System and method for detecting abnormal machine behavior using hyperbolic embeddings

Abstract

A computer-implemented method for detecting anomalies in machine operation based on signals indicating the operation of a machine performing a task includes the step of collecting hyperbolic embeddings of the signals indicating the machine operation. The hyperbolic embeddings are in hyperbolic space. The method further includes the steps of performing anomaly detection of machine operation based on the hyperbolic embeddings to determine an anomaly score, and rendering the anomaly score. Machine operation is controlled based on the rendered anomaly score.

Description

Technical Field

[0001] The present disclosure generally relates to signal processing and control applications, and more specifically to detecting anomalies by analyzing signals indicative of the operation of a machine performing a task.

Background Art

[0002] Diagnosis and monitoring of the operating performance of machines are beneficial for a wide variety of applications. Physical machines comprise several components that cooperate to perform tasks. During operation, the machines and / or their components operate in some predefined manner, and specific observations regarding the operation of such machines serve to distinguish whether the machines and / or their subcomponents are operating normally or abnormally. An example of such an observation is an acoustic or sound signal created by the machine during operation. For example, some electromechanical moving parts create a specific type of acoustic signal distinguishable from the acoustic signals created in the event of any immediate or near-future failure indicating abnormal operation. In some scenarios, automatic diagnosis of a machine based on deep learning-based techniques can be performed to detect abnormal signals in the operation of a machine performing a task.

[0003] While deep learning-based technologies available in this field offer better insights into the analysis of anomalous signals, conventional approaches in this area of ​​focus have significant limitations. To achieve the desired anomaly detection accuracy, all conventional solutions typically convert audio signals into high-dimensional embeddings that reside in a so-called high-dimensional space. Such spaces remain far less dimensional than the audio measurements themselves, but are still high-dimensional because they have a number of dimensions that are not easily manipulated by humans. The appropriate number of dimensions required for optimal performance must be balanced between conflicting constraints: it must be large enough to accurately capture the diversity of anomalous behavior, but small enough to maintain the computational load of the controlled method and avoid capturing unnecessary spurious patterns that do not actually exhibit anomalous behavior. The main drawback of existing anomaly detection systems based on audio / visual representations of machine behavior is that the features under analysis are complex and abundant for realistic anomaly detection accuracy. In these available solutions, such high-dimensional embeddings can exceed the computational time and memory requirements for systems such as embedded systems, limiting the computing power that can be incorporated into various factory automation settings.

[0004] Therefore, it is necessary to overcome the above-mentioned problems associated with detecting abnormal machine behavior through signal processing. More specifically, it is necessary to develop feasible and efficient methods and systems for detecting signal fragments indicating abnormal machine behavior. [Overview of the project]

[0005] Various embodiments of this disclosure disclose systems and methods for detecting abnormal behavior of a machine performing a task. In addition, an objective of some embodiments is to perform abnormal sound detection using deep learning techniques.

[0006] Furthermore, the objective of some embodiments is to provide unsupervised anomaly sound detection by training a model that can detect anomalies when only data from normal operating conditions is available for training the model parameters. Typical applications include machine sound condition monitoring and diagnosis in applications such as predictive maintenance and factory automation.

[0007] Automated machine diagnostics can be used to detect abnormal signals using training data corresponding to the normal operating state of the machine. Such anomaly detection based on training data is an unsupervised approach. For example, unsupervised anomaly sound detection would be suitable for detecting certain types of anomalies, such as abrupt transient disturbances or impact sounds that can be detected based on rapid changes over time. Furthermore, these approaches can capture much more complex anomaly patterns in the data, i.e., patterns that may be too subtle or complex for human operators to detect, explain, and / or diagnose.

[0008] Some exemplary embodiments are based on the understanding that, in order to achieve the desired anomaly detection accuracy, one approach may be to transform the audio signal into a high-dimensional embedding existing in a so-called high-dimensional space. Such a space is characterized as high-dimensional because, although its dimensions are still far smaller than those of the audio measurement itself, the number of dimensions is still too large to be easily manipulated by humans. Some exemplary embodiments also recognize that the appropriate number of dimensions required for optimal performance must be balanced between conflicting constraints: it must be large enough to accurately capture the diversity of anomalous behavior, but small enough to maintain the computational load of the controlled method and avoid capturing unnecessary spurious patterns that do not actually exhibit anomalous behavior. For example, an embedding network incorporates features extracted from a signal containing one or a combination of audio and video signals into a space with 128 dimensions to adequately capture the diversity of anomalies and achieve satisfactory performance. However, some exemplary embodiments are based on the recognition that such high-dimensional embeddings may exceed the computation time and memory requirements for systems such as embedded systems, thus limiting the computing power that can be incorporated into various factory automation settings.

[0009] Several exemplary embodiments recognize that unsupervised anomaly detection remains a complex and challenging task. Some exemplary embodiments recognize that one of the main shortcomings of existing anomaly detection algorithms is their inability to handle domain shifts. For example, in the context of sound detection, acoustic characteristics may change between normal data collected for training and normal data collected during inference due to factors such as different background noise and different operating voltages. This failure is generally caused by algorithms that cannot distinguish between anomaly signal changes caused by anomalies and normal signal changes caused by domain shifts. Some exemplary embodiments also recognize that one possible solution to building an effective method resilient to these task complexities is to increase the dimensionality of the vector quantities manipulated within the algorithm to detect anomalies. However, such increases can lead to further computational time and memory requirements for systems such as embedded systems, limiting the computational power that can be incorporated into various factory automation settings. In addition, increased dimensionality also increases the risk of capturing spurious behavior rather than representative behavior. Therefore, some exemplary embodiments are aimed at providing alternative methods for processing feature embeddings that are feasible on smaller embedded devices due to requiring fewer resources, and that minimize false behavior in anomaly detection tasks.

[0010] Some exemplary embodiments are based on the recognition that hyperbolic space exhibits a far greater ability to encode hierarchical relationships than Euclidean space when using its embeddings for classification. A necessary consequence of this property is that the distance of a given embedding from the hyperbolic space origin encodes the concept of classification certainty, naturally mapping normal-value class samples to the spatial edges and outliers near the origin. Thus, some exemplary embodiments are based on the recognition that, when attempting to identify unseen anomalous data, hyperbolic embeddings generated by a deep neural network pre-trained to classify short-time Fourier transform frames computed from normal machine sounds are superior to Euclidean embeddings.

[0011] Several exemplary embodiments are based on recognizing the advantages of hyperbolic space in the context of anomaly detection in machine operation. This new and further property gives hyperbolic space advantages over Euclidean space that were not previously recognized. Specifically, encoding a signal into hyperbolic space requires fewer dimensions than encoding the same signal into Euclidean space in order to convey enough information to detect anomalies with practical reliability. For example, in the context of sound anomaly detection, 128-dimensional encoding of an audio signal in Euclidean space conveys the same information as two-dimensional encoding of the same signal in hyperbolic space.

[0012] In some embodiments, this property is discovered empirically in the context of sound anomaly detection. Therefore, it was initially unclear whether this novel property was a property of hyperbolic space or a property of the acoustic signal. However, several further experiments justify the conclusion that it is a property of hyperbolic space rather than an audio signal. Thus, in addition to, or instead of, anomaly sound detection using hyperbolic embedding, some embodiments use hyperbolic embedding of different types of signals that indicate machine operation for anomaly detection.

[0013] Therefore, some exemplary embodiments are directed toward providing systems, methods, and programs for performing unsupervised anomaly detection using embeddings from a trained neural network architecture having a hyperbolic embedding layer, using embeddings generated from test samples to generate anomaly scores.

[0014] To achieve the above objectives and advancements, several exemplary embodiments provide systems, methods, and computer program products for detecting anomalies in machines operating to perform tasks.

[0015] Some exemplary embodiments provide an anomaly detection system for detecting anomalies in the operation of a machine based on signals indicating the operation of the machine performing a task. The system comprises at least one processor and a memory storing instructions. When the instructions are executed by the at least one processor, they cause the anomaly detection system to collect a hyperbolic embedding of the signals indicating the operation of the machine. The hyperbolic embedding lies in a hyperbolic space defined by a model of n-dimensional hyperbolic geometry in which multiple points of hyperbolic geometry lie in an n-dimensional open subset of a real coordinate space of dimension n. The processor is also configured to perform the detection of anomalies in the operation of the machine based on the hyperbolic embedding, determine an anomaly score, and render the anomaly score.

[0016] In several other exemplary embodiments, a computer-implemented method is provided for detecting anomalies in the operation of a machine based on signals indicating the operation of the machine performing a task. The method includes the step of collecting a hyperbolic embedding of the signals indicating the operation of the machine. The hyperbolic embedding lies in a hyperbolic space defined by a model of an n-dimensional hyperbolic geometry in which multiple points of the hyperbolic geometry lie in an n-dimensional open subset of a real coordinate space of dimension n. The method further includes the steps of performing the detection of anomalies in the operation of the machine based on the hyperbolic embedding to determine an anomaly score, and rendering the anomaly score.

[0017] In several other exemplary embodiments, a non-temporary computer-readable medium is provided which stores computer-executable instructions for performing a method for detecting abnormalities in the operation of a machine.

[0018] According to some exemplary embodiments, the measured values ​​of the signal or features extracted from the measured values ​​are processed using an embedding neural network to create a Euclidean embedding of the signal into Euclidean space. The Euclidean embedding is then projected onto hyperbolic space to create a hyperbolic embedding. The hyperbolic embedding is processed by a classification neural network to create at least a portion of the anomaly scores.

[0019] In some exemplary embodiments, the classification neural network is trained using training data generated from the normal operation of the machine. The normal operation of the machine is defined against the machine's standard operational performance data. For example, the machine's standard operational performance data may define values ​​for one or more machine parameters. According to some exemplary embodiments, the embedded neural network is trained together with the classification neural network such that the weights of the embedded neural network and the weights of the classification neural network are interdependent of each other.

[0020] According to some exemplary embodiments, the anomaly score is a function of the combination of the distance between the hyperbolic embedding and the origin of the hyperbolic space and the positive classification rate returned by the classification neural network. A control system operably connected to the anomaly detection system can control the operation of the machine based on the rendered anomaly score.

[0021] The embodiments disclosed herein will be further described with reference to the following drawings. The drawings shown are not necessarily drawn to scale; instead, emphasis is placed on explaining the principles of the embodiments disclosed herein as a whole.

Brief Description of the Drawings

[0022] [Figure 1A] FIG. is a block diagram of an anomaly detection system for a machine according to some exemplary embodiments. [Figure 1B] FIG. is a flowchart showing a method for detecting anomalies in a machine according to some exemplary embodiments. [Figure 2A] FIG. is a block diagram of a training configuration of an anomaly detection system according to some exemplary embodiments. [Figure 2B] FIG. is a block diagram of a runtime configuration of a trained anomaly detection system according to some exemplary embodiments. [Figure 3] FIG. is a diagram showing an exemplary control application that utilizes the anomaly detection system of FIG. 1A according to some exemplary embodiments. [Figure 4] FIG. is a diagram showing a runtime configuration of the anomaly detection system of FIG. 2B that operates on an acoustic signal according to some exemplary embodiments. [Figure 5A] FIG. is a diagram showing a modified architecture of a classifier backbone utilized by the anomaly detection system of FIG. 4 according to some exemplary embodiments. [Figure 5B] FIG. is a diagram showing an exemplary structure of an embedded neural network utilized by the anomaly detection system of FIG. 4 according to some exemplary embodiments. [Figure 5C] FIG. is a diagram showing an exemplary structure of a classification neural network utilized by the anomaly detection system of FIG. 4 according to some exemplary embodiments. [Figure 6] FIG. is a diagram showing a two-dimensional Poincaré sphere projection of a hyperbolic space according to some exemplary embodiments. [Figure 7]A block diagram of a training configuration of an exemplary anomaly detection system that uses an acoustic signal for anomaly detection, according to some exemplary embodiments. [Figure 8A] A block diagram of an exemplary runtime configuration of the trained anomaly detection system of FIG. 7, according to some exemplary embodiments. [Figure 8B] A block diagram of another exemplary runtime configuration of the trained anomaly detection system of FIG. 7, according to some exemplary embodiments. [Figure 9] A diagram showing a hardware configuration of an exemplary anomaly detection system, according to some exemplary embodiments.

MODE FOR CARRYING OUT THE INVENTION

[0023] The above drawings depict the embodiments disclosed herein, but as described in the description, other embodiments are also contemplated. This disclosure presents the exemplary embodiments as representative rather than limiting. Those skilled in the art can devise numerous other modifications and embodiments that fall within the scope and spirit of the principles of the embodiments disclosed herein.

[0024] The following description provides only exemplary embodiments and is not intended to limit the scope, applicability, or configuration of the present disclosure. Rather, the following description of the exemplary embodiments provides those skilled in the art with an enabling description of how to implement one or more exemplary embodiments. It is contemplated that various changes may be made in terms of the function and arrangement of the elements without departing from the spirit and scope of the disclosed subject matter as set forth in the appended claims.

[0025] Specific details are given in the following description to ensure a full understanding of the embodiments. However, those skilled in the art will understand that embodiments can be carried out without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in the form of block diagrams so as not to obscure the embodiments with unnecessary details. In other examples, well-known processes, structures, and technologies may be shown without unnecessary details to avoid obscuring the embodiments. Furthermore, similar reference numbers and names in different drawings may refer to similar elements.

[0026] Furthermore, individual embodiments may be described as processes shown as flowcharts, flow diagrams, data flow diagrams, structural diagrams, or block diagrams. While flowcharts may describe operations as sequential processes, many of these operations can be performed in parallel or simultaneously. The order of these operations may also be reordered. A process may terminate when its operations are complete, but it may have additional steps that are not discussed or included in the diagrams. Moreover, not all operations in any specifically described process occur in all embodiments. A process may correspond to a method, function, procedure, subroutine, subprogram, etc. If a process corresponds to a function, the termination of the function may correspond to the function returning to the calling function or main function.

[0027] Furthermore, embodiments of the disclosed subject matter may be implemented at least partially manually or automatically. Manual or automatic implementation may be performed, or at least assisted, by using a machine, hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. If implemented with software, firmware, middleware, or microcode, the program code or code segments for performing the required tasks may be stored in a machine-readable medium. A processor may perform the required tasks.

[0028] Automatically detecting malfunctioning equipment, or anomaly detection, is an essential task in modern industrial society. Anomaly detection in the context of a machine performing a task can be considered the detection of several observational results that provide attributes and inferences different from those of a machine performing normal operation. While accurately predicting when a failure or anomaly will occur is difficult, there have been numerous attempts to automate the task of detecting anomalies that have occurred using machine-related measurements. Several signals indicating machine operation can be of various types, such as audio, video, and vibration signals. The choice of signals for detecting anomaly operation varies depending on the type of machine, the environment in which the machine operates, and the measurement mechanisms available to acquire such signals. In some scenarios, anomaly sound detection is advantageous because it can be adapted to a wide variety of machines and operations.

[0029] For physical machines, such as those with moving parts, sound or acoustic signals are a good indicator of abnormal behavior. Performing anomaly detection from sound, i.e., anomaly sound detection, is also attractive due to factors such as sensor cost and the ability to measure signals without line of sight. However, training a neural network to predict anomalies using conventional methods requires anomaly samples, which is a difficult task because anomalies do not occur when desired and are often rare. Secondly, intentionally causing malfunctions in a machine to produce abnormal behavior is not practically feasible due to the potential damage to the machine resulting from such malfunctions. Furthermore, since anomalies depend on several operating parameters of the machine, generating a rich set of training data that can respond to all types of anomalies is a time-consuming and often impossible task. Therefore, whether sound-based or not, realistic anomaly detection designs are hindered by the difficulty of collecting anomaly samples and are further affected by issues such as the rarity of anomaly occurrence or the costs associated with intentionally causing anomalies, beyond the cost of labeling. Therefore, the unsupervised approach is particularly noteworthy in this field.

[0030] For example, unsupervised anomaly detection techniques are useful when only data from normal operating conditions (i.e., non-anomalous machine operation) is available for training model parameters. Common approaches to unsupervised anomaly detection include those based on architectures such as autoencoders, in which a model trained only on normal data to reconstruct inputs should exhibit large reconstruction errors when presented with anomaly examples at inference. Another class of approaches, called surrogate task models, train a model of normality using alternative supervised training tasks and then predict anomalies by measuring deviations from normality. Exemplary surrogate tasks include (1) outlier exposure, using sounds known to be quite different from the machine under consideration as synthetic anomalies; (2) predicting / classifying metadata (e.g., machine instance) or attributes (e.g., operational load); or (3) learning to predict what extensions (e.g., time stretching or pitch shifting) will be applied to an audio clip.

[0031] Predicting or classifying metadata (e.g., machine instance) or attributes (e.g., operational load) of an audio clip involves learning to accurately classify these metadata or attributes for normal machine sounds by learning representations of these metadata or attributes through the distribution of learned embeddings (i.e., output vectors from the last hidden layer). The anomaly detector is built on this “normal” embedding distribution, which is learned only from normal machine sounds, and uses the relative position of the embedding of an unseen sample from the “normal” embedding distribution to determine the possible state (normal or anomalous) of the sample. For example, the distance of an embedding to the K-th nearest normal value neighborhood in the embedding space can be used as a criterion for setting a distance threshold, and if this distance threshold is exceeded, the sample is considered anomalous.

[0032] Several exemplary embodiments recognize that hyperbolic spaces exhibit a far greater ability to encode hierarchical relationships than Euclidean spaces when their embeddings are used for classification. Some exemplary embodiments are based on recognizing a novel property of hyperbolic spaces relevant to the context of anomaly detection in machine behavior. This novel property gives hyperbolic spaces advantages over Euclidean spaces that were previously unrecognized. Specifically, encoding a signal into hyperbolic space requires fewer dimensions than encoding the same signal into Euclidean space to convey enough information to detect anomalies with realistic reliability. For example, in the context of sound anomaly detection, 128-dimensional encoding of an audio signal in Euclidean space conveys the same information as two-dimensional encoding of the same signal in hyperbolic space.

[0033] Therefore, some exemplary embodiments attempt to train and analyze embeddings as vectors in hyperbolic space rather than as general vectors in Euclidean space. The hyperbolic neural network approach is indeed appealing because, since the volume increases exponentially as a function of the distance to the spatial origin, tree structures can be incorporated into hyperbolic space with arbitrary low distortion. Tree nodes are positioned closer to the origin the shallower the space, and further away the deeper the space. Thus, it can more naturally accommodate the hierarchical aspects of many speech tasks and datasets. Exemplary embodiments described herein attempt to leverage such aspects, including the use of hyperbolic neural networks. A particularly appealing aspect in the context of surrogate task anomaly detection methods is the behavior of the inevitable consequences of embeddings in hyperbolic space, which shows that, when information is hierarchically organized in space, the distance of the embedding to the origin represents something akin to the concept of certainty with respect to the input features. Some exemplary embodiments consider the merits of replacing Euclidean space with hyperbolic space for learning embeddings in surrogate task-based methods that contribute to a simple and effective detection method.

[0034] If anomalies are quantized as scores, and these scores are then aggregated based on two strong parameters that indicate the underlying anomaly, then the detected anomalies are reasonably reliable. The first of the two strong parameters is the probability that the signal belongs to a particular type, and the second is the geodesic distance of the signal's embedding from the origin in hyperbolic space. Thus, the calculated anomaly score is reliable enough to design control solutions and / or some control actions for maintenance in light of the detected anomalies. Therefore, several exemplary embodiments provide techniques for controlling one or more aspects of a machine in which anomalies are detected. These and several other advantages and improvements will be demonstrated through the following description.

[0035] Figure 1A is a block diagram of a machine anomaly detection system 100 according to some exemplary embodiments. The anomaly detection system 100 is communicatively coupled to a signal processing system 10 that processes an input signal 102. The input signal 102 corresponds to observations about the operation of the machine. In this regard, the input signal 102 may be a continuous signal captured by one or more sensors in the environment in which the machine operates. In some exemplary embodiments, the input signal 102 may be a mixture of different signals, such as signals from different sources, signals captured by different sensors, or signals of different modalities. According to some exemplary embodiments, the signal may be one or a combination of audio signals, video signals, vibrations, etc. The signal processing system 10 includes appropriate circuitry and software for processing the input signal so that it is decomposed into segments of appropriate size. In this regard, the signal processing circuitry 10 may sample a threshold portion of the input signal at regular intervals and convert the sampled portion into a time-frequency representation. According to some exemplary embodiments, the sampled portion of the input signal 102 is converted into a Short-Time Fourier Transform (STFT) frame for further processing.

[0036] The anomaly detection system 100 includes an embedding neural network 20 that creates a Euclidean embedding corresponding to each segment of an input signal 102. In this regard, if the input signal 102 is an acoustic signal, the neural network 20 maps the values ​​of the time-frequency bin to a higher-dimensional embedding (i.e., a higher-dimensional embedding in Euclidean space). For example, if the input signal 102 is an acoustic signal, the embedding neural network 20 generates the embedding as a vector of numbers for each frame of the sound, where each frame generally has a fixed duration. For example, in a two-dimensional Euclidean embedding, the network 20 outputs two numbers for two two-dimensional coordinates (e.g., x and y). In a more dimensional Euclidean embedding, the network 20 outputs a number of numbers for each frame equal to the number of dimensions present. In general, there are no simple and straightforward intrinsic geometric properties to utilize in Euclidean space. For example, the origin of space (x=0 and y=0) is not particularly significant; it can be moved without affecting the system. Therefore, the anomaly detection system 100 includes a Euclidean-to-hyperbolic converter 25 that converts Euclidean embeddings to hyperbolic embeddings. In this regard, the converter 25 projects each Euclidean embedding into hyperbolic space. The advantage of using hyperbolic embeddings over Euclidean embeddings lies in the fact that, in the context of sound anomaly detection, 128-dimensional coding of a sound signal in Euclidean space conveys the same information as 2-dimensional coding of the same signal in hyperbolic space, thus requiring fewer resources and time for processing.

[0037] Hyperbolic embeddings are provided to a classification neural network 30 of the anomaly detection system 100. According to some exemplary embodiments, the classification neural network 30 is a linear classifier in hyperbolic space that learns to organize the embeddings so that auxiliary aspects of the machine's non-anomalous behavior can be identified through some geometric criteria based on their respective locations. Due to the geometric properties of hyperbolic space, when considered in terms of Euclidean space, the “linear” boundary (i.e., hyperplane) is an arc rather than a line in two dimensions, and “curved” rather than “flat” in higher dimensions. According to some exemplary embodiments, the classification neural network 30 is a deep neural network trained to classify hyperbolic embeddings as belonging to a particular attribute type. In this regard, for each hyperbolic embedding, the classification neural network 30 calculates the probability that the corresponding segment of the input signal 102 belongs to a particular attribute type. In some exemplary embodiments, the attribute type may be defined in relation to the machine on which the anomaly detection is being performed. For example, the attribute type may represent identifiers such as the machine's model number, machine type identifier, execution speed, or execution mode. The goal of the classification neural network 30 is to find geometric boundaries between different types in hyperbolic space, and it is pre-trained to learn, as frequently and accurately as possible, the boundaries where a particular attribute type of the machine is detected. Depending on the location of the corresponding embeddings for the trained type boundaries, the classification neural network 30 outputs the probability that a particular signal segment of the input signal 102 belongs to a different known attribute type.

[0038] The anomaly detection system 100 also includes an anomaly scoring system 40 for quantifying the degree of anomaly detected in the input signal 102. In this regard, the anomaly scoring system 40 calculates an anomaly score 104 for each segment (or each hyperbolic embedding) of the input signal 102 and outputs an anomaly score for the input signal 102 based on the aggregation of the anomaly scores for the individual embeddings. According to some exemplary embodiments, anomaly scoring is performed based on probabilities output by the classification neural network 30. According to some exemplary embodiments, those probabilities and / or the inherent geometric aspects of the embedding's location in the embedding space are aggregated to generate an anomaly score 104 for each hyperbolic embedding. The higher this score, the more likely the signal segment is to have anomalous data. The anomaly detection system 100 outputs the anomaly score 104 as a measure of the underlying anomaly in the machine's operation.

[0039] According to some exemplary embodiments, an anomaly score 104 may be output for each segment of the input signal 102. The anomaly score 104 may be rendered to a display device via the output port of the anomaly detection system or to the machine's control system 50 for further processing. The control system 50 may be located inside or outside the anomaly detection system 100. The control system 50 compares the anomaly score 104 to a threshold and initiates several control actions. The threshold may be defined in relation to the machine's operating state, the type of machine, or some other appropriate parameter for determining whether the machine's operation is abnormal or not. If the anomaly score 104 is greater than or equal to the threshold, the control system 50 may output one or more control commands 106 to control the machine's operation. The control commands 106 may include triggering an alert or alarm, changing one or more operating parameters of the machine, initiating maintenance actions on the machine, or a combination thereof.

[0040] In this way, the anomaly detection system 100 can detect abnormalities in the machine's operation without requiring significant resources and time for computation. Furthermore, since the neural networks 20 and 30 are trained only using training data corresponding to the machine's normal operation, the anomaly detection system 100 can classify any data that deviates from the machine's normal operation as abnormal.

[0041] Figure 1B is a flowchart of a machine anomaly detection method 150 according to several exemplary embodiments. Method 150 may be performed by the anomaly detection system 100 of Figure 1A. The method includes a step 152 of receiving an input signal, such as an input signal 102 that captures data indicating the operation of a machine performing a task. The input signal may be received via an input port or interface of the anomaly detection system 100. The received input signal is converted to a time-frequency representation in the manner described with reference to Figure 1A (154). The values ​​of the time-frequency bins are mapped to higher-dimensional embeddings in Euclidean space using an embedding neural network, and each higher-dimensional embedding is then projected into hyperbolic space (156). According to some embodiments, all time-frequency bins may be mapped collectively to a higher-dimensional embedding, for example, as one embedding per frame. For each hyperbolic embedding, in the manner described with reference to Figure 1A, the probability that the signal segment corresponding to a particular hyperbolic embedding belongs to a particular source type or attribute type is determined (158). An anomaly score is calculated for each segment of the input signal based on the determined probabilities and, optionally, on some geometric properties of the hyperbolic embedding in hyperbolic space (160). The calculated anomaly scores can then be rendered to a display device or controller for further processing and action.

[0042] Figure 2A is a block diagram of a training configuration for an anomaly detection system according to some exemplary embodiments. The objective of the training process is to provide an embedded neural network 220, such as a deep neural network, trained to classify normal machine signals or data fed in as Short-Time Fourier Transform (STFT) frames. When attempting to identify previously unseen anomaly data of a machine, the hyperbolic embeddings generated by such an embedded neural network 220 are superior to Euclidean embeddings. Referring to Figure 2A, training data 202A, 202B, ..., 202N from respective signal sources 1, 2, ..., N are fed into the embedded neural network 220 to generate hyperbolic embeddings 222. Sources 1, 2, ..., N may correspond to different machines or different sections of the same machine. According to some exemplary embodiments, the training data 202A, 202B, ..., 202N correspond to normal operating data of multiple machines, i.e., when the machines are operating without any anomalies. In some exemplary embodiments, the training data 202A, 202B, ... 202N correspond to normal operation data of multiple sections or parts of the same machine. In some other exemplary embodiments, the training data 202A, 202B, ... 202N correspond to normal operation data of multiple machines and combinations of normal operation data of multiple sections or parts of one or more machines among the multiple machines. In this regard, the training data 202A, 202B, ... 202N may be obtained from appropriate sensors or a data repository compiled using such sensors that capture signals such as sound, image, vibration, voltage, and current associated with the multiple machines. For example, in some exemplary embodiments, these signals may be one or a combination of acoustic and video signals. In some exemplary embodiments, these signals may include measurements of vibration caused by the operation of the machine. In some exemplary embodiments, these signals may include one or a combination of voltages and currents that control the machine, as well as torques generated by the machine.

[0043] The classification neural network 230 is trained in an unsupervised manner to classify auxiliary aspects of the training data (e.g., machine type, model number, execution speed, execution mode). The classifier neural network 230 is also trained to learn to organize hyperbolic embeddings 222 so that these auxiliary aspects can be identified through some geometric criteria based on their respective locations in hyperbolic space. Thus, classification is learned from embeddings in hyperbolic geometry. The embedding neural network 220 generates embeddings as numerical vectors for each segment of the input training signals from the training data 202A...202N. The embeddings are fed into the classifier neural network 230, which finds geometric boundaries 232 between different types of sources. These different types are defined through various attributes, each representing one or more characteristics of each of the signal sources 1,2,...N.

[0044] Geometric boundaries can allow us to interpret hyperbolic embeddings as classification probabilities where the auxiliary aspects of the training data samples associated with those embeddings are of a specific type (or value). Geometric boundaries can be determined by finding the boundary that minimizes a loss or cost function as much as possible. The cost function may, on the one hand, correspond to the cross-entropy loss between the classification probabilities interpreted from the embedding positions relative to the geometric boundary, and on the other hand, correspond to the true type (or value) of the auxiliary aspects of the training data samples associated with those embeddings. Finding the geometric boundary that minimizes the loss function can be done using a stochastic gradient optimization algorithm. A stochastic gradient optimization algorithm may be an adaptive optimization algorithm. An adaptive optimization algorithm may be a Riemann-Adam optimization algorithm, i.e., an algorithm that generalizes the Adam optimization algorithm to a Riemannian manifold (as seen in the space) that covers a hyperbolic space.

[0045] Various examples are presented in the training process to optimize both the embedding network and the classifier network so that the classifier 230 learns as frequently and accurately as possible the boundary where the type of signal source is detected. In some exemplary embodiments, the output of the classifier network 230 can be checked using a validation set. According to some exemplary embodiments, the embedding neural network is trained jointly with the classifier neural network so that the weights of the embedding neural network and the weights of the classifier neural network are interdependent of each other.

[0046] This document describes an exemplary training process and configuration for unsupervised anomaly detection, where the input signal is an acoustic signal provided as a file. Each input signal file is processed using STFT with a Han window of 1024 samples and a hop size of 256 samples, resulting in 313 frames from which the amplitude is obtained. In each epoch, the network is trained using blocks of 32 consecutive STFT frames from each file, i.e., 6000 blocks of size 32 × 1025 randomly selected for each file. These blocks are grouped into batches of size 32. During testing, the amplitude STFT of the input file is divided into overlapping blocks of 32 frames with a step size of 1 frame, and then, for each of these blocks, a logit is collected as the basis for the score of the input file.

[0047] Figure 2B is a block diagram of the runtime configuration of the trained anomaly detection system of Figure 2A in some exemplary embodiments. During inference, the goal is to determine whether a signal segment contains data that differs from the data of normal machine operation. Signals 252 from known sources are fed into a trained embedding neural network 260, which outputs an embedding 262 in hyperbolic space. A trained classifier neural network 270 outputs a probability 272 that each signal segment of signal 252 belongs to a known source type. For example, these probabilities can be expressed as decimals or percentages. In some exemplary embodiments, the relative position 264 of each hyperbolic embedding with respect to the origin of hyperbolic space is obtained from the embedding neural network 260, and an aggregater 275 aggregates the relative embedding positions 264 and the corresponding probabilities 272 to calculate an anomaly score 204. Some exemplary embodiments are based on the recognition that embeddings into hyperbolic space can be performed together with classification to leverage separate or joint training.

[0048] Figure 3 shows an exemplary control application utilizing the anomaly detection system of Figure 1A, according to several exemplary embodiments. The neural network of the anomaly detection system 100 can be considered to be trained in the manner described with reference to Figure 2A. Furthermore, Figure 3 is described in the context of acoustic signals, i.e., the input signal to the anomaly detection system 100 is assumed to be an audio signal, but it can be intended that the control example of Figure 3 may be performed with other appropriate types of input signals. Since the machine 302 can perform tasks such as manipulating or rotating objects, several sounds may be produced by the machine during operation due to the machine's movements. A transducer, such as a microphone 304, captures the audio signal produced by the machine during operation and transmits the captured signal 306 to the anomaly detection system 100. In some exemplary embodiments, the audio signal 306 may be a mixture of sounds produced by several sources in the environment of the machine 302. The anomaly detection system 100 performs the anomaly detection method of Figures 1B and 2B and outputs an anomaly score 308 for each audio signal. An anomaly score 308 may be provided to the controller 320 of machine 302 to perform fault analysis based on the calculated anomaly score 308. For example, the controller 320 may check the anomaly score 308 of each signal using a threshold, and if the anomaly score is greater than or equal to the threshold, the controller 320 may generate one or more control commands 322 aimed at correcting an underlying fault in machine 302. For example in this regard, the controller 320 may change the mode of operation of machine 302. In addition or optionally, in some exemplary embodiments, the controller 320 may generate one or more alarms 324 based on the results of the fault analysis. In addition or optionally, in some exemplary embodiments, the anomaly score 308 may be rendered on a display device 310.

[0049] Here, several operational and structural aspects of the anomaly detection system are described assuming the input signal is an acoustic signal. However, it is intended that such a description is non-limiting and can be extended to other types of input signals without departing from the scope of this disclosure. Unsupervised abnormal sound detection system

[0050] Figure 4 shows an exemplary runtime configuration of the anomaly detection system of Figures 1A and 2B operating on an acoustic signal. An STFT frame 402 of the acoustic signal is provided to a classifier backbone 404, which is a deep neural network that outputs an embedding 406, which is a vector of dimension N in Euclidean space. The Euclidean embedding 286 is then passed through a fully-connected (FC) classifier block 410 to output a class logit 418. The classifier block 410 leverages hyperbolic space to first map / project the Euclidean embedding 406 onto a Poincaré sphere with a corresponding exponential map (412) to generate a hyperbolic embedding 414, and then uses a hyperbolic polynomial regression layer 416 to determine the class or type logit 418 as an anomaly score.

[0051] As a baseline, the system is trained using a standard multinomial regression classifier block ("Euclidean"), i.e., we have identity mappings, and the logit layer, which functions as the anomaly scoring system 40 in Figure 1A, is simply a fully connected layer with n output channels corresponding to n machines or n sections per machine. Alternatively, the system can be trained using an ArcFace classifier block ("Arcface"). For this as well, we have identity mappings. In addition, the logit layer differs between training and testing to motivate the appearance of desired margins in the (Euclidean) embedding space during training. For all systems, training is performed on low-dimensional, e.g., N=2 embedding spaces and high-dimensional, e.g., N=128 embedding spaces of the acoustic signal for experimental evaluation purposes. In practice, training may be performed on only one specific dimension. Embedded neural networks and classification neural networks for processing acoustic signals

[0052] Here, we will describe in detail some examples of the classifier backbone 404 and classifier block 410 in Figure 4. In particular, Figure 5A is a diagram showing a modified architecture of the classifier backbone used by the anomaly detection system of Figure 4 in some exemplary embodiments, and Figure 5B is a diagram showing an exemplary structure of the classifier backbone used by the anomaly detection system of Figure 4 in some exemplary embodiments. Figure 5C is a diagram showing an exemplary structure of the classification neural network used by the anomaly detection system of Figure 4 in some exemplary embodiments.

[0053] For example, signal 252 could be an acoustic signal from a machine. In this embodiment, the classifier backbone, i.e., the embedded neural network 220 in Figure 2A, could be an adaptive version of the MobileFaceNet architecture, for example, as shown in Figure 5A. Referring to Figure 5A, the main required modification is for the global depth convolution (linear GDC) layer, where the kernel size is changed from 7x7 to 1x33. All convolutions are two-dimensional. dw-Conv means depth convolution. For each layer, the expansion coefficient (t), number of channels (c), number of iterations (n), and stride (s) are shown. All convolutions except the last linear layer use parametric rectified linear units (PreLU) as nonlinearity.

[0054] Figure 5B shows an exemplary structure of a classifier backbone 404, i.e., an embedded neural network, used by the anomaly detection system of Figure 4, in some exemplary embodiments. In some embodiments, the embedded neural network may receive input data samples 502. For example, the input data samples 502 may correspond to blocks of short-time Fourier transform frames from an audio signal. The input data samples may be processed by a two-dimensional convolutional layer 504, which may output hidden data samples, sometimes also referred to as hidden vectors. The convolutional layer 504 may include a set of three operations, which may be a convolution, a batch normalization, and a point-direction parametric rectification linear operation.

[0055] The hidden data samples output from layer 504 may be further processed by a two-dimensional depth-direction convolution layer 506, which may output another hidden vector. The depth-direction convolution layer 506 may include a series of three operations, which may be a depth-direction convolution, a batch normalization operation, and a point-direction parametric rectification linear operation.

[0056] The hidden data samples output by layer 506 may be further processed by a series of N bottleneck layers 508. The input to the first bottleneck layer 508A may be the second hidden data sample. For each of the nth bottleneck layers (i.e., from the layer after 508a to 508N), the hidden data samples input to the layer operation may be the hidden data samples output from the previous ((n-1)th) bottleneck layer operation. The hidden data samples output by the nth bottleneck 508N may be further processed by a two-dimensional convolutional layer 510, which may output another hidden data sample. Layer 510 may include a series of three operations, which may be a convolution operation, a batch normalization operation, and a point-direction parametric rectification linear operation.

[0057] The hidden data samples output from layer 510 may be further processed by a two-dimensional global depth-direction convolution layer 512, which outputs another hidden data sample. The depth-direction convolution layer 512 may include a series of two operations, which may be a depth-direction convolution operation and a batch normalization operation.

[0058] These hidden data samples can be further processed by a linear layer 514. The linear layer 514 may output embedded data samples 516. The linear layer includes a set of two operations, which may be a linear operation and a batch normalization operation.

[0059] Each convolution and depth convolution in the sequence of operations described above may be characterized by a set of parameters. These parameters may be a pair of integers corresponding to the span of the two-dimensional convolution kernel, an integer corresponding to the stride of the convolution stride, and an integer corresponding to the number of channels. The linear operations described above may further be characterized by a set of parameters. These parameters may be a single integer corresponding to the number of channels (which in some embodiments may be called the embedding dimension).

[0060] Each bottleneck layer operation may process the input (hidden) data samples 522 using a set of operations. Each bottleneck layer 508 may be characterized by a set of parameters. These parameters may be an integer corresponding to the bottleneck stride, an integer corresponding to the number of bottlenecks in the output channel, and an integer corresponding to the bottleneck expansion coefficient. The input data samples 522 may be processed by a two-dimensional convolutional layer 524, which may output internal hidden data samples. The convolutional layer 524 may include a set of three operations, which may be a convolution operation, a batch normalization operation, and a point-direction parametric rectification linear operation.

[0061] The internally hidden data sample output by layer 524 may be further processed by a two-dimensional depth-direction convolution layer 526, which may output another internally hidden data sample. The depth-direction convolution layer 526 may also include a set of three operations, which could be a depth-direction convolution, a batch normalization operation, and a point-direction parametric rectification linear operation. This internally hidden data sample may be further processed by a two-dimensional convolution layer 528, which may output another internally hidden data sample. The convolution layer 528 may include a set of two operations, which could be a convolution operation and batch normalization.

[0062] In some embodiments, the two-dimensional depth convolution operation described above may be configured to have a stride of 1 or 2, depending on the preferred configuration of the bottleneck layer operation. When the stride is set to 1 and the input hidden data sample and the last internal hidden data sample have the same number of channels, this last internal hidden data sample may be added to the input hidden data sample of the bottleneck layer operation to form another internal hidden data sample. The last internal hidden data sample to be computed may be the output data sample 530 of the bottleneck layer operation. Each convolution operation and depth convolution operation may have a kernel size of 1 × 1 span. The first convolution layer 524 may output an internal hidden data sample whose dimension may be the dimension of the input hidden data sample of the block multiplied by the bottleneck expansion coefficient. The depth convolution layer 526 may take an internal hidden data sample as input whose dimension may be the dimension of the input hidden data sample of the bottleneck block multiplied by the bottleneck expansion coefficient, and output an internal hidden data sample whose dimension may be the dimension of the input data sample to the bottleneck block multiplied by the bottleneck expansion coefficient. The depth-direction convolution operation may further have a stride corresponding to the bottleneck stride. The second convolution operation may take as input an internal hidden data sample whose dimension may be the dimension of the input hidden data sample of the bottleneck block multiplied by the bottleneck expansion coefficient, and output an internal hidden data sample whose dimension may be the number of bottlenecks in the output channel. Each of the above convolution operations, depth-direction convolution operations, and linear operations do not necessarily require the use of biasing components. The hidden data samples and embedded data samples can be represented as coordinates in Euclidean space.

[0063] According to some exemplary embodiments, the parameters and the number of bottleneck layers 508, as defined with reference to Figure 5B, may be configured following the information shown in Figure 5A.

[0064] Figure 5C shows an exemplary structure of a classification neural network 410 used by the anomaly detection system of Figure 4, according to some exemplary embodiments. In some embodiments, the input embedding data sample 552 is represented as coordinates in Euclidean space. The number of coordinates may correspond to the embedding dimension (or size). This embedding data sample may be further processed by a mapping function operation 554, which may output another embedding data sample 556, represented here as coordinates in a mapped space. This embedding data sample may be further processed by a logit function operator 558, which may output a logit data sample 560. This logit data sample may represent the probability that the input embedding data sample belongs to one of several predefined classes. The number of probabilities may be equivalent to the number of predefined classes.

[0065] In some embodiments, the mapping function operation 554 may correspond to a mapping between coordinates in Euclidean space and coordinates in hyperbolic space. In some other embodiments, the mapping between coordinates in Euclidean space and coordinates in hyperbolic space may be approximated by a mapping between coordinates in Euclidean space and coordinates in Poincaré sphere space. In this context, the output embedding data samples from the mapping function operation 554 may be embedding data samples represented as coordinates in hyperbolic space, or optionally in Poincaré sphere space. In some embodiments, the logit function operation 558 may correspond to a hyperbolic multiple linear regression function operation. hyperbolic space

[0066]

number

[0067]

number

[0068] Figure 6 shows a two-dimensional (2D) Poincaré sphere 600 representing hyperbolic space according to several exemplary embodiments. In the Poincaré sphere model of hyperbolic geometry, lines are represented as arcs whose ends are perpendicular to the boundary of the disk (and diameters are also allowed). Two non-intersecting arcs correspond to parallel lines, intersecting orthogonal arcs correspond to perpendicular lines, and arcs intersecting on the boundary are a pair of limiting lines. Each hyperbolic embedding is visually indicated at a specific location on the Poincaré sphere as a point, a cross, or some kind of visual marker. The location of a hyperbolic embedding on the Poincaré sphere is calculated based on the audio signal spectrogram and the values ​​of a corresponding set of TF bins, such as a set of time-frequency (TF) bins belonging to a given frame. The information associated with each hyperbolic embedding includes information about the original sound corresponding to the set of TF bins. Each value on the Poincaré sphere contains some information about the original sound or input audio signal. Advantageously, hyperbolic embeddings preserve hierarchical information associated with the original sound, while Euclidean embeddings do not. Hierarchical representations in Euclidean space are not very compact and require a lot of memory. On the other hand, representing hierarchies on a Poincaré sphere is advantageous because hyperbolic representations represent hierarchies compactly, requiring less memory and not being computationally extensive. Furthermore, it has the advantage of obtaining a visual representation of sound. Compared to non-hierarchical networks, it can use lower-dimensional embeddings.

[0069]

number

[0070]

number

[0071] A common practice in deep learning approaches for classification tasks is to designate vectors generated by some hidden layer, such as the deepest hidden layer, as embeddings. Embeddings obtained from neural networks trained for general classification are often found to be useful alternative representations of the input data, and since these embeddings become better representations for classification, it is expected that the distribution of these embeddings will begin encoding high-level features of the data. Therefore, they are often utilized in downstream tasks different from the original classification task.

[0072] Hyperbolic spaces possess geometric properties that make them particularly attractive in their respective contexts. In particular, their volume increases exponentially with distance from the origin, unlike Euclidean spaces where volume increases polynomially. As a result, it is possible to incorporate tree structures into hyperbolic spaces with arbitraryly low distortion. Simultaneously, the geodesic distance between two points behaves similarly to the distance between two nodes in a tree. Therefore, hierarchical features can be expected to be effectively encoded in these spaces. At the same time, high-level aspects of many common datasets can be expected to exhibit natural hierarchies. Thus, several exemplary embodiments have recognized the merits in many applications of mapping embeddings generated by deep neural networks into hyperbolic spaces before performing geometric equivalents of polynomial regression in spaces using hyperplanes, such as the space shown in Figure 6. Note that all vectors in the network other than these mapped embeddings reside in Euclidean space, so this type of approach is usually hybrid.

[0073] According to some exemplary embodiments, each hyperbolic embedding corresponds to a vector representing a unique attribute type associated with the machine. These unique attribute types may include the machine identifier, model, brand, year of manufacture, machine class, machine section identifier, and so on. Model training

[0074]

number

[0075] In hyperbolic models, some exemplary embodiments use the Riemann-Adam optimizer. In baseline cases, the standard PyTorch 1.10 Adam optimizer may be used. In other cases, 10 -4 This learning rate can be used as the default parameter, and each system can be trained over 1000 epochs, with checkpoints every 25 epochs. Scoring

[0076]

number

[0077]

number

[0078]

number

[0079] Figure 7 is a block diagram of a training configuration 700 of an exemplary anomaly detection system that uses an acoustic signal for anomaly detection, according to several exemplary embodiments. The sound signal 702 is decomposed into frames and fed into the embedding neural network 720, which creates Euclidean embeddings. These Euclidean embeddings are transformed by a transformer 723 to obtain a mapping of the Euclidean embeddings in hyperbolic space. The hyperbolic space is shown as a Poincaré sphere 725. For example, considering three sources of sound contributing to the sound signal 702, the corresponding embeddings (dark, shaded, and unshaded) for each source type are shown in hyperbolic space. That is, in the training stage shown in Figure 7, sounds from three different types of machines that function correctly are shown separately. However, the number of sources available for learning may be fewer or more. For each frame of sound (a frame generally has a fixed duration), the embedding neural network 720 generates an embedding (i.e., a vector of numbers) in Euclidean space. Using a Euclidean-hyperbolic transformer, the embeddings are ultimately positioned within hyperbolic space. The hyperbolic embeddings are then fed into a classifier neural network 730, which finds geometric boundaries 737 between different types of machines, as shown within hyperbolic space 735. In this case, the classifier is a "linear" classifier in hyperbolic space. Due to the geometric properties of the space, "linear" boundaries (i.e., hyperplanes) are arcs rather than lines, as shown, for example, as 737 in Figure 7. Each of the embeddings 732, 734, and 737 belongs to a unique source type. Various examples of the training process can be presented, optimizing both the embedding network and the classifier network so that the classifier learns, as often as accurately as possible, the boundaries where machine types are detected.

[0080] Figure 8A is a block diagram of an exemplary runtime configuration 800 of the trained anomaly detection system of Figure 7, according to some exemplary embodiments. During the inference phase (runtime), a sound 802 from a specific machine is presented to the trained anomaly detection system. For the specific machine, the type and other metadata or attributes may be known in advance, but it is unknown whether the machine is functioning correctly. An embedding for the sound fragment is computed by the embedding neural network 820, and the converter 825 converts the Euclidean embedding into a hyperbolic embedding 827. The hyperbolic embedding is fed into the classifier network 830, which outputs a probability 832 that the specific sound fragment belongs to a different known type of machine with specific metadata or attributes, depending on the embedding's position relative to the trained type boundary in hyperbolic space. In some exemplary embodiments, an anomaly score may be calculated relying solely on the probabilities 832 computed by the classifier network 830, in particular the computed probability 834 that the sound fragment belongs to a known type of machine. Such systems benefit from hyperbolic geometry because it is a more "natural" space for positioning embeddings and for the classifier to learn a good set of hyperplanes. Therefore, the proposed hyperbolic system shown in Figure 8A can 1) generate more reliable anomaly scores, and secondarily, 2) achieve the same performance with fewer dimensions, saving computation time and memory.

[0081] However, to improve the reliability of the anomaly score, some exemplary embodiments generate the anomaly score by aggregating the probabilities 832,834 and / or the inherent geometric aspects of the embedding location in the embedding space. In this regard, some exemplary embodiments recognize that, unlike Euclidean space, the origin (center) of hyperbolic space has inherent meaning. Some embodiments are based on the recognition that, in addition to, or instead of, the location of the hyperbolic embedding relative to the origin of the hyperbolic space, provides further clues relevant to anomaly detection. This is because the classification result depends not only on the data but also on the training of the classifier, while the location relative to the origin depends on the quality of the data and can be independent of the classification. Such an indicator of the quality of the input data can be advantageous because the training is primarily on normal data. In particular, anomaly indicators that depend on the distance to the origin of the embedding space are not available in Euclidean space.

[0082] Referring to Figure 8B, several exemplary embodiments utilize the correlation between the embedding distance 829 to the origin of the Poincaré sphere 827 and the concept of certainty (informed by the training data and classification task), placing data with high "certainty" closer to the edge of hyperbolic space (the periphery of the Poincaré sphere 827). Thus, Figure 8B proposes viewing the geodesic distance 836 of each embedding from the origin as a measure of its anomalousness, with embeddings closer to the origin being considered more anomalous. Accordingly, the anomalousness detection system shown in Figure 8B calculates an anomalousness score 804 using an aggregater 880 to aggregate the geodesic distance 836 of each embedding relative to the origin and the corresponding probability 834 of the embedding calculated by the classifier 830. The higher the score, the more likely the sound fragment is to be anomalous.

[0083] Figure 9 shows the hardware configuration of an exemplary anomaly detection system 900 according to several exemplary embodiments. The anomaly detection system 900 may include a processor 902, a memory 904, an output interface 906, and an input interface 908. The processor 902 may be communicatively coupled to the memory 904, the output interface 906, and the input interface 908. In some embodiments, the output interface 906 may be coupled to a display device.

[0084] The processor 902 may include appropriate logic, circuits, and / or interfaces that can be configured to execute program instructions associated with various operations performed by the anomaly detection system 900. The processor 902 may include one or more specialized processing units and can be implemented as an integrated processor or group of processors that collectively execute the functions of one or more of these specialized processing units. The processor 902 can be implemented based on a number of processor technologies known in the art. Examples of implementations of the processor 902 may be x86-based processors, graphics processing units (GPUs), reduced instruction set computing (RISC) processors, application-specific integrated circuit (ASIC) processors, complex instruction set computing (CISC) processors, microcontrollers, central processing units (CPUs), and / or other computing circuits.

[0085] Memory 904 may include appropriate logic, circuitry, and / or interfaces that can be configured to store program instructions executed by processor 902. Memory 904 may further be configured to store trained neural networks, such as embedded neural networks or embedded classifiers. Without departing from the scope of this disclosure, trained neural networks, such as embedded neural networks and embedded classifiers, may also be stored in a database. Exemplary implementations of memory 904 may include, but are not limited to, random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), hard disk drive (HDD), solid-state drive (SSD), CPU cache, and / or secure digital (SD) card.

[0086] The output interface 906 may include appropriate logic, circuitry, interfaces, and / or code that can be configured to receive inputs and provide outputs based on the received inputs. The output interface 906 may include various input and output devices that can be configured to communicate with the processor 902. For example, the anomaly detection system 900 may receive user input via the input interface 908 to select a region or hyperplane in hyperbolic space. Examples of the input interface 908 include, but are not limited to, a touchscreen, keyboard, mouse, joystick, or microphone.

[0087] The input interface 908 may also include appropriate logic, circuitry, interfaces, and / or code that can be configured to facilitate the processor 902's communication with databases and / or other communication devices via the communication network 910. The input interface 908 can be implemented by using various known techniques for supporting wireless communication of the anomaly detection system 900 via the communication network 910. The input interface 908 may include, for example, an antenna, a radio frequency (RF) transceiver, one or more amplifiers, a tuner, one or more oscillators, a digital signal processor, a coder-decoder (CODEC) chipset, a subscriber identity module (SIM) card, a local buffer circuit, and the like.

[0088] The input interface 908 may be configured to communicate via a communication network 910. The network 910 includes wired or wireless communication with networks such as the Internet, intranet, wireless network, cellular telephone network, wireless local area network (LAN), or metropolitan area network (MAN). Wireless communication may be configured to use one or more of several communication standards, protocols, and technologies, such as Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), Wideband Code Division Multiple Access (W-CDMA), Long Term Evolution (LTE), Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Bluetooth®, Wireless Fidelity (Wi-Fi)® (IEEE 802.11a, IEEE 802.11b, IEEE 802.11g, or IEEE 802.11n, etc.), Voice Over Internet Protocol (VoIP), Light Fidelity (Li-Fi), or Worldwide Interoperability for Microwave Access (Wi-MAX).

[0089] The following description provides only exemplary embodiments and is not intended to limit the scope, applicability, or configuration of the present disclosure. Rather, the following description of exemplary embodiments provides a practical description for realizing one or more exemplary embodiments for those skilled in the art. The subject matter to be considered is a variety of modifications that may be made in terms of the function and arrangement of the elements without departing from the spirit and scope of the subject matter disclosed as described in the appended claims.

[0090] Specific details are given in the following description to ensure a full understanding of the embodiments. However, those skilled in the art will understand that embodiments can be carried out without these specific details. For example, systems, processes, and other elements in the disclosed subject matter may be shown as components in the form of block diagrams so as not to obscure the embodiments with unnecessary details. In other examples, well-known processes, structures, and technologies may be shown without unnecessary details to avoid obscuring the embodiments. Furthermore, similar reference numbers and names in different drawings refer to similar elements.

[0091] Furthermore, individual embodiments may be described as processes shown as flowcharts, flow diagrams, data flow diagrams, structural diagrams, or block diagrams. While flowcharts may describe operations as sequential processes, many of these operations can be performed in parallel or simultaneously. The order of these operations may also be reordered. A process may terminate when its operations are complete, but it may have additional steps that are not discussed or included in the diagrams. Moreover, not all operations in any specifically described process occur in all embodiments. A process may correspond to a method, function, procedure, subroutine, subprogram, etc. If a process corresponds to a function, the termination of the function may correspond to the function returning to the calling function or main function.

[0092] Furthermore, embodiments of the disclosed subject matter may be implemented at least partially manually or automatically. Manual or automatic implementation may be performed, or at least assisted, by using a machine, hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. If implemented with software, firmware, middleware, or microcode, the program code or code segments for performing the required tasks may be stored in a machine-readable medium. A processor may perform the required tasks.

[0093] The various methods or processes outlined herein may be encoded as software executable on one or more processors utilizing any one of a variety of operating systems or platforms. In addition, such software may be written using any of several 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. Generally, the functionality of program modules may be combined or distributed as required in various embodiments. The embodiments described above in this disclosure can be implemented in any of a number of ways. For example, these embodiments may be implemented using hardware, software, or a combination thereof. If implemented in software, the software code can run on any suitable processor or array of processors, whether provided on a single computer or distributed across multiple computers. Such a processor may be implemented as an integrated circuit with one or more processors in an integrated circuit component. However, the processor may be implemented using circuitry of any suitable format.

[0094] Embodiments of the present disclosure may be embodied as the example provided. The actions performed as part of this method may be ordered in any suitable manner. Thus, embodiments may be constructed in which the actions are performed in an order different from the example, which may include performing several actions simultaneously, even if they are shown as sequential actions in the exemplary embodiments.

[0095] While this disclosure has been described with reference to certain preferred embodiments, it should be understood that various other adaptations and modifications may be made within the spirit and scope of this disclosure. Therefore, it is an aspect of the appended claims to cover all such variations and modifications within the true spirit and scope of this disclosure.

Claims

1. An anomaly detection system for detecting anomalies in the operation of a machine based on signals indicating the operation of the machine performing a task, At least one processor, The system comprises a memory in which instructions are stored, and when an instruction is executed by the at least one processor, the anomaly detection system... The machine is made to collect the hyperbolic embedding of the signals indicating the operation of the machine, the hyperbolic embedding is in hyperbolic space, and the instruction is further executed by the at least one processor, the anomaly detection system, Based on the hyperbolic embedding, the detection of the abnormality in the operation of the machine is performed, and an abnormality score is determined. An anomaly detection system that performs the rendering of the aforementioned anomaly score.

2. The aforementioned processor further, The measured values ​​of the signal or features extracted from the measured values ​​are processed using a neural network to create a Euclidean embedding of the signal into Euclidean space. The anomaly detection system according to claim 1, configured to project the Euclidean embedding onto the hyperbolic space to create the hyperbolic embedding.

3. The aforementioned processor further, The aforementioned signal is processed using an embedded neural network to create the hyperbolic embedding of the signal. The anomaly detection system according to claim 1, configured to process the hyperbolic embedding using a classification neural network to generate at least a portion of the anomaly score.

4. The anomaly detection system according to claim 3, wherein the classification neural network is trained using training data generated from the normal operation of the machine, and the normal operation of the machine is defined against the machine's standard operational performance data.

5. The anomaly detection system according to claim 3, wherein the embedded neural network is trained jointly with the classification neural network such that the weights of the embedded neural network and the weights of the classification neural network are interdependent of each other.

6. The anomaly detection system according to claim 3, wherein the anomaly score is a function of a combination of the distance between the hyperbolic embedding and the origin of the hyperbolic space and the positive classification rate returned by the classification neural network.

7. The anomaly detection system according to claim 1, wherein the hyperbolic space is approximated by a projection onto one of the Poincaré spheres or Poincaré disks, and the anomaly score is a function of the distance between one of the hyperbolic embeddings and one of the origins of the Poincaré spheres or Poincaré disks.

8. The anomaly detection system according to claim 1, wherein each hyperbolic embedding corresponds to a vector indicating a unique attribute type associated with the machine.

9. The anomaly detection system according to claim 1, wherein the signal indicating the operation of the machine is an audio signal created in the operating environment of the machine, the source of the audio signal and the operating environment are characterized by a set of attributes, and each hyperbolic embedding corresponds to a vector indicating a unique attribute from the set of attributes.

10. The aforementioned processor further, The signal is divided into multiple segments, The anomaly detection system according to claim 1, configured to generate one of the hyperbolic embeddings for each of the plurality of segments.

11. In order to determine the abnormal score, the processor further: For each hyperbolic embedding, calculate the probability that the corresponding segment belongs to a specific attribute type. The calculated probabilities are aggregated, An anomaly detection system according to claim 10, configured to determine the anomaly score based on the aggregated probabilities.

12. The anomaly detection system according to claim 1, wherein the signal includes one of an acoustic signal and a video signal or a combination thereof.

13. The anomaly detection system according to claim 1, wherein the signal includes a measurement of vibration of the machine caused by the operation of the machine.

14. The anomaly detection system according to claim 1, wherein the signal includes a measured value of one or a combination of the voltage and current used to control the machine, and the torque generated by the machine.

15. A control system operably connected to the anomaly detection system described in claim 1, wherein the control system is configured to control the operation of the machine based on the anomaly score rendered by the anomaly detection system.

16. The control system according to claim 15, wherein the control system is configured to change the mode of operation of the machine based on the abnormal score.

17. The anomaly detection system according to claim 1, further comprising a display device configured to display the rendered anomaly score.

18. A computer-implemented method for detecting an anomaly in the operation of a machine based on signals indicating the operation of the machine performing a task, The method includes the step of collecting a hyperbolic embedding of the signal indicating the operation of the machine, wherein the hyperbolic embedding lies in hyperbolic space, and the method further includes The steps include: performing the detection of the abnormality in the operation of the machine based on the hyperbolic embedding and determining an abnormality score; A computer-implemented method comprising the step of rendering the aforementioned anomaly score.

19. The steps include dividing the signal into multiple segments, For each of the multiple segments, the step of generating one hyperbolic embedding from among the hyperbolic embeddings, For each hyperbolic embedding, the step of calculating the probability that the corresponding segment belongs to a specific attribute type, The steps include: aggregating the calculated probabilities, The method according to claim 18, further comprising the step of determining the anomaly score based on the aggregated probabilities.

20. A non-temporary computer-readable storage medium containing a processor-executable program for performing a method for detecting an anomaly in the operation of a machine based on signals indicating the operation of the machine performing a task, wherein the method is: The method includes the step of collecting a hyperbolic embedding of the signal indicating the operation of the machine, wherein the hyperbolic embedding lies in hyperbolic space, and the method further includes The steps include: performing the detection of the abnormality in the operation of the machine based on the hyperbolic embedding and determining an abnormality score; A non-temporary computer-readable storage medium, comprising the step of rendering the aforementioned abnormal score.

Images