WO2023029229A1 - Device state detection method and related apparatus - Google Patents

Abstract

In a device state detection method and a related apparatus provided in the present application, a data processing device enables an auto-encoder to be combined with a plurality of recurrent neural networks, respectively reconstructs, by means of the auto-encoder, a plurality of state feature sets of a device to be detected, and according to differences between the plurality of state feature sets and the respective reconstructed feature sets, determines a health state of the device to be detected. Because features to be reconstructed of the auto-encoder are obtained by respectively performing feature mining by the plurality of recurrent neural networks on a plurality of encoding sequences having different feature scales, device health state-related information contained in a device state sequence can be fully explored, so as to achieve the purpose of improving the detection precision.

Description

Equipment status detection method and related device

Cross References to Related Applications

This application claims the priority of a Chinese patent application with application number 202111035618.8 and titled "Equipment Status Detection Method and Related Devices" filed with the China Patent Office on September 6, 2021, the entire contents of which are incorporated herein by reference.

technical field

The present application relates to the field of machine learning, and in particular, relates to a device status detection method and a related device.

Background technique

Since the health status of industrial equipment is of great significance to the safe production and economic benefits of enterprises, a data-driven monitoring method for industrial equipment health status is proposed. In this method, the sensor is used to collect the equipment state sequence of industrial equipment, the black box is modeled based on the deep neural network, and end-to-end training is carried out, and the industrial equipment health status contained in the massive data is excavated from the spatial dimension and the time dimension. information.

However, the inventors have found that the related technology does not make full use of the features within the local perception field of view of the data to be analyzed.

Contents of the invention

Embodiments of the present application provide a device state detection method and a related device.

In some exemplary embodiments, the present application provides a device state detection method, which can be applied to a data processing device, and the data processing device can be configured with a first detection model, and the first detection model can include a first encoder , a first decoder and a plurality of recurrent neural networks, the method may include:

Acquiring a plurality of state feature sets of the device to be detected, the plurality of state feature sets may have a preset order, and each state feature set may have respective state information of a plurality of components of the device to be detected;

Through the feature extraction of the multiple state feature sets by the first encoder from the spatial dimension, multiple coded sequences with different feature scales can be obtained, wherein the multiple coded sequences can correspond to different cyclic nerves network, each of the encoding sequences may include a plurality of encoding features in the preset order, and the plurality of encoding features may respectively correspond to different sets of the state features;

For each of the coded sequences, feature extraction is performed on the coded sequence from the time dimension through the corresponding cyclic neural network to obtain a sequence to be reconstructed, wherein the sequence to be reconstructed may include a plurality of features to be reconstructed , the plurality of features to be reconstructed may respectively correspond to different sets of state features;

For each state feature set, input all the features to be reconstructed in the state feature set to the first decoder for reconstruction to obtain a reconstructed feature set of the state feature set;

According to the difference between the plurality of state feature sets and the respective reconstructed feature sets, the health status of the device to be detected can be determined.

In some exemplary implementations, the first encoder may include multiple convolutional layers, and the first encoder performs feature extraction on the multiple state feature sets from the spatial dimension to obtain multiple A coding sequence, which can include:

For each state feature set, feature extraction is performed sequentially on the state feature set through the plurality of convolutional layers to obtain a plurality of coding features of the state feature set, wherein the plurality of state feature sets The coding features may respectively have different feature scales, and may be respectively obtained from different convolutional layers;

The multiple coded sequences can be obtained by classifying the respective coded features of the multiple state feature sets according to feature scales.

In some exemplary implementations, the first decoder may include a plurality of deconvolution layers respectively corresponding to a plurality of cyclic neural networks, and for each of the state feature sets, all the state feature sets are treated as The reconstructed feature is input to the first decoder for reconstruction, and the reconstructed feature set of the state feature set is obtained, which may include:

According to the corresponding relationship between the plurality of cyclic neural networks and the plurality of deconvolution layers, all the features to be reconstructed in the state feature set are respectively input to the corresponding deconvolution layers for reconstruction, and the obtained The reconstructed feature set of the state feature set.

In some exemplary implementations, obtaining multiple state feature sets of the device to be detected may include: obtaining a first state sequence of the device to be detected, wherein the first state sequence may include multiple state characteristics of the device to be detected state data of each component;

splitting the first state sequence into a plurality of data segments;

For each of the data segments, obtain the covariance matrix of the data segments;

The covariance matrix of all the data segments is used as a plurality of state feature sets of the device to be detected.

In some exemplary implementations, determining the health status of the device to be detected according to the difference between the multiple state feature sets and the respective reconstructed feature sets may include:

Obtaining the first mean square error between the plurality of state feature sets and the respective reconstruction feature sets;

If the first mean square error is greater than a first threshold, there is an abnormality in the device to be detected.

In some exemplary embodiments, the data processing device may also be configured with a second detection model of the target component, the second detection model may include a second encoder and a second decoder, and the method may further include:

obtaining a second state sequence of the target component, wherein the second state sequence includes state data of the target component;

performing feature extraction on the second state sequence by the second encoder to obtain first features to be reconstructed;

Acquiring the second features to be reconstructed of the second state sequence, wherein the second features to be reconstructed include one of time domain features, frequency domain features and time-frequency domain features of the second state sequence or combination;

inputting the concatenated features of the first feature to be reconstructed and the second feature to be reconstructed into the second decoder for reconstruction to obtain a reconstruction sequence of the second state sequence;

A health state of the target component is determined based on a difference between the second state sequence and the reconstructed sequence.

In some exemplary implementations, determining the health state of the target component according to the difference between the second state sequence and the reconstruction sequence may include:

obtaining a second mean square error between the second state sequence and the reconstructed sequence;

If the second mean square error is greater than a second threshold, there is an abnormality in the target component.

In some exemplary implementations, the embodiment of the present application may provide a device state detection apparatus, which is applied to a data processing device, and the data processing device may be configured with a first detection model, and the first detection model may include a first An encoder, a first decoder, and a plurality of recurrent neural networks, the device state detection device may include:

A feature acquisition module configured to acquire a plurality of state feature sets of the device to be detected, the plurality of state feature sets may have a preset order, and each state feature set may have a plurality of components of the device to be detected status information;

The feature extraction module is configured to perform feature extraction on the plurality of state feature sets from the spatial dimension through the first encoder to obtain a plurality of encoding sequences with different feature scales, wherein the plurality of encoding sequences Can respectively correspond to different cyclic neural networks, each of the encoding sequences can include a plurality of encoding features in the preset order, and the plurality of encoding features can respectively correspond to different sets of the state features;

The feature extraction module is further configured to, for each of the coded sequences, perform feature extraction on the coded sequence from the time dimension through the corresponding cyclic neural network to obtain a sequence to be reconstructed, wherein the to-be-reconstructed sequence The reconstruction sequence may include multiple features to be reconstructed, and the multiple features to be reconstructed may respectively correspond to different sets of state features;

The feature reconstruction module is configured to, for each of the state feature sets, input all the features to be reconstructed in the state feature set to the first decoder for reconstruction, and obtain the reconstruction of the state feature set Construct feature set;

A state detection module configured to determine the health state of the device to be detected according to the difference between the plurality of state feature sets and the respective reconstructed feature sets.

In some exemplary implementations, the embodiments of the present application may provide a data processing device, the data processing device may include a processor and a memory, the memory may store a computer program, and the computer program may be executed by the processor During execution, the device status detection method described above is realized.

In some exemplary implementations, the embodiments of the present application may provide a computer storage medium, where the computer storage medium may store a computer program, and when the computer program is executed by a processor, implement the device state detection method.

Compared with related technologies, the present application has at least the following beneficial effects:

In the device state detection method and related devices used in this embodiment, the data processing device combines the autoencoder with multiple recurrent neural networks, and reconstructs the multiple state feature sets of the device to be detected through the autoencoder, and According to the difference between multiple state feature sets and their respective reconstruction feature sets, the health status of the device to be detected is determined; since the features to be reconstructed of the autoencoder are encoded by multiple cyclic neural networks for multiple encoding sequences of different feature scales The feature mining is carried out separately, so the information related to the equipment health status contained in the equipment status sequence can be fully explored, so as to achieve the purpose of improving the detection accuracy.

Description of drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the accompanying drawings used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, so It should be regarded as a limitation on the scope, and those skilled in the art can also obtain other related drawings based on these drawings without creative work.

FIG. 1 is a schematic structural diagram of a data processing device provided in an embodiment of the present application;

FIG. 2 is one of the schematic flow diagrams of the device state detection method provided by the embodiment of the present application;

FIG. 3 is a schematic structural diagram of the first detection model provided by the embodiment of the present application;

FIG. 4 is the second schematic flow diagram of the device status detection method provided by the embodiment of the present application;

FIG. 5 is a schematic structural diagram of a second detection model provided in an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an apparatus for detecting device status provided by an embodiment of the present application.

Icons: 120-memory; 130-processor; 140-communication device; 201-feature acquisition module; 202-feature extraction module; 203-feature reconstruction module; 204-state detection module.

Detailed ways

To make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments It is a part of the embodiments of this application, not all of them. The components of the embodiments of the application generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations.

Accordingly, the following detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the application. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of this application.

It should be noted that like numerals and letters denote similar items in the following figures, therefore, once an item is defined in one figure, it does not require further definition and explanation in subsequent figures.

In the description of the present application, it should be noted that the terms "first", "second", "third" and so on are only used to distinguish descriptions, and should not be understood as indicating or implying relative importance. Furthermore, the term "comprises", "comprises" or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, but also includes elements not expressly listed. other elements of or also include elements inherent in such a process, method, article, or device. Without further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article or apparatus comprising said element.

It should be understood that the operations of the flowcharts may be performed out of order, and steps that have no logical context may be performed in reverse order or concurrently. In addition, those skilled in the art may add one or more other operations to the flowchart or remove one or more operations from the flowchart under the guidance of the content of the present application.

In related technologies, the equipment state sequence collected by the sensor of industrial equipment is used, the black box is modeled based on the deep neural network, and end-to-end training is carried out, and the industrial equipment health contained in the massive data is excavated from the spatial dimension and the time dimension. status information.

However, in the process of feature extraction based on the device state sequence, the deep neural network pays more attention to the feature information in the global field of view, while ignoring the feature information in the local perception field of view. Therefore, the device state sequence is not fully explored. Information related to the health status of the device contained in the .

In view of this, an embodiment of the present application provides a device state detection method applied to a data processing device. In this method, the data processing device combines an autoencoder with multiple cyclic neural networks, and reconstructs the multiple state feature sets of the device to be detected through the autoencoder, and according to the multiple state feature sets and their respective reconstructed The difference between the structural feature sets determines the health status of the device to be detected; since the features to be reconstructed of the autoencoder are obtained by feature mining of multiple encoding sequences of different feature scales by multiple cyclic neural networks, it can be Fully explore the information related to the equipment health status contained in the equipment status sequence.

Wherein, the device status detection method is applicable to different types of devices to be detected. Exemplarily, the device to be detected may be an industrial robot, a machine tool, a gate control device, a train, an unmanned aerial vehicle, and the like.

Moreover, the data processing device can select different types of devices according to the type of the device to be detected. In some implementations, the data processing device may be a server communicatively connected to the device to be detected. The type of this server can be, but not limited to, Web (website) server, FTP (File Transfer Protocol, file transfer protocol) server, data processing server etc. Also, the server can be a single server or a group of servers. Server groups can be centralized or distributed (for example, the servers can be a distributed system). In some embodiments, the server may be local or remote relative to the user terminal. In some embodiments, the server can be implemented on a cloud platform; only as an example, the cloud platform can include private cloud, public cloud, hybrid cloud, community cloud (Community Cloud), distributed cloud, inter-cloud (Inter-Cloud), Multi-Cloud, etc., or any combination of them. In some embodiments, a server may be implemented on an electronic device having one or more components.

In some other implementation manners, the data processing device may also be a user terminal communicatively connected to the device to be detected. The specific type of the user terminal may be, but not limited to, a mobile terminal, a tablet computer, a laptop computer, or a built-in device in a motor vehicle, etc., or any combination thereof. In some embodiments, the mobile terminal may include smart home devices, wearable devices, smart mobile devices, virtual reality devices, or augmented reality devices, etc., or any combination thereof. In some embodiments, smart home devices may include smart lighting devices, control devices for smart electrical devices, smart monitoring devices, smart TVs, smart cameras, or walkie-talkies, etc., or any combination thereof. In some embodiments, wearable devices may include smart bracelets, smart shoelaces, smart glasses, smart helmets, smart watches, smart clothing, smart backpacks, smart accessories, etc., or any combination thereof. In some embodiments, the smart mobile device may include a smart phone, a personal digital assistant (Personal Digital Assistant, PDA), a game device, a navigation device, or a point of sale (Point of Sale, POS) device, etc., or any combination thereof.

This embodiment also provides a schematic structural diagram of the data processing device. As shown in FIG. 1 , the data processing device includes a memory 120 , a processor 130 , and a communication device 140 .

The components of the memory 120 , the processor 130 and the communication device 140 are directly or indirectly electrically connected to each other to realize data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

Wherein, the memory 120 can be, but not limited to, random access memory (Random Access Memory, RAM), read-only memory (Read Only Memory, ROM), programmable read-only memory (Programmable Read-Only Memory, PROM), Erasable Programmable Read-Only Memory (EPROM), Electric Erasable Programmable Read-Only Memory (EEPROM), etc. Wherein, the memory 120 is used to store a program, and the processor 130 executes the program after receiving an execution instruction.

The communication device 140 is used to send and receive data through the network. The network may include a wired network, a wireless network, an optical fiber network, a telecommunication network, an intranet, the Internet, a local area network (Local Area Network, LAN), a wide area network (Wide Area Network, WAN), a wireless local area network (Wireless Local Area Networks, WLAN), metropolitan area network (Metropolitan Area Network, MAN), wide area network (Wide Area Network, WAN), public switched telephone network (Public Switched Telephone Network, PSTN), Bluetooth network, ZigBee network, or near field communication ( Near Field Communication, NFC) network, etc., or any combination thereof. In some embodiments, a network may include one or more network access points. For example, a network may include wired or wireless network access points, such as base stations and/or network switching nodes, through which one or more components of the service request processing system may connect to the network to exchange data and/or information.

The processor 130 may be an integrated circuit chip with signal processing capabilities, and the processor may include one or more processing cores (for example, a single-core processor or a multi-core processor). For example only, the above-mentioned processor may include a central processing unit (Central Processing Unit, CPU), an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), an application specific instruction set processor (Application Specific Instruction-set Processor, ASIP), graphics processing Unit (Graphics Processing Unit, GPU), Physical Processing Unit (Physics Processing Unit, PPU), Digital Signal Processor (Digital Signal Processor, DSP), Field Programmable Gate Array (Field Programmable Gate Array, FPGA), Programmable Logic Device (Programmable Logic Device, PLD), controller, microcontroller unit, reduced instruction set computer (Reduced Instruction Set Computing, RISC), or microprocessor, etc., or any combination thereof.

Based on the above-mentioned relevant introduction about the data processing device, the device status detection method running in the data processing device will be described in detail below. In this embodiment, a neural network model combining an autoencoder and multiple cyclic neural networks is called a first detection model, where the autoencoder includes a first encoder and a first decoder. As shown in FIG. 2 , it is a schematic flowchart of the device state detection method, and each step of the method will be described in detail below in conjunction with FIG. 2 . As shown in Figure 2, the method may include:

Step S101, acquiring multiple state feature sets of the device to be detected.

Wherein, the multiple state feature sets have a preset order, and each state feature set has state information of multiple components of the device to be detected.

In this example, considering that when the device to be detected is working, multiple components need to cooperate with each other, therefore, the state data of multiple components has a specific linkage relationship, and the linkage relationship will be reflected in the form of data distribution. However, this embodiment is based on this finding, and uses the first detection model to discover the linkage relationship contained in the corresponding state data of multiple components, so as to determine the health status of the device to be detected.

As an implementation, considering that the state data of multiple components has a high-dimensional feature space and requires high computational complexity, the data processing device can obtain the first state sequence of the device to be detected, wherein the first The status sequence may include status data for each component of the device to be tested.

The data processing device splits the first state sequence into multiple data segments; and for each data segment, obtains the covariance matrix of the data segment; finally, uses the covariance matrix of all the data segments as multiple data segments of the device to be detected State feature set.

The above-mentioned industrial robot is used as the equipment to be tested for exemplary description below. It is assumed that the industrial robot can include 6 degrees of freedom, and each degree of freedom corresponds to a driving motor. The data processing device periodically collects the state of the six driving motors synchronously through the sensor, and obtains the respective state data of the six driving motors, which can be expressed as:

In the formula, _M1 represents the state sequence of the first driving motor,

Indicates the state data collected by the first driving motor at time t; the representation of the corresponding state sequences of other driving motors is the same as that of the first driving motor, and will not be described in this embodiment.

The types of state data collected in this example may include at least one of the current, rotational speed, and rotation angle of the driving motors during operation; and, the types of state data between the driving motors may be all or partially the same. For example, the types of state data of the first driven motor may include current, rotational speed, and rotation angle; while the types of state data of the second driven motor may include current and rotational speed. Therefore, those skilled in the art may make adaptive adjustments as needed, and this embodiment of the present application does not make specific limitations.

Then, for each driving motor, the data processing device may intercept the state sequence of the driving motor through a time window to obtain multiple sequence fragments of the driving motor. However, it should be understood that, compared with the method of averaging and segmenting the state sequence, this embodiment adopts the sliding window method to intercept the state sequence, which can not only make the intercepted sequence fragments relatively continuous, but also avoid the influence of the interception scale on the data. distribution, and more sequence fragments can be obtained.

In addition, in order to keep the dimensions of the sequence segments the same among different driving motors, in this embodiment, a sliding window of the same scale may be selected for each driving motor. Finally, for all the sequence fragments of the driving motors, the data processing device classifies the sequence fragments at the same sequence position into one category to obtain multiple data fragments.

Assuming that the scale of the sliding window is 3 state data, and the step length of the sliding window is 2 state data, then the first data segment can be expressed as:

The second piece of data can be represented as:

By analogy, other data segments can be obtained, which will not be repeated in this embodiment. In the formula,

represents the first sequence segment of the first driven motor,

represents the second sequence fragment of the first drive motor;

represents the first sequence segment of the second drive motor,

Indicates the second sequence segment of the second drive motor; the meanings of other sequence segments can be obtained by analogy, which will not be described in detail in this embodiment.

Therefore, each of the above data fragments includes the respective state data of a plurality of drive motors. In order to facilitate the first detection model to discover the health state information contained therein, the data processing device uses the covariance matrix between the sequence fragments in the data fragments as a state feature set. It should be understood that the covariance matrix carries the state information of each driving motor, including the autocorrelation among state data of a single driving motor and the correlation among multiple driving motors.

In step S102, feature extraction is performed on multiple state feature sets from the spatial dimension by the first encoder to obtain multiple encoded sequences with different feature scales.

Wherein, multiple encoding sequences may respectively correspond to different cyclic neural networks, each encoding sequence may include multiple encoding features in a preset order, and multiple encoding features may respectively correspond to different state feature sets.

As a possible implementation manner, the first encoder may include multiple convolutional layers. For each state feature set, the data processing device sequentially extracts features from the state feature set through multiple convolutional layers to obtain multiple coding features of the state feature set; wherein, the multiple coding features of the state feature set can have different The feature scales of , and are obtained from different convolutional layers.

Finally, the data processing device classifies the coding features of the multiple state feature sets according to the feature scale to obtain multiple coding sequences.

Continue to take the above-mentioned industrial robot as an example for illustration. As shown in Figure 3, the encoder may include 4 convolutional layers, and it is assumed that the number of state feature sets of the industrial robot is 5. For the convenience of description, these five state feature sets are denoted as state feature set A, state feature set B, state feature set C, state feature set D, and state feature set E. For each state feature set, the data processing device can input it to a plurality of convolutional layers connected in series to obtain 4 coding features of the state feature set.

Taking the state feature set A as an example, in the process of processing the state feature set A through multiple convolutional layers of the encoder, the data processing device not only inputs the encoded features output by each convolutional layer to the adjacent next convolutional layer , and also copied a copy of the encoding features output by each convolutional layer to construct multiple encoding sequences.

Therefore, the data processing device can obtain 4 coding features of different feature scales through the state feature set A, that is, the above 5 state feature sets can obtain 5*4=20 coding features in total.

Since different feature scales correspond to different convolutional layers, the data processing device classifies the above 20 coding features according to the feature scales, and divides the coding features from the same convolutional layer into one category; therefore, Four encoding sets can be obtained, and each encoding set includes five encoding features, corresponding to different state feature sets.

And considering that there is a preset order among different state feature sets, the preset order corresponds to the interception order of the data segments, and also reflects the collection order of the state data. Therefore, for each coding set, the data processing device sorts the coding features corresponding to the state feature set in order to obtain the corresponding coding sequences, that is, 4 coding sets can obtain 4 coding sequences.

Step S103, for each coded sequence, extract the features of the coded sequence from the time dimension through the corresponding cyclic neural network, and obtain the sequence to be reconstructed.

Wherein, the sequence to be reconstructed may include multiple features to be reconstructed, and the multiple features to be reconstructed may respectively correspond to different state feature sets.

Exemplarily, the above-mentioned multiple recurrent neural networks may be four LSTM (Long Short-term Memory, long short-term memory) networks in FIG. 3 . For each coded sequence, the data processing device sequentially inputs the 5 coded features in the coded sequence into the corresponding LSTM network in a preset order to obtain 5 features to be reconstructed. Among them, the five features to be reconstructed correspond to different state feature sets of industrial robots.

Step S104, for each state feature set, input all the features to be reconstructed in the state feature set to the first decoder for reconstruction, and obtain the reconstructed feature set of the state feature set.

As an implementation manner, the first decoder may include multiple deconvolution layers respectively corresponding to multiple cyclic neural networks. According to the corresponding relationship between multiple cyclic neural networks and multiple deconvolution layers, the data processing device inputs all the features to be reconstructed in the state feature set to the corresponding deconvolution layers for reconstruction, and obtains the state feature set The reconstruction feature set of .

Exemplarily, continuing to refer to the first decoder in FIG. 3 , it may include four sequentially connected deconvolution layers, each corresponding to a different LSTM network. Taking the state feature set A of an industrial robot as an example, the four LSTM networks will respectively output a feature to be reconstructed of the state feature set A, that is, a total of 4 features to be reconstructed can be obtained. The data processing device can respectively input the four features to be reconstructed into the corresponding deconvolution layers for reconstruction, so as to obtain the reconstructed feature set corresponding to the state feature set A. Similarly, state feature set B, state feature set C, state feature set D, and state feature set E can all obtain their own reconstructed feature sets.

In order to keep the features from being lost, the data processing equipment will splice the two and input them to the next deconvolution layer. As shown in Figure 3, the input features of the deconvolution layer include the output features of the adjacent previous deconvolution layer and the features to be reconstructed corresponding to the output of the LSTM network.

Step S105, according to the difference between the plurality of state feature sets and the respective reconstruction feature sets, determine the health status of the device to be detected.

It should be noted that the first detection model is obtained by training the sample state feature set when the device to be detected is working normally. Therefore, assuming that the device to be detected is not abnormal, the first detection model can separate the multiple state features Perform reconstruction so that the difference between multiple state feature sets and their respective reconstructed feature sets does not exceed the set first threshold; on the contrary, if the device to be detected is abnormal, it is difficult for the first detection model to separate the multiple state features performing reconstruction; making the difference between the plurality of state feature sets and the respective reconstructed feature sets greater than a set first threshold.

Therefore, as an implementation, the data processing device can obtain the first mean square error between multiple state feature sets and their respective reconstructed feature sets; if the first mean square error is greater than the first threshold, the device to be detected exists abnormal.

Based on the above design, the data processing device combines the autoencoder with multiple recurrent neural networks, and reconstructs the multiple state feature sets of the device to be detected through the autoencoder, and reconstructs the The difference between the structural feature sets determines the health status of the device to be detected; since the features to be reconstructed of the autoencoder are obtained by feature mining of multiple encoding sequences of different feature scales by multiple cyclic neural networks, it can be Fully explore the information related to the health status of the equipment contained in the equipment status sequence to achieve the purpose of improving the detection accuracy.

In addition, it is worth noting that most of the state data collected from industrial equipment is high-frequency time series data, and the cost of labeling data is extremely high. Therefore, according to the characteristics of industrial equipment state data, the first detection model provided in this embodiment can be based on The way of self-monitoring realizes the detection of the health status of industrial equipment.

In this embodiment, not only the overall state of the device to be detected is detected, but also each component is detected separately. In the following, any target component selected from multiple components is taken as an example to describe in detail. In this embodiment, in order to detect the state of the target component, the data processing device is further configured with a second detection model for the target component, where the model may include a second encoder and a second decoder.

As shown in Figure 4, based on the second detection model above, the device state detection method may also include:

Step S106, acquiring the second state sequence of the target component.

Wherein, the second status sequence includes status data of the target component. Exemplarily, continuing to take the above-mentioned industrial robot as an example, assuming that one of the drive motors of the industrial robot is used as the target component, the working state data of the drive motor may include any one of current, rotational speed, and rotation angle. Assuming that the type of the state data is current, the data processing device periodically collects the current of the driving motor when it is working, and uses a sliding window to intercept the collected current data to obtain the second state sequence of the driving motor.

In order to better analyze the distribution of the state data corresponding to the target component, in this embodiment, the state data is obtained by standardizing the original state data of the target component.

As an optional normalization method, the data can be normalized using the z-score (zero-mean normalization) method. In this method, the original data is scaled so that the scaled data falls within the interval with a mean of 0 and a standard deviation of 1. The expression of the z-score method is as follows:

In the formula, x _i represents the i-th original state data, N represents the total amount of original state data, μ is the average value of the original state data, σ is the standard deviation of the original state data, and z represents the standardized state data.

It should be noted that after the z-score standardization of the original state data, the distribution of the data itself has not changed, but the distribution range of the normalized state data is basically the same, mainly in the interval [-2,2] ; so that the subsequent algorithms can focus more on analyzing the distribution of the state data itself.

Step S107, performing feature extraction on the second state sequence by the second encoder to obtain the first feature to be reconstructed.

In this embodiment, as shown in Figure 5, in order to explore the feature information of the state data in the time dimension, the second encoder can select an LSTM network and a PCA (Principal Component Analysis, principal component analysis) model, while the second decoder can Select the LSTM network.

Among them, the working principle of the LSTM network is to input the time series of the state number of the target component in sequence, and then update its hidden state, expressed as h _t = LSTM(h _t-1 , x _t ), in the last step of a time series (denoted as t ₂ ), the LSTM network contains all the information of the previous sequence (also known as the context vector), namely

Therefore, it can be used to explore the characteristic information contained in the data from the time dimension.

Step S108, acquiring the second features to be reconstructed of the second state sequence.

The second feature to be reconstructed includes one or a combination of time domain features, frequency domain features, and time-frequency domain features of the second state sequence.

Time domain features:

Can include _rms x rms , root square amplitude x _sra , peak-to-peak x _ppv , crest factor x _cf , margin index x _cf , skewness index x _skf , kurtosis index x _kf , form factor x _shf , pulse factor x _if and information entropy x _en . The time-domain feature vector F _t can be obtained by splicing the above 10 kinds of parameters together, and the corresponding expression is:

F _t = {x _rms , x _sra , _xppv , x _cf , x _mf , x _skf , x _kf , x _shf , x _if , x _en }

The above 10 parameters will be explained in detail by taking the second state sequence of the industrial robot as an example in combination with the above industrial robot again. Since the second state sequence of the industrial robot includes current data when the industrial robot is working, the meanings of the parameters are as follows:

The effective value, that is, the root mean square value of the current data, mainly describes the effective power of the current.

The square root amplitude is used to describe the overall amplitude of current vibration and reflects the true level of current vibration.

The peak-to-peak value indicates the difference between the maximum value and the minimum value of the current data in the second state sequence, and is mainly used to describe the span range of the current.

Crest factor, which is used to describe the shock condition present in the current.

The margin index is used to describe the wear of the machine corresponding to the current, and is more sensitive to impact faults.

The skewness index, the statistical average of the third-order moment of the vibration signal, is mainly used to describe the asymmetry of the current.

The kurtosis index, the fourth-order moment statistical average of the current data, is mainly used to describe the magnitude of the impact on the current.

The form factor is used to represent the original shape properties of the current data waveform and has nothing to do with the amplitude.

The pulse factor is used to indicate the impact of the current. Although the sensitivity is not as good as the kurtosis index, it can complement the kurtosis index.

Information entropy, used to describe the degree of uncertainty of the current.

Frequency domain features:

According to Parseval's theorem (Parseval's theorem): whether it is a real signal or a complex signal (that is, the state data of the target component in this embodiment), the integral of the square of the signal amplitude is equal to the square of the modulus of the signal's energy equal to the signal spectral density. The corresponding expression can be expressed as:

Among them, E represents the signal energy, x(t) represents the signal time threshold, and X(f) represents the signal frequency threshold. Therefore, in this embodiment, the data processing device first obtains the spectrogram of the second state sequence through a Fast Fourier Transform (FFT) method, and then takes the frequency axis of the spectrogram as the time axis to integrate the spectrogram. In this way, the frequency domain feature xf _en of the second state sequence is obtained.

Time-frequency domain features:

The data processing equipment uses EMD (Empirical Mode Decomposition, empirical mode decomposition) method and STFT (short-time Fourier Transform, short-time Fourier transform) method to perform time-frequency analysis on the second state sequence, and obtain the time-frequency domain characteristics .

First, the data processing device obtains n IMFs (Intrinsic Mode Function, intrinsic mode function) of the correlation between the second state sequence and the fault of the target component by the EMD method; then, the energy, energy, and The four types of eigenvalues are variance, skewness index and kurtosis index. Finally, the standard deviation of instantaneous frequency and the signal-to-noise ratio of instantaneous frequency are obtained by using the STFT method.

The above eigenvalues are concatenated to obtain a 4n+2-dimensional time-frequency domain eigenvector F _tf , and the corresponding expression is:

In the formula,

represents the energy of n IMFs,

Indicates the variance of n IMFs,

Identify n IMF skewness indicators,

Indicates the kurtosis index of n IMFs, σ _std is the standard deviation of the instantaneous frequency, and SNR indicates the signal-to-noise ratio of the instantaneous frequency.

Step S109 , input the concatenated features of the first feature to be reconstructed and the second feature to be reconstructed into the second decoder for reconstruction to obtain a reconstructed sequence of the second state sequence.

As shown in FIG. 5 , it is a schematic diagram of a possible structure of the second detection model. The data processing device inputs the second state sequence to the LSTM encoding network and the PCA model to obtain the first feature to be reconstructed; the above-mentioned time domain feature, frequency domain feature and time-frequency domain feature are used as the second feature to be reconstructed; A feature to be reconstructed and a second feature to be reconstructed are input to the LSTM decoding network to obtain a reconstructed sequence of the second state sequence.

Step S110, according to the difference between the second state sequence and the reconstruction sequence, determine the health state of the target component.

It should be noted that, similar to the first detection model, the second detection model is also trained in advance through the sample state sequence of the target component when it works normally. Therefore, assuming that the target device is not abnormal, the second detection model can use the first The second state sequence is reconstructed so that the second mean square error between the second state sequence and the reconstructed sequence is not greater than the second threshold; on the contrary, if the target component is abnormal, it is difficult for the second detection model to carry out the second state sequence Reconstructing; making a second mean square error between the second state sequence and the reconstructed sequence greater than a second threshold.

In addition, the above-mentioned first detection model and the second detection model are complementary to each other, and any abnormality detected by any one of the detection models is regarded as an abnormality in the device to be detected.

In this embodiment, the influence of the error factor is also considered, and a threshold range with a second threshold is provided to improve the fault tolerance, and the expression is as follows:

In the formula, X represents the sample state sequence,

represents the reconstructed sequence of sample state sequences,

represents the mean variance between the two,

represents the standard deviation between the two,

Indicates the range of the second threshold. When the mean square error between the second state sequence and the reconstructed sequence is within the range, the target component is normal, otherwise it is abnormal.

And when an abnormality occurs in the target component is detected, the abnormality score when an abnormality occurs is calculated by the following expression

Used to measure the severity of anomalies:

In the formula,

Indicates the mean square error between the second state sequence and the reconstructed sequence,

is a symbolic function, when

When it is less than 0, the value of this function is 1, when

When equal to 0, the value of the function is 0, when

When greater than 0, the function evaluates to 1.

Based on the same inventive concept as the device status detection method, this embodiment also provides related devices of the method.

This embodiment may also provide a device state detection device, which is applied to a data processing device. The device state detection device may include at least one functional module that can be stored in a memory in the form of software. As shown in Figure 6, from the functional division, the device status detection device may include:

The feature acquisition module 201 is configured to acquire multiple state feature sets of the device to be detected, the multiple state feature sets have a preset order, and each state feature set has status information of multiple components of the device to be detected.

In this embodiment, the feature acquisition module 201 is configured to implement step S101 in FIG. 2 . For a detailed description of the feature acquisition module 201 , refer to the detailed description of step S101 .

The feature extraction module 202 is configured to perform feature extraction on multiple state feature sets from the spatial dimension through the first encoder to obtain multiple encoding sequences with different feature scales, wherein the multiple encoding sequences correspond to different cycles In the neural network, each encoding sequence includes multiple encoding features with a preset order, and the multiple encoding features correspond to different state feature sets.

The feature extraction module 202 is further configured to perform feature extraction on the encoded sequence from the time dimension through the corresponding cyclic neural network for each encoded sequence to obtain the sequence to be reconstructed, wherein the sequence to be reconstructed includes a plurality of Features, multiple features to be reconstructed correspond to different state feature sets.

In this embodiment, the feature extraction module 202 is configured to implement steps S102-S103 in FIG. 2 . For a detailed description of the feature extraction module 202, refer to the detailed description of steps S102-S103.

The feature reconstruction module 203 is configured to, for each state feature set, input all features to be reconstructed in the state feature set to the first decoder for reconstruction, and obtain a reconstructed feature set of the state feature set.

In this embodiment, the feature reconstruction module 203 is configured to implement step S104 in FIG. 2 . For a detailed description of the feature reconstruction module 203 , refer to the detailed description of step S104 .

The state detection module 204 is configured to determine the health state of the device to be detected according to the difference between the multiple state feature sets and the respective reconstructed feature sets.

In this embodiment, the state detection module 204 is configured to implement step S105 in FIG. 2 . For a detailed description of the state detection module 204 , refer to the detailed description of step S105 .

It should be noted that the device status detection device may also include other software function modules for implementing other steps or sub-steps of the device status detection method; of course, the above-mentioned feature acquisition module 201, feature extraction module 202, feature reconstruction module 203 And the state detection module 204 can also be used to implement other steps or sub-steps of the device state detection method; those skilled in the art can make appropriate adjustments according to different module division standards, which are not specifically limited in this embodiment.

This embodiment may also provide a data processing device. The data processing device may include a processor and a memory, and the memory stores a computer program. When the computer program is executed by the processor, the device state detection method is implemented.

This embodiment may also provide a computer storage medium, where a computer program may be stored in the computer storage medium, and when the computer program is executed by a processor, the device status detection method described above may be implemented.

To sum up, in the device state detection method and related devices provided by the embodiment of the present application, the data processing device combines the autoencoder with multiple recurrent neural networks, and uses the autoencoder to combine multiple state feature sets of the device to be detected Reconstruct separately, and determine the health status of the device to be detected according to the difference between multiple state feature sets and their respective reconstructed feature sets; since the features to be reconstructed of the autoencoder are analyzed by multiple cyclic neural networks for different features Multiple coding sequences of the scale are obtained by feature mining respectively. Therefore, the information related to the health status of the equipment contained in the equipment status sequence can be fully explored to achieve the purpose of improving the detection accuracy.

In the embodiments provided in this application, it should be understood that the disclosed devices and methods may also be implemented in other ways. The device embodiments described above are only illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functions and possible implementations of devices, methods and computer program products according to multiple embodiments of the present application. operate. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code that includes one or more Executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks in succession may, in fact, be executed substantially concurrently, or they may sometimes be executed in the reverse order, depending upon the functionality involved. It should also be noted that each block of the block diagrams and/or flowchart illustrations, and combinations of blocks in the block diagrams and/or flowchart illustrations, can be implemented by a dedicated hardware-based system that performs the specified function or action , or may be implemented by a combination of dedicated hardware and computer instructions.

In addition, each functional module in each embodiment of the present application may be integrated to form an independent part, each module may exist independently, or two or more modules may be integrated to form an independent part.

If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

The above are just various implementations of the present application, but the scope of protection of the present application is not limited thereto. Anyone skilled in the art can easily think of changes or substitutions within the technical scope disclosed in the present application. All should be covered within the scope of protection of this application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Industrial Applicability

This application discloses a device state detection method and related devices. The data processing device combines an autoencoder with multiple cyclic neural networks, and reconstructs multiple state feature sets of the device to be detected through the autoencoder, and according to multiple The difference between each state feature set and their respective reconstruction feature sets determines the health status of the device to be detected; since the features to be reconstructed of the autoencoder are respectively processed by multiple cyclic neural networks for multiple encoding sequences of different feature scales Therefore, the information related to the health status of the equipment contained in the equipment status sequence can be fully explored to achieve the purpose of improving the detection accuracy.

In addition, it can be understood that the device status detection method and related devices of the present application are reproducible and can be applied in various industrial applications. For example, the device state detection method of the present application can be applied to various data processing devices.

Claims (10)

1. A device state detection method, characterized in that it is applied to a data processing device, the data processing device is configured with a first detection model, and the first detection model includes a first encoder, a first decoder, and a plurality of cyclic neurons network, the method comprising:

   Acquiring a plurality of state feature sets of the device to be detected, the plurality of state feature sets having a preset order, each of the state feature sets having respective state information of a plurality of components of the device to be detected;

   Using the first encoder to perform feature extraction on the plurality of state feature sets from the spatial dimension to obtain a plurality of encoding sequences with different feature scales, wherein the plurality of encoding sequences correspond to different cyclic neural networks respectively, Each of the encoding sequences includes a plurality of encoding features in the preset order, and the plurality of encoding features respectively correspond to different sets of the state features;

   For each of the coded sequences, feature extraction is performed on the coded sequence from the time dimension through the corresponding cyclic neural network to obtain a sequence to be reconstructed, wherein the sequence to be reconstructed includes a plurality of features to be reconstructed, The plurality of features to be reconstructed respectively correspond to different sets of the state features;

   For each state feature set, input all the features to be reconstructed in the state feature set to the first decoder for reconstruction to obtain a reconstructed feature set of the state feature set;

   The health state of the device to be detected is determined according to the difference between the plurality of state feature sets and the respective reconstructed feature sets.
2. The device state detection method according to claim 1, wherein the first encoder includes a plurality of convolutional layers, and the plurality of state feature sets are respectively characterized from the spatial dimension by the first encoder Extraction, to obtain multiple coded sequences with different feature scales, including:

   For each state feature set, feature extraction is performed sequentially on the state feature set through the plurality of convolutional layers to obtain a plurality of coding features of the state feature set, wherein the plurality of state feature sets The coding features respectively have different feature scales and are respectively obtained from different convolutional layers;

   Classify the respective coding features of the multiple state feature sets according to feature scales to obtain the multiple coding sequences.
3. The device state detection method according to claim 1 or 2, wherein the first decoder includes a plurality of deconvolution layers respectively corresponding to the plurality of cyclic neural networks, and for each of the A state feature set, inputting all features to be reconstructed in the state feature set to the first decoder for reconstruction, and obtaining a reconstructed feature set of the state feature set, including:

   According to the corresponding relationship between the plurality of cyclic neural networks and the plurality of deconvolution layers, all the features to be reconstructed in the state feature set are respectively input to the corresponding deconvolution layers for reconstruction, and the obtained The reconstructed feature set of the state feature set.
4. The device state detection method according to any one of claims 1 to 3, wherein said obtaining multiple state feature sets of the device to be detected comprises:

   Acquiring a first state sequence of the device to be tested, wherein the first state sequence includes state data of a plurality of components of the device to be tested;

   splitting the first state sequence into a plurality of data segments;

   For each of the data segments, obtain the covariance matrix of the data segments;

   The covariance matrix of all the data segments is used as a plurality of state feature sets of the device to be detected.
5. The device state detection method according to any one of claims 1 to 4, wherein the device to be detected is determined according to the difference between the plurality of state feature sets and their respective reconstruction feature sets health status, including:

   Obtaining the first mean square error between the plurality of state feature sets and the respective reconstruction feature sets;

   If the first mean square error is greater than a first threshold, there is an abnormality in the device to be detected.
6. The device state detection method according to any one of claims 1 to 5, wherein the data processing device is further configured with a second detection model of the target component, and the second detection model includes a second encoder and a second decoder, the method further comprising:

   obtaining a second state sequence of the target component, wherein the second state sequence includes state data of the target component;

   performing feature extraction on the second state sequence by the second encoder to obtain first features to be reconstructed;

   Acquiring the second features to be reconstructed of the second state sequence, wherein the second features to be reconstructed include one of time domain features, frequency domain features and time-frequency domain features of the second state sequence or combination;

   inputting the concatenated features of the first feature to be reconstructed and the second feature to be reconstructed into the second decoder for reconstruction to obtain a reconstruction sequence of the second state sequence;

   A health state of the target component is determined based on a difference between the second state sequence and the reconstructed sequence.
7. The device state detection method according to claim 6, wherein the determining the health state of the target component according to the difference between the second state sequence and the reconstruction sequence comprises:

   obtaining a second mean square error between the second state sequence and the reconstructed sequence;

   If the second mean square error is greater than a second threshold, there is an abnormality in the target component.
8. A device state detection device, characterized in that it is applied to a data processing device, and the data processing device is configured with a first detection model, and the first detection model includes a first encoder, a first decoder, and a plurality of cyclic neurons network, the device status detection device includes:

   A feature acquisition module configured to acquire a plurality of state feature sets of the device to be detected, the plurality of state feature sets having a preset order, each of the state feature sets having respective states of a plurality of components of the device to be detected information;

   The feature extraction module is configured to perform feature extraction on the plurality of state feature sets from the spatial dimension through the first encoder to obtain a plurality of encoding sequences with different feature scales, wherein the plurality of encoding sequences Corresponding to different recurrent neural networks, each of the encoding sequences includes a plurality of encoding features in the preset order, and the plurality of encoding features respectively correspond to different sets of the state features;

   The feature extraction module is further configured to, for each of the coded sequences, perform feature extraction on the coded sequence from the time dimension through the corresponding cyclic neural network to obtain a sequence to be reconstructed, wherein the to-be-reconstructed sequence The reconstruction sequence includes a plurality of features to be reconstructed, and the plurality of features to be reconstructed respectively correspond to different sets of state features;

   The feature reconstruction module is configured to, for each of the state feature sets, input all the features to be reconstructed in the state feature set to the first decoder for reconstruction, and obtain the reconstruction of the state feature set Construct feature set;

   A state detection module configured to determine the health state of the device to be detected according to the difference between the plurality of state feature sets and the respective reconstructed feature sets.
9. A data processing device, characterized in that the data processing device includes a processor and a memory, the memory stores a computer program, and when the computer program is executed by the processor, any one of claims 1 to 7 is realized. A device state detection method described in one item.
10. A computer storage medium, characterized in that the computer storage medium stores a computer program, and when the computer program is executed by a processor, the device state detection method according to any one of claims 1 to 7 is implemented.

Images