A Smart Prediction Method and Device Based on Mamba Deep Learning Network

Abstract

This application discloses an intelligent prediction method and apparatus based on a Mamba deep learning network, relating to the field of intelligent prediction of equipment performance. The method includes: acquiring and preprocessing sensor monitoring data from industrial equipment; generating time window samples from the preprocessed sensor monitoring data using a sliding time window strategy; constructing a Mamba-GRU parallel network, which includes a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer; training the Mamba-GRU parallel network based on the time window samples to obtain a performance prediction model; and predicting the remaining service life and health status of industrial equipment based on the performance prediction model. This application effectively combines the complementarity of long-term degradation trends and short-term dynamic features, improving the quality of multi-scale feature fusion and thus enhancing prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of intelligent prediction of equipment performance, and in particular to an intelligent prediction method and apparatus based on Mamba deep learning network. Background Technology

[0002] As a key technology in the field of equipment failure prediction and health management (PHM), equipment RUL prediction or health status prediction technology has received increasing attention in recent years. Existing methods are mainly divided into two categories: physical model-based and data-driven. With the development of IoT and sensor technologies, data-driven methods, especially deep learning-based methods, have become the mainstream methods in current equipment RUL prediction or health status prediction. However, traditional deep learning models have significant shortcomings when processing time-series data. For example, while recurrent neural network (RNN)-based methods can capture local features, their ability to model long-term dependencies is limited, leading to insufficient prediction accuracy. Transformer-based methods, while capable of handling long-sequence dependencies, suffer from high computational complexity and slow inference speed.

[0003] To reduce computational complexity and improve model inference speed, Gu et al. first proposed the Mamba model in December 2023. The Mamba model, through a structured state-space model (SSM) and a selective scanning mechanism, exhibits linear time complexity advantages in long sequence modeling, and can more efficiently capture complex dependencies in sequence data. Since its introduction, it has attracted widespread attention from researchers. However, while the Mamba model can effectively capture global degradation trends, it still has limited ability to capture local temporal features. Therefore, how to effectively combine the complementarity of long-term degradation trends and short-term dynamic features to improve the multi-scale feature fusion quality and prediction accuracy of the Mamba model is a pressing technical challenge and a cutting-edge research topic in the field of PHM (Prognostics and Health Management). Summary of the Invention

[0004] The purpose of this application is to provide an intelligent prediction method and device based on the Mamba deep learning network, which can effectively combine the complementarity of long-term degradation trends and short-term dynamic features to improve the quality of multi-scale feature fusion and thus improve prediction accuracy.

[0005] To achieve the above objectives, this application provides the following solution:

[0006] Firstly, this application provides an intelligent prediction method based on a Mamba deep learning network, the intelligent prediction method based on a Mamba deep learning network comprising:

[0007] Acquire sensor monitoring data from industrial equipment;

[0008] The sensor monitoring data of the industrial equipment is preprocessed;

[0009] A time window sample is generated by using a sliding time window strategy to preprocess the sensor monitoring data.

[0010] A parallel Mamba-GRU (Gated Recurrent Unit) network is constructed, which includes: a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer.

[0011] The Mamba-GRU parallel network is trained based on the time window samples to obtain a performance prediction model;

[0012] The remaining useful life (RUL) or health status of industrial equipment is predicted based on the performance prediction model.

[0013] Optionally, the sensor monitoring data of the industrial equipment is preprocessed, specifically including the following steps:

[0014] Remove abnormal data from the sensor monitoring data of the industrial equipment;

[0015] Wavelet transform is used to reduce noise in the sensor monitoring data after removing outlier data.

[0016] The sensor monitoring data after wavelet transform denoising is normalized.

[0017] Optionally, the linear marker layer specifically adopts the following formula:

[0018] ;

[0019] in, express t The output of the linear labeling layer at time intervals, express t The input to the linear labeling layer is defined by the time interval, and "Batch" indicates batch operations. and The parameters representing the linear labeling layer, L Indicates the length of the input sequence. d Representing feature dimension, R It represents the set of real numbers.

[0020] Optionally, the Mamba module specifically adopts the following formula:

[0021] ;

[0022] ;

[0023] in, Represents the state matrix; and Indicates projection parameters; x ( t )for t The output of the linear labeling layer at time intervals, This indicates the hidden state of the input in the Mamba module at the current moment. y ( t ) indicates the output of the Mamba module. This represents the hidden state in the Mamba module at the previous time step. R Represents the set of real numbers. m Represents the state dimension. L Indicates the length of the input sequence.

[0024] Optionally, the GRU module specifically adopts the following formula:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] in, for t The door to reset time. It is the sigmoid activation function. For the GRU module t The hidden state at time -1 Hidden state The weight matrix, x ( t )for t The output of the linear labeling layer at time intervals, Input to reset the door x ( t The weight matrix of ) for t The door is constantly being updated. Input to update the gate x ( t The weight matrix of ) To update the hidden state in the door The weight matrix, For the GRU module t The candidate hidden states at time t, where tanh(·) is the tanh activation function, and ☉ is the Hadamard product. To reset the weight matrix of the hidden state after the gate is processed, Hidden state The weight matrix, For the GRU module t The output at each moment, i.e., in the GRU module t The hidden state at any given moment.

[0030] Optionally, the data processing procedure in the feature fusion module is as follows:

[0031] The long-term degradation trend extracted by the Mamba module is concatenated with the short-term dynamic features captured by the GRU module to obtain the concatenated features.

[0032] The spliced ​​features are calculated using a scaled dot product attention mechanism to obtain two feature scores;

[0033] The dynamic weight distribution is obtained by normalizing the two feature scores using the softmax function;

[0034] The long-term degradation trend and short-term dynamic features are weighted and summed according to the dynamic weight distribution to achieve dynamic fusion and obtain fused features.

[0035] Optionally, the wavelet transform noise reduction of the sensor monitoring data after removing outlier data specifically adopts the following formula:

[0036] ;

[0037] ;

[0038] in, j Indicates scale. k Indicates translation. t Indicates time, f ( t () represents the original signal. c j,k These are the coefficients obtained after discrete wavelet transform. These are discrete wavelet basis functions. For the mother wavelet function, It is the reconstructed signal.

[0039] Secondly, this application provides an intelligent prediction device based on a Mamba deep learning network, the intelligent prediction device for industrial equipment performance comprising:

[0040] The data acquisition module is used to acquire monitoring data from industrial equipment sensors.

[0041] The preprocessing module is used to preprocess the sensor monitoring data of the industrial equipment.

[0042] The time window sample generation module is used to generate samples from the preprocessed sensor monitoring data based on a sliding time window strategy to obtain time window samples.

[0043] The Mamba-GRU parallel network construction module is used to construct a Mamba-GRU parallel network, which includes: a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer.

[0044] The training module is used to train the Mamba-GRU parallel network based on the time window samples to obtain a performance prediction model;

[0045] The performance prediction module is used to predict the RUL or health status of industrial equipment based on the performance prediction model.

[0046] Optionally, the sensor monitoring data of the industrial equipment is preprocessed, specifically including the following steps:

[0047] Remove abnormal data from sensor monitoring data;

[0048] Wavelet transform is used to reduce noise in the sensor monitoring data after removing outlier data.

[0049] The sensor monitoring data after wavelet transform denoising is normalized.

[0050] Optionally, the linear marker layer specifically adopts the following formula:

[0051] ;

[0052] in, express t The output of the linear labeling layer at time intervals, express t The input to the linear labeling layer is defined by the time interval, and "Batch" indicates batch operations. and The parameters representing the linear labeling layer, L Indicates the length of the input sequence. d Representing feature dimension, R It represents the set of real numbers.

[0053] According to the specific embodiments provided in this application, this application has the following technical effects:

[0054] This application provides an intelligent prediction method and device based on the Mamba deep learning network. Through the parallel structure of the Mamba and GRU modules, it achieves complementary capture of long-term degradation trends and short-term dynamic features, overcoming the representation bottleneck of traditional models. A scaled dot product attention mechanism is introduced to compute the concatenated features, achieving adaptive fusion of multi-scale features and effectively suppressing the interference of noise and operating condition disturbances, ultimately achieving efficient fusion and high-precision prediction. Specifically, the linear time complexity of the Mamba module reduces computational costs, and combined with the local feature capture capability of the GRU module, it significantly improves prediction accuracy while ensuring computational efficiency. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart illustrating an intelligent prediction method based on a Mamba deep learning network, provided as an embodiment of this application;

[0057] Figure 2 A schematic diagram of the sliding time window processing procedure provided in this application;

[0058] Figure 3 A schematic diagram of the overall structure of the Mamba-GRU parallel network provided in this application;

[0059] Figure 4 This is a schematic diagram of the linear marker layer provided in this application;

[0060] Figure 5 This is a schematic diagram of the structure of the Mamba module provided in this application;

[0061] Figure 6 A schematic diagram of the GRU module provided in this application;

[0062] Figure 7 A schematic diagram illustrating the prediction performance of the FD001 test set provided in this application;

[0063] Figure 8 A schematic diagram illustrating the prediction performance of the FD003 test set provided in this application;

[0064] Figure 9 This is a schematic diagram illustrating the predicted RUL increment effect for a single engine provided in this application; wherein, Figure 9(a) in the figure is a schematic diagram of the RUL prediction for engine No. 24 in the FD001 test set; Figure 9 (b) in the diagram is a schematic diagram of the RUL prediction for engine number 76 in the FD001 test set; Figure 9 (c) in the diagram is a schematic diagram of the RUL prediction for engine No. 39 in the FD003 test set; Figure 9 (d) in the figure is a schematic diagram of the RUL prediction for engine No. 99 in the FD003 test set. Detailed Implementation

[0065] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0066] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0067] In one exemplary embodiment, such as Figure 1 As shown, an intelligent prediction method based on a Mamba deep learning network is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method includes steps 101 to 106. Wherein:

[0068] Step 101: Acquire sensor monitoring data from industrial equipment.

[0069] Step 102: Preprocess the sensor monitoring data of the industrial equipment.

[0070] Specifically, the preprocessing includes feature selection based on sensor signal stability, wavelet transform denoising, and Min-Max normalization, as detailed below:

[0071] S1021: Feature selection based on sensor signal stability, i.e., removing abnormal data from sensor monitoring data.

[0072] In practice, under single operating conditions, sensors exhibit different patterns of change over time. It can be observed that some sensors do not change significantly over time. Obviously, these values ​​are not related to the device's RUL and should be removed.

[0073] S1022: Discrete wavelet transform denoising, which is to perform wavelet transform denoising on sensor monitoring data after removing abnormal data.

[0074] In practice, the signals acquired by the sensors are affected by noise. To reduce the impact of noise on the device's RUL prediction results, this application uses Discrete Wavelet Transform (DWT) to denoise the original signal. The output sensor signal is processed using the following formula.

[0075] (1)

[0076] (2)

[0077] in, j Indicates scale. k Indicates translation. t Indicates time, f ( t () represents the original signal. c j,k These are the coefficients obtained after discrete wavelet transform. These are discrete wavelet basis functions. For the mother wavelet function, It is the reconstructed signal.

[0078] S1023: Perform min-max normalization on the sensor monitoring data to determine the normalized observation data, that is, perform normalization on the sensor monitoring data after wavelet transform noise reduction.

[0079] In practical applications, when the sensor monitoring data collected by the sensors are of different sizes and ranges, directly using them may have an adverse effect on the stability and convergence of the model during the training process. To solve this problem, this application adopts the min-max normalization method to scale all selected sensor values ​​to a uniform range of [0,1], as specifically defined in equation (3) below.

[0080] (3)

[0081] in, Indicates the first i The normalized measurement values ​​of each sensor are located between [0,1]. Indicates the first i The first sensor j One measurement value, and They represent the first i The minimum and maximum values ​​of the data monitored by each sensor.

[0082] Step 103: Generate samples from the preprocessed sensor monitoring data based on the sliding time window strategy to obtain time window samples.

[0083] While step-by-step input to the model is simple and computationally efficient, it cannot fully capture long-term dependencies and complex dynamic behaviors. To address this limitation, this application employs a time-window input strategy, providing additional contextual information for each time step. This enables the deep learning algorithm to extract higher-order features from the data, thereby improving the accuracy and robustness of RUL predictions. Specifically, the steps include the following:

[0084] S1031: Based on the time window input strategy, the sensor monitoring data is segmented using a sliding window mechanism to construct time window samples. The sliding time window is as follows: Figure 2 As shown, where, n Indicates the number of sensors. S This indicates the sliding step size of the sliding time window. T Indicates the length of the time window. N This represents the total length of the sensor monitoring data. The time window generates samples as follows: ,in, This represents the first moment within the sliding time window, and the sample label represents the last moment within the time window. The corresponding true RUL value. To ensure that each sample in the test set has a complete time window, the window length is... T Must meet ,in, Indicates the first test set q The lifecycle length of each engine. , This indicates the total number of engines in the test set.

[0085] S1032: Use the time window samples as input samples for the Mamba-GRU parallel network.

[0086] Step 104: Construct a Mamba-GRU parallel network, which includes: a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer.

[0087] like Figure 3 As shown, the Mamba-GRU parallel network connects Mamba and GRU modules in parallel, making full use of the advantages of both modules to achieve complementary capture of time series features. This allows the model to learn performance degradation features in time series more comprehensively, ultimately improving the model's performance.

[0088] In addition, a feature fusion module was designed, employing a scaled dot product attention mechanism to calculate the correlation between features, obtaining two feature scores. These scores are then normalized using a softmax function to obtain a dynamic weight distribution, which is used to adaptively fuse multi-scale degenerate features. Specifically, the Mamba-GRU parallel network includes a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer. Only the Mamba and GRU modules are connected in parallel; the others are sequentially connected. The following sections will describe each module in detail:

[0089] Linear labeling layer

[0090] Linear tokenization processes the input time series data through a linear layer, such as... Figure 4 As shown, the normalized time series format is performed through a single linear layer in equation (4).

[0091] (4)

[0092] in, express t The output of the linear labeling layer at time intervals, express t The input to the linear labeling layer is defined by the time interval, and "Batch" indicates batch operations. and The parameters representing the linear labeling layer, L Indicates the length of the input sequence. d Representing feature dimension, R It represents the set of real numbers.

[0093] Mamba module

[0094] The Mamba module, based on SSM, combines the advantages of RNN and Convolutional Neural Network (CNN) by employing a hybrid of recursive and convolutional operations to reduce computational complexity to linear levels, significantly improving the efficiency of long sequence processing. The state-space model represents a one-dimensional function or sequence (as...). x ( t Mapping ∈ R to hidden state Finally, output .in, m Indicates the size of the state dimension. L Given the sequence length, this process can be represented by linear ordinary differential equations (5) and (6):

[0095] (5)

[0096] (6)

[0097] in, State matrix , and Indicates projection parameters; x ( t )for t The output of the linear labeling layer at time intervals, This indicates the hidden state of the input in the Mamba module at the current moment. y ( t ) indicates the output of the Mamba module. This represents the hidden state in the Mamba module at the previous time step. R Represents the set of real numbers. m Represents the state dimension. L Indicates the length of the input sequence.

[0098] However, state-space models do not use discrete sequences; instead, they take continuous sequences as input and predict the output sequence. Mamba discretizes continuous systems, making them more suitable for deep learning scenarios. Specifically, it introduces a time-scale parameter and uses fixed discretization rules to discretize the continuous parameters. A and B Transform into discrete parameters and The commonly used discretization rule is zero-order preservation, which is defined as shown in equations (7) and (8):

[0099] (7)

[0100] (8)

[0101] in, Indicates the time scale parameter. I Represents the identity matrix. A , B After discretization, the SSM-based model can be computed in two ways: linear recursion or global convolution, defined by the following equation:

[0102] (9)

[0103] (10)

[0104] (11)

[0105] (12)

[0106] in, L and Representing the input sequences respectivelyx And the length of the structured convolution kernel.

[0107] The Mamba module combines the basic blocks of most SSM architectures, such as H3, with the Gated MLP commonly found in modern neural networks, such as Transformer. Figure 5 As shown.

[0108] GRU module

[0109] The GRU module is a type of neural network derived from RNNs. Compared to traditional RNNs, the GRU module can not only extract time-series information but also effectively solve the vanishing and exploding gradient problems found in traditional RNNs. The core idea of ​​the GRU module is to introduce a gating mechanism, which includes an update gate and a reset gate. These two gating units control the flow of information in the network, allowing the network to selectively forget or update certain information. This enables the GRU module to better handle time dependencies and reduce the number of parameters that need to be trained. The GRU module calculation formula is as follows:

[0110] (13)

[0111] (14)

[0112] (15)

[0113] (16)

[0114] in, for t The door to reset time. It is the sigmoid activation function. For the GRU module t The hidden state at time -1 Hidden state The weight matrix, x ( t )for t The output of the linear labeling layer at time intervals, Input to reset the door x ( t The weight matrix of ) for t The door is constantly being updated. Input to update the gate x ( t The weight matrix of ) To update the hidden state in the door The weight matrix, For the GRU module tThe candidate hidden states at time t, where tanh(·) is the tanh activation function, and ☉ is the Hadamard product. To reset the weight matrix of the hidden state after the gate is processed, Hidden state The weight matrix, For the GRU module t The output at any given time. The structure of the GRU module is as follows: Figure 6 As shown.

[0115] Feature fusion module

[0116] The feature fusion module uses a dynamic feature fusion method based on a scaled dot product attention mechanism, specifically including the following steps:

[0117] Step a: First, the long-term degradation trend extracted by the Mamba module is concatenated with the short-term dynamic features captured by the GRU module to obtain the concatenated features;

[0118] Step b: Subsequently, a scaled dot product attention mechanism is used to calculate the concatenated features, resulting in two feature scores: the original attention score is generated by calculating the dot product of the query vector and the key vector, and then divided by the scaling compensation factor of the feature dimension. This achieves temperature regulation and effectively suppresses gradient vanishing in high-dimensional spaces. m For the state dimension;

[0119] Step c: Finally, the two feature scores are normalized using the softmax function to obtain a dynamic weight distribution. The outputs of the two branches are then weighted and summed according to the weights to achieve dynamic fusion and obtain the fused features. This attention mechanism allows the model to flexibly adjust its dependence on the two branches based on the input data, enhancing the model's expressive power and adaptability.

[0120] Fully connected layer

[0121] The fully connected layers implement degenerate feature mapping and RUL prediction, and regression prediction is performed through three fully connected layers. First, the input features are mapped by the ReLU activation function; then, they are processed by a second layer of Dropout regularization; finally, the RUL predicted value is output.

[0122] Step 105: Train the Mamba-GRU parallel network based on the time window samples to obtain the lifetime prediction model.

[0123] Specifically, the above training process includes the following steps:

[0124] S1051: Divide the input features into a training set and a test set.

[0125] S1052: Input the features and RUL from the training set into the Mamba-GRU parallel network to train the Mamba-GRU parallel network.

[0126] S1053: Input the features in the test set into the trained Mamba-GRU parallel network to determine the RUL prediction value.

[0127] S1054: Compare the predicted RUL value with the actual RUL to determine the lifetime comparison result.

[0128] S1055: Based on the lifetime comparison results, select performance evaluation indicators to assess the prediction accuracy of the trained Mamba-GRU parallel network.

[0129] The root mean square error (RMSE) and score function are used to evaluate the model's predictive performance.

[0130] S1056: Based on the prediction accuracy, adjust the trained Mamba-GRU parallel network and use the adjusted Mamba-GRU parallel network as the RUL prediction model.

[0131] Based on the results of error distribution and robustness analysis, the performance of the Mamba-GRU parallel network in handling different working conditions and complex degradation modes is summarized, and the model design is further optimized.

[0132] Step 106: Predict the RUL of the industrial equipment based on the performance prediction model.

[0133] The Mamba-GRU parallel network constructed in this application was numerically tested on the publicly available C-MAPSS (Commercial Modular Aero-Propulsion System Simulation) engine dataset provided by NASA. The configuration of model parameters directly determines the model's learning ability, generalization ability, and ability to capture degradation patterns in time-series data. The training process in this application systematically optimizes the model's hyperparameters using a grid search method. Within a predefined parameter range, key hyperparameters such as window size, batch size, Dropout rate, regularization coefficient, and learning rate were systematically evaluated through training and validation with different combinations. The optimal configuration was selected based on the model's accuracy in predicting RUL (Rate-Induced Limits). This process effectively balances training efficiency and generalization performance, ensuring the model's accuracy and stability in identifying degradation patterns in time-series data. Taking the FD001 subset dataset as an example, Table 1 shows the selected hyperparameters and their ranges; bold values ​​indicate the best-performing configuration. These results demonstrate that optimized parameter settings significantly contribute to the improvement of model performance.

[0134] Table 1

[0135]

[0136] The size of the time window determines the length of historical information the model can capture. A larger window provides more historical context, enabling the model to better identify long-term trends. Experimental results show that a time window of 25 yields the best performance. Batch size primarily affects training efficiency and the stability of parameter updates. Experiments indicate that the optimal batch size is 256. Dropout is a regularization technique that reduces overfitting by randomly discarding neurons during training. Experiments show that a Dropout rate of 0.3 results in the best model performance, which is highly consistent with the model's complexity and the characteristics of the dataset. Furthermore, the learning rate significantly impacts training stability. Experiments show that a learning rate of 0.005 often leads to convergence problems, while a learning rate of 0.001 achieves the best predictive performance. Finally, the optimal parameters for the proposed Mamba-GRU parallel network are set as follows: time window length... T The training batch size is 256, the dropout rate is 0.3, the learning rate is 0.001, and the Adam optimizer and mean square error (MSE) loss function are used.

[0137] The test set consists of four subsets with 100, 259, 100, and 248 engines respectively. Experiments were conducted using the FD001 and FD003 subsets, and the RUL prediction results for these two subsets are shown below. Figure 7 and Figure 8 As shown in the figure, when the actual RUL is small, the predicted RUL is very close to the actual value with minimal error. Even when the actual RUL is large, the predicted value remains highly consistent with the actual value.

[0138] To simulate incremental RUL prediction during actual aircraft operation, this experiment not only performed overall RUL prediction for all engines in the test set, but also performed time-series RUL incremental prediction for individual engines in different subsets of the dataset. The prediction results are as follows: Figure 9 As shown. Among them, Figure 9 Part (a) in the diagram is a schematic diagram of the RUL prediction for engine number 24 in the FD001 test set; Figure 9 Part (b) is a schematic diagram of the RUL prediction for engine No. 76 in the FD001 test set; Figure 9 Part (c) is a schematic diagram of RUL prediction for engine No. 39 in the FD003 test set; Figure 9 Part (d) in the diagram is a schematic diagram of the RUL prediction for engine No. 99 in the FD003 test set.

[0139] As the engine's lifespan nears its end, prediction accuracy improves significantly. This is because degradation characteristics become more pronounced as failure approaches, and the degradation trend exhibits stronger linearity. This allows the prediction model to learn and capture these characteristics more effectively, resulting in a predicted RUL curve that more closely approximates the true linear degradation curve.

[0140] The experimental results comparing the Mamba-GRU parallel network proposed in this application with existing deep learning models are shown in Table 2.

[0141] Table 2

[0142]

[0143] As shown in Table 2, on the FD001 subset, the RMSE of the Mamba-GRU parallel network proposed in this application is second only to CNN-Transformer, and its score is significantly better than other methods. On the FD003 subset, the RMSE and score of the Mamba-GRU parallel network proposed in this application are both significantly better than existing methods. Compared with the five methods listed in Table 2, the RMSE is reduced by 18.15%, 16.39%, 14.04%, 12.29%, and 7.91%, respectively, and the score is reduced by 49.3%, 46.6%, 44.0%, 42.1%, and 70.2%, respectively. Experimental results show that the network constructed in this paper has good performance on the C-MAPSS dataset, further demonstrating the superiority of the Mamba-GRU parallel network proposed in this application in predicting the RUL of industrial equipment.

[0144] Based on the same inventive concept, this application also provides a Mamba deep learning network-based intelligent prediction device for implementing the aforementioned intelligent prediction method based on the Mamba deep learning network. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations in one or more embodiments of the Mamba deep learning network-based intelligent prediction device provided below can be found in the limitations of the RUL prediction method for industrial equipment described above, and will not be repeated here.

[0145] In one exemplary embodiment, a smart prediction device based on a Mamba deep learning network is provided, comprising:

[0146] The data acquisition module is used to acquire monitoring data from industrial equipment sensors.

[0147] The preprocessing module is used to preprocess the sensor monitoring data of the industrial equipment.

[0148] The time window sample generation module is used to generate samples from the preprocessed sensor monitoring data based on a sliding time window strategy to obtain time window samples.

[0149] The Mamba-GRU parallel network construction module is used to construct a Mamba-GRU parallel network, which includes: a linear labeling layer, a Mamba module, a GRU module, a feature fusion module, and a fully connected layer.

[0150] The training module is used to train the Mamba-GRU parallel network based on the time window samples to obtain a performance prediction model;

[0151] The performance prediction module is used to predict the RUL or health status of industrial equipment based on the performance prediction model.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An intelligent prediction method based on Mamba deep learning network, characterized in that, The intelligent prediction method based on the Mamba deep learning network comprises: obtaining industrial equipment sensor monitoring data; preprocessing the industrial equipment sensor monitoring data; generating samples based on the preprocessed sensor monitoring data based on a sliding time window strategy to obtain time window samples; constructing a Mamba-GRU parallel network, the Mamba-GRU parallel network comprising: a linear mark layer, a Mamba module, a GRU module, a feature fusion module, and a full connection layer; training the Mamba-GRU parallel network based on the time window samples to obtain a performance prediction model; predicting the remaining useful life or health state of the industrial equipment based on the performance prediction model; the data processing process in the feature fusion module is as follows: concatenating the long-period degradation trend extracted by the Mamba module and the short-time dynamic features captured by the GRU module to obtain concatenated features; calculating the concatenated features using a scaled dot-product attention mechanism to obtain two feature scores; normalizing the two feature scores using a softmax function to obtain a dynamic weight distribution; weighting and summing the long-period degradation trend and the short-time dynamic features according to the dynamic weight distribution to realize dynamic fusion and obtain fused features. 2.The intelligent prediction method based on Mamba deep learning network according to claim 1, wherein, The preprocessing of the industrial equipment sensor monitoring data comprises the following steps: removing abnormal data in the industrial equipment sensor monitoring data; performing wavelet transform denoising on the sensor monitoring data after removing abnormal data; performing normalization processing on the sensor monitoring data after wavelet transform denoising. 3.The intelligent prediction method based on Mamba deep learning network according to claim 1, wherein, The linear mark layer specifically uses the following formula: ； wherein, denotes t denotes the output of the linearly marked layer at time, denotes t denotes the input of the linearly marked layer at time, Batch denotes a batch operation, and denotes the parameters of the linearly marked layer, L denotes the input sequence length, d denotes the feature dimension, R denotes the set of real numbers. 4.The intelligent prediction method based on Mamba deep learning network according to claim 1, wherein, The Mamba module specifically uses the following formula: ； ； in, Represents the state matrix; and Indicates projection parameters; x ( t )for t The output of the linear labeling layer at time intervals, This indicates the hidden state of the input in the Mamba module at the current moment. y ( t () indicates the output of the Mamba module. This represents the hidden state in the Mamba module at the previous time step. R Represents the set of real numbers. m Represents the state dimension. L Indicates the length of the input sequence. 5.The intelligent prediction method based on Mamba deep learning network according to claim 1, wherein, The GRU module specifically uses the following formula: ； ； ； ； in, for t The door to reset time. It is the sigmoid activation function. For the GRU module t The hidden state at time -1 Hidden state The weight matrix, x ( t )for t The output of the linear labeling layer at time intervals, Input to reset the door x ( t The weight matrix of ) for t The door is constantly being updated. Input to update the gate x ( t The weight matrix of ) To update the hidden state in the door The weight matrix, For the GRU module t The candidate hidden states at time t, where tanh(·) is the tanh activation function, and ☉ is the Hadamard product. To reset the weight matrix of the hidden state after the gate is processed, Hidden state The weight matrix, For the GRU module t The output at each moment, i.e., in the GRU module t The hidden state at any given moment. 6.The intelligent prediction method based on Mamba deep learning network according to claim 2, wherein, The wavelet transform denoising of the sensor monitoring data after removing abnormal data specifically uses the following formula: ； ； wherein, j denotes a scale, k denotes a translation, t denotes a time, f t is the original signal, c j,k is the coefficient obtained after discrete wavelet transform, is the discrete wavelet base function, is the mother wavelet function, is the reconstructed signal.​ 7. An intelligent prediction device based on Mamba deep learning network, characterized in that, The intelligent prediction device based on the Mamba deep learning network comprises: a data acquisition module for obtaining industrial equipment sensor monitoring data; a preprocessing module for preprocessing the industrial equipment sensor monitoring data; a time window sample generation module for generating samples based on the preprocessed sensor monitoring data based on a sliding time window strategy to obtain time window samples; a Mamba-GRU parallel network construction module for constructing a Mamba-GRU parallel network, the Mamba-GRU parallel network comprising: a linear mark layer, a Mamba module, a GRU module, a feature fusion module, and a full connection layer; a training module for training the Mamba-GRU parallel network based on the time window samples to obtain a performance prediction model; a performance prediction module for predicting the RUL or health state of the industrial equipment based on the performance prediction model; the data processing process in the feature fusion module is as follows: concatenating the long-period degradation trend extracted by the Mamba module and the short-time dynamic features captured by the GRU module to obtain concatenated features; calculating the concatenated features using a scaled dot-product attention mechanism to obtain two feature scores; The two feature scores are normalized by a softmax function to obtain a dynamic weight distribution; The long-period degradation trend and the short-time dynamic feature are weighted and summed according to the dynamic weight distribution, dynamic fusion is realized, and a fusion feature is obtained. 8.The intelligent prediction device based on Mamba deep learning network according to claim 7, wherein, The industrial equipment sensor monitoring data are preprocessed, specifically including the following steps: Abnormal data in the sensor monitoring data are removed; The sensor monitoring data after removing the abnormal data are subjected to wavelet transform denoising; The sensor monitoring data after wavelet transform denoising are subjected to normalization processing. 9.The intelligent prediction device based on Mamba deep learning network according to claim 7, wherein, The linear mark layer specifically adopts the following formula: ； where, represents t represents the output of the linear marker layer at time, represents t represents the input of the linear marker layer at time, Batch represents a batch operation, and represents the parameters of the linear marker layer, L represents the input sequence length, d represents the feature dimension, R represents the set of real numbers.

Images