Adaptive timestamp encoding enhanced complex equipment missing state monitoring data time series interpolation method

Abstract

The application discloses a kind of complex equipment incomplete state monitoring data time sequence interpolation methods of adaptive timestamp coding enhancement, it is related to industrial sensing data processing technical field, the method includes: the online incomplete state monitoring time sequence data obtained is standardized, then online timestamp sequence data is normalized, then utilize timestamp prior feature coding module output online timestamp prior coding feature, realize the effective use of timestamp information in incomplete state monitoring data in interpolation task;Then utilize the timestamp-local interpolation model joint framework trained, obtain incomplete state monitoring data online interpolation result;The scheme realizes the complementary advantages between locality time sequence interpolation model and timestamp probabilistic interpolation model, effectively improves the accuracy and robustness of complex equipment state monitoring data time sequence interpolation, can be widely applied to the data enhancement link in various complex equipment operation state analysis tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial sensor data processing, in particular to a complex equipment incomplete state monitoring data time sequence interpolation method with adaptive timestamp encoding enhancement. BACKGROUND

[0002] In the field of industrial intelligent operation and maintenance in the era of big data, complex equipment state monitoring data is widely used in predicting equipment performance, diagnosing faults, and predicting life, etc. The time sequence integrity of the state monitoring data is a key prerequisite for ensuring the accurate execution of these tasks. However, in actual industrial scenarios, due to the combined effects of sensor failure, communication interference, equipment failure downtime, and other factors, the collected complex equipment state monitoring data often has some missing, which interferes with the accurate perception and judgment of the equipment state, and thus affects the effectiveness of equipment operation decision and predictive maintenance. Therefore, effective interpolation of incomplete complex equipment state monitoring data not only restores the time sequence continuity of the data, but also improves the data quality, thereby enhancing the reliability of fault diagnosis and performance prediction, improving equipment safety, and ensuring the efficient operation of complex equipment. Among the many interpolation methods, the interpolation method based on deep learning can more effectively model and fully utilize the nonlinear correlation characteristics in the multivariate state monitoring data, has the advantages of strong representation ability and suitability for high-dimensional data, and has been widely used in complex equipment state monitoring data time sequence interpolation in recent years.

[0003] The existing mainstream deep learning time sequence interpolation method usually continuously intercepts a specified local time window size of incomplete state monitoring data in the time dimension as input, infers the missing data based on the modeling of the correlation characteristics between the visible data and the missing data, and completes the time sequence interpolation. Although this kind of local interpolation method has achieved good results, it still does not fully utilize the time stamp sequence data in the complex equipment state monitoring data that does not have a missing problem. On the one hand, most interpolation methods only limit the use of time stamps to position encoding or calculating missing time lag matrices within the local time window corresponding to a single sample to indicate the relative position relationship of input data features in the time dimension, lack of mining of the time sequence characteristics inherent in the time stamp sequence data itself, and cannot directly model the macro-scale periodic time sequence correlation pattern beyond the local window range. On the other hand, some existing methods embed and encode the time stamp information based on prior knowledge to obtain several dominant periods of complex equipment time sequence data, and fuse it with the input time sequence data features to achieve improved interpolation performance, but it highly depends on domain prior knowledge and is difficult to generalize to different industrial fields. In summary, the lack of automatic and effective use of time stamp information makes it difficult for existing interpolation methods to adapt to the time sequence interpolation problem of complex equipment state monitoring data with uneven data missing and insufficient observable data in some samples in the absence of prior knowledge. SUMMARY

[0004] The application aims to provide a complex equipment incomplete state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement, which can effectively utilize the timestamp information in complex equipment state monitoring data and improve the precision and robustness of the interpolation method.

[0005] To achieve the above object, the application provides the following solutions.

[0006] The application provides a complex equipment incomplete state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement, comprising the following steps.

[0007] Obtaining online incomplete state monitoring data of complex equipment; the online incomplete state monitoring data comprises online incomplete state monitoring time sequence data and corresponding online timestamp sequence data of the complex equipment; the online incomplete state monitoring time sequence data is the monitoring value of a plurality of state monitoring variables of the complex equipment at a plurality of continuous sampling steps.

[0008] Standardizing the online incomplete state monitoring time sequence data to obtain preprocessed online incomplete state monitoring time sequence data.

[0009] After normalizing the online timestamp sequence data, outputting online timestamp prior coding features by using a timestamp prior feature coding module; the parameters of the timestamp prior feature coding module are obtained based on an unsupervised frequency domain prior mining model.

[0010] Inputting the preprocessed online incomplete state monitoring time sequence data and the online timestamp prior coding features into a trained timestamp-local interpolation model joint framework to obtain online interpolation results of the incomplete state monitoring data; the timestamp-local interpolation model joint framework is obtained based on a timestamp probability interpolation model and a local time sequence interpolation model.

[0011] According to the specific embodiments provided by the application, the following technical effects are disclosed.

[0012] The application provides a complex equipment incomplete state monitoring data time sequence interpolation method enhanced by adaptive timestamp coding, in which, for the obtained online incomplete state monitoring data of the complex equipment, the online incomplete state monitoring time sequence data is first standardized to obtain preprocessed online incomplete state monitoring time sequence data, and then the online timestamp sequence data is normalized, and then the timestamp prior feature coding module is used to output online timestamp prior coding features; the parameters of the timestamp prior feature coding module are obtained based on an unsupervised frequency domain prior mining model, so that the timestamp information in the incomplete state monitoring data is effectively utilized in the interpolation task; then, the preprocessed online incomplete state monitoring time sequence data and the online timestamp prior coding features are input into the trained timestamp-local interpolation model joint framework to obtain online incomplete state monitoring data interpolation results; the timestamp-local interpolation model joint framework is obtained based on a timestamp probabilistic interpolation model and a local time sequence interpolation model, so that the advantages of the local time sequence interpolation model and the timestamp probabilistic interpolation model are complementary, the accuracy and robustness of the complex equipment state monitoring data time sequence interpolation are effectively improved, and the method can be widely applied to the data enhancement link in various complex equipment operation state analysis tasks. The above scheme of the application realizes adaptive feature coding of the complex equipment macro-scale inherent periodicity time sequence mode information corresponding to the timestamp information without introducing any external prior knowledge, and effectively enhances the interpolation performance of the existing local time sequence interpolation model through the design of the joint framework, and is suitable for the time sequence interpolation of complex equipment state monitoring data in various complex equipment state monitoring data missing scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0014] Figure 1 A flowchart of a complex equipment incomplete state monitoring data time sequence interpolation method enhanced by adaptive timestamp coding provided by an embodiment of the present application.

[0015] Figure 2 A flowchart of training and optimizing parameters of a timestamp-local interpolation model joint framework in a complex equipment incomplete state monitoring data time sequence interpolation method enhanced by adaptive timestamp coding provided by an embodiment of the present application.

[0016] Figure 3A flow chart of step B1 in a complex equipment missing state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement provided by an embodiment of the present application.

[0017] Figure 4 A flow chart of step B4 in a complex equipment missing state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement provided by an embodiment of the present application.

[0018] Figure 5 A flow chart of step B42 in a complex equipment missing state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement provided by an embodiment of the present application.

[0019] Figure 6 A technical concept schematic diagram of a complex equipment missing state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement provided by an embodiment of the present application. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0021] The above purposes, features and advantages of the present application will be more obvious and easy to understand. The present application will be described in further detail below with the drawings and specific embodiments.

[0022] A complex equipment missing state monitoring data time sequence interpolation method with adaptive timestamp coding enhancement provided by an embodiment of the present application is shown in a flow chart as Figure 1 The method comprises the following steps:

[0023] A1, obtaining online missing state monitoring data of complex equipment; the online missing state monitoring data comprises online missing state monitoring time sequence data and corresponding online timestamp sequence data of the complex equipment; the online missing state monitoring time sequence data is monitoring values of a plurality of state monitoring variables of the complex equipment at a plurality of continuous sampling steps.

[0024] A2, normalizing the online missing state monitoring time sequence data to obtain preprocessed online missing state monitoring time sequence data.

[0025] A3, after normalizing the online timestamp sequence data, outputting online timestamp prior coding features by using a timestamp prior feature coding module; parameters of the timestamp prior feature coding module are obtained based on an unsupervised frequency domain prior mining model.

[0026] A4, inputting the pretreated online incomplete condition monitoring time series data and online timestamp priori encoding features into the trained timestamp-local interpolation model joint framework to obtain an online interpolation result of the incomplete condition monitoring data.

[0027] Specifically, before step A3, the adaptive timestamp encoding enhanced complex equipment incomplete condition monitoring data time series interpolation method provided in this embodiment further includes a process of training and optimizing parameters of the timestamp-local interpolation model joint framework, as shown in Figure 2 The process includes the following steps:

[0028] B1, according to offline monitoring data of the complex equipment under historical scenarios, interpolation data training set, interpolation data validation set and interpolation data test set are constructed. The interpolation data training set is used to train the timestamp-local interpolation model joint framework; the interpolation data validation set and the interpolation data test set are used to verify and optimize the trained timestamp-local interpolation model joint framework and evaluate the performance. In this embodiment, as shown in Figure 3 The process includes the following steps:

[0029] B11, initial incomplete condition monitoring data of the complex equipment is obtained; the initial incomplete condition monitoring data includes offline incomplete condition monitoring time series data and corresponding offline timestamp sequence data of the complex equipment; the offline incomplete condition monitoring time series data are monitoring values of a plurality of state monitoring variables of the complex equipment under a plurality of continuous sampling steps in historical scenarios.

[0030] B12, after the offline timestamp sequence data is normalized, non-overlapping data division is performed according to the initial incomplete condition monitoring data to obtain initial training data, initial validation data and initial test data.

[0031] B13, the offline incomplete condition monitoring time series data in the initial training data, the initial validation data and the initial test data are respectively standardized to obtain pretreated initial training data, pretreated initial validation data and pretreated initial test data.

[0032] B14, perform a sliding window operation on the preprocessed initial training data, the preprocessed initial validation data and the preprocessed initial test data respectively with the same size of local time window to construct an imputation data training set, an imputation data validation set and an imputation data test set; any sample in the imputation data training set, the imputation data validation set and the imputation data test set includes the offline incomplete local state monitoring time series data, the high-missing-rate incomplete local state monitoring time series data obtained by performing the manual masking operation on the offline incomplete local state monitoring time series data and the local timestamp sequence data.

[0033] B2, input the high-quality offline monitoring data under the historical scenario into the frequency domain prior mining model to output the prior core frequency information and the prior phase information corresponding to each state monitoring variable on the basis of obtaining the clustering label of the state monitoring variable. Specifically, in the embodiment, the offline incomplete state monitoring time series data in the preprocessed initial validation data is taken as the high-quality offline monitoring data under the historical scenario, which is input into the frequency domain prior mining model to output the prior core frequency information and the prior phase information corresponding to each state monitoring variable on the basis of obtaining the clustering label of the state monitoring variable.

[0034] B3, determine the parameters of the timestamp prior feature encoding module based on the prior core frequency information and the prior phase information corresponding to each state monitoring variable.

[0035] B4, train the timestamp-local imputation model joint framework by using the imputation data training set to obtain the trained timestamp-local imputation model joint framework. In the embodiment, as shown in FIG. 4, step B4 specifically includes the following steps: Figure 4

[0036] B41, input the local timestamp sequence data in the imputation data training set into the timestamp prior feature encoding module to obtain offline timestamp prior encoding features.

[0037] B42, input the offline timestamp prior encoding features into the timestamp probabilistic imputation model to obtain a local prior distribution mean estimate value, a local prior distribution standard deviation estimate value and a timestamp imputation result.

[0038] B43, input the high-missing-rate incomplete local state monitoring time series data in the imputation data training set into the locality time series imputation model to obtain a locality imputation result. In the embodiment, step B43 specifically includes the following steps:

[0039] B431, construct a missing value mask by detecting missing data position information of the high-missing-rate incomplete local state monitoring time series data.

[0040] ​B432, inputting the high-missing-rate incomplete local state monitoring time series data and the missing value mask into the locality time series imputation model to obtain a locality imputation result.

[0041] B44, determining a final imputation result based on the timestamp imputation result and the locality imputation result.

[0042] B45, constructing a joint loss function based on the high-missing-rate incomplete local state monitoring time series data, the local prior distribution mean estimate value, the local prior distribution standard deviation estimate value, and the final imputation result.

[0043] B46, training the timestamp-locality imputation model joint framework based on gradient back propagation according to the joint loss function to obtain a trained timestamp-locality imputation model joint framework.

[0044] In an exemplary embodiment, the timestamp probabilistic imputation model comprises a distribution encoding module, a distribution aggregation module, and a reparameterization sampling module; the distribution encoding module comprises a plurality of parameter-independent distribution encoders; the distribution encoders are used to process state monitoring variables corresponding to cluster labels; the distribution encoders comprise a plurality of parameter-independent distribution estimation networks, each corresponding to each core frequency component of the state monitoring variables.

[0045] In this embodiment, as shown in Figure 5 , step B42 specifically comprises the following steps:

[0046] B421, inputting the offline timestamp prior encoding features into the corresponding distribution encoders to output distribution parameters corresponding to each core frequency component; the distribution parameters include mean estimate values and log variance estimate values.

[0047] B422, using the distribution aggregation module to aggregate the mean estimate values and the log variance estimate values corresponding to each core frequency component respectively to output local prior distribution mean estimate values and local prior distribution log variance estimate values.

[0048] B423, transforming the local prior distribution log variance estimate values into local prior distribution standard deviation estimate values, and randomly sampling an auxiliary noise quantity from a standard normal distribution.

[0049] B424, sampling a timestamp imputation result based on the local prior distribution mean estimate values, the local prior distribution standard deviation estimate values, and the auxiliary noise quantity.

[0050] After obtaining the trained timestamp-locality imputation model joint framework through the above process, the following steps are further included:

[0051] B5, the trained timestamp-local interpolation model joint framework is verified and optimized and performance evaluation is conducted by using the interpolation data verification set and the interpolation data test set. The way each sample in the interpolation data verification set and the interpolation data test set inputs the trained timestamp-local interpolation model joint framework is the same as that of each sample in the interpolation data training set, and the final interpolation result output by each sample is compared with the offline missing local state monitoring time series data in each sample to calculate the interpolation error, thereby realizing the verification optimization and performance evaluation of the interpolation model joint framework.

[0052] The above embodiments describe the adaptive timestamp encoding enhanced complex equipment missing state monitoring data time series interpolation method provided by the present application from the perspective of practical application. The following describes the above method provided by the present application from the perspective of theoretical derivation, and the overall technical concept is as shown in Figure 6

[0053] A gas turbine is a complete power machinery taking continuous flow gas as working medium, which is composed of a compressor, a combustion chamber, a gas turbine and other components, has the characteristics of complex structure and high performance coupling between sub-components, is known as the "crown of the pearl" in manufacturing industry, and is a very representative complex equipment. With the development of advanced sensing technology, Internet of Things and other technologies, a large amount of gas turbine state monitoring data including output power, blade passage temperature, exhaust pressure, combustion chamber shell internal pressure, main shaft speed and the like can be collected and analyzed. However, due to the influence of extreme operating environment such as high temperature and high pressure, and complex and variable operating conditions, the collected gas turbine state monitoring data generally has missing phenomenon. Time series interpolation of the missing gas turbine state monitoring data is beneficial to improve the integrity and continuity of the gas turbine operating state data, and provides a reliable data basis for equipment operating performance analysis and predictive maintenance.

[0054] In this embodiment, through cooperation with related enterprises, historical state monitoring data of a certain type of gas turbine in a real service state within a time interval is obtained, which includes the value of each data point itself and the corresponding timestamp information. Due to sensor failure, data transmission and storage failure and other reasons, part of the state monitoring data has missing problem. First, the following data preprocessing steps of the missing gas turbine state monitoring data are needed:

[0055] The missing gas turbine state monitoring data is arranged in sequence in time and variable dimension to obtain the initial missing state monitoring data Z of the complex equipment raw = {X raw , TS raw}. X raw represents the initial value of each state monitoring time series data, part of which is missing, and the size is T full × C, wherein T​full denotes the total time step number of the continuous uniform sampling process, C = 199 denotes the total number of state monitoring variables, TS raw denotes the time stamp sequence data of the initial incomplete state monitoring time series data, with a size of T full × 1, without missing phenomenon.

[0056] Since in the present embodiment, the encoding of the time stamp sequence data is determined adaptively by a learnable model, without directly using prior time period information such as year, month, day, etc., to unify the input of the time stamp prior feature encoding model, it is necessary to normalize the time stamp sequence data in the initial incomplete state monitoring data of the complex equipment. That is, after converting each time stamp into a long integer original time stamp value, according to the following formula, a unified translation and scaling are performed, so that the sampling frequency f s after normalization becomes 1:

[0057]

[0058] wherein, denotes the normalized time stamp information, and denote the minimum and second minimum values in all original time stamp values, respectively, and s denotes the difference value of the original time stamp values corresponding to adjacent time steps.

[0059] After completing the normalization of the time stamp sequence data, the initial state monitoring time series data value X raw and the normalized time stamp information are sequentially divided into non-overlapping data according to the proportion of 5:1:1 in the time dimension, to obtain the initial training data initial validation data initial test data The basic information of each type of state monitoring data after uniform division is shown in Table 1 below.

[0060] Table 1 Basic information of initial incomplete state monitoring data of gas turbine

[0061]

[0062] In addition, according to statistics, the initial missing rate of each type of data is less than 5%, and the data completeness is relatively high. In particular, the relative completeness of the validation data is conducive to ensuring the accuracy of the core frequency related information of each state monitoring variable learned by the subsequent frequency domain prior mining model.

[0063] Then, the mean μ and standard deviation σ of the time series data of each state monitoring variable in the initial training data feature sequence are calculated, both with a size of 1 × C, and Perform independent standardization operations on each channel in the time dimension, convert the monitoring data of each channel into a standard normal distribution with a mean of 0 and a standard deviation of 1, and obtain three sets of preprocessed time series data feature sequences

[0064] Then T s =15 is the local time window size, L=15 is the step size, and the three sets of preprocessed time series data feature sequences are respectively and three sets of normalized timestamp sequences Perform sliding window operation simultaneously to construct interpolation data training set Z train , imputation data validation set Z val and the imputed data test set Z test Three sample sets. At this time, each sample in the above three sample sets contains a set of actual, i.e., offline incomplete local state monitoring time series data. (Size is T s ×C) and a set of local time stamp series data (Size is T s ×1).

[0065] In order to solve the problem of model verification and evaluation, it is necessary to specify appropriate artificial missing types and perform artificial masking operations on the actual incomplete local state monitoring time series data in each sample set. The results of the visual analysis show that the actual missing pattern of the gas turbine state monitoring time series data in this example is mainly random discrete point missing, with occasional local continuous block missing in the variable and time dimensions. In order to comprehensively evaluate the performance of the model under different missing rates and missing patterns, four missing types are designed here, and their relevant information is summarized in Table 2 below. The subsequent training, verification, and testing of the interpolation model are carried out separately under each group of missing types to comprehensively evaluate the time series interpolation method proposed in this application.

[0066] Table 2 Basic information of four missing types

[0067]

[0068] A set of missing types is selected from Table 2, and the actual incomplete local state monitoring time series data of each sample in the sample set are analyzed according to their missing patterns and missing parameters. Perform artificial masking operations to obtain high-missing rate incomplete local state monitoring time series data X local (Size is T s ×C), and also as part of the sample. Finally, the interpolation data training set sample , imputation data validation set samples , imputed data test set samples All included in the time dimension alignment Three parts of data.

[0069] After the preprocessing of the gas turbine incomplete condition monitoring data described above, the effectiveness of the adaptive timestamp encoding enhanced complex equipment incomplete condition monitoring data time series interpolation method provided in the present application can be verified.

[0070] First, for a segment of high-quality historical condition monitoring data (as defined in the foregoing embodiment step B2, the preprocessed initial verification data offline incomplete condition monitoring time series data can be used as high-quality offline monitoring data under historical scenarios), fixed large size time window W L = 256 is randomly sampled along the time dimension to obtain L = 5 long sequence time series samples , where the size of a single sample is W L × C. Then, the data of each channel in each sample is respectively subjected to discrete Fourier transform to calculate the amplitude spectrum and phase spectrum of each time series data in the frequency domain. Since the high-quality initial verification data may still belong to incomplete data, the above Fourier transform problem can be regarded as a non-uniform Fourier transform with time domain non-uniformity and frequency domain uniformity, as shown in the following formula:

[0071]

[0072] t n = n o / W L , (0≤n o <N o ).

[0073] wherein X(k) represents the discrete Fourier transform result of the incomplete time series data x(n), N o represents the total number of non-missing data in the incomplete time series data, and n o represents the relative position index of a specific non-missing data point in the specified sequence x(n) in the time dimension.

[0074] The non-uniform fast Fourier transform algorithm based on interpolation can be used to solve the frequency domain expression of each incomplete time series data, and then calculate the normalized one-sided amplitude spectrum c k and one-sided phase spectrum φ k , as shown in the following formula:

[0075] X(k) = NUFFT(x(n)).

[0076]

[0077]

[0078] wherein c k and φ kBoth discard the negative frequencies and the DC component, and have size L x W L / 2 x C, W L / 2 = W L / 2 - 1, X re and X im denote the real and imaginary parts of the frequency domain feature, respectively, and X(k | 1:(W L / 2 - 1)) denotes the subsequence of X(k) consisting of the elements with index 1 through index (W L / 2 - 1).

[0079] The normalized one-sided amplitude spectrum c k of each sample is taken, and the average one-sided amplitude spectrum feature c k is obtained by averaging, which has size W L / 2 x C. Based on the average one-sided amplitude spectrum feature, the amplitude spectrum sparsification mask M A is constructed, which has size W L / 2 x C, where the positions corresponding to the Z = 32 frequency components with the largest amplitudes in each variable take value 1, and the rest take value 0. The amplitude spectrum sparsification mask is multiplied with the average one-sided amplitude spectrum feature element by element, and the sparsified amplitude feature H A-s is obtained, so as to realize the zeroing of the amplitudes of the noise frequencies, as shown in the following formula:

[0080]

[0081] After the Z-Score standardization operation is performed on the sparsified amplitude feature H A-s in the frequency dimension, the principal component analysis (PCA) is used to reduce the dimensionality thereof, and the compressed amplitude feature is obtained, which has size M L x C, where M L = 4 is the number of low-dimensional principal features reserved in the principal component analysis. Then, the number of clusters G = 10 is specified, the K-means clustering is performed on each condition monitoring variable based on the compressed amplitude feature , and the condition monitoring variable clustering label Y L (having size C) and the cluster center condition monitoring variable index I C (having size G) are obtained.

[0082] Considering that the total number of condition monitoring variables of complex equipment is large, and there is a significant local shared correlation pattern among many variables, in order to simplify the analysis process, it is assumed here that all condition monitoring variables under the same cluster share the same set of core frequency components, and the differences in the periodic prior time features among these variables can be expressed as the phase differences corresponding to each core frequency. Under this assumption, the TOP-K selection strategy is used to select the K = 4 frequencies with the largest amplitudes in the sparsified amplitude feature corresponding to each cluster center variable as the class center frequencies and share it in the variables under the same cluster, and finally get the prior core frequency information F c Meanwhile, the phase under each core frequency component in each cluster center variable is regarded as the phase origin, and the difference between each variable and the corresponding single-sided phase spectrum φ k of the cluster center variable is calculated in each sample, and the average is taken between samples to obtain the average phase feature . The phase value corresponding to the class center frequency position is selected to obtain the prior phase information Φ c , whose size is KxC.

[0083] The timestamp prior feature encoding module takes the normalized timestamp data at any time step T (size Txl) as input, takes the prior core frequency information F c and the prior phase information Φ c corresponding to each state monitoring variable as auxiliary parameters, respectively constructs the encoding function for sine position encoding and cosine position encoding, and stacks the two encoding results to output the timestamp prior encoding feature H TS , whose size is TxlKxC. The above encoding stacking process is as follows.

[0084]

[0085] H TS (i,:,j,k)=[H sin (i,j,k),H cos (i,j,k)]

[0086] where 0≤i<T,0≤j<K,0≤k<C

[0087] where [·] is the concatenation operation.

[0088] The timestamp prior feature encoding module f s (·) can be expressed as follows:

[0089]

[0090] In the scheme of the present application, it is assumed that under the condition of known prior core frequency and corresponding phase information, the value probability of a specific state monitoring variable c corresponding to a single timestamp TS at the corresponding time point is denoted as a state random variable X TS,c , which is subject to a normal distribution with variable independence and time sequence independence:

[0091]

[0092] where, respectively represent the prior core frequency information and the prior phase information corresponding to the state monitoring variable c.

[0093] The timestamp probabilistic interpolation model f d (·) takes the timestamp prior encoding feature as input, and outputs the probability distribution parameter estimation value of the specified state monitoring variable value at the specified time point, that is, the single-point prior distribution mean estimation value and the single-point prior distribution standard deviation estimation value , as shown in the following formula:

[0094]

[0095] Here represents the normalized timestamp data at a single time step.

[0096] According to the estimated distribution parameters, the variable is independently and time series independently sampled, and the sampling results of all variables in the specified time step range are combined, that is, the single interpolation result of the timestamp probabilistic interpolation model is obtained.

[0097] To achieve the above functions, the timestamp probabilistic interpolation model is composed of three parts: distribution encoding module, distribution aggregation module and reparameterization sampling module.

[0098] (1) Distribution encoding module:

[0099] The distribution encoding module includes G groups of parameter-independent distribution encoders, and each group of distribution encoders processes variables under a specific cluster label. A group of distribution encoders includes K parameter-independent distribution estimation networks, each corresponding to each core frequency component. Each distribution estimation network is a single-hidden-layer multilayer perceptron structure, with input dimension i=2, hidden layer dimension h=4, output dimension o=2, and GELU function as the hidden layer activation function. Its function is to transform the sine and cosine position encoding results corresponding to each core frequency component into distribution parameter component estimation values. The estimated distribution parameters include component mean and component log variance

[0100]

[0101] , wherein represents the timestamp prior encoding feature at a specific time step t and a specific variable c, and its size is 2×K. g is the cluster label value corresponding to the variable c. represents the component distribution parameter estimation value splicing vector corresponding to the kth core frequency component. Linear (g),1 (·) and Linear (g),2 (·) represent two parameter-independent linear layers corresponding to the cluster label value g, and GELU(·) represents the GELU nonlinear activation function, represents a truncated two-dimensional matrix all data with the second dimension index k.

[0102] Stacking the component distribution parameter estimation vectors obtained based on each core frequency component, a distribution parameter component estimation feature is obtained As the output of the module, its size is 2 x K.

[0103] (2) Distribution aggregation module:

[0104] The distribution encoding module assumes that the single-point component value of the state monitoring variable under the condition of known single core frequency prior information obeys a normal distribution, and estimates the distribution parameters thereof. In the distribution aggregation module, it is assumed that the value of the state monitoring variable under the condition of known all core frequency prior information can be obtained by combining the state monitoring variable component values corresponding to each core frequency according to a combination operation g(·,...,·) containing addition or multiplication, etc. The distribution thereof still maintains a normal distribution, as shown in the following formula.

[0105]

[0106] wherein, respectively represent the prior core frequency information and the prior phase information corresponding to the i-th core frequency of the state monitoring variable c.

[0107] Under the idealized assumption, the single-point prior distribution mean estimation value of the state monitoring variable under the condition of known all core frequency prior information and the single-point prior distribution log-variance estimation value can be respectively expressed as a nonlinear combination of the component mean estimation value and the component log-variance estimation value corresponding to each core frequency component. Here, two parameter-independent single-output single-layer fully connected neural networks are used to approximate the aggregation operation of each distribution parameter estimation value.

[0108]

[0109] wherein, g is the cluster label value corresponding to the variable c. In the distribution aggregation module, the parameters of the neural network layer are also independent under different variable clusters. Aggregation(·,...,·,) represents an aggregation operation, Linear (g),μ (·) and Linear (g),-2logσ (·) respectively represent two parameter-independent linear layers corresponding to the mean estimation value and the log-variance estimation value under the cluster label value g, and GELU(·) represents the use of a GELU nonlinear activation function.

[0110] (3) Re-parameterization sampling module:

[0111] The time stamp probabilistic imputation model outputs not only the distribution parameter estimation value of a specific variable at a specific time point, but also a specific imputation value for constructing the final imputation result.

[0112] Here, the state random variable X TS,c is sampled once to generate the numerical result of time stamp imputation. To ensure the correct back propagation during the subsequent optimization of the time stamp probabilistic imputation model, the reparameterization technique is introduced to convert the random sampling process into a deterministic operation. First, the single-point prior distribution log variance estimation value is transformed to obtain the single-point prior distribution standard deviation estimation value Then, an auxiliary noise quantity ε is randomly sampled from the standard normal distribution, and the time stamp imputation result obtained by single sampling can be calculated as as shown in the following formula:

[0113]

[0114] The input specifies the normalized time stamp data at a time step T , and based on the time stamp prior feature encoding module and the time stamp probabilistic imputation model, the prior distribution mean estimation value and the prior distribution standard deviation estimation value transformed from the prior distribution log variance estimation value can be calculated. Both have a size of TxC. Based on each set of distribution parameter estimation values, TxC independent reparameterization sampling is performed, and the results of the sampling are sequentially spliced and stacked to obtain the time stamp imputation result of the multivariate time series data , which has a size of TxC.

[0115] The local time series imputation model of the present embodiment can directly select any one of the existing technologies to specify the incomplete local state monitoring time series data X s under a specified local time window size T local as input, and the state monitoring time series data local imputation result as output, and can use the gradient back propagation algorithm for optimization of the deep learning time series imputation model, wherein the size of X local and is T s xC.

[0116] In this embodiment, a self-attention-based time series completion model (SAITS) is selected as the local temporal imputation model. The model performs additional artificial missing operations on the input data during the training process, while considering the time series completion and time series reconstruction tasks, and has good performance in imputation accuracy and imputation efficiency. It is worth noting that although the position encoding of the time series features is also performed in this model, it is only a relative position encoding within a local window, rather than a feature encoding considering the global unified position information in the timestamp.

[0117] Both the timestamp probabilistic imputation model and the local temporal imputation model can achieve the imputation of complex equipment state monitoring data. However, the characteristics of the two models have complementary relationship. Specifically, the timestamp probabilistic imputation model can fully utilize the prior periodic time series pattern features obtained by the frequency domain prior mining model analysis, is suitable for modeling stable, large time scale, and strong periodic time series variation rules, can provide a basic estimate for the value of each state monitoring variable at a specific time point, has strong globality, and is less affected by the missing pattern and local missing rate. However, due to the limited model representation ability and the use of multiple idealized assumptions, when facing complex equipment incomplete state monitoring data with complex multi-element local dependence relationship, the imputation accuracy achieved by the model is often difficult to directly meet the actual requirements.

[0118] The local temporal imputation model can fully capture various complex time series dependence relationships and cross-variable dependence relationships of complex equipment state monitoring data within a specific local time window, is suitable for modeling local, dynamic, and fine-grained time series variation rules, and has the advantages of strong model representation ability and high upper limit of imputation accuracy. However, the imputation effect of this kind of model is greatly affected by the local missing rate and missing pattern of the input data, and has the problem of insufficient generalization performance. For example, when dealing with samples with high local missing rate or abnormal missing pattern, it is often difficult to fully and accurately capture the local dependence relationship, resulting in unsatisfactory imputation effect.

[0119] The role of the timestamp-local imputation model joint framework is to adaptively fuse the imputation results of the two imputation models in a suitable way to ensure the imputation accuracy while improving the stability of the imputation performance.

[0120] The input of the joint framework is incomplete local state monitoring time series data X local and the corresponding local timestamp prior encoding feature , wherein the local timestamp prior encoding feature is obtained by transforming the local timestamp sequence through the timestamp prior feature encoding module. In addition, by detecting the incomplete local state monitoring time series data X localmissing data position information, a missing value mask M local (size T s ×C), where the missing value corresponds to the element value 0, and the rest of the position value is 1.

[0121] The above joint framework completes the process of interpolation as follows:

[0122] 1) input the incomplete local state monitoring time series data X local and the missing value mask M local into the locality time series interpolation model to obtain the locality interpolation result of the state monitoring time series data input the timestamp priori encoding feature into the timestamp probability interpolation model to obtain the local prior distribution mean estimate value obtained by the local prior distribution log variance estimate value transformed local prior distribution standard deviation estimate value and the timestamp interpolation result of the state monitoring time series data

[0123] 2) introduce an adaptive weighted fusion module to calculate an adaptive fusion threshold α fusion , size T s ×C. The final interpolation result of the state monitoring time series data based on the adaptive fusion threshold, the timestamp interpolation result and the locality interpolation result are weighted and fused to obtain, as shown in the following formula:

[0124]

[0125] where, ⊙ represents element-by-element multiplication.

[0126] In the above process, the adaptive weighted fusion module aims to adaptively analyze the relative difference between the two types of model output values, the missing state information, and the uncertainty information of the timestamp probability interpolation model, and to assign appropriate fusion weights to each time point of each variable. Since for a random variable subject to normal distribution, the width of a specific confidence interval is positively correlated with the size of the standard deviation, the uncertainty information generated by the timestamp probability interpolation model can be approximately represented by the local prior distribution standard deviation estimate value . In addition, the missing state information can be represented by the missing value mask M local . Thus, the locality interpolation result of the state monitoring time series data , the timestamp interpolation result of the state monitoring time series data , the local prior distribution standard deviation estimate value , and the missing value mask M localStacked to construct the adaptive weighted fusion module input feature H fusion , whose size is 4×T s ×C, as shown below:

[0127]

[0128] Among them, [·,·] represents the splicing operation.

[0129] Input feature H to the adaptive weighted fusion module fusion Perform deformation operation to obtain transposed fusion input features Its dimensions are C×4×T s Then, P = 2 one-dimensional convolutional layers with different convolution kernel sizes are used in parallel to extract and fuse the multi-scale temporal dependencies of the input features within each state monitoring variable. To simplify the model structure, the parameters of each convolution kernel are shared among different variables.

[0130]

[0131] in, Represents the output of the pth convolutional layer, whose size is C×D×T s , D = 32 is the fusion hidden layer representation dimension. Conv1d(·) represents a one-dimensional convolution operation, ω (0) ,ω (1) ,...,ω (P-1) Represents the learnable convolution kernel weights of different sizes, whose sizes are; D×4×K (0) ,D×4×K (1) ,...,D×4×K (P-1) , in this case, [K (0) ,K (1) ]=[3,5]. b (0) ,b (1) ,...,b (P-1) This indicates a learnable bias. Each convolutional layer pads the boundaries with 0 to ensure that the length of the input and output features in the time dimension remains unchanged.

[0132] The outputs of the multi-scale convolutional layers are summed, and then the features of each hidden layer are fused using a fusion convolutional layer with a convolution kernel size of 1. The Sigmoid function is activated to limit the range of the output value to [0, 1], and the fusion weighted output feature H is obtained. fusion_out , whose dimensions are C×1×T s The fusion weight output features are rearranged and dimensionally compressed to obtain the final adaptive fusion threshold α fusion (Size is T s ×C).

[0133]

[0134] Among them, ω F Indicates the learnable weights of the fused convolutional layer, whose size is 1×D×1, b F is the learnable bias of the fused convolutional layer.

[0135] The above introduces the respective processing flows of the frequency domain prior mining model, timestamp prior feature encoding module and timestamp-local interpolation model joint framework of this application. Before the above scheme is actually applied online, corresponding training is required before online application can be carried out. The application process of the above scheme provided in this embodiment can be divided into two links: offline and online. The offline link is only used to realize the adaptive analysis of some prior parameters and the offline training of each model. After the offline training is completed, it is no longer necessary to execute; and each module in the online link participates in the online application of the interpolation method.

[0136] Specifically, the above offline training process is divided into two stages: the timestamp prior information adaptive unsupervised learning stage and the timestamp-local interpolation model joint framework parameter optimization training stage.

[0137] In the adaptive unsupervised learning stage of timestamp prior information, the preprocessed complex equipment verification time series data feature sequence is converted into It is regarded as high-quality historical condition monitoring data and input into the frequency domain priori mining model. After random sampling, spectrum analysis, data dimension reduction, unsupervised feature clustering, TOP-K selection and other links, the condition monitoring variable clustering label Y is obtained. L Based on the output of the prior core frequency information F corresponding to each variable c and prior phase information Φ c , and then determine the parameters of the timestamp prior feature encoding module. This stage is mainly based on the adaptive implementation of the unsupervised machine learning algorithm. Once the obtained parameters are determined, they remain unchanged and do not participate in the gradient backpropagation process in the next stage.

[0138] In the parameter optimization training phase of the timestamp-local interpolation model joint framework, the interpolation data training set Z is first train The samples in are taken as input, and the high missing rate incomplete local state monitoring time series data X local and based on local timestamp sequence The local timestamp prior encoding feature obtained by transformation Input into the timestamp-local interpolation model joint framework to calculate the final interpolation results of the condition monitoring time series data , local prior distribution mean estimate , local prior distribution standard deviation estimate Then, the joint loss function is calculated by the following formula

[0139]

[0140] The joint loss function is composed of the interpolated main loss and prior distribution loss The interpolation main loss in this example uses a form similar to that in the SAITS algorithm. The only difference is that the final interpolation result after fusion is Replaces the original local interpolation result As the final output of the algorithm.

[0141] The prior distribution loss The negative log-likelihood loss and distribution standard deviation regularization loss Based on the distribution weight β, we get:

[0142]

[0143] The negative log-likelihood distribution loss Based on the principle of maximum likelihood estimation, its goal is to improve the rationality of the distribution parameter estimation in the timestamp probabilistic interpolation model. Assume that the observation data corresponding to a specific time step t in a specific variable c is At the same time, a set of distribution parameter estimates are given based on the timestamp probabilistic interpolation model and The corresponding likelihood function can be expressed as:

[0144]

[0145] Among them TS t,c Represents the timestamp information corresponding to the data. Then the negative log-likelihood function can be expressed as:

[0146]

[0147] Discarding the last constant term, we get the negative log-likelihood loss for a single data point, as shown below:

[0148]

[0149] The negative log-likelihood loss on the entire input data scale is obtained by calculating the negative log-likelihood loss on the non-missing positions of the condition monitoring time series data and taking the average:

[0150]

[0151] Among them, M local Indicates the input state monitoring time series data X local The corresponding missing value mask.

[0152] The standard deviation of the distribution regularization loss By further restricting the local prior distribution standard deviation estimate The L1 norm of t is used to reduce the uncertainty of the timestamp probabilistic interpolation model. Assume that the standard deviation estimate of the single-point prior distribution corresponding to a specific time step t in a specific variable c is , then the distribution standard deviation regularization loss under a single data point is:

[0153]

[0154] In order to make the distribution standard deviation regularization loss play a joint optimization effect with the negative log-likelihood distribution loss, the distribution standard deviation regularization loss on the entire input data scale is calculated in a similar way to the negative log-likelihood distribution loss:

[0155]

[0156] In the above loss function, the hyperparameters α and β are optimized based on the hyperparameter tuning algorithm.

[0157] After determining the loss function, the joint framework can be trained based on gradient backpropagation. The learner in the training process uses Adams, and the basic learning rate is optimized based on the hyperparameter tuning algorithm.

[0158] The joint framework adopts a multi-round training mode. After each round of training, the interpolation data validation set Z val The high missing rate of each sample in the incomplete local state monitoring time series data X local and local timestamp sequence After transformation, it is input into the joint framework to obtain the final time series interpolation result At the same time, based on the actual incomplete local state monitoring time series data and high missing rate incomplete local state monitoring time series data X local Construct indicator mask , whose size is T s ×C, where where is the observable value and in X local The corresponding position elements of missing values ​​in are set to 1, and the remaining position elements are set to 0. On this basis, the mean square error (MSE) is used to calculate the interpolation error as the validation loss

[0159]

[0160] The verification loss is used to analyze the overfitting degree of the model, and the early stopping strategy with Patience of 10 is combined to determine the termination condition of the training round to optimize and improve the training effect. To prevent the early stopping condition from being met, the maximum number of training rounds is set to 100.

[0161] The interpolation data test set Z is tested based on the interpolation data test set Z test The interpolation model based on adaptive timestamp coding trained in the embodiment can be tested and evaluated. The interpolation data test set Z test The input and output mode of each sample in the interpolation data test set Z val is the same as the interpolation data test set Z The evaluation indicators used in the test are the mean absolute error (MAE) and the mean relative error (MRE). The calculation of the indicators is similar to the verification loss and is also based on the construction of the indicator mask

[0162] Table 3 compares the test results of the interpolation method proposed in this example and the interpolation method using only the local temporal interpolation model (SAITS) under four types of artificial missing data.

[0163]

[0164] In Table 3, each experimental condition is independently repeated 5 times using different random seeds for model training and testing, and the average value of the test results is recorded. In addition, the hyperparameters of the two types of models are automatically determined based on the training set and the validation set using the random search hyperparameter tuning algorithm under the P1 missing type, and the number of tuning experiments is set to 100. The hyperparameters of the same model under different missing types are consistent.

[0165] After completing the offline training and evaluation process described above, when performing an online interpolation task, first, the online incomplete state monitoring data s of the complex equipment under T consecutive sampling steps is obtained , and the corresponding online timestamp sequence data is obtained , then the online incomplete state monitoring data is standardized to obtain the preprocessed online incomplete state monitoring data , the online timestamp sequence data is normalized and transformed by the timestamp prior feature encoding module to obtain the online timestamp prior encoding feature , and then the two are input into the trained timestamp-local interpolation model joint framework to output the online interpolation result of the incomplete state monitoring data

[0166] By analyzing the experimental data, it can be seen that the average interpolation error of the proposed method under different missing patterns and missing rates is lower than that of the method using only the local temporal interpolation model, which shows that the proposed method has the advantages of high interpolation accuracy and strong adaptability. At the same time, the interpolation accuracy of the proposed method in the high missing rate scene is significantly improved compared with the local interpolation model, which shows that the proposed method has strong robustness and can better adapt to the scene of highly incomplete state monitoring data of complex equipment under sudden severe working conditions. In summary, the proposed method designs an adaptive timestamp encoding and timestamp probabilistic interpolation model, and combines it with the existing advanced local temporal interpolation model, which realizes the overall improvement of the interpolation performance on the incomplete state monitoring data of the gas turbine.

[0167] The technical features of the above embodiments can be combined in any manner. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present disclosure.

[0168] The principles and implementation modes of the present application are described by using specific examples in this paper, and the above embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the present application should not be understood as a limitation.

Claims

1. A method for temporal interpolation of incomplete state monitoring data of complex equipment enhanced by adaptive timestamp coding, characterized in that: include: Obtain online defect status monitoring data for complex equipment; The online defect state monitoring data includes online defect state monitoring time series data of complex equipment and corresponding online time stamp sequence data; The online incomplete state monitoring time series data is the monitoring values ​​of several state monitoring variables of complex equipment under multiple continuous sampling steps; Standardizing the online defect state monitoring time series data to obtain preprocessed online defect state monitoring time series data; After normalizing the online timestamp sequence data, the online timestamp priori coding features are outputted using a timestamp priori feature coding module; The parameters of the timestamp prior feature encoding module are obtained based on an unsupervised frequency domain prior mining model; Inputting the pre-processed online incomplete state monitoring time series data and the online timestamp prior coding features into a trained timestamp-local interpolation model joint framework to obtain an online interpolation result of the incomplete state monitoring data; the timestamp-local interpolation model joint framework is a joint framework obtained based on a timestamp probabilistic interpolation model and a local time series interpolation model; training the timestamp-local interpolation model joint framework using an interpolation data training set to obtain a trained timestamp-local interpolation model joint framework, specifically comprising: Inputting the local timestamp sequence data in the interpolation data training set into the timestamp prior feature encoding module to obtain offline timestamp prior encoding features; Inputting the offline timestamp prior coding feature into a timestamp probabilistic interpolation model to obtain a local prior distribution mean estimate, a local prior distribution standard deviation estimate, and a timestamp interpolation result; Inputting the incomplete local state monitoring time series data with a high missing rate in the interpolation data training set into the local time series interpolation model to obtain a local interpolation result; Determining a final interpolation result based on the timestamp interpolation result and the locality interpolation result; Constructing a joint loss function based on the incomplete local state monitoring time series data with a high missing rate, the local prior distribution mean estimate, the local prior distribution standard deviation estimate, and the final interpolation result; According to the joint loss function, the timestamp-local interpolation model joint framework is trained based on gradient back propagation to obtain a trained timestamp-local interpolation model joint framework; the joint loss function is shown in the following formula: in, is the joint loss function, X local For high missing rate incomplete local state monitoring time series data, is the final interpolation result, is the estimated value of the local prior distribution mean, is the estimated standard deviation of the local prior distribution, is the interpolation main loss function, is the prior distribution loss function.

2. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 1 is characterized in that: Before using the time stamp prior feature encoding module to output the online time stamp prior encoding feature, the adaptive time stamp encoding enhanced complex equipment defective state monitoring data time series interpolation method further includes: Based on offline monitoring data in historical scenarios of complex equipment, an interpolation data training set, an interpolation data validation set, and an interpolation data test set are constructed; the interpolation data training set is used to train the timestamp-local interpolation model joint framework; the interpolation data validation set and the interpolation data test set are used to verify, optimize, and evaluate the performance of the trained timestamp-local interpolation model joint framework; The high-quality offline monitoring data in historical scenarios is input into the frequency domain priori mining model. Based on the cluster labels of the state monitoring variables, the priori core frequency information and priori phase information corresponding to each state monitoring variable are output; Determining parameters of a timestamp priori feature encoding module based on the priori core frequency information and the priori phase information corresponding to each state monitoring variable; The timestamp-local interpolation model joint framework is trained using the interpolation data training set to obtain a trained timestamp-local interpolation model joint framework.

3. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 2 is characterized in that: Based on the offline monitoring data of historical scenarios of complex equipment, the interpolation data training set, interpolation data validation set and interpolation data test set are constructed, including: Acquire initial defective state monitoring data of complex equipment; the initial defective state monitoring data includes offline defective state monitoring time series data of the complex equipment and corresponding offline timestamp sequence data; the offline defective state monitoring time series data is the monitoring values ​​of several state monitoring variables of the complex equipment under multiple continuous sampling steps in historical scenarios; After normalizing the offline timestamp sequence data, non-overlapping data division is performed based on the initial incomplete state monitoring data to obtain initial training data, initial verification data, and initial test data; Performing standardization processing on the offline incomplete state monitoring time series data in the initial training data, the initial verification data, and the initial test data, respectively, to obtain preprocessed initial training data, preprocessed initial verification data, and preprocessed initial test data; Using local time windows of the same size, sliding window operations are performed simultaneously on preprocessed initial training data, preprocessed initial verification data, and preprocessed initial test data to construct interpolated data training sets, interpolated data verification sets, and interpolated data test sets; any sample in the interpolated data training sets, interpolated data verification sets, and interpolated data test sets includes offline incomplete local state monitoring time series data, high-missing rate incomplete local state monitoring time series data obtained by artificially masking the offline incomplete local state monitoring time series data, and local timestamp series data.

4. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 3 is characterized in that: The offline incomplete state monitoring time series data in the preprocessed initial verification data is used as high-quality offline monitoring data in the historical scenario and input into the frequency domain prior mining model. Based on the clustering labels of the state monitoring variables, the prior core frequency information and prior phase information corresponding to each state monitoring variable are output.

5. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 1 is characterized in that: The timestamp probabilistic interpolation model includes: a distribution encoding module, a distribution aggregation module, and a reparameter sampling module; the distribution encoding module includes a plurality of parameter-independent distribution encoders; the distribution encoders are used to process state monitoring variables corresponding to cluster labels; the distribution encoders include a plurality of parameter-independent distribution estimation networks, each corresponding to each core frequency component of the state monitoring variable; The offline timestamp prior coding feature is input into the timestamp probabilistic interpolation model to obtain the local prior distribution mean estimate, the local prior distribution standard deviation estimate and the timestamp interpolation result, specifically including: Inputting the offline timestamp priori coding feature into the corresponding distribution encoder, and outputting the distribution parameters corresponding to each core frequency component; the distribution parameters include the mean estimate and the logarithmic variance estimate; Utilizing the distribution aggregation module, respectively aggregating the mean estimate and the logarithmic variance estimate corresponding to each core frequency component, and outputting a local prior distribution mean estimate and a local prior distribution logarithmic variance estimate; Transforming the local prior distribution logarithmic variance estimate into a local prior distribution standard deviation estimate, and randomly sampling from a standard normal distribution to obtain an auxiliary noise amount; Based on the local prior distribution mean estimate, the local prior distribution standard deviation estimate and the auxiliary noise amount, sampling is performed to obtain a timestamp interpolation result.

6. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 1 is characterized in that: Inputting the incomplete local state monitoring time series data with high missing rate in the interpolation data training set into the local time series interpolation model to obtain the local interpolation results, specifically including: By detecting the missing data location information of the time series data with high missing rate and incomplete local state monitoring, a missing value mask is constructed; The high missing rate incomplete local state monitoring time series data and the missing value mask are input into a local time series interpolation model to obtain a local interpolation result.

7. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 1 is characterized in that: The final interpolation result is determined according to the following formula: in, is the final interpolation result, is the local interpolation result, α fusion is the adaptive fusion threshold, is the timestamp interpolation result, and ⊙ represents element-by-element multiplication.

8. The method for temporal interpolation of complex equipment incomplete state monitoring data enhanced by adaptive timestamp coding according to claim 7 is characterized in that: The adaptive fusion threshold is obtained through the following process: The local interpolation results, timestamp interpolation results, local prior distribution standard deviation estimate and missing value mask are stacked to construct the input features of the adaptive weighted fusion module; The input features of the adaptive weighted fusion module are deformed to obtain transposed fused input features. Then, a multi-scale convolutional layer is used to extract and fuse the multi-scale temporal dependencies of the input features within each state monitoring variable. The multi-scale convolution layer is a one-dimensional convolution layer with multiple convolution kernels of different sizes; The outputs of the multi-scale convolutional layers are summed, and the summed results are sequentially passed through the fusion convolutional layer and the activation function layer to output the fusion weighted output features; Rearrange and dimensionally compress the fusion weight output features to obtain an adaptive fusion threshold.

Images