Denoising auto-encoder and multi-entropy fused engine running state evaluation method

Abstract

The invention aims to provide a denoising auto-encoder and multi-entropy fusion engine running state evaluation method, which comprises the following steps of: acquiring health state data and current moment state data of a to-be-evaluated engine through data acquisition at an online stage; obtaining a multi-feature index entropy by using the health state data and the state data at the current moment, and stacking the multi-feature index entropy; in the off-line stage, denoising self-encoder training is carried out, and a trained decoder is obtained; obtaining health state features after dimension reduction and current return features after dimension reduction through the trained decoder and the multi-feature index entropy; and obtaining a health index of the to-be-evaluated engine according to the health state feature after the dimension reduction and the current return feature after the dimension reduction. Because the health index value after multi-entropy fusion contains more engine operation state characteristics, the method provided by the invention has high feasibility and effectiveness.

Description

Technical Field

[0001] The present invention relates to the field of engine operating state assessment, and in particular to an engine operating state assessment method combining denoising autoencoder and multi-entropy fusion. Background Art

[0002] As one of the core components of the equipment that works continuously, the engine is subject to interference from a variety of uncontrollable factors during its operation while bearing loads and completing high-intensity operation tasks. In addition, since the engine often works in complex, changeable and even harsh environmental conditions, it is prone to a variety of degradation forms and damage. Therefore, in order to ensure the normal operation of mechanical equipment, complete production tasks on time, maximize corporate profits and protect the lives of employees, it is necessary to regularly inspect and maintain the engine to promptly discover and resolve potential problems. Specifically, common forms of engine failure include fracture failure, fatigue failure, and wear failure of components.

[0003] At present, the analysis of engine health status usually adopts the vibration signal analysis method. This method can effectively evaluate the health status of the engine by monitoring and collecting relevant information about the engine operation status, combined with in-depth analysis and research. In this process, the linear analysis method plays a key role. In recent years, many methods for distinguishing the health status of the engine have been proposed and applied in practical research. These methods are not only highly accurate, but also highly efficient. Commonly used methods include temperature detection method, acoustic signal method, oil sample analysis method, etc. The main advantage is that the core characteristics of the engine operation status, such as temperature, sound or vibration signals, can be obtained through professional equipment without disassembling the mechanical equipment. Through these methods, the health status of the engine can be quickly and accurately evaluated, and the complex disassembly and assembly process of mechanical equipment can be simplified. Among the many methods, the monitoring method based on vibration signals has become one of the core methods in the field of engine health management research due to its advantages such as simple signal acquisition and significant fault characteristics.

[0004] At the same time, the operating status assessment method can also monitor the operating status of the engine in real time, collect key data, and conduct in-depth evaluation and diagnosis. At the same time, it predicts possible engine failures and formulates reasonable and effective repair and maintenance strategies for enterprises. This can not only significantly improve the reliability and safety of the system, but also ensure the comprehensive benefits of the enterprise. However, there are still deficiencies in the current research on the state assessment method of the engine. The main problems include the following aspects: First, the diagnosis, evaluation and prediction of the engine health status involve multiple disciplines, which require both deep theoretical support and significant practical application value. In practical applications, the early stage of the engine health status evaluation method is relatively simple, mainly relying on the intuitive judgment and experience accumulation of experts or technicians, and usually only a simple classification of the operating status can be made, that is, normal or faulty. However, with the continuous increase in the complexity of engine functions and structures, this simple evaluation method has been difficult to meet the diversified needs of modern complex systems. Secondly, the amount of feature data such as vibration signals and temperature signals generated during the operation of the engine is large, the information dimension is high and the redundancy is serious. In traditional feature extraction methods, it is easy to cause the omission of important feature information or the calculation time is too long, and it is difficult to ensure the accuracy and globality of feature extraction.

[0005] In view of the above problems, the present invention provides an engine operating state evaluation method combining denoising autoencoder and multi-entropy fusion on the basis of ensuring the accuracy of the engine operating state evaluation method. Summary of the invention

[0006] The object of the present invention is to provide an engine operating state assessment method combining a denoising autoencoder and multi-entropy fusion, so as to at least solve one technical problem in the prior art.

[0007] The technical solution of the present invention is:

[0008] A method for evaluating engine operating status by fusion of a denoising autoencoder and multi-entropy, comprising:

[0009] Through online data collection, the health status data and current status data of the engine to be evaluated are obtained;

[0010] Using the health status data and the current state data, a multi-feature index entropy is obtained and stacked;

[0011] Perform denoising autoencoder training in the offline stage to obtain a trained decoder;

[0012] Obtaining the reduced-dimensional health status feature and the reduced-dimensional current refund feature through the trained decoder and the multi-feature index entropy;

[0013] The health index of the engine to be evaluated is obtained according to the reduced-dimensional health status feature and the reduced-dimensional current return feature.

[0014] The denoising autoencoder training is performed in the offline stage to obtain a trained decoder, including:

[0015] Acquiring a full-life vibration acceleration signal of the engine to be evaluated, and using the full-life vibration acceleration signal as a sample;

[0016] Extract the first N frequency points of the data in the sample;

[0017] Get the entropy of multiple feature indicators and stack them;

[0018] Training the denoising autoencoder;

[0019] When the loss value stops decreasing, the training is completed, and a trained decoder that can effectively extract features and has dimensionality reduction capabilities is obtained.

[0020] Obtaining the full-life vibration acceleration signal of the engine to be evaluated, and taking the first 2048 frequency points of each data set as training samples;

[0021] Calculating characteristic indices of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy for the intercepted training samples;

[0022] The DAE is trained using the training set using Adam as the optimizer.

[0023] The loss function when training DAE using the training set is:

[0024]

[0025]

[0026] Among them, k is the number of feature samples contained in the input data; y (i) is the eigenvalue of the i-th sample after reconstructing the feature sample; x (i) is the eigenvalue of the i-th sample of the input data; λ is the L2 regularization coefficient; m is the number of weights in the neural network; w i is the i-th weight in the neural network;

[0027] The characteristic index calculation of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy for the intercepted training samples includes:

[0028] The steps for obtaining the relative entropy are as follows:

[0029]

[0030] The steps for obtaining the approximate entropy are as follows:

[0031] Define the time series X(i), construct a vector given the pattern dimension q, that is: X(n) = [x(i), x(n+1), ..., x(n+q-1)]; where: X(n) is a q-dimensional vector; n = 1, 2, ..., i-q+1;

[0032] Define d[X(n), X(m)] as the maximum value of the range of the distance between vectors X(n) and X(m), that is:

[0033]

[0034] Given a similarity tolerance r (r>0), obtain the ratio of the maximum value of the distance range to i-q+1, perform a logarithmic operation on it, and then calculate its average value for all n, that is:

[0035]

[0036] Where: It is the ratio of the maximum value of the distance range to i-q+1;

[0037] Let q = q + 1, we get and C q+1 (r), for the actual sequence, the approximate entropy is:

[0038] A ApEn (q, r, i) = φ q (t)-φ q+1 (r);

[0039] The steps for obtaining the permutation entropy are as follows:

[0040] The original time series is reconstructed in phase space and transformed into the following matrix Y with j rows and q columns, namely:

[0041]

[0042] Where: q is the embedding dimension; t is the delay factor;

[0043] Rearrange the elements of each row in the matrix Y in ascending order, and let the index of each element after arrangement be λ α (α=1,2,…,q), by sorting, the index of each row in the matrix can be expressed by the following formula:

[0044] S(w)=(λ1,λ2,…λ q ), w=1, 2,...j, j≤q!

[0045] Where: There are q! different arrangements of the q-dimensional phase space mapping;

[0046] Get the permutation entropy of the time series as:

[0047] Where: P w The probability of each arrangement being present divided by the total number of times q! different arrangements appear is:

[0048] The steps for obtaining the fuzzy entropy are as follows:

[0049] The original time series is reconstructed in phase space to obtain the time series Y, that is:

[0050] Y=[x(i),x(i+1),...,x(i+q-1),]-x0(i);

[0051]

[0052] Where: q is the embedding dimension; x0(i) is the mean;

[0053] The distance between the two decomposed time series Y(i) and Y(j) is defined as Right now:

[0054]

[0055] The fuzzy membership function is introduced and the similarity between two time series is calculated using the fuzzy function, namely:

[0056]

[0057] Where: r is the similarity tolerance; i, j = 1, 2, ..., K-m + 1, and i ≠ j; K is the length of the time series;

[0058] The fuzzy entropy of the original time series is:

[0059]

[0060] Where: q (r) is the definition function;

[0061] The steps to obtain discrete entropy are as follows:

[0062] Transform the original time series into a normal cumulative distribution function and map the series to the new domain y i (i=1, 2,...,N):

[0063]

[0064] Where: μ and σ are the mean and standard deviation of the time series x(i) respectively;

[0065] Use the embedding dimension q and the number of categories c to construct all possible discrete models The total number of model categories is c q , y i All data in are automatically assigned to integers from 1 to c and are converted to Constructed as vector group

[0066] make With the potential model Matching, use the following formula to count the frequency of each potential model appearing in the original time series:

[0067]

[0068] Where: * is the total number of occurrences of each potential model;

[0069] Get the discrete entropy of the original time series:

[0070]

[0071] The steps to obtain symbolic dynamics entropy are as follows:

[0072] Binarize the original vibration signal X(i) with a sequence length of N to obtain the symbol sequence h i :

[0073]

[0074] Where: is the sample mean;

[0075] In determining the symbol sequence h i After that, in order to determine the symbol sequence length L and the delay factor t, it is necessary to i Truncated to a short symbol sequence {h k}:

[0076] h k =(h k ,h k+t , ..., h k+(L-1)t );

[0077] Where: k = 1, 2, ..., N-(L-1)t;

[0078] The truncated short symbol sequence {h k} is converted into decimal, and the dynamic characteristics of the original time series are characterized by statistically analyzing the number of occurrences of each symbol code, and the symbol dynamic entropy is finally obtained as:

[0079]

[0080] Where: A SDEn (L) is the symbol dynamics entropy of a short symbol sequence with length L; P m，L Represents the probability corresponding to each symbol encoding in the encoding; M is the total number of different encodings that appear.

[0081] The decoder completed through training and the multi-feature index entropy obtain the reduced-dimensional health status feature and the reduced-dimensional current refund feature, including:

[0082] The data at the initial moment in the data set acquired during the online data collection is stacked after being calculated with multiple entropy feature indicators, and input into the trained decoder to obtain the features of the engine in a healthy state;

[0083] Obtaining multi-entropy feature indicators of data at the current moment, and inputting them into the trained decoder to obtain degradation state characteristics of the engine in the current state;

[0084] The Bray-Curtis distance between the healthy state feature and the degraded state feature is obtained, and the index data is used as the engine operation state HI.

[0085] The obtaining of the Bray-Curtis distance between the healthy state feature and the degraded state feature includes:

[0086] The calculation method is as follows:

[0087]

[0088] Where: x pk is the kth feature under the health status feature; x qk is the kth feature under the degraded state feature.

[0089] Get the HI value at the current moment, and then insert it into the HI sequence in chronological order;

[0090] The current HI sequence is filtered using a moving average filter.

[0091] Get the root mean square X of the HI sequence at the current moment RMS , and compare the current HI value with the failure threshold (4X RMS ) is compared. If the failure threshold is reached, the engine has reached the degradation stage, and the HI value at this time is set as the failure threshold of the engine.

[0092] The formula for calculating the root mean square is as follows:

[0093]

[0094] Where: N is the number of HI samples at the current moment; x i is the HI value at the current moment.

[0095] The minimization loss function J of the denoising autoencoder is:

[0096]

[0097] Where: x is the original input data; To reconstruct sample data; is the reconstruction error of a single sample;

[0098] It is represented by the mean square error between the input layer and the reconstruction layer, that is:

[0099] Wherein, N1 is the number of samples of HI value obtained at the end of the operation time.

[0100] The beneficial effects of the present invention include at least:

[0101] The method of the present invention first acquires the health status data and the current status data of the engine to be evaluated through data collection in the online stage; then, the health status data and the current status data are used to obtain the entropy of multiple feature indicators and stack them; then, a denoising autoencoder is trained in the offline stage to obtain a trained decoder; and the health status features after dimensionality reduction and the current refund features after dimensionality reduction are obtained through the trained decoder and the entropy of multiple feature indicators; finally, the health indicators of the engine to be evaluated are obtained according to the health status features after dimensionality reduction and the current refund features after dimensionality reduction; the method of the present invention solves the problems of unclear representation of single feature indicator information and omissions in the feature screening process. The problem of information is solved, and the Bray-Curtis distance is used to construct the health index during the operation of the engine. Taking into account the problem of inaccurate selection of the failure threshold, the root mean square value jump moment of the health index is regarded as the failure moment to construct the failure threshold; moreover, the proposed operating state evaluation method and failure threshold method are highly sensitive to slight changes in the early operating state of the engine; when the operating state of the engine changes or an early fault occurs, these changes can be intuitively reflected by the sudden increase of the health index value; finally, the health index value after multi-entropy fusion contains more engine operating state characteristics, so the engine operating state evaluation method proposed by the present invention has high feasibility and effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 This is the structure diagram of the denoising autoencoder;

[0103] Figure 2 Construct a flow chart for HI;

[0104] Figure 3 Construct a flow chart for failure thresholds. DETAILED DESCRIPTION

[0105] The present application is further described below in conjunction with the accompanying drawings.

[0106] The present invention conducts in-depth research on the more critical operating status assessment link in engine health management, mainly to solve the problems of a large amount of environmental noise interference in the selected data set and the poor effectiveness of single feature information. The present invention proposes an engine health status assessment method based on vibration signal analysis. The adopted method and model structure play a vital role in the characterization of the engine operating status. Specifically, first, the multi-entropy index of the vibration signal is extracted, and the multi-information features are input into the denoising autoencoder for information fusion and dimensionality reduction, so as to construct a health index; secondly, the failure threshold is constructed according to the jump time of the root mean square value in the vibration signal.

[0107] Moreover, in view of the problems in the prior art, in order to ensure the effectiveness of the extracted engine operating state degradation features, the present invention comprehensively considers the multi-domain feature information of the signal such as the time domain and frequency domain in the feature extraction process. However, the selection and extraction of multi-domain feature indicators often significantly increase the calculation amount of the model, further exacerbating the problems of redundant information and computational complexity. To this end, the present invention proposes an improved method, which uses a denoising autoencoder to stack and reduce the dimension of multi-entropy feature indicators. The main advantages of this method include: effectively reducing the complexity of high-dimensional feature information and reducing the interference of redundant features on model performance; improving the effectiveness of feature information extraction and model operation efficiency, and achieving faster and more accurate health status assessment; while eliminating redundant features, retaining key feature information, so that the engine operating state assessment can cover global features and avoid the omission of important features. Through the above-mentioned improvement scheme, the present invention significantly improves the operating efficiency of the model on the basis of ensuring the accuracy of the engine operating state assessment method, and provides scientific and reliable technical support for engine health management.

[0108] The method provided by the present invention is described in detail below with reference to the accompanying drawings.

[0109] The operation process of the present invention is as follows: constructing a health index (HI) through two stages, offline and online. In the offline stage, a model is trained to obtain an encoder that can effectively extract features by training the minimum loss function; in the online stage, the HI is constructed using the trained model.

[0110] Specifically, the present embodiment discloses an embodiment of an engine operating state assessment method integrating a denoising autoencoder and multi-entropy, comprising: acquiring health state data and current state data of the engine to be assessed through online data collection; obtaining multi-feature indicator entropy using the health state data and the current state data, and stacking the data; training the denoising autoencoder in an offline stage to obtain a trained decoder; obtaining health state features and current refund features after dimensionality reduction through the trained decoder and the multi-feature indicator entropy; and obtaining health indicators of the engine to be assessed based on the health state features and current refund features after dimensionality reduction.

[0111] Among them, the denoising autoencoder (DAE) is an unsupervised neural network that contains implicit features and is mainly composed of an input layer, a processing layer, a hidden layer, and a reconstruction layer. Its basic model is as follows: Figure 1 In the operation of DAE, x i represents the original input data, and yi is the dimension-reduced data obtained through encoding learning, such as Figure 1 ;

[0112] The core idea of ​​DAE is as follows: given a training sample set N, for each input raw data x(x∈N), according to q D The distribution is processed by adding noise. This process actually clears the nodes of certain dimensions to zero with a certain probability, and then uses these node-cleared data for training. Through this process, the feature y reduced to t dimensions is learned. Using these features y, the original k-dimensional input data is reconstructed During the entire training process, DAE optimizes the model by adjusting model parameters to minimize the loss function.

[0113] For the sample set N, the minimization loss function J of DAE can be expressed as:

[0114]

[0115] Where: x is the original input data; To reconstruct sample data; is the reconstruction error of a single sample.

[0116] It can be expressed by calculating the mean square error between the input layer and the reconstruction layer, that is:

[0117]

[0118] Under the learning condition of optimizing the reconstruction error, the internal mapping of DAE has strong robustness, which significantly reduces the sensitivity of the model to small random perturbations. This enables DAE to not only effectively reduce the dimensionality of data, but also has a certain noise reduction capability, thereby improving the stability and performance of the model.

[0119] The construction process of HI is as follows Figure 2 As shown, the specific steps are as follows:

[0120] (1) Obtain the vibration acceleration signal of the engine throughout its life cycle.

[0121] (2) Take the first 2048 frequency points of each data set as training samples for the model.

[0122] (3) The characteristic indicators of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy are calculated for the intercepted training samples.

[0123] (4) Select Adam as the optimizer and use the training set to train DAE. The loss function is:

[0124]

[0125] Where: k is the number of feature samples contained in the input data; y (i) is the eigenvalue of the i-th sample after reconstructing the feature sample; x (i) is the eigenvalue of the i-th sample of the input data; λ is the L2 regularization coefficient; m is the number of weights in the neural network; w i is the i-th weight in the neural network.

[0126] The loss function consists of two parts: smooth L1 loss function and L2 regularization. Compared with the L1 loss function, the smooth L1 loss function has significant advantages. It can not only accelerate the convergence of the model, but also is not easily disturbed by outliers and abnormal values. The introduction of L2 regularization can effectively solve the problem that may arise when the weight matrix is ​​too large: that is, a small change in the input data causes a large fluctuation in the eigenvalue. By applying L2 regularization, the generalization performance of the encoder can be significantly improved, making it have stable performance.

[0127] (5) When the loss value stops decreasing, the training is considered complete, and an encoder that can effectively extract features and has dimensionality reduction capabilities is obtained.

[0128] Figure 2 The specific operations for obtaining HI in the online stage are as follows:

[0129] (1) The data at the initial moment of the dataset is stacked after calculating the multi-entropy feature index and input into the trained DAE to obtain the characteristics of the engine in a healthy state.

[0130] (2) Calculate the multi-entropy feature indicators of the current data and input these indicators into the trained DAE to obtain the degradation state characteristics of the engine in the current state.

[0131] (3) Calculate the Bray-Curtis distance between the healthy state characteristics and the degraded state characteristics, and use this index data as the engine's operating state HI.

[0132] Bray-Curtis distance is an important indicator used in the field of ecology to evaluate the similarity between biological communities. It comprehensively considers the presence or absence of species and the relative abundance of species. In the calculation process, Bray-Curtis distance not only reflects the presence or absence of species, but also reflects the relative quantity difference between species. When applied to the construction of engine HI value, Bray-Curtis distance can effectively measure the degree of difference in characteristic values ​​between healthy state characteristics and degraded state characteristics. In addition, Bray-Curtis distance has the advantage of normalized difference, which can eliminate the influence of different characteristic dimensions.

[0133] The Bray-Curtis distance ranges from 0 to 1. When the index data is closer to 0, it means that the difference between the two features is smaller, which reflects that the engine is in a healthier state. On the contrary, when the data is closer to 1, it means that the difference between the two features is larger, indicating that the engine is gradually approaching failure. The specific calculation method is as follows:

[0134]

[0135] Where: x pk is the kth feature under the health status feature; x qk is the kth feature under the degraded state feature.

[0136] For the HI values ​​established by the present invention, especially some HIs without clear physical meaning, there is no standard to determine their failure thresholds. Therefore, in many studies, it is assumed that the root mean square (RMS) value of the data sample can reflect the degree of rolling engine degradation, and the significant amplitude jump of RMS is considered to be the state change moment of the engine life. Under the actual trend, the growth at this time is very rapid. The present invention uses the RMS jump moment as the state change time of the engine, and uses the amplitude of HI at this moment as the failure threshold of the engine.

[0137] According to the above content, the moment when RMS jumps to 4 times the normal value is regarded as the failure moment, and the HI value at this time is the failure threshold. In order to reduce the interference of redundant information on the model and improve robustness, the present invention uses a moving average filter to flatten the current HI value, with a window size of 5. The specific process is as follows Figure 3 As shown, the specific steps are as follows:

[0138] (1) Get the HI value at the current moment and then insert it into the HI sequence in chronological order.

[0139] (2) In order to obtain a smoother HI sequence, a moving average filter is used to filter the current HI sequence.

[0140] (3) Calculate the root mean square X of the HI sequence at the current moment RMS , and compare the current HI value with the failure threshold (4X RMS ) is compared. If the failure threshold is reached, the engine has reached the degradation stage. The HI value at this time is set as the failure threshold of the engine. The calculation formula of the root mean square is as follows:

[0141]

[0142] Where: N is the number of HI samples at the current moment; x i is the HI value at the current moment.

[0143] As for the characteristic index calculation of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy for the intercepted training samples described in this embodiment, the specific steps are as follows.

[0144] Entropy is a thermophysics concept used to characterize the amount of information. In recent years, many experts and scholars have introduced entropy features into the field of feature extraction, and have realized the datafication of operating status by calculating various entropy values ​​of vibration signals. Entropy is a dimensionless indicator that characterizes the complexity of time series. The larger the entropy value, the greater the complexity of the disorder of the signal. The degree of signal complexity is of great significance in engine status assessment and fault diagnosis. It can resist the interference of environmental factors and has good application in the field of feature extraction.

[0145] Relative Entropy (ReEn) is also called information divergence, divergence entropy, etc. ReEn refers to the asymmetric measure of the difference in probability distribution between two time series M and N. For a continuous random variable or two probability distributions of discrete random variables, its calculation method is as follows:

[0146]

[0147] The calculation method of Approximate Entropy (ApEn) is to calculate the self-similarity of the time series pattern and use the change of the approximate entropy value to achieve the purpose of identifying the change of the signal sequence. It has the advantages of less data calculation and can meet the needs of most time series. The result is effective. The calculation method is as follows:

[0148] (1) Define the time series X(i), and construct a vector given the pattern dimension q, that is:

[0149] X(n)=[x(n),x(n+1),...,x(n+q-1)];

[0150] Where: X(n) is a q-dimensional vector; n=1,2,…,i-q+1.

[0151] (2) Define d[X(n), X(m)] as the maximum value of the range of the distance between vectors X(n) and X(m), that is:

[0152]

[0153] (3) Given a similarity tolerance r (r>0), calculate the ratio of the maximum distance range to i-q+1, perform a logarithmic operation on it, and then find its average value for all n, that is:

[0154]

[0155] Where: It is the ratio of the maximum value of the distance range to i-q+1.

[0156] (4) Let q = q + 1 and repeat the above steps to obtain and C q+1 (r), for the actual sequence, i cannot approach infinity, so its approximate entropy can be simplified to:

[0157] A ApEn (q, r, i) = φ q (r)-φ q+1 (r);

[0158] Permutation Entropy (PmEn) improves computational efficiency and anti-interference capability by introducing the idea of ​​permutation, and has a higher sensitivity to mutations in the signal. The calculation method is as follows:

[0159] (1) Reconstruct the original time series into a phase space matrix Y with j rows and q columns, namely:

[0160]

[0161] Where: q is the embedding dimension; t is the delay factor (the value must be a positive integer, if t = 1, the sequence is defined the same as the approximate entropy sequence).

[0162] (2) Rearrange the elements of each row in the matrix in ascending order, and let the index of each element after arrangement be λ α (α=1,2,…,q) By sorting, the index of each row in the matrix can be expressed by the following formula:

[0163] S(w)=(λ1,λ2,…λ q ), w=1,2,…,j,j≤q!;

[0164] Where: There are q! different arrangements of the q-dimensional phase space mapping.

[0165] (3) The permutation entropy of the time series is obtained as:

[0166]

[0167] Where: P w The probability of each arrangement being present divided by the total number of q! different arrangements being present is q!

[0168] Fuzzy Entropy (FuEn) introduces a fuzzy membership function based on sample entropy. After introducing this idea, the algorithm can more directly characterize the intrinsic characteristics of signal information. The calculation method is as follows:

[0169] (1) Reconstruct the original time series in phase space to obtain the time series Y, that is:

[0170] Y=[x(i),x(i+1),...,x(i+q-1),]-x0(i);

[0171]

[0172] Where: q is the embedding dimension; x0(i) is the mean.

[0173] (2) Define the distance between the two decomposed time series Y(i) and Y(j) as Right now:

[0174]

[0175] (3) Introduce the fuzzy membership function and use the fuzzy function to calculate the similarity between two time series, that is:

[0176]

[0177] Where: r is the similarity tolerance. i,j=1,2,…,K-m+1, and i≠j; K is the length of the time series.

[0178] (4) The fuzzy entropy of the original time series is:

[0179]

[0180] Where: q (r) is the defined function.

[0181] Dispersion Entropy (DisEn) is an indicator used to measure the irregularity or variability of time series signals. It quantifies the complexity of signals by analyzing the distribution and degree of discreteness of values. The probability of the appearance of potential models in discrete statistical signals is calculated based on the theory of secure encryption algorithms. Since it was proposed, it has been applied by many scholars to the feature extraction of complex nonlinear signals. The results show that the algorithm has a strong feature capture capability. The calculation method is as follows:

[0182] (1) Transform the original time series into a normal cumulative distribution function and map the series to a new domain y i (i=1, 2,...,N):

[0183]

[0184] Where: μ and σ are the mean and standard deviation of the time series x(i) respectively.

[0185] (2) Use the embedding dimension q and the number of categories c to construct all possible discrete models The total number of model categories is c q , y i All data in are automatically assigned to integers from 1 to c and are converted to Constructed as vector group

[0186]

[0187] (3) Make With the potential model Matching, use the following formula to count the frequency of each potential model appearing in the original time series:

[0188]

[0189] Where: * is the total number of times each potential model appears.

[0190] (4) The final calculated discrete entropy of the original time series is:

[0191]

[0192] Symbolic Dynamic Entropy (SDEn) is a method for analyzing time series data. It is based on the concept of symbolic dynamics and is used to study the behavior of nonlinear dynamic systems. The basic idea is to convert continuous time series data into discrete symbol sequences and then calculate the entropy of these symbol sequences. This method can be used to study the complexity, uncertainty and confusion of time series data. The calculation method is as follows:

[0193] (1) Binarize the original vibration signal X(i) with a sequence length of N to obtain the symbol sequence h i :

[0194]

[0195] Where: is the sample mean.

[0196] (2) After determining the symbol sequence h i After that, in order to determine the symbol sequence length L and the delay factor t, it is necessary to i Truncated to a short symbol sequence {h k}:

[0197] h k =(h k ,h k+t , ..., h k+(L-1)t );

[0198] Where: k = 1, 2, ..., N-(L-1)t; the determination of L directly affects the description of the signal. The smaller L is, the more it can highlight the local characteristics of the signal, and vice versa. The selection of the delay factor t currently mostly relies on experience. This paper determines L = 3, t = 1.

[0199] (3) The truncated short symbol sequence {h k} is converted into decimal, and the dynamic characteristics of the original time series are characterized by statistically analyzing the number of occurrences of each symbol code. Finally, the symbol dynamic entropy can be calculated as:

[0200]

[0201] Where: A SDEn (L) is the symbol dynamics entropy of a short symbol sequence with length L; P m，L Represents the probability corresponding to each symbol encoding in the encoding; M is the total number of different encodings that appear.

[0202] The above disclosures are only several specific implementation scenarios of the present invention, but the present invention is not limited thereto, and any changes that can be thought of by a person skilled in the art should fall within the protection scope of the present invention. The above serial numbers of the present invention are only for description and do not represent the advantages and disadvantages of the implementation scenarios.

Claims

1. A method for evaluating engine operating status by fusion of denoising autoencoder and multi-entropy, characterized in that: include: Through online data collection, the health status data and current status data of the engine to be evaluated are obtained; Using the health status data and the current status data, a multi-feature index entropy is obtained and stacked; Perform denoising autoencoder training in the offline stage to obtain a trained decoder; Obtaining the reduced-dimensional health status feature and the reduced-dimensional current refund feature through the trained decoder and the multi-feature index entropy; The health index of the engine to be evaluated is obtained according to the reduced-dimensional health status feature and the reduced-dimensional current return feature.

2. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 1 is characterized in that: The denoising autoencoder training is performed in the offline stage to obtain a trained decoder, including: Acquiring a full-life vibration acceleration signal of the engine to be evaluated, and using the full-life vibration acceleration signal as a sample; Extract the first N frequency points of the data in the sample; Get the entropy of multiple feature indicators and stack them; Training the denoising autoencoder; When the loss value stops decreasing, the training is completed, and a trained decoder that can effectively extract features and has dimensionality reduction capabilities is obtained.

3. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 2 is characterized in that: Obtaining the full-life vibration acceleration signal of the engine to be evaluated, and taking the first 2048 frequency points of each data set as training samples; Calculating characteristic indices of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy for the intercepted training samples; The DAE is trained using the training set using Adam as the optimizer.

4. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 3 is characterized in that: The loss function when training DAE using the training set is: Among them, k is the number of feature samples contained in the input data; y (i) is the eigenvalue of the i-th sample after reconstructing the feature sample; x (i) is the eigenvalue of the i-th sample of the input data; λ is the L2 regularization coefficient; m is the number of weights in the neural network; w i is the i-th weight in the neural network.

5. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 3 is characterized in that: The characteristic index calculation of relative entropy, approximate entropy, permutation entropy, fuzzy entropy, discrete entropy and symbolic dynamics entropy for the intercepted training samples includes: The steps for obtaining the relative entropy are as follows: The steps for obtaining the approximate entropy are as follows: Define the time series X(i), construct a vector given the pattern dimension q, that is: X(n)=[x(n), x(n+1), ..., x(n+q-1)]; where: X(n) is a q-dimensional vector; n=1, 2, ..., i-q+1; Define d[X(n), X(m)] as the maximum value of the range of the distance between vectors X(n) and X(m), that is: Given a similarity tolerance r, r>0, obtain the ratio of the maximum value of the distance range to i-q+1, perform a logarithmic operation on it, and then calculate its average value for all n, that is: Where: It is the ratio of the maximum value of the distance range to i-q+1; Let q = q + 1, and we get and C q+1 (r), for the actual sequence, the approximate entropy is: A ApEn (q,r,i)=φ q (r)-φ q+1 (r); The steps for obtaining the permutation entropy are as follows: The original time series is reconstructed in phase space and transformed into the following matrix Y with j rows and q columns, namely: Where: q is the embedding dimension; t is the delay factor; Rearrange the elements of each row in the matrix Y in ascending order, and let the index of each element after arrangement be λ α (α=1,2,…,q), by sorting, the index of each row in the matrix can be expressed by the following formula: S(w)=(λ1,λ2,…λ q ),w=1,2,…,j,j≤q!; Where: There are q! different arrangements of the q-dimensional phase space mapping; Get the permutation entropy of the time series as: Where: P w The probability of each arrangement being present divided by the total number of times q! different arrangements appear is: The steps for obtaining the fuzzy entropy are as follows: The original time series is reconstructed in phase space to obtain the time series Y, that is: Y=[x(i),x(i+1),...,x(i+q-1),]-x0(i); Where: q is the embedding dimension; x0(i) is the mean; The distance between the two decomposed time series Y(i) and Y(j) is defined as Right now: The fuzzy membership function is introduced and the similarity between two time series is calculated using the fuzzy function, namely: Where: r is the similarity tolerance; i, j = 1, 2, ..., K-m + 1, and i ≠ j; K is the length of the time series; The fuzzy entropy of the original time series is: Where: q (r) is the definition function; The steps to obtain discrete entropy are as follows: Transform the original time series into a normal cumulative distribution function and map the series to the new domain y i (i=1, 2,...,N): Where: μ and σ are the mean and standard deviation of the time series x(i) respectively; Use the embedding dimension q and the number of categories c to construct all possible discrete models The total number of model categories is c q , y i All data in are automatically assigned to integers from 1 to c and are converted to Constructed as vector group make With the potential model Matching, use the following formula to count the frequency of each potential model appearing in the original time series: Where: * is the total number of occurrences of each potential model; Get the discrete entropy of the original time series: The steps to obtain symbolic dynamics entropy are as follows: Binarize the original vibration signal X(i) with a sequence length of N to obtain the symbol sequence h i : Where: is the sample mean; In determining the symbol sequence h i After that, in order to determine the symbol sequence length L and the delay factor t, it is necessary to i Truncated to a short symbol sequence {h k }: h k =(h k ,h k+t ,...,h k+(L-1)t ); Where: k = 1, 2, ..., N-(L-1)t; The truncated short symbol sequence {h k } is converted into decimal, and the dynamic characteristics of the original time series are characterized by statistically analyzing the number of occurrences of each symbol code, and the symbol dynamic entropy is finally obtained as: Where: A SDEn (L) is the symbol dynamics entropy of a short symbol sequence with length L; P m，L Represents the probability corresponding to each symbol encoding in the encoding; M is the total number of different encodings that appear.

6. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 1 is characterized in that: The decoder completed through training and the multi-feature index entropy obtain the reduced-dimensional health status feature and the reduced-dimensional current refund feature, including: The data at the initial moment in the data set acquired during the online data collection is stacked after being calculated with multiple entropy feature indicators, and input into the trained decoder to obtain the features of the engine in a healthy state; Obtaining multi-entropy feature indicators of data at the current moment, and inputting them into the trained decoder to obtain degradation state characteristics of the engine in the current state; The Bray-Curtis distance between the healthy state feature and the degraded state feature is obtained, and the index data is used as the engine operation state HI.

7. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 6 is characterized in that: The obtaining of the Bray-Curtis distance between the healthy state feature and the degraded state feature includes: The calculation method is as follows: Where: x pk is the kth feature under the health status feature; x qk is the kth feature under the degraded state feature.

8. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 6 is characterized in that: Get the HI value at the current moment, and then insert it into the HI sequence in chronological order; The current HI sequence is filtered using a moving average filter. Get the root mean square X of the HI sequence at the current moment RMS , and compare the current HI value with the failure threshold 4X RMS By comparison, if the failure threshold is reached, the engine has reached the degradation stage, and the HI value at this time is set as the failure threshold of the engine.

9. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 8 is characterized in that: The formula for calculating the root mean square is as follows: Where: N is the number of HI samples at the current moment; x i is the HI value at the current moment.

10. The engine operating state assessment method of denoising autoencoder and multi-entropy fusion according to claim 1 or 2, characterized in that: The minimization loss function J of the denoising autoencoder is: Where: x is the original input data; To reconstruct sample data; is the reconstruction error of a single sample; It is represented by the mean square error between the input layer and the reconstruction layer, that is: Wherein, N1 is the number of samples of HI value obtained at the end of the operation time.

Images