generate a second set of control inputs (I t2 ) may require a second set of control inputs (I t1) may be, but are not limited to, electrical current, tactile input, audio input, etc., and a second set of outputs (O t1 Examples of the outputs 212 may be, but are not limited to, rotational displacement, friction, audio output, etc. Similarly, in the third operation, tools T1 and T3 provide a second set of outputs (O t3 ) to generate a third set of control inputs (I t3 ) may be required. A third set of control inputs (I t1 ) can be, but are not limited to, pressure, temperature, vibration, etc., and a third output set (O t1 Examples of ) can be, but are not limited to, thermal displacement, friction, etc.

[0068] The combination unit 234 shown in FIG. 2B receives a first set of control inputs (I t1 ) and a first set of output variables (O t1 ) in the sequence of extension states 244 corresponding to the first control step (t1). t1 ) may be generated. Additionally, the combination unit 234 may generate a second set of control inputs (I t2 ) and a second set of output state variables 240 representing t2 ) in the sequence of extension states 244 corresponding to the second control step (t2). t2 Similarly, the combination unit 234 may form a third set of control inputs (I t3 ) and a third set of output state variables 240 representing t3 ) in the sequence of expansion states 244 corresponding to the third control step (t3). t3 ) may be formed.

[0069] The first generation unit 236 shown in FIG. 2B generates a first expansion state (E t1 ) based on the first query (Q t1 ) corresponding to the second control step (t2). t2 ) based on the second query (Q t2 Similarly, the first generation unit 236 may generate a third expansion state (E t3 ) based on the third query (Q t3 ) may be generated.

[0070] The attention model 222 calculates the first query (Q) corresponding to the first control step (t1). t1 ) to generate a first coded query (C t1 ) corresponding to the second control step (t2). t2 ) to generate a second coded query (C t2 Similarly, the attention model 222 may generate a third query (Q t3 ) to generate the third query (C t3 ) may be generated.

[0071] It will be appreciated that a three-step manufacturing process is contemplated herein, however the manufacturing process may be carried out in multiple steps.

[0072] FIG. 4 illustrates a block diagram 400 for learning a dictionary 402 during normal operation of the system 102 according to some embodiments of the present disclosure.

[0073] The processor 218 is configured to learn the dictionary 402 during training of the attention model 222. The dictionary 402 may be in the form of a dictionary matrix having pairs of a set of keys 248 and a set of first values ​​250 stored during normal operation. To that end, the anomaly detector 202 receives a sequence 404 of extended states of normal operation from the system 102. The anomaly detector 202 includes a second generation unit 406 configured to generate the set of keys 248 from the sequence 404 of extended states of normal operation. The second generation unit 406 may extract features of state variables in the sequence 404 of extended states of normal operation and convert them into a vector format as the set of keys 248. Each element of a key in the set of keys 248 defines a second entity of normal operation. Furthermore, each key in the set of keys 248 defines an extended state of normal operation. Each key is a vector, and each element of the vector is a value and / or another vector having a value of a corresponding state variable. It will be understood that the processor 218 performs the functions of the second generation unit 406. The second generation unit 406 then transmits the set of keys 248 to the attention model 222 and the dictionary 402. The attention model 222 converts the set of keys 248 into the set of first values ​​250, as described above with reference to FIG. 2C. The set of first values ​​250 is then transmitted to the dictionary 402. The dictionary 402 may then be stored in the memory 220 of the anomaly detector 202. One or more columns of the dictionary 402 may include extrema of the set of first values ​​250 in the latent space. The dictionary 402 may be trained end-to-end, such that all normal samples of normal behavior are trained to lie within a predetermined convex subspace. The dictionary columns may also be trained using a neural network that takes weighted linear combinations of the first values ​​as inputs and reconstructs them as convex combinations of the dictionary columns, a neural variant of a classical dictionary learning and sparse coding formulation, or a classical prototype analysis formulation. The normal samples represent samples of the first set of values ​​250, and thus represent embeddings of the sequence of extended states 404 of normal behavior in the latent space.

[0074] FIG. 5 shows a schematic diagram 500 of a predetermined convex subspace 502 according to some embodiments of the present disclosure.

[0075] The processor 218 is configured to generate a predetermined convex subspace 502 in the latent space based on the first set of values ​​250 and the dictionary 402. To do so, the processor 218 uses the dictionary 402 to determine all normal samples of normal behavior and extreme values ​​of the normal samples. The processor 218 is further configured to generate a predetermined convex subspace 502, the boundary of which is determined by the extreme values ​​of the normal samples of normal behavior. In other words, the extreme values ​​of the first set of values ​​250 are normal sample 504, normal sample 506, normal sample 508, normal sample 510, normal sample 512, and normal sample 514. These normal samples are on the boundary of the predetermined convex subspace 502. The first set of values ​​250 is an embedding of the sequence of extended states 404 of normal behavior in the latent space, so that the predetermined convex subspace 502 is generated in the latent space. The predetermined convex subspace may be a convex hull, and the convexity of the convex hull includes all normal samples of normal behavior.

[0076] The processor 218 is further configured to detect an anomaly in the current operation of the system 102 when at least one coded query in the set of coded queries 252 is outside a predetermined convex subspace 502 in the latent space. To that end, the processor 218 is configured to determine a position of a current sample of the current operation relative to the predetermined convex subspace 502. The current sample represents a sample of the set of coded queries 252, which in turn represents an embedding of the sequence of extended states 244 in the latent space. The samples of the set of coded queries 252 as shown in FIG. 5 are current sample 516, current sample 518, current sample 520, current sample 522, current sample 524, current sample 526, and current sample 528. The positions of all current samples are determined relative to the predetermined convex subspace 502 to determine whether all current samples are within the predetermined convex subspace 502. 5, processor 218 detects an anomaly in the current operation because current sample 526 and current sample 528 are outside the predetermined convex subspace 502. A decision process for all current samples may be performed for each control step of the manufacturing process, and processor 218 determines at which control step an anomaly is detected. Thus, by learning dictionary 402 based on normal samples and placing the current sample within predetermined convex subspace 502, whose convexity allows bounding the distribution of normal samples within the convex hull of the dictionary elements, samples outside the distribution can be easily characterized.

[0077] Although the shape of the predetermined convex subspace 502 is considered here to be hexagonal, it will be understood that the predetermined convex subspace 502 may have any polygonal shape depending on the nature of the extreme values ​​of the first set of values ​​250 defined in the dictionary 402.

[0078] FIG. 6 illustrates a flowchart 600 illustrating the interaction between the system 102 for detecting anomalies in current operation and the anomaly detector 202 according to an embodiment of the present disclosure.

[0079] In block 602, the anomaly detector 202 is configured to acquire electrical signals corresponding to the sequence of control inputs 208 and the sequence of outputs 212 from the system 102 via the network. To that end, the sensors 208 record the signals of the sequence of control inputs 208, such as voltage, current, pressure, force, etc., applied to the actuators 206 of the system 102. In addition, the sensors 208 record the signals of the sequence of outputs 212 produced by the actuators 206 of the system 102 in response to the sequence of control inputs 208. The sensors 210 transmit the recorded signals to the transducers 214, which convert the sensor signals into corresponding electrical signals and transmit them to the anomaly detector 202. The sequence of control inputs 208 and the sequence of outputs 212 of the system 102 define a sequence of extended states 244 of the system 102 during current operation.

[0080] In block 604, the processor 218 is configured to detect anomalies in the current operation of the system 102 based on the sequence of augmented states 244 of the system. To do so, the processor 218 controls the attention model 222 stored in the memory 220 of the anomaly detector to encode the sequence of augmented states 244 in a latent space. The processor 218 further detects anomalies based on the encoded sequence of augmented states 244 of normal operation in the latent space and the embedded sequence of augmented states. Details of anomaly detection by the anomaly detector are described in the flowchart shown in FIG. 7.

[0081] At block 606, results 224 of the anomaly detection 202 are output by the anomaly detector 202 and displayed on a user interface 226. The results 224 may be transmitted to the user interface 226 via a wired or wireless network. Furthermore, an operator of the system 102 may modify one or more parameters of the actuator 206 based on the displayed results 224 to avoid the anomaly.

[0082] At block 608, the anomaly detector 202 is configured to generate a control command 228 for the actuator 206 based on the detection of the anomaly in the current operation. The control command 228 may be generated to control one or more devices of the actuator 206 to avoid damage to the system 102.

[0083] In block 610, the anomaly detector 202 is configured to send the generated control command 228 to the system 102. In one embodiment, the control command 228 may be sent to a motor of the actuator 206 to change the speed of a tool of the system 102. In another embodiment, the control command 228 may be sent to a rheostat to change the resistance of a tool of the system 102. In another embodiment, the control command 228 may be sent to a temperature controller to change the temperature of a tool of the system 102. In another embodiment, the control command 228 may be sent to a power device to stop or suspend operation of a tool of the system 102. For example, upon detecting a fault in a transmission line or a failure of a circuit breaker in the transmission line, the control command 228 may be sent immediately to a power device to stop or suspend operation of one or more tools coupled to the transmission line.

[0084] FIG. 7 illustrates a flowchart 700 of an exemplary method for anomaly detection in the current operation of the system 102 .

[0085] In block 702, the anomaly detector 202 is configured to collect a feedforward signal 216A and a feedback signal 216B as current inputs 216 from the system 102. The feedforward signal 216A represents a sequence 208 of control inputs provided to the actuators 206. The feedback signal 216B represents a sequence 212 of outputs from the system 102 caused by a plurality of actuators operated based on the sequence 208 of control inputs.

[0086] At block 704, the processor 218 of the anomaly detector 202 is configured to combine the input state variables 240 extracted from the feedforward signal 216A and the output state variables 242 extracted from the feedback signal 216B to form a sequence 244 of extended states of the system 102. To do so, the processor 218 extracts the input state variables 240 from the feedforward signal 216A and the output state variables 242 from the feedback signal 216B. The input state variables 240 indicate a sequence 208 of control inputs provided to the actuators 206. The output state variables 242 indicate a sequence 212 of outputs of the system 102 in response to the sequence 208 of control inputs. The processor 218 further combines the input state variables 240 and the output state variables 242 to form a sequence 244 of extended states of the system 102.

[0087] At block 706, the processor 218 is configured to send the sequence of augmentation states 244 to the attention model 222 as a set of queries 246. To that end, the processor 218 extracts magnitude, frequency, phase, etc. from each augmentation state in the sequence of augmentation states 244 to generate the set of queries 246.

[0088] At block 708, the processor 218 is configured to generate an encoded set of queries 252 from the set of queries 246 using the attention model 222. To do so, the attention model 222 compares each of the set of queries 246 with the set of keys 248 to generate a plurality of attention weights. The attention model 222 further encodes each of the set of queries 246 as a weighted linear combination of the first set of values ​​250 weighted by a corresponding attention weight from the plurality of attention weights. Each of the set of queries 246 is encoded to generate an encoding of a respective augmented state from the sequence of augmented states 244 in latent space.

[0089] At block 710, the processor 218 is configured to detect anomalies in the current operation of the system 102 based on the coded query set 252. To that end, the processor 218 compares the coded query set 252 with the first value set 250 using a predetermined convex subspace 502 in the latent space. The coded query set 252 and the first value set 250 are both embedded samples in the latent space. Thus, the predetermined convex subspace 502 is generated from the first value set 250, and the samples of the coded query set 252 are positioned relative to the predetermined convex subspace 502 to determine whether the samples of the coded query set 252 are within the predetermined convex subspace 502. If any of the samples is outside the predetermined convex subspace 502, an anomaly corresponding to that sample is flagged.

[0090] At block 712, the anomaly detector 202 is configured to generate anomaly detection results 224. In one embodiment, the generated results 224 may be output to a user interface 226. Additionally or alternatively, the processor 218 may generate control commands 228 based on the anomaly detection results 224 and transmit the generated results 224 to the system 102.

[0091] FIG. 8 illustrates a block diagram 800 of an attention model 222 based on a multi-head attention mechanism according to an embodiment of the present disclosure.

[0092]

number

[0093] FIG. 9 illustrates a block diagram 900 of a use case for controlling the system 102 using the anomaly detector 202 according to an embodiment of the present disclosure.

[0094] System 102 depicts a factory automation scenario in which multiple robots 902, 904, and 906 manipulate an object 908 moving on a first conveyor belt 910. For example, robot 902 moves object 908 from position 912 on first conveyor belt 910 to position 914 on structure 916. Additionally, robot 904 moves object 908 from position 914 to position 918 on a second conveyor belt 920. Additionally, robot 906 moves object 908 from position 914 to position 922, stacking object 908 on another object 924.

[0095] The anomaly detector 202 receives a current input corresponding to the current operation of the system. The current input defines a sequence of extended states 244 of the current operation. An attention model 222 in the anomaly detector executes an attention mechanism to encode the sequence of extended states 244 into a latent space. The anomaly detector 202 detects anomalies in the current operation based on a set of encoded queries 252. The anomaly detector 202 further generates control commands 228 based on the detection of the anomaly and sends the control commands 228 to the system 102. In this manner, the system 102 is controlled by the anomaly detector 202 to avoid anomalies in the current operation.

[0096] The above description provides exemplary embodiments only and is not intended to limit the scope, applicability, or configuration of the disclosure. Rather, the following description of exemplary embodiments will provide those skilled in the art with an enabling description for implementing one or more exemplary embodiments. Various changes are contemplated that may be made in the function and arrangement of elements without departing from the spirit and scope of the disclosed subject matter as set forth in the appended claims.

[0097] Specific details are provided in the following description to provide a thorough understanding of the embodiments. However, it will be understood by those skilled in the art that the embodiments may be practiced without these specific details. For example, systems, processes, and other elements of the disclosed subject matter may be shown as components in block diagram form in order to avoid obscuring the embodiments in unnecessary detail. In other instances, well-known processes, structures, and techniques may be shown without unnecessary detail in order to avoid obscuring the embodiments. Furthermore, like reference numbers and names in the various drawings indicate like elements.

[0098] Also, particular embodiments may be described as a process that is depicted as a flowchart, a flow diagram, a data flow diagram, a structure diagram, or a block diagram. While a flowchart may describe operations as a sequential process, many of the operations may be performed in parallel or simultaneously. Additionally, the order of operations may be rearranged. A process may be terminated when its operations are completed, but may have additional steps not discussed or included in the diagram. Moreover, not all operations in any specifically described process may occur in all embodiments. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, etc. When a process corresponds to a function, the termination of the function may correspond to a return of the function to the calling function or the main function.

[0099] Furthermore, embodiments of the disclosed subject matter may be implemented, at least in part, either manually or automatically. The manual or automatic implementation may be performed or at least assisted by machine, hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments to perform the necessary tasks may be stored on a machine-readable medium. A processor(s) may perform the necessary tasks.

[0100] The various methods or processes outlined herein may be coded as software executable on one or more processors using any one of a variety of operating systems or platforms. Additionally, such software may be written using any of a number of suitable programming languages ​​and / or programming or scripting tools, and may be compiled as executable machine language code or intermediate code that runs on a framework or virtual machine. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.

[0101] Each embodiment is described as a process, which is depicted as a flowchart, flow diagram, data flow diagram, structure diagram, or block diagram. While a flowchart depicts operations as a sequential process, many of the operations can be performed in parallel or simultaneously. Additionally, the order of operations may be rearranged. A process may be terminated when its operations are completed, but may have additional steps not discussed or included in the diagram. Moreover, not all operations in any process specifically described may occur in all embodiments. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, etc. When a process corresponds to a function, the termination of the function may correspond to a return of the function to the calling function or the main function.

[0102] Furthermore, embodiments of the disclosed subject matter may be implemented, at least in part, either manually or automatically. The manual or automatic implementation may be performed or at least assisted by machine, hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segments to perform the necessary tasks may be stored on a machine-readable medium. A processor(s) may perform the necessary tasks.

[0103] Numerous modifications and other embodiments of the disclosures described herein will come to mind to one skilled in the art to which the disclosures described herein pertain having the benefit of the teachings presented in the above description and the associated drawings. It is to be understood that the disclosure is not limited to the particular embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Moreover, while the above description and the associated drawings describe exemplary embodiments in the context of certain illustrative combinations of elements and / or functions, it should be recognized that various combinations of elements and / or functions may be provided by alternative embodiments without departing from the scope of the appended claims. In this regard, combinations of elements and / or functions other than those expressly described above are also contemplated, for example, as may be set forth in some of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

Claims

1. 1. An anomaly detector for controlling a system including a plurality of actuators assisting one or more tools in performing one or more tasks, comprising: at least one processor; and a memory having instructions stored thereon, the instructions, when executed by the at least one processor, causing the anomaly detector to: collecting feedforward signals indicative of a sequence of control inputs to the plurality of actuators at a plurality of control steps; collecting feedback signals indicative of a sequence of outputs of the system produced by the plurality of actuators operated based on the sequence of control inputs at corresponding control steps of the plurality of control steps; combining input state variables derived from the feedforward signal and output state variables derived from the feedback signal to form a sequence of extended states of the system; sending the sequence of augmented states as a set of queries to an attention model; each augmentation state in the sequence of augmentation states corresponds to a respective query in the set of queries to the attention model; The attention model is comparing each of the set of submitted queries to a set of keys to generate a plurality of attention weights; configured to encode each of the set of queries as a weighted linear combination of a first set of values ​​weighted with a corresponding attention weight from the plurality of attention weights to generate an encoding of each extension state from the sequence of extension states in a latent space; each key in the set of keys is associated with a normal extended state of normal operation of the system; each first value in the set of first values ​​is an encoding of a corresponding key in the set of keys into the latent space; each attention weight of the plurality of attention weights is generated by comparing the plurality of first entities and the plurality of second entities at a respective state by combining similarity distances between a first entity of a plurality of first entities of a query of the set of queries and a corresponding second entity of a plurality of second entities of a key of the set of keys; The similarity distance for comparing the first entity and the second entity is obtained from a mathematical space used to describe the first entity and the second entity, and the anomaly detector further comprises: Detecting anomalies in the current operation of the system based on the set of encoded queries; and generating an anomaly detection result.

2. The anomaly detector of claim 1 , wherein during training of the anomaly detector, the at least one processor is further configured to learn a dictionary containing extreme values ​​of the first set of values.

3. The at least one processor further comprises: configured to generate a predetermined convex subspace in the latent space based on the first set of values ​​and the dictionary, wherein a boundary of the predetermined convex subspace is determined based on the extreme values ​​of the first set of values; and the at least one processor further 3. The anomaly detector of claim 2, configured to detect the anomaly in the current operation of the system when at least one coded query of the set of coded queries is outside the predetermined convex subspace of the latent space.

4. The at least one processor further comprises: encoding each key in the set of keys to a respective first value in the set of first values ​​in the latent space during training of the anomaly detector; The anomaly detector of claim 1 , configured to store the set of keys and the set of first values ​​in the memory.

5. The anomaly detector of claim 1 , wherein the state variables extracted from the feedforward signal correspond to at least one physical entity provided to the system.

6. The anomaly detector of claim 1 , wherein the state variable extracted from the feedback signal corresponds to at least one output of the system sensed by a sensor.

7. 2. The anomaly detector of claim 1, wherein each of the normal and current operations of the system includes at least one of an assembly operation performed by synchronized movement of various robotic manipulators, moving an object through a series of locations, or machining a workpiece with various tools in a predetermined sequential or partially overlapping order.

8. 2. The anomaly detector of claim 1, wherein the at least one processor is further configured to send a first extension state of the sequence of extension states for a first control step of the plurality of control steps of the current operation of the system to the attention model as a first query of the set of queries.

9. the at least one processor is further configured to compare the first query to a first key of the set of keys for the first control step; The anomaly detector of claim 8 , wherein the first key is associated with the normal operation of the system for the first control step.

10. 2. The anomaly detector of claim 1, wherein the attention model is based on a multi-head attention mechanism, and the at least one processor is further configured to encode the set of queries into the latent space based on the multi-head attention mechanism.

11. The anomaly detector of claim 1 , wherein the at least one processor is further configured to output results of the anomaly detection to a user interface.

12. The at least one processor further comprises: generating a control command based on the result of the abnormality detection; The anomaly detector of claim 1 , configured to send the control command to the system.

13. 1. An anomaly detection method, comprising: collecting feedforward signals indicative of a sequence of control inputs to a plurality of actuators at a plurality of control steps; collecting feedback signals indicative of a sequence of system outputs produced by the plurality of actuators operated based on the sequence of control inputs at corresponding control steps of the plurality of control steps; combining input state variables derived from the feedforward signal and output state variables derived from the feedback signal to form a sequence of extended states of the system; sending the sequence of augmented states as a set of queries to an attention model; each augmentation state in the sequence of augmentation states corresponds to a respective query in the set of queries to the attention model; The attention model is comparing each of the set of submitted queries to a set of keys to generate a plurality of attention weights; configured to encode each of the set of queries as a weighted linear combination of a first set of values ​​weighted with a corresponding attention weight from the plurality of attention weights to generate an encoding of each extension state from the sequence of extension states in a latent space; each key in the set of keys is associated with a normal extended state of normal operation of the system; each first value in the set of first values ​​is an encoding of a corresponding key in the set of keys into the latent space; each attention weight of the plurality of attention weights is generated by comparing the plurality of first entities and the plurality of second entities at a respective state by combining similarity distances between a first entity of a plurality of first entities of a query of the set of queries and a corresponding second entity of a plurality of second entities of a key of the set of keys; The similarity distance for comparing the first entity and the second entity is obtained from a mathematical space used to describe the first entity and the second entity, and the anomaly detection method further comprises: Detecting anomalies in the current operation of the system based on the set of encoded queries; and generating an anomaly detection result.

14. The anomaly detection method of claim 13 , further comprising learning a dictionary containing extreme values ​​of the first set of values ​​during training of the anomaly detector.

15. generating a predetermined convex subspace in the latent space based on the first set of values ​​and the dictionary, wherein a boundary of the predetermined convex subspace is determined based on the extreme values ​​of the first set of values; 15. The anomaly detection method of claim 14, further comprising detecting the anomaly in the current operation of the system when at least one coded query of the set of coded queries is outside the predetermined convex subspace of the latent space.

16. encoding each key of the set of keys to a respective first value of the set of first values ​​in the latent space during training of the anomaly detector; 14. The anomaly detection method of claim 13, further comprising: storing the set of keys and the first set of values ​​in the memory.

17. 14. The anomaly detection method of claim 13, further comprising sending a first extension state of the sequence of extension states for a first control step of the plurality of control steps of the current operation of the system to the attention model as a first query of the set of queries.

18. 20. The anomaly detection method of claim 17, further comprising comparing the first query to a first key of the set of keys for the first control step, the first key being associated with the normal operation of the system for the first control step.

19. 14. The anomaly detection method of claim 13, wherein the attention model is based on a multi-head attention mechanism, and the at least one processor is further configured to encode the set of queries into the latent space based on the multi-head attention mechanism.

20. A non-transitory computer-readable medium having computer-executable instructions stored thereon, the computer-executable instructions, when executed by a computer, causing the computer to perform an operation of collecting feedforward signals indicative of a sequence of control inputs to a plurality of actuators at a plurality of control steps; collecting feedforward signals indicative of a sequence of control inputs to a plurality of actuators at a plurality of control steps; collecting feedback signals indicative of a sequence of system outputs produced by the plurality of actuators operated based on the sequence of control inputs at corresponding control steps of the plurality of control steps; combining input state variables derived from the feedforward signal and output state variables derived from the feedback signal to form a sequence of extended states of the system; sending the sequence of augmented states as a set of queries to an attention model; each augmentation state in the sequence of augmentation states corresponds to a respective query in the set of queries to the attention model; The attention model is comparing each of the set of submitted queries to a set of keys to generate a plurality of attention weights; configured to encode each of the set of queries as a weighted linear combination of a first set of values ​​weighted with a corresponding attention weight from the plurality of attention weights to generate an encoding of each extension state from the sequence of extension states in a latent space; each key in the set of keys is associated with a normal extended state of normal operation of the system; each first value in the set of first values ​​is an encoding of a corresponding key in the set of keys into the latent space; each attention weight of the plurality of attention weights is generated by comparing the plurality of first entities and the plurality of second entities at a respective state by combining similarity distances between a first entity of a plurality of first entities of a query of the set of queries and a corresponding second entity of a plurality of second entities of a key of the set of keys; The similarity distance for comparing the first entity and the second entity is obtained from a mathematical space used to describe the first entity and the second entity, and the operations further include: Detecting anomalies in the current operation of the system based on the set of encoded queries; and generating an anomaly detection result.

Images