CN114298443A - Predictive maintenance method, device and electronic equipment for industrial equipment based on health state index - Google Patents

Abstract

The application provides a health state index-based industrial equipment predictive maintenance method, a health state index-based industrial equipment predictive maintenance device and electronic equipment. And then, importing the health state index corresponding to the time point after the degradation point into a prediction model obtained by pre-training for fitting to obtain an extension curve, comparing the predicted health state data of each time point on the extension curve with a preset threshold value, and determining the time point with the same data as the failure time point. In the scheme, by using a pre-trained reconstruction model and a pre-trained prediction model, the determination of the degradation point and the prediction of the data can be accurately realized by learning the characteristics of the operating data, and the method can be suitable for predictive maintenance based on a small amount of data.

Description

Predictive maintenance method, device and electronic equipment for industrial equipment based on health state index

technical field

The present application relates to the technical field of industrial equipment management, and in particular, to a method, device and electronic equipment for predictive maintenance of industrial equipment based on a state of health index.

Background technique

Industrial equipment maintenance is of great significance to the economic benefits of manufacturing. The method of preventive maintenance, that is, regular maintenance, can fully prevent the economic losses caused by the downtime of the robot during the production process, but it increases the workload of the operation and maintenance personnel and causes the waste of a large number of parts and components, which increases the maintenance cost of the robot. Therefore, the industry is exploring predictive maintenance techniques for industrial equipment, in order to perform maintenance when equipment performance is at a minimum, thereby saving maintenance costs.

In the prior art, for the predictive maintenance of industrial equipment, multi-sensors are mainly used to collect multi-source data, and a single or composite health state indicator is constructed to predict the failure of industrial equipment. However, in the prior art, this method of setting multiple sensors to collect data for prediction is mainly applicable to large-scale equipment. When performing predictive maintenance on small equipment such as industrial robots, it becomes impractical to carry a large number of sensors due to the constraints of the actual working environment and the cost of the process itself and the sensors to be performed. Therefore, for small equipment, it is very important to accurately implement predictive maintenance without setting up a large number of sensors to collect multi-source data.

SUMMARY OF THE INVENTION

The objects of the present application include, for example, to provide a method, device and electronic device for predictive maintenance of industrial equipment based on a state of health index, which can accurately realize the determination of degradation points and the prediction of data.

The embodiments of the present application can be implemented as follows:

In a first aspect, the present application provides a predictive maintenance method for industrial equipment based on a state of health index, the method comprising:

Obtaining the operation data at each time point during the operation of the device under test, importing the operation data into the reconstruction model obtained by pre-training, and obtaining the reconstruction data corresponding to the operation data;

A health state index is obtained according to the reconstruction data and the operation data, and a degradation point in a time point is determined according to the health state index, and the degradation point represents the time point when the health state of the device under test begins to degrade;

The health state index corresponding to the time point after the degradation point is imported into the prediction model obtained by pre-training to fit the health state index, and an extension curve is obtained based on the fitting curve, and each of the extension curves is obtained. The predicted health state index at the time point is compared with a preset threshold, and a time point at which the predicted health state index and the preset threshold are the same is determined as a failure time point.

Health state index In an optional implementation manner, the step of importing the operating data into a pre-trained reconstruction model to obtain reconstructed data corresponding to the operating data includes:

For the obtained operation data of the device under test at multiple consecutive time points, intercept the operation data according to the preset step size and the preset window length, and obtain the operation data in multiple time windows;

For the operation data in each time window, zoom the operation data into a preset range;

Extract the time-domain eigenvectors, frequency-domain eigenvectors, and time-frequency-domain eigenvectors of the scaled operating data;

The time-domain feature vector, the frequency-domain feature vector, and the time-frequency-domain feature vector are imported into a pre-trained reconstruction model to obtain reconstruction data corresponding to the operating data.

In an optional implementation manner, the preset window length is greater than the preset step size.

In an optional implementation manner, the step of obtaining a health state index according to the reconstruction data and operation data, and determining a degradation point in a time point according to the health state index, includes:

Obtaining the difference data between the reconstruction data and the operation data, and using the difference data as a health state index;

The health state index is compared with the health state threshold, and the time point corresponding to the health state index starting to deviate from the health state threshold is determined as a degradation point.

In an optional implementation manner, the step of determining the time point corresponding to the health state index starting to deviate from the health state threshold as the degradation point includes:

obtaining a time point in the health state index that begins to deviate from the health state threshold;

It is detected whether the health state indices corresponding to the set number of time points after the time point all deviate from the health state threshold, and if they all deviate, the time point is determined to be a degradation point.

In an optional implementation manner, the method further includes the step of obtaining the reconstructed model based on the pre-built neural network model training, the neural network model includes an encoder and a decoder, and the step includes:

collecting sample data, where the sample data includes data corresponding to multiple consecutive time points;

Importing the sample data into the encoder for encoding processing to obtain characteristic data;

Importing the feature data and sample data into the decoder for fusion and decoding processing to obtain sample reconstruction data;

After adjusting the model parameters of the encoder and the decoder based on the loss function constructed according to the sample data and the sample reconstruction data, the training continues until the preset requirements are met, and the reconstruction model is obtained.

In an optional embodiment, the health state threshold is obtained in the following manner:

calculating the difference between the sample data and the sample reconstructed data;

Calculate the difference mean and the difference standard deviation based on the difference;

The health state threshold is obtained according to the difference mean and the difference standard deviation.

In an optional embodiment, the step of importing the health state index corresponding to the time point after the degradation point into the prediction model obtained by pre-training to obtain the prediction data in the next prediction period includes:

Obtain the health state index corresponding to the time point after the degradation point of the device under test;

Divide the health state index into time windows according to time series to obtain health state indices in multiple time windows;

Normalize the health status index within each time window;

Extracting data features of the normalized health state index, and importing the data features into a pre-trained prediction model to fit the health state index.

In a second aspect, the present application provides a device for predictive maintenance of industrial equipment based on a state of health index, the device comprising:

an acquisition module, configured to acquire operation data at various time points during the operation of the device under test, import the operation data into a reconstruction model obtained by pre-training, and obtain reconstruction data corresponding to the operation data;

A determination module, configured to obtain a health state index according to the reconstructed data and operation data, and determine a degradation point in a time point according to the health state index, where the degradation point indicates that the health state of the device under test begins to degrade time point;

The prediction module is used to import the health state index corresponding to the time point after the degradation point into the prediction model obtained by pre-training to fit the health state index, and obtain an extension curve based on the fitting curve. The predicted health state index at each time point in the extension curve is compared with a preset threshold, and a time point at which the predicted health state index and the preset threshold are the same is determined as a failure time point.

In a third aspect, the present application provides an electronic device, comprising one or more storage media and one or more processors in communication with the storage media, wherein the one or more storage media stores machine-executable instructions executable by the processor, When the electronic device is running, the processor executes the machine-executable instructions to perform the method steps described in any one of the preceding embodiments.

The beneficial effects of the embodiments of the present application include, for example:

The present application provides a method, device and electronic device for predictive maintenance of industrial equipment based on a health state index. By acquiring the operation data at various time points during the operation of the equipment to be tested, and importing the operation data into a pre-trained reconstruction model, the following results are obtained: For the reconstructed data corresponding to the operation data, the health state index is obtained according to the reconstructed data and the operation data, and the degradation point is determined according to the health state index. Then import the health state index corresponding to the time point after the degradation point into the pre-trained prediction model for fitting and obtain the extension curve, compare the predicted health state index at each time point on the extension curve with the preset threshold, and compare the two The consistent time point is determined as the failure time point. In this solution, the pre-trained reconstruction model and prediction model can be used to accurately determine the degradation point and predict the data by learning the characteristics of the running data, which can be applied to the predictive maintenance based on a small amount of data.

Description of drawings

In order to illustrate the technical solutions of the embodiments of the present application more clearly, the following drawings will briefly introduce the drawings that need to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore do not It should be regarded as a limitation of the scope, and for those of ordinary skill in the art, other related drawings can also be obtained according to these drawings without any creative effort.

1 is a flowchart of a predictive maintenance method provided by an embodiment of the present application;

2 is a schematic diagram of a fitting curve and an extension curve provided by an embodiment of the present application;

FIG. 3 is a flowchart of the sub-steps included in step S101 in FIG. 1;

4 is a schematic diagram of intercepting time window data in an embodiment of the present application;

FIG. 5 is a schematic diagram of processing a reconstructed model provided in an embodiment of the present application;

6 is a flowchart of a reconstruction model training method provided by an embodiment of the present application;

FIG. 7 is a flowchart of the sub-steps included in step S102 in FIG. 1;

FIG. 8 is a flowchart of the sub-steps included in step S1022 in FIG. 7;

FIG. 9 is a flowchart of the sub-steps included in step S103 in FIG. 1;

10 is a structural block diagram of an electronic device provided by an embodiment of the present application;

FIG. 11 is a block diagram of functional modules of an apparatus for predictive maintenance of industrial equipment based on a state of health index provided by an embodiment of the present application.

Icons: 110 - storage medium; 120 - processor; 130 - industrial equipment predictive maintenance device based on health state index; 131 - acquisition module; 132 - determination module; 133 - prediction module; 140 - communication interface.

Detailed ways

In order 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 described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments It is a part of the embodiments of the present application, but not all of the embodiments. The components of the embodiments of the present application generally described and illustrated in the drawings herein may be arranged and designed in a variety of different configurations.

Thus, the following detailed description of the embodiments of the application provided in the accompanying drawings is not intended to limit the scope of the application as claimed, but is merely representative of selected embodiments of the application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

It should be noted that like numerals and letters refer to like items in the following figures, so 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 features in the embodiments of the present application may be combined with each other without conflict.

Please refer to FIG. 1 , which is a flowchart of an industrial equipment predictive maintenance method based on a health state index provided by an embodiment of the present application. The method steps defined in the related process of the predictive maintenance method can be performed by an electronic device with data analysis and processing functions. realized by the device. The electronic device may be a computer device, or may be a server where the platform for maintaining the relevant functions of the industrial device is located. The specific flow shown in FIG. 1 will be described in detail below.

S101: Acquire operation data at each time point during the operation of the device under test, import the operation data into a pre-trained reconstruction model, and obtain reconstruction data corresponding to the operation data.

S102: Obtain a health state index according to the reconstruction data and the operation data, and determine a degradation point in a time point according to the health state index.

S103: Import the health state index corresponding to the time point after the degradation point into a prediction model obtained by pre-training to fit the health state index, and obtain an extension curve based on the fitting curve. The predicted health state index at each time point is compared with the preset threshold, and the time point at which the predicted health state index and the preset threshold are the same is determined as the failure time point.

In this embodiment, the device to be tested may be an industrial device, such as an industrial robot or the like. From the time when the device under test is put into use, generally, the performance state of the device under test is normal during the initial period of time. However, as the time of use increases, the performance state of the device under test may begin to degrade and eventually fail.

In this embodiment, for the device under test that is put into use, the current time point of predictive maintenance can be used as a node, and the operation data of each time point in the operation process of the device under test before the node is obtained. The time point may be a set sampling point, for example, an interval of 1 minute, 1 hour, etc. is not limited to be a sampling point.

The obtained operation data may include dynamic operation data and static operation data, wherein the dynamic operation data may include real-time current, torque, shaft angular position and other data during the operation of the device under test. The static operating data may include the parameters of the device under test, such as the number of axes, degrees of freedom, etc.

In this embodiment, the electromechanical control data inside the industrial equipment is used to implement predictive maintenance, and the current data and the like play an important role in control.

In this embodiment, a reconstruction model obtained by pre-training may also be obtained, and the reconstruction model is obtained by pre-training based on sample data. The reconstruction model can reflect the health state of the device in the form of data change characteristics by learning the characteristics of the change of the health state of the device embodied by the sample data. Therefore, in this embodiment, for the device under test, the operation data of the device under test can be imported into the reconstruction model obtained by pre-training, so as to output the reconstruction data corresponding to the operation data.

It can be seen from the above that the reconstructed data can reflect the situation related to the health state of the device under test during operation. The health state-related situation can be reflected in the difference between the reconstructed data and the operational data. That is, the reconstructed data can be understood as characteristic data that fits the normal health state, so the difference between the operation data and the reconstructed data can reflect the health state of the device under test.

In this embodiment, the health state index is obtained according to the reconstruction data and the operation data. The health state index is a series of data in time series, that is, it includes the health state index corresponding to each time point. Based on the analysis of the health state index corresponding to each time point, the degradation point in the time point can be determined.

In this embodiment, the degradation point represents the time point when the health state of the device under test begins to degrade. That is, it can be understood that at each time point before the degradation point, the operation data of the device under test is data belonging to the normal health state, and at each time point after the degradation point, the operation of the device under test begins to decline. The phenomenon. However, recession does not mean failure, and it generally takes a period of time from the beginning of the operation of the device under test to the failure state. The operating data from the time point of the beginning of the recession can provide an effective data prediction basis for the prediction of the failure point.

In this embodiment, a prediction model may be obtained by pre-training, and the prediction model may be obtained by pre-training based on sample data. The sample data may be related data after the degradation point of the sampled device. Therefore, the prediction model can learn the relevant features of the data after the degradation point, so as to accurately predict the failure point based on the learned relevant features.

Therefore, in this embodiment, for the device under test, after determining the degradation point in the operation of the device under test, the health state index corresponding to the time point after the degradation point can be imported into the prediction model for prediction. The prediction model can fit the health state data, and on the basis of obtaining the fitting curve, it can be further extended based on the fitting curve to obtain the extension curve. The extension curve also includes the predicted health state index at multiple time points.

A preset threshold can be set, and based on the preset threshold, it can be determined whether the predicted health state index indicates that the device under test is in a failure state. For example, on the prediction curve, if the predicted health state index reaches the same condition as the preset threshold, the corresponding time point may be determined as the failure time point.

For example, please refer to Figure 2. For example, the time point 20-06-06 is the determined degradation point, and the data collected from the degradation point in the time period from 20-06-06 to 21-11-28 is after the degradation point health status index at each time point. The prediction model can fit the health state index in this time period, and then obtain the fitting curve in the time period from 20-06-06 to 21-11-28 as shown in the figure.

Based on the curve trend of the fitted curve, an extension curve can be obtained by extension, for example, the extension curve in the time period from 21-11-28 to 23-05-22 in the figure. Wherein, the constructed extension curve may also be a plurality of extension curves within a certain error range.

The value represented by the horizontal dotted line in the figure can be the preset threshold value, and it can be determined as failure at the time point when the extension curve intersects the preset threshold value, that is, the time point when the predicted health state index is consistent with the preset threshold value point in time. That is, the device under test is predicted to fail at that point in time.

The predictive maintenance method provided in this embodiment, using the pre-trained reconstruction model and prediction model, can accurately determine the degradation point and predict the data by learning the characteristics of the running data, and can be applied to the situation based on a small amount of data. predictive maintenance.

In this embodiment, the acquired operation data is simple data such as current and torque of the device to be tested, and it is difficult to reflect the characteristics of the data in terms of time series. In order to facilitate model learning or obtain the characteristics of time series data, before importing the running data into the reconstructed model, certain processing can be performed on the running data. Referring to FIG. 3 , in this embodiment, when the operating data is processed based on the reconstruction model, the following methods can be used:

S1011 , with respect to the acquired operation data of the device under test at multiple consecutive time points, intercept the operation data according to a preset step size and a preset window length to obtain operation data within multiple time windows.

S1012, for the operation data in each time window, zoom the operation data into a preset range.

S1013 , extract the time-domain feature vector, the frequency-domain feature vector, and the time-frequency-domain feature vector of the scaled operating data.

S1014: Import the time-domain feature vector, the frequency-domain feature vector, and the time-frequency-domain feature vector into a pre-trained reconstruction model to obtain reconstruction data corresponding to the operating data.

In this embodiment, formal definition processing may be performed first for the acquired operation data, and for each time point t , the data of various types of operation data at this time point t may be represented as the following form:

Among them, m represents the number of components to be tested in the device under test, and the health state at this time point t can be represented by h _{k ,} which can indicate that the health state of the device under test at this time point is a normal state or a fault state. For example, when h _k is 1, it indicates that the device under test is in a normal state, and when h _k is 0, it indicates that the device under test is in a fault state. This is only for illustration, and this embodiment is not limited thereto.

In this way, the operation data of the device under test at multiple consecutive time points in time series can be obtained. On this basis, in order to facilitate the processing of data by the model, the running data can be divided into data segments within multiple time windows for input into the model. In this embodiment, the operation data may be intercepted according to the preset step size and the preset window length, so as to obtain the operation data in multiple time windows.

In order to ensure that the intercepted operation data is generally continuous, in this embodiment, the preset window length may be greater than the preset step size. In this way, in the two adjacent time windows, the data of the last part of the previous time window overlaps with the data of the previous part of the latter time window, which can effectively ensure that the running data in different time windows is continuously chopped. As shown in FIG. 4 , for example, the preset step size may be 2, and the preset window length may be 7.

In this embodiment, in order to better analyze the nature of the distribution of various types of operation data, the operation data may also be standardized, and the operation data in each time window may be scaled within a preset range.

In one possible implementation, the run data can be normalized using the z-score method (zero-mean normalization). The preset range may be an interval range with a mean value of 0 and a standard deviation of 1. That is, make the scaled running data fall within an interval with a mean of 0 and a standard deviation of 1.

In this embodiment, scaling processing can be performed on the running data according to the following scaling formula.

表示缩放前运行数据的平均值，

表示缩放前运行数据的标准差。where x is the running data before scaling, z is the running data after scaling, N is the total number of running data,

represents the mean of the running data before scaling,

Represents the standard deviation of the running data before scaling.

In this embodiment, z-score normalization may be performed on each type of operating data in the above-mentioned manner. The distribution of the data itself has not changed in the scaled running data, but the data distribution interval after scaling can remain basically the same, and the data can be mainly distributed in the [-2, 2] interval. By scaling the running data, subsequent models can focus more on analyzing the distribution of the data itself.

In order to further enable the model to analyze and obtain the distribution of the operating data from multiple dimensions, in this embodiment, features in multiple dimensions of the scaled operating data can be extracted, including, for example, a time-domain feature vector, a frequency-domain feature vector, and a time-domain feature vector. Frequency domain eigenvectors.

In this embodiment, for the time domain feature extraction, feature analysis can be directly performed on the operating data of each time window. Multiple time-domain features can be extracted mainly from the time-domain perspective, including RMS, root square amplitude, peak-to-peak value, crest factor, margin index, skewness index, kurtosis index, form factor, impulse factor, information entropy, correlation Sex coefficient. The multiple time-domain features are spliced into a vector, and the time-domain feature vector is obtained as follows:

In addition, frequency domain feature extraction can also be performed. It can be known from Baševal's theorem that whether it is a real signal or a complex signal, the integral of the square of the signal amplitude is equal to the energy of the signal and equal to the square of the modulus of the signal spectral density. The formula can be expressed as follows:

Among them, E represents the signal energy, x ( t ) represents the signal time domain value, and X ( f ) represents the signal frequency domain value.

Therefore, the energy of the high frequency discrete signal can be obtained by accumulating the squares of each value in the running data, as follows:

Among them, xf _e represents a characteristic component as a whole, and e represents an energy signal, which corresponds to rms and sra in the above-mentioned x _rms and x _sra . f ( i ) represents the i -th value of a time-domain signal such as x _rms . The above formula can be understood as the frequency domain feature component on the left side of the equal sign, which is equal to the sum of the modulo squares of the corresponding time domain feature components.

In this way, the frequency domain feature vector can be obtained as follows:

、瞬时频率的信噪比SNR2个特征值，将它们进行拼接，得到信号4n+2维的时频域特征：On the basis of the above, time-frequency domain feature analysis can be performed. In this embodiment, the time-frequency analysis can be performed by using the EDM method, the short-time Fourier transform method, and the like. First, the n eigenmode functions (IMFs) related to the fault of the equipment under test in each time window are obtained through EDM. Then use EDM to take energy xtf _e , variance xtf _sd , skewness index xtf _sf and kurtosis index xtf _kf for the n IMFs screened by 4 types of eigenvalues, and use STFT to get the standard deviation of the instantaneous frequency

, the signal-to-noise ratio SNR of the instantaneous frequency is two eigenvalues, and they are spliced to obtain the 4n+2-dimensional time-frequency domain characteristics of the signal:

Among them, the above xtf as a whole represents a component.

The time-domain feature vector, frequency-domain feature vector, and time-frequency-domain feature vector obtained above are imported into the reconstruction model to obtain reconstruction data corresponding to the running data.

In this embodiment, the reconstructed model is obtained by pre-training based on sample data, and the predictive maintenance method provided in this embodiment further includes the step of obtaining a reconstructed model by pre-training based on the constructed neural network model, wherein the neural network model Can be an LSTM model. The neural network model includes an encoder and a decoder, as shown in Figure 5. Please refer to Figure 6 in conjunction with the steps of pre-training to obtain a reconstructed model, which can be implemented in the following ways:

S201. Collect sample data, where the sample data includes data corresponding to multiple consecutive time points.

S202, import the sample data into the encoder for encoding processing to obtain characteristic data.

S203, importing the feature data and sample data into the decoder for fusion and decoding processing to obtain sample reconstruction data.

S204: Continue training after adjusting the model parameters of the encoder and the decoder based on the loss function constructed according to the sample data and the sample reconstruction data, and obtain the reconstruction model when a preset requirement is met.

In this embodiment, the sample data may be operation data corresponding to continuous time points during the operation of the industrial equipment. Similarly, the sample data can be processed according to the above-mentioned preprocessing, scaling processing, time-domain feature extraction, frequency-domain feature extraction, and time-frequency-domain feature extraction.

。该细胞状态也可被称为上下文向量(Context Vectors)，而解码器通过上下文向量来重构编码器的输入。解码器和编码器一样，也是一个LSTM单元。解码器中每一步的输入是上一步的预测或者是上一步的标签，解码器更新隐状态可以描述为

。The sample data after the above processing is imported into the encoder of the neural network model for encoding processing to obtain characteristic data. In this embodiment, the structures of the encoder and the decoder are each an LSTM unit. LSTM can take a period of time series data as input, and then update its hidden state until the last step of the time series, denoted as t2, the cell state generated by LSTM contains all the information of the previous sequence, namely

. The cell states may also be referred to as Context Vectors, and the decoder uses the context vectors to reconstruct the encoder's input. The decoder, like the encoder, is also an LSTM unit. The input of each step in the decoder is the prediction of the previous step or the label of the previous step, and the updated hidden state of the decoder can be described as

.

The feature data obtained by the encoder is combined with the sample data, including the time domain features, frequency domain features and time-frequency domain features of the sample data, and the decoder performs fusion and decoding processing to obtain sample reconstruction data.

A loss function can be constructed based on sample data and sample reconstruction data, and the constructed loss function can be as follows:

表示样本数据，

表示样本重构数据，t ₁和t ₂分别表示时间序列数据的开始时间点和结束时间点。in,

represents sample data,

represents the sample reconstruction data, and t ₁ and t ₂ represent the start time point and end time point of the time series data, respectively.

The training of the encoder and the decoder can have multiple iterations. After each iteration, the function value of the above-mentioned loss function can be calculated, and the model parameters of the encoder and the decoder can be adjusted to continue the training. When the number of iterations reaches the set maximum number, or the loss function reaches convergence and no longer decreases, or when the iteration time reaches the set maximum duration, it can be determined that the preset requirements are met, and the result obtained by the neural network model at this time can be obtained. the reconstructed model.

The above is the process of pre-training to obtain a reconstructed model. When using the reconstructed model to reconstruct the operating data of the device under test and determine the degradation point, first use the reconstructed model to obtain the reconstructed data corresponding to the operating data, and then use the reconstructed model to obtain the reconstructed data corresponding to the operating data. The data obtains a health state index, and the degradation point in the time point is determined according to the health state index. Referring to FIG. 7, in this embodiment, the step of determining the degradation point may be implemented in the following manner:

S1021: Obtain difference data between the reconstructed data and the operation data, and use the difference data as a health state index.

S1022: Compare the health state index with a health state threshold, and determine a time point corresponding to the health state index starting to deviate from the health state threshold as a degradation point.

In this embodiment, the health state index may be the difference between the actual operating data and the data reconstructed by the reconstructed model (considered normal). Therefore, as the reconstruction error increases, it means that the operating state deviates from the normal state.

In this embodiment, a health state threshold may be preset as a criterion for judging whether the health state of the device under test is abnormal. The health state threshold may be set based on the relevant data in the process of constructing the reconstruction model based on the sample data in advance. In this embodiment, the health state threshold can be constructed in the following manner:

Calculate the difference between the sample data and the sample reconstructed data, calculate the difference average and the difference standard deviation based on the difference, and obtain the health state threshold according to the difference average and the difference standard deviation.

In this embodiment, the specific calculation formula of the health state threshold may be as follows:

Among them, mean means taking the mean value, std means taking the standard deviation, and ‖‖ ₂ means calculating the L2 norm.

When the health state index deviates from the health state threshold, that is, when the difference between the operating data and the reconstructed data exceeds the health state threshold, the corresponding time point may be determined as a degradation point.

Considering that in the actual processing process, there may be some mutation points in the running data, resulting in some mutation points in the obtained health state index. If the health state index corresponding to the mutation point has a mutation in the data characteristics, it may cause the health state index at the corresponding time point to deviate from the health state threshold, thereby being misjudged as a degradation point. Therefore, referring to FIG. 8 , in this embodiment, in the above step of determining the degradation point, it can be implemented in the following manner:

S10221: Acquire a time point in the health state index that starts to deviate from the health state threshold.

S10222: Detect whether the health state indices corresponding to the set number of time points after the time point all deviate from the health state threshold, if all deviate from the health state threshold, execute the following step S10223, if not, execute the following step S10224 .

Step S10223, determining that the time point is a degradation point.

Step S10224, it is determined that the time point is not a degradation point.

In this embodiment, if the corresponding health state index starts to deviate from the health state threshold from a certain time point, then it can be determined as 5 time points, 10 time points, etc., which are not limited. The health status index corresponding to each subsequent time point can be obtained, and then it is detected whether each health status index deviates from the health status threshold. If each health status index deviates from the health status threshold, it indicates that there is a long time period. of data consistently deviates from the health state threshold and is not an accidental deviation due to a mutation in the data. Therefore, in this case, the above-mentioned time point at which the deviation from the state-of-health threshold is started can be determined as the degradation point.

However, if the health state index of the set number of time points after it does not all deviate from the health state threshold from the time point when it starts to deviate from the health state threshold, it means that the health state index at the above time point may only be due to data mutation resulting deviation. Therefore, in this case, it can be determined that the above time point is not a degradation point.

In a possible implementation manner, the Laida's rule may be used to remove abnormal points in a series of health state indices in time series, that is, to remove health state indices with data mutation. Thus, the actual degradation point at which the device under test begins to degrade can be found, and subsequent failure points can be predicted based on the health state index after the real degradation point.

Referring to FIG. 9 , in this embodiment, when performing data fitting prediction based on the health state index corresponding to the time point after the degradation point, the following methods can be used:

S1031 , acquiring a health state index corresponding to a time point after the degradation point of the device under test.

S1032: Divide the health state index into time windows according to time series to obtain health state indices in multiple time windows.

S1033, normalize the health state index in each time window.

S1034, extracting data features of the normalized health state index, and importing the data features into a pre-trained prediction model to fit the health state index.

In this embodiment, for the health state index at each time point after the degradation point of the device under test, the health state index may be intercepted according to a certain window length and a certain interception step size. Wherein, similarly, in order to ensure the continuity of the health state index of each intercepted time window, the window length may be greater than the interception step size.

For the intercepted health state index in each time window, the health state index may be normalized to a certain uniform value range according to the above-mentioned scaling processing method for the operation data. So that the prediction model can focus on the distribution characteristics of the data itself.

Data features of the normalized health state index may be extracted, and the data features may include, for example, time domain features, frequency domain features, time-frequency domain features, and the like. The data features of the health state index are imported into the prediction model obtained by pre-training, and the prediction model can fit the health state index to obtain a fitting curve. Then, the failure time point is determined based on the extension curve of the fitted curve.

In this embodiment, the prediction model may be obtained by training a neural network model constructed in advance based on sample data. For example, the health state index corresponding to the time point after the degradation point of the industrial equipment is used as the sample data, and the neural network model may be a GRU (Gate Recurrent Unit) network model.

On this basis, the predicted health index obtained each time is compared with a preset threshold, which may be set based on known operating conditions of industrial equipment of the same type running the same process as the equipment to be tested. When the predicted health state index is consistent with the preset threshold, it can be considered that the corresponding time point is the failure time point.

The remaining service life of the device under test can be determined based on the failure time point. For example, the time point when the prediction model is used to fit the state of health index is a node, and the time period from the node to the predicted failure time point , which is the remaining service life of the device under test.

The overall flow of the predictive maintenance method provided in this embodiment is described below.

In this embodiment, sample data may be collected in advance, and the sample data may be data corresponding to multiple consecutive time points. The sample data can be real-time current, torque, and shaft angular position values during the operation of the industrial equipment, as well as the body parameters of the industrial equipment.

Data preprocessing can be performed on the sample data, for example, using a certain step size and intercepting the sample data according to a certain window size to obtain sample data in multiple windows. In addition, data standardization can be performed, for example, the sample data in each window is scaled to a preset range, such as a range of a certain mean and a certain standard deviation.

Then perform feature extraction on the scaled data, including time-domain feature extraction, frequency-domain feature extraction, and time-frequency-domain feature extraction.

The time domain features, frequency domain features, and time-frequency domain features obtained based on the sample data are imported into the constructed neural network model to train the neural network model to obtain a reconstructed model. During the training process, the loss function constructed based on the difference information between the input and the output can be used as a training guide, and the training is stopped when the iteration meets certain requirements.

In the process of training the reconstructed model, the health state threshold may also be constructed based on the difference between the sample data input into the reconstructed model and the sample reconstructed data output by the reconstructed model. This state of health threshold can subsequently be used to determine the degradation point of the industrial equipment.

On this basis, the health state index can be obtained based on the sample reconstruction data and sample data obtained by the reconstruction model, and then the time point when the health state index begins to degenerate for the first time is found as the degradation point.

Based on the health state index after the degradation point, the deep model of the GRU network can be trained to obtain a prediction model.

On this basis, in the actual application stage, for the device under test, the operation data of the device under test can be obtained. The above-mentioned data preprocessing, time window extraction processing, scaling processing, as well as time domain feature processing, frequency domain feature processing, and time-frequency domain feature processing, etc. are performed on the operating data.

Then, the above-mentioned time-domain features, frequency-domain features, and time-frequency-domain features are imported into the reconstruction model to obtain corresponding reconstruction data. The health state index of the device under test is obtained by combining the reconstructed data and the sample data of the device under test.

When performing predictive maintenance on the device under test, the operation data of the device under test can be imported into the reconstruction model to obtain the corresponding reconstruction data. The health state index can be obtained according to the reconstruction data and the operation data. By analyzing and processing the health state index, a degradation point can be obtained that can characterize the degeneration of the health state of the device under test.

Import the health state index after the degradation point into the prediction model and fit the health state index to obtain a fitting curve. An extension curve is obtained by extending based on the fitted curve, and each time point on the extension curve has a corresponding predicted health state index.

Each predicted health state index is compared with a preset threshold, and when the predicted health state index is the same as the preset threshold, the corresponding time point is determined as the predicted failure time point. The connection point between the fitting curve and the extension curve is obtained, that is, the difference between the time point when the prediction model is used for prediction and the predicted failure time point, which is the predicted remaining service life of the equipment under test.

The predictive maintenance method provided in this embodiment adopts the health state index as an indicator of the predictive maintenance of industrial equipment based on the health state index, which reduces the dependence of the predictive maintenance on various sensors and reduces the actual performance of the predictive maintenance technology. application cost.

In addition, the reconstruction model including the encoder and the decoder is used to output the reconstructed data and then construct the health state index. The value of the health state index can be extracted from the time series signal, which reduces the cost of building the model, and furthermore accurately determines the degradation point on the basis of It can provide data basis for the accurate prediction of subsequent failure points.

When predicting the failure point, the features of the time series basis are extracted by GRU deep learning, and the prediction of the health state index and the calculation of the remaining service life are performed to improve the accuracy of the health state monitoring.

Please refer to FIG. 10 , which is a schematic diagram of an exemplary component of an electronic device according to an embodiment of the present application. The electronic device may include a storage medium 110 , a processor 120 , an industrial equipment predictive maintenance device 130 based on a state of health index, and a communication interface 140 . In this embodiment, both the storage medium 110 and the processor 120 are located in the electronic device and are separately provided. However, it should be understood that the storage medium 110 may also be independent of the electronic device, and may be accessed by the processor 120 through a bus interface. Alternatively, the storage medium 110 may also be integrated into the processor 120, for example, may be a cache and/or a general purpose register.

The industrial equipment predictive maintenance device 130 based on the health state index can be understood as the above-mentioned electronic equipment, or the processor 120 of the electronic equipment, and can also be understood as implementing the above-mentioned electronic equipment independently of the above-mentioned electronic equipment or the processor 120 under the control of the electronic equipment Software functional modules for predictive maintenance methods.

As shown in FIG. 11 , the above-mentioned apparatus 130 for predictive maintenance of industrial equipment based on the state of health index may include an acquisition module 131 , a determination module 132 and a prediction module 133 . The functions of each functional module of the industrial equipment predictive maintenance device 130 based on the health state index will be described in detail below.

The obtaining module 131 is configured to obtain the operation data at each time point in the operation of the device under test, import the operation data into the reconstruction model obtained by pre-training, and obtain the reconstruction data corresponding to the operation data.

It can be understood that the obtaining module 131 may be configured to execute the above-mentioned step S101, and for the detailed implementation of the obtaining module 131, reference may be made to the above-mentioned content related to the step S101.

A determination module 132, configured to obtain a health state index according to the reconstruction data and the operation data, and determine a degradation point in a time point according to the health state index, where the degradation point indicates that the health state of the device under test begins to appear Degenerate time point.

It can be understood that the determining module 132 may be configured to execute the above-mentioned step S102, and for the detailed implementation of the determining module 132, reference may be made to the above-mentioned content related to the step S102.

The prediction module 133 is used to import the health state index corresponding to the time point after the degradation point into a prediction model obtained by pre-training to fit the health state index, and obtain an extension curve based on the fitting curve, The predicted health state index at each time point in the extension curve is compared with a preset threshold, and a time point at which the predicted health state index and the preset threshold are the same is determined as a failure time point.

It can be understood that the prediction module 133 can be used to execute the above-mentioned step S103, and for the detailed implementation of the prediction module 133, please refer to the above-mentioned content related to the step S103.

In a possible implementation manner, the above acquisition module 131 may be used for:

For the obtained operation data of the device under test at multiple consecutive time points, intercept the operation data according to the preset step size and the preset window length, and obtain the operation data in multiple time windows;

For the operation data in each time window, zoom the operation data into a preset range;

Extract the time-domain eigenvectors, frequency-domain eigenvectors, and time-frequency-domain eigenvectors of the scaled operating data;

The time-domain feature vector, the frequency-domain feature vector, and the time-frequency-domain feature vector are imported into a pre-trained reconstruction model to obtain reconstruction data corresponding to the operating data.

In a possible implementation manner, the preset window length is greater than the preset step size.

In a possible implementation manner, the above determination module 132 may be used to:

Obtaining the difference data between the reconstruction data and the operation data, and using the difference data as a health state index;

The health state index is compared with the health state threshold, and the time point corresponding to the health state index starting to deviate from the health state threshold is determined as a degradation point.

In a possible implementation manner, the above determination module 132 may be used to:

obtaining a time point in the health state index that begins to deviate from the health state threshold;

It is detected whether the health state indices corresponding to the set number of time points after the time point all deviate from the health state threshold, and if they all deviate, the time point is determined to be a degradation point.

In a possible implementation manner, the apparatus 130 for the predictive maintenance of industrial equipment based on the health state index further includes a building module for obtaining the reconstructed model based on a pre-built neural network model training, where the neural network model includes Encoders and Decoders, this building block can be used to:

collecting sample data, where the sample data includes data corresponding to multiple consecutive time points;

Importing the sample data into the encoder for encoding processing to obtain characteristic data;

Importing the feature data and sample data into the decoder for fusion and decoding processing to obtain sample reconstruction data;

After adjusting the model parameters of the encoder and the decoder based on the loss function constructed according to the sample data and the sample reconstruction data, the training continues until the preset requirements are met, and the reconstruction model is obtained.

In a possible implementation manner, the apparatus 130 for the predictive maintenance of industrial equipment based on the state of health index further includes an obtaining module for obtaining the state of health threshold, and the obtaining module may be used for:

calculating the difference between the sample data and the sample reconstructed data;

Calculate the difference mean and the difference standard deviation based on the difference;

The health state threshold is obtained according to the difference mean and the difference standard deviation.

In a possible implementation manner, the above prediction module 133 may be used to:

Obtain the health state index corresponding to the time point after the degradation point of the device under test;

Divide the health state index into time windows according to time series to obtain health state indices in multiple time windows;

Normalize the health status index within each time window;

Extracting data features of the normalized health state index, and importing the data features into a pre-trained prediction model to fit the health state index.

For the description of the processing flow of each module in the apparatus and the interaction flow between the modules, reference may be made to the relevant descriptions in the foregoing method embodiments, which will not be described in detail here.

Further, the embodiments of the present application further provide a computer-readable storage medium, where the computer-readable storage medium stores machine-executable instructions, and when the machine-executable instructions are executed, implement the predictive maintenance method provided by the foregoing embodiments.

Specifically, the computer-readable storage medium can be a general-purpose storage medium, such as a removable disk, a hard disk, etc., when the computer program on the computer-readable storage medium is executed, the above-mentioned predictive maintenance method can be executed. For the processes involved when the computer-readable storage medium and its executable instructions are executed, reference may be made to the relevant descriptions in the foregoing method embodiments, which will not be described in detail here.

To sum up, the method, device and electronic device for the predictive maintenance of industrial equipment based on the health state index provided by the embodiments of the present application obtain the operation data at various time points during the operation of the equipment under test, and import the operation data into pre-training to obtain the result. According to the reconstruction model, the reconstruction data corresponding to the operation data is obtained, the health state index is obtained according to the reconstruction data and the operation data, and the degradation point is determined according to the health state index. Then import the health state index corresponding to the time point after the degradation point into the pre-trained prediction model for fitting and obtain the extension curve, compare the predicted health state index at each time point on the extension curve with the preset threshold, and compare the two The consistent time point is determined as the failure time point. In this solution, the pre-trained reconstruction model and prediction model can be used to accurately determine the degradation point and predict the data by learning the characteristics of the running data, which can be applied to the predictive maintenance based on a small amount of data.

The above are only specific embodiments of the present application, but the protection scope of the present application is not limited to this. Any person 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 subject to the protection scope of the claims.

Claims (10)

1. A method for predictive maintenance of industrial equipment based on a state of health index, wherein the method comprises: Obtaining the operation data at each time point during the operation of the device under test, importing the operation data into the reconstruction model obtained by pre-training, and obtaining the reconstruction data corresponding to the operation data; A health state index is obtained according to the reconstruction data and the operation data, and a degradation point in a time point is determined according to the health state index, and the degradation point represents the time point when the health state of the device under test begins to degrade; The health state index corresponding to the time point after the degradation point is imported into the prediction model obtained by pre-training to fit the health state index, and an extension curve is obtained based on the fitting curve, and each of the extension curves is obtained. The predicted health state index at the time point is compared with a preset threshold, and a time point at which the predicted health state index and the preset threshold are the same is determined as a failure time point. 2 . The method for predictive maintenance of industrial equipment based on health state index according to claim 1 , wherein the operation data is imported into a pre-trained reconstruction model to obtain a reconstruction corresponding to the operation data. 3 . Data steps, including: For the obtained operation data of the device under test at multiple consecutive time points, intercept the operation data according to the preset step size and the preset window length, and obtain the operation data in multiple time windows; For the operation data in each time window, zoom the operation data into a preset range; Extract the time-domain eigenvectors, frequency-domain eigenvectors, and time-frequency-domain eigenvectors of the scaled operating data; The time-domain feature vector, the frequency-domain feature vector, and the time-frequency-domain feature vector are imported into a pre-trained reconstruction model to obtain reconstruction data corresponding to the operating data. 3 . The method for predictive maintenance of industrial equipment based on a state of health index according to claim 2 , wherein the preset window length is greater than the preset step size. 4 . 4 . The method for predictive maintenance of industrial equipment based on health state index according to claim 1 , wherein the health state index is obtained according to the reconstruction data and the operation data, and the time is determined according to the health state index. 5 . Steps for degenerating points in points, including: Obtaining the difference data between the reconstruction data and the operation data, and using the difference data as a health state index; The health state index is compared with the health state threshold, and the time point corresponding to the health state index starting to deviate from the health state threshold is determined as a degradation point. 5 . The method for predictive maintenance of industrial equipment based on the state of health index according to claim 4 , wherein the step of determining the time point corresponding to the state of health index starting to deviate from the state of health threshold value as a degradation point. 6 . ,include: obtaining a time point in the health state index that begins to deviate from the health state threshold; It is detected whether the health state indices corresponding to the set number of time points after the time point all deviate from the health state threshold, and if they all deviate, the time point is determined to be a degradation point. 6 . The method for predictive maintenance of industrial equipment based on health state index according to claim 4 , wherein the method further comprises the step of obtaining the reconstructed model based on a pre-built neural network model training. The network model includes an encoder and a decoder, and this step includes: collecting sample data, where the sample data includes data corresponding to multiple consecutive time points; Importing the sample data into the encoder for encoding processing to obtain characteristic data; Importing the feature data and sample data into the decoder for fusion and decoding processing to obtain sample reconstruction data; After adjusting the model parameters of the encoder and the decoder based on the loss function constructed according to the sample data and the sample reconstruction data, the training continues until the preset requirements are met, and the reconstruction model is obtained. 7. The method for predictive maintenance of industrial equipment based on a state of health index according to claim 6, wherein the state of health threshold is obtained in the following manner: calculating the difference between the sample data and the sample reconstructed data; Calculate the difference mean and the difference standard deviation based on the difference; The health state threshold is obtained according to the difference mean and the difference standard deviation. 8 . The method for predictive maintenance of industrial equipment based on health state index according to claim 1 , wherein the health state index corresponding to the time point after the degradation point is imported into a pre-trained prediction model to obtain 8 . The step of fitting the health state index to the health state index includes: Obtain the health state index corresponding to the time point after the degradation point of the device under test; Divide the health state index into time windows according to time series to obtain health state indices in multiple time windows; Normalize the health status index within each time window; Extracting data features of the normalized health state index, and importing the data features into a pre-trained prediction model to fit the health state index. 9. A device for predictive maintenance of industrial equipment based on a state of health index, wherein the device comprises: an acquisition module, configured to acquire operation data at various time points during the operation of the device under test, import the operation data into a reconstruction model obtained by pre-training, and obtain reconstruction data corresponding to the operation data; A determination module, configured to obtain a health state index according to the reconstructed data and operation data, and determine a degradation point in a time point according to the health state index, where the degradation point indicates that the health state of the device under test begins to degrade time point; The prediction module is used to import the health state index corresponding to the time point after the degradation point into the prediction model obtained by pre-training to fit the health state index, and obtain an extension curve based on the fitting curve. The predicted health state index at each time point in the extension curve is compared with a preset threshold, and a time point at which the predicted health state index and the preset threshold are the same is determined as a failure time point. 10. An electronic device, characterized by comprising one or more storage media and one or more processors in communication with the storage media, wherein the one or more storage media stores machine-executable instructions executable by the processor, when When the electronic device is running, the processor executes the machine-executable instructions to perform the method steps of any one of claims 1-8.

Images