CN118036435A

CN118036435A - Method, device and storage medium for predicting remaining usable life of compressor

Info

Publication number: CN118036435A
Application number: CN202410024651.8A
Authority: CN
Inventors: 朱汪友; 刘保侠; 朱喜平; 李刚; 拜禾; 张盟; 赵洪亮; 谷思宇; 刘白杨; 王猛
Original assignee: China Oil and Gas Pipeline Network Corp; National Pipe Network Group North Pipeline Co Ltd
Current assignee: China Oil and Gas Pipeline Network Corp; National Pipe Network Group North Pipeline Co Ltd
Priority date: 2024-01-08
Filing date: 2024-01-08
Publication date: 2024-05-14

Abstract

The invention relates to a method, a device and a storage medium for predicting the residual service life of a compressor, wherein the method comprises the following steps: acquiring a time sequence data sample of a compressor vibration signal; extracting and fusing the characteristics of the time sequence data samples to obtain comprehensive characteristics; mapping the comprehensive features into text features, and determining word frequency features of the text features; predicting the running states of the compressor at different time points in the future by using a state prediction model based on word frequency characteristics; predicting the remaining life of the compressor at a corresponding point in time using a life prediction model with delay prediction constraints based on the operating conditions; performing linear regression on the percentage value of the residual life to obtain the total life of the compressor; and determining the actual residual usable life of the compressor according to the total life and the current working time. The invention can reflect the early failure of the compressor, can more accurately reflect the fine change of data, and can effectively represent the degradation trend in the later period of the life of the compressor; and the accuracy of equipment life prediction can be effectively improved.

Description

Method, device and storage medium for predicting remaining usable life of compressor

Technical Field

The present invention relates to the field of compressor technologies, and in particular, to a method and apparatus for predicting a remaining useful life of a compressor, and a storage medium.

Background

The large-scale compressor is used as an indispensable energy conversion device in basic industry, and has wide application in the fields of civil engineering, ships, water conservancy, petroleum, chemical engineering, machinery, smelting, refrigeration, mine ventilation and the like. Along with the improvement of society and the improvement of scientific technology level, industry has higher and higher requirements on the structural design, the safety and the reliability, the production efficiency and the like of the compressor, so that the development degree of the compressor industry directly influences the economic benefit and the operation safety of the related industrial field. Accurate residual life prediction can avoid compressor failure, and is the key of a maintenance mode from passive to active.

The conventional compressor life prediction method generally predicts the residual life of the compressor by establishing a mathematical and physical model, but is difficult to construct a specific model in the face of various working states and complicated internal parts. In recent years, neural networks have rapidly developed, and various CNNs, RNNs and neural networks based on attention mechanisms have emerged. These neural networks are highly popular in the image and natural language processing fields, but they are difficult to apply directly to the problem of residual life prediction of compressors. If the degradation indexes are different in performance in different time periods, the model life prediction result based on the neural network is often unstable; the lifetime prediction problem is less tolerant to delay prediction, whereas the conventional loss function is tolerant to delay prediction and advance prediction.

Therefore, how to improve the stability of the model life prediction result and the tolerance to delay prediction is a problem to be solved.

Disclosure of Invention

The invention aims to solve the technical problems existing in the prior art and provides a method and a device for predicting the residual usable life of a compressor and a storage medium.

In order to solve the above technical problems, an embodiment of the present invention provides a method for predicting a remaining usable life of a compressor, including:

acquiring time sequence data samples of a compressor vibration signal, and preprocessing the time sequence data samples;

performing feature extraction and feature fusion on the preprocessed time sequence data samples to obtain comprehensive features representing the health state of the compressor;

mapping the comprehensive features into text features, and determining word frequency features of the text features;

Predicting the running states of the compressor at different time points in the future by using a state prediction model based on the word frequency characteristics;

Predicting a remaining life of the compressor at a corresponding point in time using a life prediction model with delay prediction constraints based on an operating state of the compressor at any of the different points in time in the future;

Performing linear regression on the percentage values of the residual life at different time points to obtain the total life of the compressor;

and determining the actual residual usable life of the compressor according to the total life of the compressor and the current running time of the compressor.

In order to solve the above technical problem, an embodiment of the present invention further provides a device for predicting a remaining usable life of a compressor, including:

the data acquisition module acquires time sequence data samples of the vibration signals of the compressor and preprocesses the time sequence data samples;

The feature extraction fusion module is used for carrying out feature extraction and feature fusion on the preprocessed time sequence data samples to obtain comprehensive features representing the health state of the compressor;

the feature conversion module is used for mapping the comprehensive features into text features and determining word frequency features of the text features;

The state prediction module is used for predicting the running states of the compressor at different time points in the future by using a state prediction model based on the word frequency characteristics;

a life prediction module for predicting a remaining life of the compressor at a corresponding point in time using a life prediction model with a delay prediction constraint based on an operating state of the compressor at any one of different points in time in the future;

The total service life determining module is used for carrying out linear regression on the percentage values of the residual service life at different time points to obtain the total service life of the compressor;

And the actual remaining available life determining module is used for determining the actual remaining available life of the compressor according to the total life of the compressor and the current running duration of the compressor.

In order to solve the above technical problem, an embodiment of the present invention further provides a device for predicting a remaining usable life of a compressor, including: the compressor residual useful life prediction method provided by the technical scheme is realized when the processor executes the program.

To solve the above technical problem, an embodiment of the present invention further provides a computer readable storage medium, including instructions, which when executed on a computer, cause the computer to execute the method for predicting remaining usable life of a compressor according to the above technical solution.

The beneficial effects of the invention are as follows: screening out characteristic parameters capable of effectively representing health states, eliminating redundant characteristic quantities, ensuring accuracy of a state identification model, reducing algorithm running time and improving running efficiency; feature fusion is carried out through a linear dimension reduction fusion algorithm, so that heterogeneous feature dimension reduction fusion is realized; the comprehensive features are mapped into text features, word frequency features of the text features are determined, the word frequency features can reflect early faults of the compressor, the word frequency features can more accurately reflect fine changes of data, degradation trends of the compressor can be better represented, and the degradation trends can be effectively represented in the later period of life of the compressor; the running states of the compressor at different time points in the future are predicted by using the state prediction model, the residual life of the compressor at the corresponding time points is predicted by using the life prediction model with delay prediction constraint based on the running states, so that the life prediction model is more biased to predict in advance, and the accuracy of equipment life prediction can be effectively improved; and performing linear fitting on the residual life of different time points to obtain the total life of the compressor, and further accurately determining the actual residual available life of the compressor according to the current working time of the compressor.

Additional aspects of the invention and advantages thereof will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

FIG. 1 is a flow chart illustrating a method for predicting remaining useful life of a compressor in accordance with an exemplary embodiment of the present invention;

FIG. 2 is a schematic diagram of a comprehensive feature-to-text feature mapping provided by an embodiment of the present invention;

FIG. 3 is a word frequency feature diagram provided in an embodiment of the present invention;

FIG. 4 is a schematic diagram of word frequency versus inverse text frequency characteristics according to an embodiment of the present invention;

FIG. 5 is a Encoder-Decoder block diagram provided by an embodiment of the present invention;

FIG. 6 is a diagram illustrating a Encoder-Decoder configuration of an attention mechanism according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of a delay prediction constraint loss function according to an embodiment of the present invention;

Fig. 8 is a flowchart of a method for predicting remaining usable life of a compressor according to another exemplary embodiment of the present invention.

Detailed Description

Other advantages and effects of the present disclosure will become readily apparent to those skilled in the art from the following disclosure, which describes embodiments of the present disclosure by way of specific examples. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the present disclosure. The disclosure may be embodied or practiced in other different specific embodiments, and details within the subject specification may be modified or changed from various points of view and applications without departing from the spirit of the disclosure. It should be noted that the following embodiments and features in the embodiments may be combined with each other without conflict. All other embodiments, which can be made by one of ordinary skill in the art without inventive effort, based on the embodiments in this disclosure are intended to be within the scope of this disclosure.

It is noted that various aspects of the embodiments are described below within the scope of the following claims. It should be apparent that the aspects described herein may be embodied in a wide variety of forms and that any specific structure and/or function described herein is merely illustrative. Based on the present disclosure, one skilled in the art will appreciate that one aspect described herein may be implemented independently of any other aspect, and that two or more of these aspects may be combined in various ways. For example, an apparatus may be implemented and/or a method practiced using any number of the aspects set forth herein. In addition, such apparatus may be implemented and/or such methods practiced using other structure and/or functionality in addition to one or more of the aspects set forth herein.

The residual life prediction of the compressor is to accurately evaluate and predict the possibility of failure of the compressor in a future period of time according to the monitored information such as equipment state, working environment, load and the like.

With the continuous improvement of production safety requirements of enterprises, it is important to monitor the state of the compressor and diagnose and predict faults of the compressor. The life prediction method is mainly classified into a test method and an analysis method. The test method obtains the required fatigue data by placing the structure in the same or similar environment as the actual situation. The analysis method is to integrate fatigue response of the material under different loads into a certain mathematical relationship by researching failure mechanism of the structural member so as to predict the fatigue life of the material under a certain specified condition. Analytical methods have greatly reduced the cost of prediction compared to experimental methods, and are therefore also an important item of research for researchers.

Compressors, as a core equipment for gas stations, cause serious economic losses once unexpected shutdowns and delays occur. To reduce unnecessary economic losses, more and more businesses are performing predictive maintenance and health management on compressors. But new problems also arise during the application process. The problem of life prediction becomes more and more difficult along with the improvement of the complexity of an industrial manufacturing system, and the compressor often bears a plurality of tasks along with the continuous update and development of manufacturing technology and production requirements, the continuous changing working environment and conditions and the complex equipment structure, so that new challenges are presented to the existing life prediction method, and the existing life prediction precision can also improve the space; in the field of residual life prediction, a predicted life being smaller than an actual life is referred to as an advanced prediction, a predicted life being larger than an actual life is referred to as a delayed prediction, and people have a higher tolerance to the advanced prediction than the delayed prediction.

Accurate residual life prediction can avoid compressor failure, and is the key of a maintenance mode from passive to active. The conventional compressor life prediction method generally predicts the residual life of the compressor by establishing a mathematical and physical model, but is difficult to construct a specific model in the face of various working states and complicated internal parts. In recent years, neural networks have rapidly developed, and various CNNs, RNNs and neural networks based on attention mechanisms have emerged. These neural networks are highly popular in the image and natural language processing fields, but they are difficult to apply directly to the problem of residual life prediction of compressors. If the data noise is too much, the neural network learning effect is not good; the degradation indexes are different in performance in different time periods, so that the model life prediction result based on the neural network is often unstable; the lifetime prediction problem is less tolerant to delay prediction, whereas the conventional loss function is tolerant to delay prediction and advance prediction.

In order to solve the above problems, embodiments of the present application provide a method, apparatus and computer-readable storage medium for predicting remaining useful life of a compressor. These embodiments will be described in detail below.

Referring first to fig. 1, fig. 1 is a flow chart illustrating a prediction of remaining usable life of a compressor according to an exemplary embodiment of the present application. The method may be specifically executed by a server, where the server may be an independent server, or may be a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and basic cloud computing services such as big data and artificial intelligence platforms, which are not limited herein.

As shown in fig. 1, in an exemplary embodiment, the method for predicting remaining usable life of a compressor may include steps S101 to S107, which are described in detail as follows:

s101, acquiring time sequence data samples of a compressor vibration signal, and preprocessing the time sequence data samples.

Specifically, for the time series data D, the dimension is (N, T, L), expressed as data D, N samples in total, each sample D _i contains T time points, and the sample D _i samples the vibration signal data of length L at time point T _i.

In an exemplary embodiment of the present application, after obtaining the time series data sample of the vibration signal of the compressor, noise reduction processing is further performed on the time series data sample. Specific noise reduction means include, but are not limited to, threshold noise reduction, wavelet noise reduction, self-encoder noise reduction, and the like.

S102, carrying out feature extraction and feature fusion on the preprocessed time sequence data samples to obtain comprehensive features representing the health state of the compressor.

According to the embodiment of the invention, the characteristic parameters capable of representing the health state of the compressor are extracted through the characteristic extraction, so that the health state of the compressor is effectively represented, redundant characteristic quantities are eliminated, the accuracy of a state identification model is ensured, the operation time of an algorithm is shortened, and the operation efficiency is improved. And the aim of heterogeneous feature fusion can be fulfilled through feature fusion.

S103, mapping the comprehensive features into text features, and determining word frequency features of the text features.

According to the embodiment of the invention, the time domain characteristics of the time sequence data samples are converted into text characteristics, so that early faults of the compressor can be reflected, and meanwhile, the degradation trend can be effectively represented in the middle and later periods of the service life of the compressor.

The embodiment of the invention converts the time sequence data in different numerical value intervals into the text characteristic change, and the change of the data is represented by the change of word frequency. Entropy features in time domain features are similar to text features, but word frequency features give different weights to different numerical intervals. In the life prediction problem, the low-frequency interval usually comprises individuals with shorter life, and reflects potential risks and abnormal events, so that the low-frequency interval in the data is often more important, and the word frequency characteristics give higher weight to the intervals, so that the subtle changes of the data are reflected more accurately, and the degradation trend of the compressor is better represented.

S104, predicting the running states of the compressor at different time points in the future by using a state prediction model based on the word frequency characteristics.

In the embodiment of the invention, the state prediction model can learn the corresponding relation between the word frequency characteristics and the running state by using a small sample learning method.

Small sample learning (Few Shot Learning, FSL) mainly solves the machine learning application problem with very small data volume. In the small sample learning method, the model acquires priori knowledge through a preposed task, and then on a new task, the model only learns a small amount of samples and carries out gradient updating in a small amount, so that the model can be quickly generalized and adapted.

S105, predicting the residual life of the compressor at a corresponding time point by using a life prediction model with delay prediction constraint based on the running state of the compressor at any time point of different time points in the future.

Likewise, in the embodiment of the invention, the life prediction model may also learn the correspondence between the running state and the remaining life by using a small sample learning method.

And S106, linearly regressing the percentage values of the residual life at different time points to obtain the total life of the compressor.

And S107, determining the actual residual available life of the compressor according to the total life of the compressor and the current running duration of the compressor.

From the above, a method for predicting the remaining useful life of a compressor based on text features and delay prediction constraints is proposed in the present embodiment. Specifically, the embodiment of the invention eliminates redundant characteristic quantity by screening out the characteristic parameters capable of effectively representing the health state so as to ensure the accuracy of a state identification model, reduce the operation time of an algorithm and improve the operation efficiency; feature fusion is carried out through a linear dimension reduction fusion algorithm, so that heterogeneous feature dimension reduction fusion is realized; the comprehensive features are mapped into text features, word frequency features of the text features are determined, the word frequency features can reflect early faults of the compressor, the word frequency features can more accurately reflect fine changes of data, degradation trends of the compressor can be better represented, and the degradation trends can be effectively represented in the later period of life of the compressor; the running states of the compressor at different time points in the future are predicted by using the state prediction model, the residual life of the compressor at the corresponding time points is predicted by using the life prediction model with delay prediction constraint based on the running states, so that the life prediction model is more biased to predict in advance, and the accuracy of equipment life prediction can be effectively improved; and performing linear fitting on the residual life of different time points to obtain the total life of the compressor, and further accurately determining the actual residual available life of the compressor according to the current working time of the compressor.

Optionally, in an exemplary embodiment of the present invention, step S102 of the embodiment shown in fig. 1 may specifically include steps S201 to S203, which are described in detail below:

S201, extracting at least one dimensionless feature and at least one dimensionless feature of the time series data sample.

The signal vibration data is usually time sequence data, namely a time domain signal, wherein the time domain signal comprises dimensional characteristics and dimensionless characteristics. The general dimensional time domain features are shown in table 1 below.

Table 1 has dimensional time domain features and formulas thereof

The peak value is used to reflect the maximum value in the data, and this feature is very sensitive to vibration signal failure but is easily affected by outliers. The peak-to-peak value is used to reflect the range of data variation, and is sensitive to vibration signal faults and is easily affected by outliers, similar to the peak value. The mean value is used to reflect the mean change of the data, and the feature is insensitive to outliers and is used to calculate other features. Root mean square, also known as effective value, is commonly used to reflect circuit power. Variance, a feature used to reflect the degree of dispersion of the data. Standard deviation, like variance, is also used to reflect the degree of dispersion of the data, unlike variance features, which unify dimensions with the original data.

According to the difference of working conditions, the size of the dimensional characteristics is changed correspondingly, and the working environment has great influence on the dimensional characteristics and shows a certain instability. Such as mean features, which are commonly used for fault detection, mean values are also commonly used for calculating other features, but the mean features do not reflect real-time dynamic changes of the sample. In engineering application, the root mean square and standard deviation are increased correspondingly along with the change of the health state of equipment until the equipment is completely disabled. The dimensional characteristic values are easy to understand and simple to calculate, so that the method is often used by researchers.

The dimensional characteristics are changed due to the change of working conditions (such as load), are easily affected by environmental interference, and have the defect of unstable performance. In contrast, dimensionless indicators can exclude the influence of these disturbance factors, and are thus widely used in the field of feature extraction. Table 2 is a general dimensionless time domain characterization.

Table 2 dimensionless time domain features and formulas thereof

Kurtosis represents the degree of smoothness of a signal, used to describe the distribution of data. The kurtosis of a normal distribution is equal to 3, and the distribution curve is flatter when the kurtosis is smaller than 3, and steeper when the kurtosis is larger than 3. Skewness is also called skewness and skewness. The skewness and the kurtosis have certain relativity, and the kurtosis factor is the ratio of the fourth-order central moment to the fourth power of the standard deviation; the skewness factor is the ratio of the third order center moment to the third power of the standard deviation. Skewness and kurtosis are both characteristics that describe the distribution of data. The peak factor is the ratio of the peak value to the effective value (RMS) of the signal and is used to detect the presence or absence of an impact in the signal. The peak is a parameter that varies greatly with time, and varies greatly from moment to moment. Since the stability of the peak value is poor and the sensitivity to the impact is also poor, the index is gradually replaced by kurtosis in the failure diagnosis. The pulse factor is the ratio of the signal peak to the rectified mean. The pulse factor may be used to detect whether an impulse is present in the signal. The margin factor is typically used to detect wear conditions of the compressor. The above dimensionless characteristics exhibit different stability and sensitivity during compressor degradation, and table 3 is a comparison of the sensitivity and stability of the dimensionless characteristics to faults.

TABLE 3 stability and sensitivity comparison of dimensionless characteristics

Pulse factors, peak factors and kurtosis are all relatively sensitive to impact-type faults, and particularly when faults occur early, they increase significantly; however, as the faults develop, these characteristics decrease, indicating that they are more sensitive to early faults and have poor stability. The stability of the effective value is better, but not sensitive to early failure signals. In order to achieve a better effect, they are often applied simultaneously. In summary, dimensional features are easy to calculate, easy to understand, but are easily affected by operating conditions, states and environments, whereas non-dimensional features are insensitive to operating environments.

Through analysis of the dimensional features and the dimensionless features in the time domain features, the dimensional features are sensitive to data change and are easily influenced by external factors, and the dimensionless features are relatively stable, but usually only perform better in a certain life cycle. For example, dimensionless features with better stability are difficult to find early faults of the compressor, while dimensionless features with higher sensitivity are easy to fail in middle and later fault diagnosis. Embodiments of the present invention thus use dimensional features in combination with non-dimensional features.

S202, performing feature fusion on the at least one dimensional feature and the at least one dimensionless feature by using a linear dimension reduction fusion algorithm to obtain comprehensive features representing the health state of the compressor.

According to the embodiment of the invention, the selected characteristics can be subjected to characteristic fusion through a PCA algorithm, so that comprehensive characteristics representing the health state of the compressor are obtained. PCA is a common linear dimension-reduction fusion algorithm, and the main idea is to map n-dimensional characteristic variables onto k dimensions, wherein each dimension of the mapped variables is a linear combination of original variables. Thus, the purpose of heterogeneous feature dimension reduction fusion is achieved. Wherein n and k are positive integers, and n is greater than k.

Optionally, in an exemplary embodiment, step S103 of the embodiment shown in fig. 1 may specifically include steps S301 to S303, which are described in detail below:

s301, dividing the value threshold of the comprehensive feature into a plurality of value range intervals, wherein each value range interval is defined as a text feature, and mapping the comprehensive feature into the text feature;

The embodiment of the invention divides the value range of the comprehensive characteristics into j intervals, wherein j is a positive integer; the number of different comprehensive features contained in each time point of each sample D _i is counted, and the j intervals are regarded as j different words wj. FIG. 2 is a schematic diagram of a comprehensive feature-to-text feature mapping, where the dimension of data D is (1,1,7) and the number of words j is 5.

Fig. 2 shows the mapping of the integrated feature [0.07,0.16,0.04,0.155,0.21,0.08,0.13] to the text feature [ d, b, e, b, a, d, e ].

S302, counting the text feature frequency of the text feature, and determining word frequency-inverse text frequency features based on the text feature frequency.

For word frequency characteristics, word frequencies of text characteristics are counted to obtain a word frequency matrix M, and the dimension of the word frequency matrix M is (T, j). And then reducing the dimension of the word frequency matrix, and constructing low-dimension word frequency features by using a matrix singular value decomposition-reconstruction method. Specifically, singular value decomposition is performed on the word frequency matrix M, and then the matrix is reconstructed by using the largest singular value, so that a low-dimensional feature moment matrix is obtained.

Specifically, for the term frequency-inverse text frequency (tf-idf) feature, tf-idf feature is a weighted term frequency matrix after dimension reduction, tf-idf contains two parts, term frequency coefficient tf and normalized inverse text frequency idf.

The word frequency coefficient tf is shown in formula (1):

Wherein M is the word frequency matrix, and L is the length of the time sequence data.

The normalized inverse text frequency idf is shown in formula (2):

Where N is the number of samples of the time series data described above, and the denominator |j:w _j∈D_i | in equation (2) is the number of samples D _i containing the word w _j. In summary, the formula of tf-idf is shown in formula (3):

S303, performing dimension reduction on the word frequency-inverse text frequency characteristic by using a matrix singular value decomposition-reconstruction method, and reconstructing a matrix by using the maximum singular value to obtain the word frequency characteristic.

After the tf-idf feature is calculated, the tf-idf matrix is subjected to dimension reduction by using a matrix singular value decomposition-reconstruction method, the largest singular value is utilized to reconstruct the matrix, the obtained low-dimension feature matrix M _tf-idf is used as word frequency feature, and the dimension of M _tf-idf is (T, 1).

Text features are shown in fig. 3 and 4. Fig. 3 shows word frequency characteristics M _tf-idf, fig. 4 shows tf-idf characteristics, and the left word frequency characteristics show degradation trend earlier (700-1000 time points) than the traditional time domain characteristics such as standard deviation and skewness. The text features provided by the embodiment of the invention convert the transformation of data distribution into the transformation of text word frequency by using a data discretization method, enhance the data representation by using the word frequency features M _tf-idf and if-idf features in the natural language processing field, and have better degradation trend compared with the traditional time domain features.

The embodiment of the invention converts the time sequence data into the word change in different numerical value intervals, and the change of the data is represented by the change of word frequency. Entropy features in time domain features are similar to text features, but tf-idf features give different weights to different numerical intervals, and in life prediction problems, low-frequency intervals in data are often more important, and tf-idf algorithms give higher weights to the intervals, so that fine changes of the data are reflected more accurately, and degradation trends of compressors are better represented.

Existing life prediction methods generally use LSTM (Long Short-Term Memory Networks, LSTM) to sense the degradation trend of a compressor, but in industrial practice, it is often necessary to predict the future operation state of the compressor, so the present invention divides the life prediction of the compressor into two parts, namely, the state prediction and the life prediction of the compressor. For the state prediction part, embodiments of the present invention use a Encoder-Decoder structure based on LSTM and use the attention mechanism to enhance the Decoder phase. For the life prediction part, the embodiment of the invention designs a delay prediction constraint loss function which is used for constraining the training process of the neural network.

Optionally, in an exemplary embodiment, step S104 of the embodiment shown in fig. 1 may specifically include steps S401 to S403, which are described in detail below:

S401, inputting the word frequency characteristics into the state prediction model; wherein the state prediction model adopts Encoder-Decoder structure, encoder and hidden layer of Decoder are composed of LSTM.

S402, the state prediction model Decoder structure calculates weighted values of the Decoder hidden layer state at the time t and all hidden layer states Encoder by using an attention mechanism aiming at the output at the time t, and uses the weighted values as semantic vectors of the Decoder at the time t.

S403, splicing the hidden layer state of the Decoder at the time t and the semantic vector of the Decoder at the time t to generate output at the time t, and taking the output as the residual running state of the compressor. Wherein t is any one of the future points in time.

Because the residual running time of the compressor is unknown, the traditional neural network model structure cannot cope with variable-length output, and the embodiment of the invention adopts a Encoder-Decoder structure to predict the residual running state of the compressor. The structure solves the problem of variable length output by encoding the variable length sequence into a fixed length vector, and continuously decoding the fixed length vector until a termination condition is reached. Encoder-Decoder structure is shown in FIG. 5. Wherein x _t and y _t represent the compressor running state vector at any time point t in the future at different time points, h represents the hidden layer state, c represents the semantic vector output by the last hidden layer node Encoder, and the Decoder predicts the output y _t-1 at the current time according to c and the output y _t'-1 at the previous time. Encoder and the hidden layer of the Decoder are both composed of LSTM.

In Encoder-Decoder structure, as the decoding time step increases, the Decoder gradually forgets the hidden layer state in the initial stage of Encoder, and the semantic vector c contains all the hidden layer states of Encoder, but as the decoding time is changed, the Decoder focuses more on state change in a short time, so that the problem of forgetting the early knowledge of the model is difficult to avoid. To improve long-term decoding, the present invention uses an attention mechanism to enhance the Decoder decoding stage.

In the original Encoder-Decoder structure, the Decoder would only use the semantic vector c calculated at the beginning at Encoder and the semantic vector calculated by itself for the subsequent decoding process. Because the semantic vector cannot completely represent the original input sequence, new input information is changed continuously, and finally the original encoded semantic vector is lost. To solve the information loss problem, the invention calculates semantic vectors by using an attention mechanism, and a Encoder-Decoder structure of the attention mechanism is shown in fig. 6. The Decoder calculates weighted values of all hidden layer states h _s of the Decoder hidden layer states h _t and Encoder at the time t (hs, ht' -1) as semantic vectors of the Decoder at the time t by using an Attention mechanism for the output y _t at the time t, t epsilon (0, 1,2 … …). By the method, each moment in the decoding stage refers to the state of all the previous moments, and the problem of information loss is solved. The attention mechanism used in the invention is Bahdanau Attention, and the calculation mode is as follows.

score(h_t,h_s)＝Vtanh(W₁h_t+W₂h_s) (4)

c_t＝∑_sa_tsh_s (6)

Score in equation (4) represents the degree of association between the state at time t and all states Encoder, h _s represents Encoder all hidden layer states, h _t represents the Decoder state at time t, V represents the hidden layer in the attention module, tanh represents the activation function, and W ₁ and W ₂ are matrices of learnable parameters; a _ts denotes a softmax activation function, S denotes each hidden layer, S' denotes a first hidden layer, and S denotes the number of hidden layers. After the correlation is calculated, the point-in-time state weights are calculated according to the equation (5) softmax () function. Finally, all states h _s of Encoder are weighted and summed according to the state weight to obtain the semantic vector c _t at the time t. The Decoder splices h _t and c _t to generate an output at time t.

Optionally, in an exemplary embodiment, step S105 of the embodiment shown in fig. 1 may specifically include steps S501 to S502, which are described in detail below:

S501, inputting the running state of the compressor at any one of different time points in the future into the life prediction model; wherein, the life prediction model adopts a double-layer LSTM structure with delay prediction constraint;

S502, determining the residual life corresponding to the running state based on the mapping relation between the running state and the residual life learned by the double-layer LSTM structure.

For the life prediction portion, embodiments of the present invention use a double layer LSTM to learn the mapping between compressor operating conditions y to remaining life. However, in the initial stage of the operation of the compressor, the remaining life of the compressor is long, and if the compressor is completely damaged at this time as a termination condition of the Decoder decoding, the calculation complexity is too high and the accuracy is poor.

In summary, when predicting the residual life of the current time point, directly predicting the current running state by using a double-layer LSTM life prediction model; for future operating conditions and remaining life of the compressor, the Decoder is used to predict the operating conditions for the next k time windows (k being the best performing value on the validation set) and to perform a linear regression on the life prediction results as a future degradation trend of the compressor.

Aiming at the problem of low tolerance of model delay prediction in industrial practice, the invention designs a loss function of delay prediction constraint. The loss function formula is shown in formula (7):

where L' represents the sum of losses of all samples, n is the number of samples, y _i represents the model predicted remaining life, y _i' represents the true remaining life of the samples, and a represents the delay prediction penalty factor. Where a is a hyper-parameter (a.gtoreq.0), the greater the delay prediction penalty is when a is greater.

The loss function is improved on the basis of the MSE, and a delay penalty term max (1, 2- (y _i-y_i')^a) is added.

And after normalization, the output layer of the neural network outputs between 0 and 1. For most samples, the delay prediction error is concentrated between [0,0.5 ]. Based on the above analysis, the present invention uses this penalty to increase the delay prediction error in this range.

Fig. 7 is a schematic diagram of a delay prediction constraint loss function. As shown in fig. 5, the penalty function proposed by the present invention penalizes the error between [0,1], the greater the delay penalty when a is greater, and when a=0, there is no delay penalty, with the effect being equal to the MSE penalty function.

Compared to directly adding penalty coefficients in the delay prediction portion of the MSE loss function, embodiments of the present invention primarily increase the error loss value in the [0,0.5] interval. If penalty coefficients are directly added to the MSE function, penalty in the [0.5,1] interval is overlarge, and errors of most samples are larger in the initial stage of model training, instability in the initial stage of training is easy to cause by a method of directly adding penalty coefficients to the MSE function, and finally, the convergence effect is poor, and the delay prediction constraint function is more stable in the initial stage of training.

In summary, the delay prediction constraint loss function of the embodiment of the invention reduces the occurrence of the situation of model delay prediction under the condition of not losing prediction precision.

Optionally, in an exemplary embodiment, step S105 of the embodiment shown in fig. 1 may specifically include steps S601 to S603, which are described in detail below:

s601, carrying out linear regression on the percentage values of the residual lives at different time points to obtain a linear equation;

s602, calculating an intersection point of the corresponding straight line and the x-axis through the straight line equation, and taking the abscissa of the intersection point as the total service life of the compressor.

Based on Encoder-Decoder structure, attention mechanism and delay prediction constraint loss function, a life prediction method is constructed. Firstly, learning a future running state by utilizing an LSTM model based on Encoder-Decoder structure, and training by adopting a free-running mode, wherein the mode carries out next training according to a result generated by the LSTM, and compared with training by directly using real data of the next time point, the training mode is closer to a real test condition, and the model has stronger noise tolerance. Meanwhile, the decoding process in the state prediction process is enhanced using an attention mechanism. Secondly, a double-layer LSTM model is used for learning the mapping between the real running state and the residual life, and a delay prediction constraint is used as a loss function to bias the model towards the advance prediction. Finally, after the state prediction model and the life prediction model are trained, carrying out life prediction on the operation data of the current time point, then predicting future operation trend by using the state prediction model, carrying out life prediction on the corresponding time point according to the operation state, carrying out linear regression on the life prediction percentage values of all the time points of the sample to obtain a straight line l=ax+b, calculating an intersection point (-b/a, 0) of the l and the x axis by a straight line equation, wherein-b/a is the total life of the sample, and subtracting the current working time length, namely the actual residual life of the sample at the current moment from the total life.

As shown in fig. 8, in another exemplary embodiment, the method for predicting remaining usable life of a compressor may include the following steps, which are described in detail as follows:

The flow of the lifetime prediction method based on text features and delay prediction constraint according to the embodiment of the present invention is shown in fig. 8. The method mainly comprises data noise reduction, feature extraction, data transformation and residual life prediction.

Firstly, carrying out signal processing on original data, eliminating noise in the original data, if the noise can be reduced by adopting a wavelet noise reduction processing mode, removing the influence of noise and interference in a system, and improving the prediction precision; secondly, in a feature extraction and fusion stage, extracting traditional degradation indexes such as margin factors, kurtosis, pulse factors and the like, carrying out feature fusion, and simultaneously carrying out Kalman filtering (namely smoothing and accumulated transformation) on the features to further remove noise; then carrying out equipment running state prediction by using Encoder-Decoder structure based on LSTM, and carrying out service life prediction by using double-layer LSTM with delay prediction constraint for the equipment running state; and finally, carrying out linear regression fitting on the residual life percentage value, carrying out fitting by utilizing a known data set by establishing a linear relation between the residual life percentage and other related variables, then predicting the future residual life percentage by using the model, and calculating the real residual life of the compressor according to the current service life of the compressor, thereby helping to make a more effective maintenance plan, optimizing the utilization rate of equipment, improving the production efficiency and reducing the maintenance cost.

An exemplary embodiment of the present application shows a compressor remaining usable life prediction apparatus, including:

The device converts the time domain characteristics of the time sequence data sample into text characteristics by applying the method for predicting the residual available life of the compressor, and determines the word frequency characteristics of the text characteristics, wherein the word frequency characteristics not only can reflect early faults of the compressor, but also can more accurately reflect fine changes of data, better represents the degradation trend of the compressor, and can effectively represent the degradation trend in the later period of the life of the compressor; the residual running state of the compressor is predicted by using the state prediction model, the residual service life of the compressor is predicted by using the service life prediction model with delay prediction constraint, so that the service life prediction model is more biased to predict in advance, the effectiveness of a method used in the service life prediction problem is proved, and the service life prediction precision of equipment can be effectively improved.

Another aspect of the present application also provides a compressor remaining usable life prediction apparatus, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing a compressor remaining useful life prediction method as before when executing the program.

Another aspect of the application also provides a computer readable storage medium comprising instructions which, when run on a computer, cause the computer to perform a method of predicting remaining useful life of a compressor as previously described.

It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the apparatus and units described above may refer to corresponding procedures in the foregoing method embodiments, which are not described herein again.

In the several embodiments provided by the present application, it should be understood that the disclosed apparatus and method may be implemented in other manners. For example, the apparatus embodiments described above are merely illustrative, e.g., the division of elements is merely a logical functional division, and there may be additional divisions of actual implementation, e.g., multiple elements or components may be combined or integrated into another system, or some features may be omitted, or not performed.

The units described as separate units may or may not be physically separate, and units shown as units may or may not be physical units, may be located in one place, or may be distributed over a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiment of the present invention.

In addition, each functional unit in the embodiments of the present invention may be integrated in one processing unit, or each unit may exist alone physically, or two or more units may be integrated in one unit. The integrated units may be implemented in hardware or in software functional units.

The integrated units, if implemented in the form of software functional units and sold or used as stand-alone products, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention is essentially or a part contributing to the prior art, or all or part of the technical solution may be embodied in the form of a software product stored in a storage medium, comprising several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods of the embodiments of the present invention. And the aforementioned storage medium includes: a usb disk, a removable hard disk, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), a magnetic disk, or an optical disk, or other various media capable of storing program codes.

The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims

1. A method for predicting remaining useful life of a compressor, comprising:

2. The method of claim 1, wherein the feature extraction and feature fusion of the preprocessed time series data samples to obtain integrated features characterizing compressor health, comprises:

extracting at least one dimensional feature and at least one non-dimensional feature of the time series data sample;

and carrying out feature fusion on the at least one dimensional feature and the at least one dimensionless feature by using a linear dimension reduction fusion algorithm to obtain comprehensive features representing the health state of the compressor.

3. The method of claim 1, wherein mapping the integrated feature to a text feature and determining a word frequency feature of the text feature comprises:

dividing the value threshold of the comprehensive feature into a plurality of value range intervals, defining each value range interval as a text feature, and mapping the comprehensive feature into the text feature;

counting the text feature frequency of the text feature, and determining word frequency-inverse text frequency features based on the text feature frequency;

And reducing the dimension of the word frequency-inverse text frequency characteristic by using a matrix singular value decomposition-reconstruction method, and reconstructing a matrix by using the maximum singular value to obtain the word frequency characteristic.

4. The method of claim 1, wherein predicting the operating state of the compressor at different points in time in the future using a state prediction model based on the word frequency characteristics comprises:

Inputting the word frequency characteristics into the state prediction model; wherein the state prediction model adopts Encoder-Decoder structure, and both Encoder and hidden layers of the Decoder are composed of LSTM;

The state prediction model Decoder structure calculates the weight values of the Decoder hidden layer state at the time t and all hidden layer states Encoder by using an attention mechanism aiming at the output at the time t, and uses the weight values as semantic vectors of the Decoder at the time t;

Splicing the hidden layer state of the Decoder at the time t with the semantic vector of the Decoder at the time t to generate the running state of the compressor at the time t; wherein t is any one of the different future time points.

5. The method of claim 1, wherein predicting remaining life of the compressor at a corresponding point in time based on an operating state of the compressor at any of the different points in time in the future using a life prediction model with a delay prediction constraint, comprising:

Inputting the operating state of the compressor at any one of different future time points into the life prediction model; wherein, the life prediction model adopts a double-layer LSTM structure with delay prediction constraint;

and determining the residual life corresponding to the running state based on the mapping relation between the running state and the residual life learned by the double-layer LSTM structure.

6. The method according to any one of claims 1 to 5, wherein said linearly regressing the percentage values of the remaining life at different points in time, yields the total life of the compressor, comprising:

Performing linear regression on the percentage values of the residual lives at different time points to obtain a linear equation;

and calculating an intersection point of the corresponding straight line and the x-axis through the straight line equation, and taking the abscissa of the intersection point as the total service life of the compressor.

7. The method according to any one of claims 1 to 5, wherein the loss function of the life prediction model is as follows:

In the formula, L' represents the loss sum of all samples, n is the number of samples, y _i represents the predicted residual life of the model, y _i' represents the real residual life of the samples, and a represents the delay prediction penalty factor, wherein a is the super-parameter (a is more than or equal to 0), and when a is larger, the delay prediction penalty is larger.

8. A compressor remaining usable life prediction apparatus, comprising:

9. A compressor remaining useful life prediction apparatus, comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for predicting remaining useful life of a compressor as claimed in any one of claims 1 to 7 when executing the program.

10. A computer readable storage medium comprising instructions which, when run on a computer, cause the computer to perform the method of predicting remaining useful life of a compressor as claimed in any one of claims 1 to 7.