Equipment anomaly detection method, computer-readable storage medium, chip and device

A neural network cascade system learns normal conditions to reconstruct sensor signals, detecting anomalies by comparing differences, enabling early prediction and prevention of machine failures.

JP7580394B2Active Publication Date: 2024-11-11CHENGDU SHENZHEN TECH CO LTD
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
JP2021560351
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2021-01-25
Publication Date
2024-11-11
Estimated Expiration
2041-01-25

AI Technical Summary

Technical Problem

Industrial and consumer machines face challenges in obtaining sufficient data to detect abnormal conditions such as machine malfunctions, making it difficult to predict these conditions effectively.

Method used

A neural network cascade system that learns normal sensor event streams to reconstruct monitored signals, comparing the difference between actual and reconstructed streams to detect anomalies by exceeding a threshold, triggering an alarm when necessary.

Benefits of technology

The system can predict abnormal operating conditions in advance, allowing for timely intervention to prevent system failures and reduce potential losses.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007580394000065
    Figure 0007580394000065
  • Figure 0007580394000066
    Figure 0007580394000066
  • Figure 0007580394000067
    Figure 0007580394000067
Patent Text Reader

Abstract

An apparatus for detecting an abnormality in equipment is provided. [Solution] The device uses an unsupervised or self-supervised neural network to learn the normal conditions of a target system (e.g., equipment or machinery). The trained neural network can reconstruct N normal sensor event streams of the target system as neural network-reconstructed versions of N normal sensor event streams under the target system's normal conditions. The equipment anomaly detection device can analyze and predict the operating status of the target system based on the neural network-reconstructed versions if a certain threshold is exceeded. When the difference between the N normal sensor event streams and the signals of the neural network-reconstructed versions exceeds a certain threshold, the equipment anomaly detection device issues an alarm signal reflecting the abnormal operating status of the target system.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to the field of computer technology, and in particular to a neural network-based equipment anomaly detection method, a computer-readable storage medium, a chip and an apparatus. [Background technology]

[0002] Industrial and domestic machines usually use motors to drive their operating mechanisms and functions. Motors and machine parts can age and wear out. Accumulated deformation of these parts can cause the machine to suddenly break down or even cause unexpected accidents and losses. Summary of the Invention [Problem to be solved by the invention]

[0003] It is difficult for end users and consumers of industrial and consumer machinery to obtain sufficient data to identify abnormal conditions, such as operational anomalies, errors or failure modes in the machinery, or to predict these abnormal conditions if data is insufficient. [Means for solving the problem]

[0004] A first embodiment of the present invention provides an equipment anomaly detection device, a cascade of neural networks for receiving and learning from N normal sensor event streams in a learning state to form a learned state of the cascade of neural networks, wherein the cascade of neural networks in the learned state receives N monitored sensor event streams as inputs from the N event stream input connections in an operating state of the cascade of neural networks, and outputs N monitored reconstructed event streams as an output of the cascade of neural networks based on the N monitored sensor event streams, the N monitored reconstructed event streams representing a reconstructed version of the N monitored sensor event streams; a detection module for calculating a difference between the N monitored sensor event streams and the N monitored reconstructed event streams, the difference between the N monitored sensor event streams and the N monitored reconstructed event streams being: threshold and a detection module that sends out a trigger signal when the threshold voltage Vcc exceeds the threshold voltage Vcc.

[0005] A second aspect of the present invention provides an anomaly detection method, comprising: using the cascade of neural networks in a training state to learn from N normal sensor event streams at N event stream input connections to form a trained state of the cascade of neural networks; using the cascade of neural networks in a trained state to receive N monitored sensor event streams as inputs from the N event stream input connections in an operational state of the cascade of neural networks and to output N monitored reconstructed event streams as an output of the cascade of neural networks based on the N monitored sensor event streams, the N monitored reconstructed event streams representing a reconstructed version of the N monitored sensor event streams; calculating, using a detection module, the difference between the N monitored sensor event streams and the N monitored reconstructed event streams; The difference between the N monitored sensor event streams and the N monitored reconstructed event streams is threshold and when the threshold voltage is exceeded, the detection module sends out a trigger signal.

[0006] The invented methods may be implemented in a chip, which may include a processor arranged to read and execute a computer program stored in a memory, thereby causing a device in which the chip is implemented to perform the inventive methods.

[0007] The methods of the present invention can be programmed as computer-executable instructions stored in a non-volatile computer-readable medium that, when loaded into a computer, directs the computer's processor to execute the methods.

[0008] The non-volatile computer-readable medium may include at least one of a hard disk, an optical disk (CD-ROM), an optical storage device, a magnetic storage device, a ROM (Read Only Memory), a PROM (Programmable Read Only Memory), an EPROM (Erasable Programmable Read Only Memory), an EEPROM (Electrically Erasable Programmable Read Only Memory), and a flash memory.

[0009] The methods of the present invention may be programmed as a computer program product capable of instructing a computer to carry out the methods of the present invention.

[0010] The methods of the present invention may be programmed as a computer program capable of instructing a computer to carry out the methods of the present invention. Effect of the Invention

[0011] In the anomaly detection system of the invention, an unsupervised or self-supervised neural network is used to learn the normal conditions of a target system (e.g., equipment or machinery). The trained neural network can reconstruct the sensing signal of the target system as the neural network's reconstructed version of the sensing signal under the normal conditions of the target system. The anomaly detection system can analyze and predict whether the operating condition of the target system exceeds a certain level or threshold based on the neural network reconstructed version. For example, when the signal difference between the N normal sensor event streams and the neural network reconstructed version exceeds a certain threshold, the anomaly detection system can issue an alarm signal reflecting the abnormal operating status of the target system. The alarm signal can be issued ahead of the time when the system failure of the target system actually occurs. In this way, personnel can have more time to deal with the anomaly of the target system to prevent substantial losses from occurring. [Brief description of the drawings]

[0012] In order to more clearly explain the technical methods in the embodiments of the present invention or the prior art, a brief description of the drawings that need to be used in the description of the embodiments or the prior art will be provided below. It should be clear that the drawings in the following description are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings on the premise that they do not perform creative work. [Figure 1] FIG. 1 is a schematic diagram of an anomaly detection system according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic diagram of a signal conversion module according to an embodiment of the present invention. [Diagram 3] FIG. 3 is a schematic diagram of a signal conversion means according to an embodiment of the present invention. [Figure 4] FIG. 4 is a schematic diagram of a signal waveform during signal processing in the signal conversion means according to the embodiment of the present invention. [Diagram 5] FIG. 5 is a schematic diagram of the connection relationship between an input event stream, a decoding network, and a detection module in an anomaly detection system according to an embodiment of the present invention. [Figure 6] FIG. 6 is a schematic diagram of an embedding network for implementing a winner-takes-all (WTA) algorithm according to an embodiment of the present invention. [Figure 7] FIG. 7 is a schematic diagram showing an embedding network for implementing a soft-winner-takes-all (sWTA) algorithm according to another embodiment of the present invention. [Figure 8] FIG. 8 is a schematic diagram illustrating a decoding network according to an embodiment of the present invention. [Figure 9] FIG. 9 is a schematic diagram of a sensing module according to an embodiment of the present invention. [Figure 10] FIG. 10 is a schematic diagram of the operation of the anomaly detection system according to an embodiment of the present invention. [Figure 11] FIG. 11 is a schematic diagram of the anomaly detection system when in a learning state. [Figure 12] FIG. 12 is a schematic diagram of the anomaly detection system in an operating state. [Figure 13] FIG. 13 is a schematic diagram of the Tetzlaff and Jonke learning rules. [Figure 14] FIG. 14 is a schematic diagram of software modules of an anomaly detection system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] The following describes the technical methods in the embodiments of the present invention more clearly and completely with the accompanying drawings, but it is obvious that the described embodiments are only a part of the embodiments of the present invention, and are not all of the embodiments. All other embodiments obtained by those skilled in the art without creative work according to the embodiments of the present invention are all within the scope of protection of the present invention.

[0014] The present invention provides an anomaly detection method and apparatus for detecting anomalies in a time domain signal of a target system (eg, a system, an apparatus, a facility, or a plant) to be monitored.

[0015] In a learning state, the equipment anomaly detection device automatically adapts and learns from normal sensor signals the normal state of a monitored target system and identifies deviations from the normal state as anomalies. The inventive device uses online learning in a coder-decoder network to learn the normal input sensor signal in the learning state and to learn a network reconstructed version that reconstructs the normal input sensor signal. In an operating state, the device detects and sends an anomaly signal associated with the monitored target system and the current input sensor signal by measuring the difference between the current input sensor signal and the network reconstructed version of the current input sensor signal.

[0016] When the current network reconfiguration version and the current input sensor signal match, the equipment anomaly detection device determines that the monitored target system is in a normal state. When the difference between the current input sensor signal and the network reconfiguration version is small, e.g., does not exceed a threshold, the equipment anomaly detection device does not generate an anomaly signal.

[0017] When a monitored target system experiences an abnormal condition, the network is unable to accurately reconstruct the input sensor signals. The device can determine a trend in the monitored target system where the difference between the current input sensor signals and the network reconstructed version becomes large, e.g., exceeds a threshold. When the difference exceeds a threshold, the device generates and sends an abnormality signal.

[0018] System Overview: Referring to Figure 1, one embodiment of the present invention provides an anomaly detection system 100. A monitored target system 901 may include one or more entities, such as one or more industrial plants, motors, equipment, facilities, or systems. The monitored target system 901 emits a vibration signal 902 that is indicative of a condition of the monitored target system.

[0019] The sensor system may include sensors 111-1 to 111-N for detecting a target system 901 to be monitored. N is a positive integer and may be two or more. For example, the sensors in the sensor system may be connected to the target system 901 to be monitored or fixed. Each of the sensors 111-1 to 111-N detects the target system to be monitored and generates a sensor signal stream representing a vibration signal of the target system to be monitored. The vibration signal 902 is detected by the sensor system 110 as a sensor signal 112 and converted by the signal conversion module 120 into a sequence of asynchronous input event streams 190. The sensor signals 112 may be transmitted by a wired or wireless connection. The embedding network 130 generates an embedding type encoded stream 139 based on the input event stream 190 and outputs it to the decoding network 140. The embedding network 130 and the decoding network 140 form a cascade of neural networks, which are trained to reconstruct a reconstructed event stream 149 as a network-reconstructed version of the input event stream 190. The detection module 150 compares the reconstructed event stream 149 with the input event stream 190 and calculates the difference between the reconstructed event stream 149 and the input event stream 190. If the difference is less than a threshold, indicating that no anomaly has been detected, the detection module 150 does not raise the trigger signal 180 accordingly. If the difference is greater than a threshold, indicating that an anomaly has been detected, the detection module 150 raises the trigger signal 180 accordingly.

[0020] An external system 200 can send signals 210 to the anomaly detection system 100 to reset, configure and control the monitored target system 901 and the system 100. The external system 200 receives the trigger signal 180 and detection condition 170 information related to the detected anomaly. The trigger signal 180 and detection condition 170 may be sent via a wired or wireless connection.

[0021] A learning configuration and control module 160 switches the system 100 between a learning state and a trained state.

[0022] The equipment anomaly detection device 101 includes the embedding network 130, the decoding network 140, the detection module 150, and a learning setting and control module 160. In one embodiment, the equipment anomaly detection device 101 may be packaged as an integrated circuit (IC), such as a single system on a chip (SOC). In another embodiment, the equipment anomaly detection device 101 may include a signal conversion module 120.

[0023] In one embodiment, the embedding network 130, the decoding network 140, the detection module 150 and the learning setting and control module 160 may be implemented as computer-executable instructions that, when executed by a processor, perform the functions of the present invention.

[0024] Signal conversion module functions: Please refer to Figures 1 and 2. Sensors 111-1 to 111-N transmit a plurality of sensor signals 112-1, 112-2, ... and 112-N to signal conversion means 122-1, 122-2, ... and 122-N in the signal conversion module 120. Each input sensor signal 112-1, 112-2, ... and 112-N is converted into one of sensor event streams 190-1, 190-2, ... and 190-N by one of the signal conversion means 122-1, 122-2, ... and 122-N. Specifically, the input sensor signal 112 includes sensor signals 112-1, 112-2, ... and 112-N. The input event stream 190 includes sensor event streams 190-1, 190-2, ... and 190-N. The signal converter 122-1 generates the sensor event stream 190-1 from the input sensor signal 112-1. Similarly, the signal converter 122-2 generates the sensor event stream 190-2 from the input sensor signal 112-2. The signal converter 122-N generates the sensor event stream 190-N from the input sensor signal 112-N.

[0025] Each of the signal conversion means 122-1 to 122-N in the signal conversion module 120 performs the following functions, and an example of an input signal passing through the signal conversion means is shown below.

[0026] 3, 4 and 5. A sensor in the sensor system 110 sends an input signal 1220 to the signal conversion means 122. The signal conversion means 122 may be a specific embodiment of each of the signal conversion means 122-1, 122-2, ... and 122-N. The input signal 1220 represents one of the sensor signals 112-1, 112-2, ... and 112-N, and an event stream 1225 represents one of the sensor event streams 190-1, 190-2, ... and 190-N.

[0027] A bandpass filter 1221 with a selected center frequency and bandwidth is used to select frequency components from the input signal 1220 and output the selected frequency components as filtered input signal 1221a. The input signal 1220 is sent to several bandpass filters, which splits the input signal 1220 into several frequency components.

[0028] The band-pass filtered signal 1221a is rectified and smoothed by a low-pass filter 1223. For example, the low-pass filter 1223 includes a rectifier 1223-1 and a signal smoothing means 1223-2. The signal smoothing means 1223-2 can be composed of one or more capacitors. The rectifier 1223-1 generates a rectified signal 1223a from the filtered input signal 1221a. The signal smoothing means 1223-2 generates a smoothed rectified signal 1223b from the rectified signal 1223a.

[0029] The smoothed rectified signal 1223b is encoded using pulse density modulation by a PDM (pulse-density modulation) encoder 1224. For example, the PDM encoder 1224 can be a sigma-delta encoder. The PDM encoder 1224 generates the encoded smoothed rectified signal 1223b as an event stream 1225. Each pulse (spike) in the event stream 1225 is sent out as an event.

[0030] Network structure and function: The signal transformation module 120 provides a plurality of input event streams 190 and encodes the input sensor signals into a plurality of event encoded signals in the selected frequency bands.

[0031] The embedding network 130 represents high-dimensional signals, such as, for example, N input event streams, in a low dimension, such as, for example, M dimensions, where M is a positive integer and M < N. The embedding network 130 reliably maps a specific mode of the input event stream 190 as a specific mode on the embedding-type encoded stream 139 at any predetermined timing. This is based on the assumption that the input event stream 190 to be monitored exists in a small number of sub-states in the normal operating state of the target system to be monitored. Such sub-states represent the signal to be monitored and the normal state of the target system to be monitored. By changing the magnitude of M in the embedding network 130, the number of possible normal sub-states can be selected.

[0032] Embedding network: In the present invention, at least two embodiments of the embedding network 130 are presented.

[0033] In one embodiment, the embedding network 130 includes a plurality of embedding neurons for implementing an online learning algorithm. In one embodiment, the embedding network 130 is a spiking neural network that includes a plurality of spiking neurons for implementing an online learning algorithm. One example of the spiking neural network presented is basically an autoencoder composed of an input layer, a hidden layer, and an output layer. The main differences from a conventional autoencoder are as follows.

[0034] The embedding network 130 is trained through unsupervised learning. The training data is assumed to be a subset of the complete data set that does not contain any anomalies. The embodiment of the present invention allows the encoder, i.e., the embedding network 130, to learn other representations of the normal input signal and the decoder, i.e., the decoding network 140, to accurately reconstruct the normal input signal from these representations. The basic principle behind this is that the autoencoder cannot learn to reconstruct an abnormal signal forever. Thus, in normal situations, the reconstruction loss is small, but when an abnormal signal is present, this loss becomes large because the autoencoder never learns to reconstruct the abnormal signal. This training process allows the detection of abnormal signals in situations where the abnormal signal is not defined.

[0035] The dimension of the output layer is twice that of the input layer, while the output layer and the input layer in a conventional autoencoder have the same dimension. This is because the reconstruction loss is the absolute difference between the input and the output. A single spiking neuron cannot perform absolute value operations, but a pair of spiking neurons can.

[0036] Referring to FIG. 6, in the embedding network 130, an input connection unit 131 having an online learning function connects input event streams 190-1 to 190-N and a set of weights W I and provide signals to a set of embedding neurons 135. y(t)=W I·x(t), where y(t) is the neuron activity of the embedding neuron 135 and x(t) is the neuron activity of the input connection unit 131. That is, the neuron activity y(t) represents the output of the embedding network and the neuron activity x(t) represents the input of the embedding network. The neuron activity y(t) of the embedding neuron 135 becomes the embedding type coded stream 139, which includes the embedding type coded streams 139-1 to 139-M.

[0037] The learning setting and control module 160 uses online learning rules to I An example of the online learning rule is disclosed in the paper "Synaptic scaling enables dynamically distinct short- and long-term memory formation" by Tetzlaff C., Kolodziejski, C., Timme, M., Tsodyks, M., and Worgotter, F., published in 2013, BMC Neuroscience 14(1), p. 415. For example, the learning setting and control module 160 applies a rule based on Tetzlaff. The rule is that the weights W I has been updated to read as follows:

[0038]

number

[0039] In one embodiment, the non-learning inhibitory connections between all embedding neurons inhibit the neuronal activity of all other neurons when a single embedding neuron is activated. This encourages the association of a particular mode of input events with the neuronal activity of a particular embedding neuron. That is, the embedding network 130 uses the lateral inhibitory connections 136a between the embedding neurons 135 of the embedding network 130 to perform a first type of WTA algorithm, for example a conventional winner-takes-all (WTA) algorithm. The lateral inhibitory connections 136 between neurons include lateral inhibitory connections 136-2 to 136-M.

[0040] The learning rule was disclosed in the paper “Feedback inhibition shapes emergent computational properties of cortical microcircuit motifs” by Jonke, Z., Legenstein, R., Habenschuss, S. and Maass, W., published in the Journal of Neuroscience 37(35), 8511-8523 in 2017.

[0041] 7. In another embodiment, the learning configuration and control module 160 employs a similar method, but now uses a Jonke-based learning rule to learn the weights W I Under this rule, the learning configuration and control module 160 weights W I has been updated to the following:

[0042]

number

[0043] In another embodiment, the embedding network 130 uses a set of inhibitory neurons 136b instead of direct inhibitory connections between the embedding neurons 135. The inhibitory neurons receive inputs from all embedding neurons 135 and send inhibitory outputs to all embedding neurons 135. Such a network design allows multiple embedding neurons to perform neuronal activity simultaneously and increases the number of embedding modes that the embedding network can learn for a given number of embedding neurons. That is, the embedding network 130 uses the set of inhibitory connections between the embedding neurons 135 of the embedding network 130 to perform a second type of WTA algorithm, such as the sWTA algorithm.

[0044] In the learning state of the system 100, the training of the embedding network 130 is activated. In the operating state of the system 100, the training of the embedding network 130 is inhibited.

[0045] Decoding network: 8, the decoding network 14 is responsible for predicting an input signal based on the embedded coded stream 139. The decoding network 140 filters the M normal embedded coded streams to generate filtered versions of the M normal embedded coded streams.

number

number

number

number

number

[0046] What can be clearly stated is that the embedding type coded stream 139 including the embedding type coded streams 139-1 to 139-M is filtered using the low-pass exponential filters 141-1 to 141-M included in the low-pass exponential filter 141, and the input event stream 190 including the input event streams 190-1 to 190-N is filtered using the low-pass exponential filters 142-1 to 142-N of the low-pass exponential filter 142. The embedded type coded stream after filtering is

number

number

[0047] The decoding network 140 uses self-supervised learning weights W O The decoding network 140 uses the neurons 145 to generate a reconstructed event stream 149 from the filtered embedded encoded stream 139.

[0048]

number

number

number

number

number

[0049] For example, in a self-supervised learning process, weights W are calculated using a ridge regression process or some other optimization method. O is adjusted to minimize the error E. In the training state of the system 100, training of the decoding network 140 is activated. In the operating state of the system 100, training of the decoding network 140 is inhibited.

[0050] Detection Modules: 9, the detection module 150 compares the reconstructed event stream 149 with the input event stream 190 after filtering through an exponential low-pass filter 152. The exponential low-pass filter 152 includes exponential low-pass filters 152-1 to 152-N. The detection module 150 includes a subtraction calculator 155, a summation calculator 156, and a comparator 157. The subtraction calculator 155 includes a plurality of subtractors 155-1 to 155-N. Each of the subtractors 155-1 to 155-N calculates an absolute difference value

number

number

number

[0051] The neuronal activity of the subtractor 155, the adder 156 and the input event stream 190 is collected as sensed condition information 170, which can be requested by the external system 200. The connections between the modules 120, 130, 140, 150 and 160 can be realized as wires, buses, cables or optical fibers.

[0052] The external system 200 can be an on-site server co-located with the monitored target system 901 or a remote cloud server separate from the monitored target system 901. The external system 200 can further include a portable device, such as a mobile phone or a handheld device, and a server.

[0053] The neural network in the present invention can be realized by a field programmable gate array (FPGA), an application-specific integrated circuit (ASIC), a tensor processing unit (TPU), a graphics processing unit (GPU), or a multi-core central processing unit (CPU).

[0054] System features: With reference to FIG. 10, the process of adapting and using the anomaly detection system 100 to monitor a signal is described in detail below.

[0055] In step S102, the system 100 is set up and switched to a learning state to receive a set of vibration signals 902 from the monitored target system 901 when the monitored target system 901 is operating in a normal operating state during step S103.

[0056] In step S104, the system 100 is reset to assign random weights W I , and the decoding network 140 has random weights W O , and start training the embedding network 130 and the decoding network 140. Disable trigger signal detection.

[0057] In step S105, the target system to be monitored is allowed to operate in a known normal operating state, which may include multiple normal operating sub-states. In step S106, the system 100 adjusts the weights W based on the learning rule during the learning state. I and weight W O The system 100 adapts and learns from the normal state by adjusting the detection module 104. When the current detection module does not indicate any abnormality in any of the normal operating sub-states of the monitored target system, the adaptation of the system 100 is also completed as a learning process. The system 100 may be configured as a duration to control the adaptation process.

[0058] In step S108, after the current learning of the system 100 is completed, the system 100 switches from the learning state to the trained state, in which the embedding network 130 and the decoding network 140 also transition from the learning state to the trained state, respectively. I and the weights W of the decoding network 140 Ois updated to the training state and fixed to the trained state. The trained state of the embedding network 130 is called the trained embedding network state of the embedding network 130. The trained state of the decoding network 140 is called the trained decoding network state of the decoding network 140.

[0059] In step S110, as described above, the system 100 in the operating state and the embedding network 130 and the decoding network 140 in the trained state are both continuously operating. The system 100 judges whether an anomaly is detected in the target system being monitored. The detection module 150 is configured to raise a detection trigger signal 180 when an anomaly is detected for a desired period of time (e.g., one minute).

[0060] In step S109, the target system 901 to be monitored operates in an operational state under the monitoring of the system 100.

[0061] In step S112, when an anomaly is detected in the monitored target system 901, the detection module 150 sends a trigger signal 180 to the external system 200, which can, for example, awaken the external system 200. The external system 200 can send a control signal 210 to the system 100 to request detailed information 170 about the anomaly. The detected status information 170 is then sent to the external system 200. The trigger signal 180, the control signal 210 and the detected condition information 170 can be transmitted via a wired or wireless connection.

[0062] Example of system behavior in learning state: 11 and 12. The learning configuration and control module 160 switches the system 100 between a learning state and a trained state. An example of the vibration signal 902, the input sensor signal 112, the input event stream 190, the embedded encoding stream 139 and the reconstructed event stream 149 is shown in Table 1 below.

[0063] [Table 1]

[0064] In a learning state, the system 100 processes the normal vibration signal 902a, the N normal sensor signal streams 112a, the N normal sensor event streams 190a, the M normal embedded encoded streams 139a, and the N normal reconstructed event streams 149a as the vibration signal 902, the input sensor signal 112, the input event stream 190, the embedded encoded stream 139, and the reconstructed event stream 149, respectively. In a learned state, the system 100 processes the monitored vibration signal 902b, the N monitored sensor signal streams 112b, the N monitored sensor event streams 190b, the M monitored embedded encoded streams 139b, and the N monitored reconstructed event streams 149b as the vibration signal 902, the input sensor signal 112, the input event stream 190, the embedded encoded stream 139, and the reconstructed event stream 149, respectively.

[0065] 11, the sensor system 110 generates N normal sensor signal streams 112a from a normal vibration signal 902a. A signal conversion module 120 for receiving, filtering, rectifying, smoothing and A / D converting converts the N normal sensor signal streams 112a into N normal sensor event streams 190a.

[0066] In the learning state of the anomaly detection system 100, the embedding network 130 and the decoding network 140 form a cascade of neural networks, which receives and learns from N normal sensor event streams 190a from the N event stream input connection parts, and is used to form a learned state of the cascade of neural networks.

[0067] The embedding network 130 is a spiking neural network that receives N normal sensor event streams 190a as inputs in a training state of the embedding network 130 from among N event stream input connections, and outputs M normal embedded coded streams 139a as outputs of the embedding network 130 based on the N normal sensor event streams 190a, where N is an integer greater than M, and the embedding network 130 learns from the embedding network 130 inputs to generate the embedding network 130 outputs, and uses spike-timing dependent plasticity (STDP) to form a trained embedding network state of the embedding network 130. The embedding network 130 calculates weights W based on a learning rule related to equation (1) or (2). I , so that it can learn from the N normal sensor event streams 190a.

[0068] The embedding network 130 learns from the input of the embedding network 130 to generate the output of the embedding network 130, and uses STDP with a symmetric learning window or STDP with an asymmetric learning window to form a trained embedding network state of the embedding network 130. Figure 13 shows the Tetzlaff learning rule as an example of a symmetric learning window and the Jonke learning rule as an example of an asymmetric learning window.

[0069] The decoding network 140 is a neural network for receiving M normal embedding-type encoded streams 139a as inputs of the decoding network 140 in a learning state, and outputting N normal reconstructed event streams 149a as outputs of the decoding network 140 based on the M normal embedding-type encoded streams 139a. The decoding network 140 learns from among the inputs of the decoding network 140 to generate the output of the decoding network 140, and forms a trained decoding network state of the decoding network 140 based on a rule, thereby reducing the difference between the N normal reconstructed event streams 149a and the N normal sensor event streams 190a. The decoding network 140 calculates weights W based on a learning rule related to equation (3): O can be updated and learned from the M successful embedded encoded streams 139a.

[0070] The decoding network 140 filters the M normal embedded encoded streams 139a to generate filtered versions of the M normal embedded encoded streams 139a.

number

number

number

number

number

number

[0071] An example of system behavior in a running state: 12, the sensor system 110 generates N monitored sensor signal streams 112b from a monitored vibration signal 902b. The signal conversion module 120 receives, filters, rectifies, smooths, and A / D converts the N monitored sensor signal streams 112b to form N monitored sensor event streams 190b.

[0072] In an operational state, the cascade of neural networks in a trained state receives N target sensor event streams 190b as inputs of the cascade of neural networks from N event stream input connections and outputs N target reconstructed event streams 149b as outputs of the cascade of neural networks based on the N target sensor event streams 190b. The N target reconstructed event streams 149b represent a reconstructed version of the N target sensor event streams 190b.

[0073] The embedding network 130 in the trained embedding network state outputs an embedded encoded stream 139b of M monitored objects based on the N monitored sensor event streams 190b.

[0074] The decoding network 140 in the trained decoding network state outputs N monitored reconstructed event streams 149b based on M monitored embedded encoded streams 139b.

[0075] The detection module 150 calculates the difference between the N monitored sensor event streams 190b and the N monitored reconstructed event streams 149b, and sends a detection trigger signal 180 when the difference between the N monitored sensor event streams 190b and the N monitored reconstructed event streams 149b exceeds a threshold.

[0076] In the trained decoding network state, the decoding network 140 filters the M monitored embedded-type encoded streams 139b to generate filtered versions of the M monitored embedded-type encoded streams 139b.

number

number

number

number

number

[0077] The detection module 150 filters the N monitored sensor event streams 190b to generate a filtered version of the N monitored sensor event streams 190b.

number

[0078] The subtraction calculator 155 includes a plurality of subtractors, selected from which a filtered version of the N monitored sensor event streams 190b is calculated.

number

number

number

number

[0079] The sum calculator 156 calculates the filtered version

number

number

[0080] The comparator 157 compares the sum with a signal difference threshold and outputs an alarm signal 180 indicating an anomaly in the N monitored sensor event streams 190b when the sum exceeds the signal difference threshold.

[0081] Please refer to FIG. 14. In one embodiment, the embedding network 130, the decoding network 140, the detection module 150 and the learning configuration and control module 160 can be implemented in the embedding network module 130a, the decoding network module 140a, the detection module 150a and the learning configuration and control module 160a as computer executable commands. The computer executable commands in the modules 130a, 140a, 150a and 160a, when loaded from the storage device 105 into the main memory 104 and executed by the CPU 102, perform the functions of the embedding network 130, the decoding network 140, the detection module 150 and the learning configuration and control module 160 in the invention. The CPU 102 uses the neural processing unit (NPU) 103 to perform the functions of the embedding network module 130a and the decoding network module 140a. The neural processing unit 103 can be implemented by an FPGA, an ASIC, a TPU or a GPU.

[0082] Those skilled in the art can understand that the description in the embodiments of the present invention and each module, module, means (unit), algorithm and process of the invention can be realized by electronic hardware or a combination of computer and electronic hardware software. Whether a function is implemented by hardware or software is determined by the application conditions and design requirements of the technical method. Those skilled in the art can adopt different forms to realize each specific application, but at the same time, such realization forms do not exceed the scope of the present invention. As those skilled in the art can understand, the operation processes of the above systems, devices and means are basically the same, so reference can be made to the operation processes of the systems, devices and means in the above embodiments. These operation processes are not described in detail for the sake of description and simplification.

[0083] It is understood that the system, device, and method of the invention in the embodiments of the present invention can be realized in other forms. The above embodiments are merely exemplary. The division of the means is merely a division based on logical functions, and other division forms may exist when realized. A plurality of means or components may be combined or integrated into other systems. Also, some features may be omitted or not considered. Meanwhile, the mutual coupling, direct coupling, or communicative coupling shown or discussed may operate in an indirect or communicative manner through some ports, devices, or means, electrically, mechanically, or in other forms.

[0084] The means of the described partitioning members may be physically separated means or not. The display means may be, but is not limited to, physical means located in one place or on multiple network means. Depending on the purpose of the embodiment, some or all of the means may be used. In addition, each functional means in each embodiment may be integrated into one processing means, may be physically independent, or may be two or more means integrated into one processing means.

[0085] If the software functionality is implemented and used and sold as a product, it can be stored in a computer readable medium. Based on this understanding, the technical methodology presented in the present invention can be implemented primarily or partially in the form of software. Or, a part of the technical methodology that is advantageous over the prior art can be implemented in the form of software. The software product in a computer stored in a storage medium includes a plurality of commands that are used in a computing device (such as a PC, a server, or a network device) to execute all or part of the steps of the invention in the embodiment of the present invention. The storage medium includes a universal serial bus (USB) hard drive, a portable memory, a read-only memory (ROM), a random access memory (RAM), a floppy disk, or other types of media that can store program code.

[0086] The equipment anomaly detection system of the invention uses an unsupervised or self-supervised neural network to learn the normal conditions of a target system (e.g., equipment or machine). The trained neural network can reconstruct the sensing signal of the target system as a neural network reconstructed version of the sensing signal under the normal conditions of the target system. The equipment anomaly detection device can analyze and predict whether the operating condition of the target system exceeds a certain threshold based on the neural network reconstructed version. When the signal difference between the N normal sensor event streams and the neural network reconstructed version exceeds a certain threshold, the equipment anomaly detection system can issue an alarm signal reflecting the abnormal operating status of the target system. The alarm signal can be issued ahead of the time when a system failure of the target system actually occurs.

[0087] The present invention has been described in conjunction with the preferred embodiment which is recognized to be the most practical, but it should be understood that the present invention is not limited to the embodiment of the invention, but is intended to cover various arrangements that may be made, without departing from the scope of the appended claims in their broadest interpretation.

[0088] The above describes the information collection method, device, system and storage device provided by the embodiments of the present invention in detail, and the principle and embodiment of the present invention are described in the present text by applying specific examples. However, the above description of the embodiments is merely used to help understand the method of the present invention and its core idea. At the same time, those skilled in the art may make changes in the specific embodiment and application scope based on the technical idea of ​​the present invention. In summary, the contents of this specification should not be understood as limitations on the present invention. [Explanation of symbols]

[0089] 100...Anomaly detection system 101...Equipment abnormality detection device 102...CPU 103...Neural Processing Unit 104...Main memory 105...Storage device 110...Sensor system 111-1~111-N...Sensors 112-1~112-N...Sensor signal 112a, 112b...sensor signal stream 120...Signal conversion module 122-1 to 122-N...signal conversion means 130...Embedding Network 131...Input connection 135... Embedding neurons 136-2~136-M...Inhibitory connection 139-1~139-M...Embedded encoded stream 140...Decoding network 141-1~141-M, 142, 142-1~142-N...Low-pass exponential filters 145...Neurons 150...Detection module 152-1~152-N...Exponential low-pass filters 155...Subtraction Calculator 155-1~155-N...Subtractor 156...Wash calculator 157...Comparator 160...Learning setting and control module 170...Detection conditions, condition information, status information, detailed information 180...Trigger signal 190-1~190-N, 1225...Event stream 200...External system 210...Control signal 901...Target System 902...Vibration signal 1220, 1221a...Input signal 1221...Bandpass filter 1223...Low pass filter 1223-1... Rectifier 1223a... rectified signal 1223b...Smoothed rectified signal 1224...PDM Encoder

Claims

1. An equipment abnormality detection device, a cascade of neural networks including a learning state and a trained state, in which in the learning state, N normal sensor event streams are received from N event stream input connections based on online learning and learned to form a trained state of the cascade of neural networks, each pulse of the event stream is sent out as an event, in the trained state, the cascade of neural networks receives N monitored sensor event streams as inputs from the N event stream input connections in an operating state of the cascade of neural networks, and outputs N monitored reconstructed event streams as an output of the cascade of neural networks based on the N monitored sensor event streams, the N monitored reconstructed event streams representing a reconstructed version of the N monitored sensor event streams; a detection module for calculating a difference between the N monitored sensor event streams and the N monitored reconstructed event streams, the detection module sending a trigger signal when a difference between the N monitored sensor event streams and the N monitored reconstructed event streams exceeds a threshold; In the equipment anomaly detection device, the cascade of neural networks includes: an embedding network for receiving the N normal sensor event streams as inputs of the learning state and outputting M normal embedded coded streams as outputs of the embedding network based on the N normal sensor event streams, where N is an integer greater than integer M; a decoding network for receiving the M normal embedding-type encoded streams as an input of the decoding network in the learning state and outputting N normal reconstructed event streams as an output of the decoding network based on the M normal embedding-type encoded streams, the decoding network learning from among the inputs of the decoding network to generate an output of the decoding network and reducing differences between the N normal reconstructed event streams and the N normal sensor event streams based on a rule to form a trained decoding network state of the decoding network; In the above-mentioned equipment anomaly detection device, the embedding network performs a first type of WTA algorithm using lateral inhibitory connections between the embedding neurons of the embedding network; or The device for detecting an anomaly in an equipment, characterized in that the embedding network executes a second type of WTA algorithm using a group of inhibitory connections between the embedding neurons of the embedding network.

2. 2. The equipment anomaly detection device according to claim 1, wherein the embedding network in the trained embedding network state outputs an embedding-type encoded stream of M monitored objects based on the sensor event streams of the N monitored objects; and An equipment anomaly detection device characterized in that the decoding network in the trained decoding network state outputs N reconstructed event streams to be monitored based on the M embedded encoded streams to be monitored.

3. 3. The device according to claim 2, wherein the decoding network filters the M normal embedded-type encoded streams to generate filtered versions of the M normal embedded-type encoded streams. [Equation 38] and filtering the N normal sensor event streams to obtain a filtered version of the N normal sensor event streams. [0039] and obtaining a filtered version of the M normal embedded encoded streams. [0040] as an output of the decoding network based on a weighted combination of the N correct reconstructed event streams, [0041] Generate [0042] An equipment anomaly detection device comprising:

4. 4. The equipment anomaly detection device according to claim 3, wherein the decoding network filters the M normal embedding-type encoded streams and the N normal sensor event streams using an exponential low-pass filter.

5. 4. The device anomaly detection device according to claim 3, wherein the decoding network filters the M monitored embedding-type encoded streams in the trained decoding network state to generate filtered versions of the M monitored embedding-type encoded streams. [0044] and filtering the N monitored sensor event streams to obtain a filtered version of the N monitored sensor event streams. [0045] and obtaining a filtered version of the M monitored embedded-coded streams. [0046] a weighted combination of N monitored reconstructed event streams as an output of the decoding network based on the weighted combination of [0047] Generate [0048] An equipment anomaly detection device comprising:

6. The device abnormality detection device according to claim 5, The detection module filters the N target sensor event streams to generate a filtered version of the N target sensor event streams. [0049] The object is to obtain The detection module detects the reconstructed event stream. [0051] and a filtered version of the monitored sensor event stream. [0052] and determining whether or not an abnormality exists by comparing the detected abnormality with the detected abnormality.

7. 7. The device according to claim 6, wherein the detection module comprises: a subtraction calculator including a plurality of subtractors, selected from the plurality of subtractors, for subtracting a filtered version of the N monitored sensor event streams; [Number 50] One of the monitoring sensor event streams and N monitoring target reconstruction event streams [0051] The subtractors calculate a difference due to reconstruction between the filtered version of the N monitored sensor event streams and one of the monitored reconstructed event streams. [0052] and the N monitored reconstructed event streams [0,53] and a subtraction calculator that calculates a difference resulting from reconstruction between the

8. 8. The device according to claim 7, wherein the detection module comprises: Filtering version [0,54] and the N monitored reconstructed event streams [0.55] a sum calculator for calculating the sum of the difference between a comparator for comparing the sum value with a signal difference threshold and outputting an alarm signal indicating an abnormality in the N monitored sensor event streams when the sum value exceeds the signal difference threshold.

9. The equipment abnormality detection device according to any one of claims 1 to 8, further comprising: and a signal conversion module for converting the N normal sensor signal streams into the N normal sensor event streams and for converting the N monitored sensor signal streams into the N monitored sensor event streams.

10. 10. The equipment abnormality detection device according to claim 9, wherein the signal conversion module includes a cascade of a band-pass filter, a low-pass filter, and a PDM encoder; The bandpass filter is used for bandpass filtering the signal received from the signal conversion module to select a frequency component; The low-pass filter is used to rectify and smooth the band-pass filtered signal; The equipment anomaly detection device, characterized in that the PDM encoder is used to encode a signal that has been rectified and smoothed using pulse density modulation.

11. 11. The equipment abnormality detection device according to claim 10, wherein the PDM encoder is a sigma-delta encoder.

12. An equipment abnormality detection method, comprising: in a learning state, the cascade of neural networks receives and learns from the N normal sensor event streams from the N event stream input connections based on online learning to form a learned state of the cascade of neural networks, and each pulse in the event stream is sent out as an event; in a trained state, the cascade of neural networks receives N monitored sensor event streams as inputs from the N event stream input connections in an operational state of the cascade of neural networks, and outputs N monitored reconstructed event streams as an output of the cascade of neural networks based on the N monitored sensor event streams, the N monitored reconstructed event streams representing a reconstructed version of the N monitored sensor event streams; calculating, using a detection module, the difference between the N monitored sensor event streams and the N monitored reconstructed event streams; and the detection module sending a trigger signal when a difference between the N monitored sensor event streams and the N monitored reconstructed event streams exceeds a threshold; In the equipment anomaly detection method, the cascade of neural networks includes an embedding network and a decoding network; The embedding network is used for receiving the N normal sensor event streams as inputs of the learning state, and outputting M normal embedded encoded streams as outputs of the embedding network according to the N normal sensor event streams, where N is an integer greater than an integer M; In the learning state, the decoding network is used to receive the M normal embedding-type encoded streams as the input of the decoding network, and output N normal reconstructed event streams as the output of the decoding network according to the M normal embedding-type encoded streams, and the decoding network learns from the input of the decoding network to generate the output of the decoding network, and reduces the difference between the N normal reconstructed event streams and the N normal sensor event streams based on a rule, thereby forming a trained decoding network state of the decoding network; In the above-mentioned equipment anomaly detection method, the embedding network performs a first type of WTA algorithm using lateral inhibitory connections between the embedding neurons of the embedding network; or The method for detecting an anomaly in an equipment, characterized in that the embedding network executes a second type of WTA algorithm using a group of inhibitory connections between the embedding neurons of the embedding network.

13. 13. The method for detecting an anomaly in an equipment according to claim 12, wherein the embedding network in the trained embedding network state outputs an embedding-type encoded stream of M monitored objects based on the sensor event streams of the N monitored objects; and An equipment anomaly detection method, characterized in that the decoding network in the trained decoding network state outputs N reconstructed event streams to be monitored based on the M embedded encoded streams to be monitored.

14. The device anomaly detection method according to claim 13, the detection module uses a subtraction calculator to calculate the difference between the N monitored sensor event streams and the N monitored reconstructed event streams; The subtraction calculator includes a plurality of subtractors, selected from the plurality of subtractors, which calculate filtered versions of the N monitored sensor event streams. [0045] One of the monitoring sensor event streams and N monitoring target reconstruction event streams [0047] The subtractors calculate a difference due to reconstruction between the filtered version of the N monitored sensor event streams and one of the monitored reconstructed event streams. [0045] and the N monitored reconstructed event streams [0047] and calculating a difference resulting from reconstruction between the

15. 15. The method for detecting an abnormality in an equipment according to claim 12, wherein the decoding network filters the M normal embedded-type encoded streams to generate filtered versions of the M normal embedded-type encoded streams. [Equation 38] and filtering the N normal sensor event streams to obtain a filtered version of the N normal sensor event streams. [0039] and obtaining a filtered version of the M normal embedded encoded streams. [0040] as an output of the decoding network based on a weighted combination of the N correct reconstructed event streams, [0041] Generate [0042] An equipment anomaly detection method comprising:

16. 16. The equipment anomaly detection method according to claim 15, wherein the decoding network filters the M normal embedding-type encoded streams and the N normal sensor event streams using an exponential low-pass filter.

17. A chip, A chip comprising the equipment anomaly detection device according to any one of claims 1 to 11.

Citation Information

Patent Citations

  • AE monitor unit

    JP1980078250A

  • Information processing method and processor

    JP2020027614A

  • Abnormality detection apparatus, abnormality detection system, and abnormality detection method

    JP2020135170A

  • Spiking neural network for probabilistic computation

    US20200026981A1

  • Platform for detecting abnormal entities and activities using machine learning algorithms

    US20200380531A1