Method for predicting industrial residual service life based on multi-rate data

Abstract

The invention provides a method for predicting the industrial residual service life based on multi-rate data, and the method comprises the steps: training a prediction model, and inputting the data of a test target domain into the prediction model to obtain the prediction output, and the process of training the prediction model comprises the steps: obtaining a source domain data sample, preprocessing the source domain data sample to obtain a source data matrix sample; performing position code embedding processing on the source data matrix sample to obtain a code matrix after position code embedding; and inputting the coding matrix into a Mamba model to obtain a prediction model. According to the method, firstly, the input sequence X in a source domain data sample and the corresponding timestamp t are obtained, missing data are processed through linear interpolation and forward / backward filling, and data continuity and integrity are ensured. Afterwards, the complemented data is connected with the residual error through a multi-layer convolutional network to ensure that the original value is not interfered, and fine-grained complementation is carried out through an encoder to optimize the data precision.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a method for predicting the remaining service life of an industry based on multi-rate data. Background Art

[0002] In contemporary industrial processes, implementing real-time health monitoring systems has become indispensable for optimizing production efficiency and mitigating potential safety risks. Within the field of prognostics and health management (PHM), remaining useful life (RUL) prediction is one of the most complex and critical tasks, undergoing extensive development and widespread application to ensure the reliable operation of industrial systems. Generally speaking, RUL prediction techniques can be categorized into three categories: physics-based methods, data-driven methods, and hybrid approaches. With the rapid development of sensing technology, intelligent instrumentation has laid a solid foundation for process data acquisition. This development enables the efficient collection of large amounts of high-quality data. This has led to the widespread adoption of data-driven RUL prediction, particularly in the field of industrial data modeling.

[0003] In industrial settings, the inherent variability of sensor sampling frequencies often leads to incomplete and asymmetric data structures. Irregular sampling patterns are a significant obstacle in data modeling. Traditional methods employing upsampling and downsampling techniques often introduce artifacts, measurement errors, or degradation of key information, limiting their practical application. To overcome these limitations, researchers have designed advanced estimation frameworks that aim to normalize sampling frequencies to improve the completeness of irregularly sampled datasets and facilitate subsequent modeling. With the rapid advancement of deep learning technology, researchers have proposed sampling interval-aware LSTM (Long Short-Term Memory) architectures, model architectures, and hierarchical sequence generation networks (SGNs). While these architectures have demonstrated effectiveness in addressing data incompleteness caused by multiple sampling rates, they remain limited in addressing the challenges of multi-rate data in highly irregular or complex unknown scenarios. In addition to irregular sampling conditions, cross-operational modeling presents another challenge in industrial settings. Traditional RUL prediction methods assume that source and target domain data have identical distribution characteristics, an assumption often not met in industrial applications. Practical constraints also make it difficult to retrain with large amounts of target domain data. Therefore, how to optimize the prediction model under the conditions of limited target domain data and abundant source domain data becomes a key challenge. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the remaining useful life of an industrial product based on multi-rate data.

[0005] The technical solution of the present invention is: a method for predicting industrial remaining useful life based on multi-rate data, including the process of training a prediction model and inputting test target domain data into the prediction model to obtain a prediction output, wherein the process of training the prediction model includes the steps of:

[0006] S1. Obtain source domain data samples, and preprocess the source domain data samples to obtain source data matrix samples;

[0007] S2, performing position coding embedding processing on the source data matrix sample to obtain a coding matrix after position coding embedding;

[0008] S3. Input the encoding matrix into the Mamba model to obtain a prediction model.

[0009] Preferably, step S1 further includes the steps of:

[0010] S11. Obtain multiple groups of input sequences X in the source domain data samples. ori ={x1,x2,…,x i ,…,x s} and its corresponding timestamp t={t1,t2,…,t i ,…,t s}, where s represents the length of the input sequence;

[0011] S12. Use linear interpolation to complete the missing data of each input sequence;

[0012] S13. The completed data is processed through a multi-layer convolutional network and residual connection, and fine-grained completion is performed through an encoder to obtain the source data matrix sample X.

[0013] Preferably, in step S12, the processing formula for filling in the missing data of the sequence using linear interpolation is:

[0014]

[0015] In the above formula, t start and t end Indicates the timestamps of the non-missing values ​​before and after the missing element, t j Represents the time to be estimated, j∈[1, 2,…, s], s is a positive integer, y j Refers to the value completed at time j, the original time j is missing, x end Refers to the most recent non-missing value after the current missing value j moments, x startRefers to the most recent non-missing value before the current missing value j moments ago.

[0016] Preferably, in step S13, the completed data is processed through a multi-layer convolutional network and residual connection formula is:

[0017]

[0018]

[0019] In the above formula, X conv Represents the output of the multi-layer convolutional neural network, M miss (·) represents the unobserved indicator matrix, N c Represents the number of layers in a multi-layer convolutional neural network, X block Refers to the data without missing values ​​after the preliminary interpolation of all missing values ​​is completed after processing by formula (1); X ori Refers to the original data containing missing values ​​before being processed by formula (1), x m,n Refers to each element in the source domain data sample, m and n are rows and columns respectively, which are used to determine the element position.

[0020] Preferably, in step S2, the encoding formula is:

[0021]

[0022] X I =X e +PE(X e ) Formula (5)

[0023] where X I represents the encoding matrix after position encoding embedding, PE(·) represents the position encoding calculation of all samples, pos∈[0,n] is the position index, i∈[0,d model ] is the dimension, Indicates frequency, X e The input source data matrix sample X is mapped into the specified dimension through linear transformation.

[0024] Preferably, in step S3, the formula for inputting the encoding matrix into the Mamba model to obtain the prediction model is:

[0025] X fused =gate⊙X c +(1-gate)⊙Memory formula (6)

[0026] In the above formula, X c For X I The dataset after normalization layer and convolution processing, Memory refers to memory information, X fused It means input Xc The features are obtained by fusing the image and memory through a dynamic gating mechanism.

[0027] Preferably, the above method further comprises the steps of:

[0028] S4. Provide a target domain fine-tuning dataset, and input the target domain fine-tuning dataset into the prediction model obtained in S3 to fine-tune the prediction model to obtain a final prediction model.

[0029] The above technical solution has the following advantages or beneficial effects: In the above method, the input sequence X and its corresponding timestamp t in the source domain data sample are first obtained, and the missing data is processed by linear interpolation and forward / backward filling to ensure data continuity and integrity. Afterwards, the completed data is processed by a multi-layer convolutional network and residual connection to ensure that the original value is not disturbed, and fine-grained completion is performed by the encoder to optimize data accuracy. The final data is input as a sample of the source data matrix to the next step. In addition, the position information of each time step is embedded in the input feature, which enhances the model's perception of sequence order. In addition, in this model, the flexible fusion of historical information and current input is achieved through a dynamic gating structure, which improves the performance of long sequence modeling and has higher accuracy and robustness. Finally, the prediction model is fine-tuned using the target domain fine-tuning dataset, further improving the accuracy of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The embodiments of the present invention will be described more fully with reference to the accompanying drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.

[0031] Figure 1 This is a flow chart of a method for predicting the remaining useful life of an industry based on multi-rate data according to the present invention. Figure 1 ;

[0032] Figure 2 This is a flow chart of a method for predicting the remaining useful life of an industry based on multi-rate data according to the present invention. Figure 2 ;

[0033] Figure 3 This is a flow chart of a method for predicting the remaining useful life of an industry based on multi-rate data according to the present invention. Figure 3 ;

[0034] Figure 4 This is a model diagram of a method for predicting the remaining useful life of an industry based on multi-rate data according to the present invention. DETAILED DESCRIPTION

[0035] The following describes in detail a method for predicting the remaining useful life of an industry based on multi-rate data in conjunction with the accompanying drawings and specific embodiments.

[0036] In the description of the present invention, it should be understood that the directions or positional relationships indicated by terms such as "upper," "lower," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," and "outside" are based on the directions or positional relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or components referred to must have and have a specific direction, be constructed and operate in a specific direction. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," etc. are merely used to distinguish the names of the various components and do not have a primary or secondary relationship. Therefore, they should not be construed as limitations on the present invention.

[0037] like Figure 1 As shown, a method for predicting the remaining useful life of an industrial product based on multi-rate data includes the process of training a prediction model using source domain data and inputting test target domain data into the prediction model to obtain a prediction output. Specifically, the process of training the prediction model includes the following steps:

[0038] S1. Obtain source domain data samples and pre-process the source domain data samples to obtain source data matrix samples to ensure the continuity and integrity of the source data matrix samples;

[0039] S2, perform position coding embedding processing on the source data matrix sample to obtain the encoding matrix after position coding embedding;

[0040] S3. Input the encoding matrix into the Mamba model to obtain a prediction model.

[0041] Specifically, since the collected source domain data is missing, the source domain data samples must first be preprocessed to ensure the continuity and integrity of the source data matrix samples. Afterwards, the source data matrix samples are position-encoded and embedded to obtain the encoding matrix after position encoding embedding. In this process, the position information of the time step is embedded into the input features, enhancing the prediction model's ability to perceive the sequence order. Finally, the encoding matrix is ​​input into the Mamba model to obtain the prediction model. Through the above method, complex long sequence data can be better processed, and the characteristics of time series data can be more accurately captured and represented, realizing complex dynamic time series data modeling. Among them, the Mamba model is a new deep learning architecture based on the Selective State Space Model (SSM).

[0042] like Figure 2 As shown, further speaking, in step S1, the process is also included:

[0043] S11. Obtain multiple sets of input sequences X in source domain data samples ori ={x1,x2,…,x i ,…,x s} and its corresponding timestamp t={t1,t2,…,t i ,…,t s}, where s represents the length of the input sequence;

[0044] S12. Use linear interpolation to complete the missing data of each input sequence to ensure data continuity;

[0045] S13. All completed data are processed through a multi-layer convolutional network and residual connection, and fine-grained completion is performed through the encoder to obtain the source data matrix sample X.

[0046] Specifically, first obtain each set of input sequence X in the source domain data sample ori and its corresponding timestamp t. Missing data is handled through linear interpolation and forward / backward filling to ensure data continuity and integrity. All completed data is then processed through a multi-layer convolutional network with residual connections to ensure that the original values ​​are not disturbed. An encoder performs fine-grained completion to optimize data accuracy. The final data serves as the source data matrix sample and is input into the next step.

[0047] Furthermore, in step S12, the processing formula for filling in the missing data of the sequence using the linear interpolation method is:

[0048]

[0049] In the above formula, t start and t end Indicates the timestamps of the non-missing values ​​before and after the missing element, t j Represents the time to be estimated, j∈[1, 2,…, s], s is a positive integer. j Refers to the value completed at time j, the original time j is missing, x end Refers to the most recent non-missing value after the current missing value j moments, x start Refers to the most recent non-missing value before the current missing value j moments ago.

[0050] Furthermore, in step S13, the completed data is processed through a multi-layer convolutional network using the following formula:

[0051]

[0052] In the above formula, X conv Represents the output of the multi-layer convolutional neural network, M miss(·) represents the unobserved indicator matrix, N c Indicates the number of layers in a multi-layer convolutional neural network. block Refers to the data without missing values ​​after the preliminary interpolation of all missing values ​​is completed after processing by formula (1); X ori Refers to the original data containing missing values ​​before being processed by formula (1), X ori The missing value positions in the formula (1) are processed one by one, using y j After interpolation, we get X block .x m,n Refers to each element in the source domain data sample, m and n are rows and columns respectively, used to determine the element position, that is, refers to the position in X ori The position in the mask matrix is ​​used to determine the 0 / 1 value of the mask matrix, so that the mask matrix can be used to directly determine which elements are originally observed and which are artificially filled. It is worth noting that if x m,n In the source domain data sample, it is unobserved, so it is filled with artificially processed data; if x m,n If the source domain data sample is in an observed state (i.e., the data already exists), then fill in the gaps based on the source data values. Ensure that if the data itself exists, try to use the source data. If the source data is missing, use manually processed data to fill in the gaps.

[0053] Furthermore, for X conv After performing residual connection processing and fine-grained completion through the encoder, the source data matrix sample X is obtained.

[0054] Furthermore, after preprocessing the source domain data samples to obtain source data matrix samples, the source data matrix samples need to be input into the RUL prediction model to achieve learning and training of the prediction model. In this process, we first enter step S2, which is to perform position encoding embedding processing on the source data matrix samples to obtain the encoding matrix after position encoding embedding. The encoding formula is:

[0055]

[0056] X I =X e +PE(X e ) Formula (5)

[0057] where X I represents the encoding matrix after position encoding embedding, PE(·) represents the position encoding calculation of all samples, pos∈[0,n] is the position index, i∈[0,d model ] is the dimension, In the above formula (4), the relative relationship between time steps is encoded using sine and cosine functions, and the position information of each time step is embedded into the input feature using formula (5), which enhances the model's ability to perceive the sequence order. e The input source data matrix sample X is mapped into the specified dimension through linear transformation, X e =W e *X+be, where W e is the parameter matrix of the M*dmodel, and be is the bias vector. Here, the X dimension is B*T*M, and when it undergoes linear changes, W is required. e Is a M*d model The trainable parameter matrix, d model Equivalent to the dimension of hidden features, the purpose is to transform X into B*T*d model dimension.

[0058] Furthermore, in step S3, the encoding matrix is ​​input into the Mamba model to obtain a prediction model. Specifically, a dynamic gating structure is used to flexibly integrate historical information with current input, as shown in the following formula:

[0059] X fused =gate⊙X c +(1-gate)⊙Memory formula (6)

[0060] In the above formula, X c For X I The dataset after normalization and convolution processing. Memory refers to memory information. Since Mamba belongs to a multi-layer architecture, the memory of the second layer is the output of the gate mechanism of the first layer. The memory of the first layer is artificially set to 0 in the specific code; X fused It means input X c The features are fused with the memory through a dynamic gating mechanism. That is, in step S3, the original Mamba model is not directly fed into the model, but into a gated Mamba model improved by the gating mechanism. This model uses a dynamic gating structure to flexibly integrate historical information with current input, improving long-sequence modeling performance and achieving higher accuracy and robustness.

[0061] like Figure 2 As shown, further speaking, the above method also includes the steps of:

[0062] S4. Provide a target domain fine-tuning dataset and input the target domain fine-tuning dataset into the prediction model obtained in S3 to fine-tune the prediction model and obtain a final prediction model. Through this method, the prediction model is fine-tuned to further improve the accuracy of the prediction model.

[0063] like Figure 4 As shown in the figure, in the above method, the input sequence X and its corresponding timestamp t in the source domain data sample are first obtained. Missing data is processed through linear interpolation and forward / backward filling to ensure data continuity and integrity. The completed data is then processed through a multi-layer convolutional network with residual connections to ensure that the original values ​​are not disturbed. Fine-grained completion is performed through the encoder to optimize data accuracy. The final data is used as the source data matrix sample and input into the next step. In addition, the position information of each time step is embedded in the input features, enhancing the model's perception of sequence order. In addition, in this model, the dynamic gating structure is used to achieve flexible fusion of historical information and current input, improving the performance of long sequence modeling and achieving higher accuracy and robustness. Finally, the prediction model is fine-tuned using the target domain fine-tuning dataset, further improving the accuracy of the prediction model.

[0064] Various changes and modifications will undoubtedly become apparent to those skilled in the art upon reading the foregoing description. Therefore, the appended claims should be construed to encompass all changes and modifications within the true intent and scope of the present invention. Any and all equivalents within the scope of the claims should be considered to be within the intent and scope of the present invention.

Claims

1. A method for predicting industrial remaining useful life based on multi-rate data, comprising the process of training a prediction model and inputting test target domain data into the prediction model to obtain a prediction output, characterized in that: The process of training the prediction model includes the following steps: S1. Obtain source domain data samples, and preprocess the source domain data samples to obtain source data matrix samples; S2, performing position coding embedding processing on the source data matrix sample to obtain a coding matrix after position coding embedding; S3. Input the encoding matrix into the Mamba model to obtain a prediction model.

2. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 1, characterized in that: Step S1 further includes the steps of: S11. Obtain multiple groups of input sequences X in the source domain data samples. ori ={x1,x2,…,x i ,…,x s } and its corresponding timestamp t={t1,t2,…,t i ,…,t s }, where s represents the length of the input sequence; S12. Use linear interpolation to complete the missing data of each input sequence; S13. The completed data is processed through a multi-layer convolutional network and residual connection, and fine-grained completion is performed through an encoder to obtain the source data matrix sample X.

3. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 2, characterized in that: In step S12, the processing formula for filling the missing data of the sequence using linear interpolation is: In the above formula, t start and t end Indicates the timestamps of the non-missing values ​​before and after the missing element, t j Represents the time to be estimated, j∈[1, 2,…, s], s is a positive integer, y j Refers to the value completed at time j, the original time j is missing, x end Refers to the most recent non-missing value after the current missing value j moments, x start Refers to the most recent non-missing value before the current missing value j moments ago.

4. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 3, characterized in that: In step S13, the completed data is processed through a multi-layer convolutional network and residual connection. The formula is: In the above formula, X conv Represents the output of the multi-layer convolutional neural network, M miss (·) represents the unobserved indicator matrix, N c Represents the number of layers in a multi-layer convolutional neural network, X block Refers to the data without missing values ​​after the preliminary interpolation of all missing values ​​is completed after processing by formula (1); X ori Refers to the original data containing missing values ​​before being processed by formula (1), x m,n Refers to each element in the source domain data sample, m and n are rows and columns respectively, which are used to determine the element position.

5. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 2, characterized in that: In step S2, the encoding formula is: X I =X e +PE(X e ) Formula (5) where X I represents the encoding matrix after position encoding embedding, PE(·) represents the position encoding calculation of all samples, pos∈[0,n] is the position index, i∈[0,d model ] is the dimension, Indicates frequency, X e The input source data matrix sample X is mapped into the specified dimension through linear transformation.

6. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 5, characterized in that: In step S3, the encoding matrix is ​​input into the Mamba model to obtain the prediction model formula: X fused =gate⊙X c +(1-gate)⊙Memory Formula (6) In the above formula, X c For X I The dataset after normalization layer and convolution processing, Memory refers to memory information, X fused It means input X c The features are obtained by fusing the image and memory through a dynamic gating mechanism.

7. The method for predicting the remaining useful life of an industry based on multi-rate data according to claim 1, characterized in that: The above method further comprises the steps of: S4. Provide a target domain fine-tuning dataset, and input the target domain fine-tuning dataset into the prediction model obtained in S3 to fine-tune the prediction model to obtain a final prediction model.

Images