A method and system for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN

Abstract

This invention belongs to the technical field of fault prediction and health management, and discloses a method and system for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN. The method includes the following steps: S1, collecting data on the equipment to be predicted from operation to failure, preprocessing and enhancing the data, labeling the remaining useful life of the data, and forming a database with a one-to-one correspondence between each window of data and its remaining useful life; S2, constructing a prediction model, training the prediction model using the data in the database to obtain a converged prediction model, and using this converged prediction model to predict the remaining useful life of the equipment to be predicted. This invention solves the problems of insufficient RTF data for mechanical equipment in RUL prediction, erroneous degradation information in noisy data, and low model prediction accuracy caused by gradient vanishing in deep time series networks.

Description

Technical Field

[0001] This invention belongs to the technical field of fault prediction and health management, and more specifically, relates to a method and system for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN. Background Technology

[0002] Failure prediction and health management (PHM) can maximize the availability, reliability, and safety of machinery by monitoring its operating status. Remaining useful life (RUL) estimation is a crucial component of PHM. Accurate RUL estimation helps improve machine maintenance planning, save maintenance costs, and increase enterprise productivity.

[0003] Accurate Recovery Usage Limit (RUL) estimation is crucial for improving the safety and reliability of large machinery. Generally, RUL prediction methods can be categorized into model-based and data-driven methods. Model-based RUL prediction methods attempt to establish mathematical models to express the degradation patterns of mechanical systems by fully understanding the failure mechanisms. Data-driven RUL prediction methods utilize historical data related to the mechanical degradation process to estimate RUL. In recent years, due to the rapid development of deep learning (DL), more and more flexible deep learning models have been used for data-driven RUL estimation. Although DL-based methods have performed well in RUL estimation, they still face the following challenges. First, DL-based methods typically require sufficient labeled training data to avoid model overfitting and ensure generalization ability. However, collecting time-series data from machine operation to failure is undoubtedly a time-consuming task, and the amount of labeled data is often very limited. Therefore, how to effectively utilize labeled information becomes a key challenge. Data augmentation (DA) techniques can generate many training samples by introducing data transformations and have been successfully applied in fields such as image processing. However, in the field of PHM (Prognostics and Image Processing), most sample data is time-series data. Due to the unique distribution of time-series data, some algorithms previously used for image processing cannot be directly applied to the time-series domain. Furthermore, RUL estimation is a regression problem, requiring the capture of more detailed information from the time-series data. Secondly, current data augmentation methods for RUL do not consider the correlations between time-series data when generating virtual data, which may affect the quality and diversity of the generated data. In addition, most deep prediction models often encounter difficulties in effectively capturing degradation information from time-series data. For example, these models frequently ignore detailed degradation information in shallow layers and introduce potential erroneous degradation information from noisy data (both the original and generated virtual data inevitably contain a certain degree of noise), thus significantly impacting the overall performance of the prediction model.

[0004] Therefore, the performance of deep models can be improved by promoting the interaction of degradation information between different layers in the network and comprehensively considering information from multiple levels (shallow detailed degradation information and deep global degradation information). Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN, solving the problems of insufficient RTF data for mechanical equipment and low model prediction accuracy caused by gradient vanishing in time series deep networks in RUL prediction.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN is provided, the method comprising the following steps: S1 collects data on the equipment to be predicted to fail, preprocesses and enhances the data to obtain processed data, uses a sliding window to extract the processed data and marks the remaining lifespan of each segment of data to form a database in which each window data corresponds one-to-one with the remaining lifespan. The data enhancement is achieved by adaptively weighting data with similarity that meets a preset threshold to generate new data. S2 constructs a prediction model, trains the prediction model using data from the database to obtain a converged prediction model, and uses the converged prediction model to predict the remaining service life of the device to be predicted.

[0007] More preferably, in step S2, the prediction model adopts an improved DenseNet model, which includes an input layer, a Dense Block layer, a Transition Block layer and a fully connected layer connected in sequence.

[0008] More preferably, the input layer employs a one-dimensional filter, the Dense Block layer comprises five stacked convolutional layers, each convolutional layer containing a one-dimensional convolution and a Tanh activation function, the Transition Block layer comprises a convolutional layer and a filter, and the output layer outputs the remaining lifetime.

[0009] More preferably, in step S1, the similarity is measured using DTW distance.

[0010] More preferably, the similarity is calculated according to the following formula:

[0011] in, It is a sequence ]and ] The squared similarity distance between them and They represent the first i Article and Section j One piece of raw RTF data, m and n They represent their lengths respectively. This indicates the maximum length in the original RTF data.

[0012] More preferably, the new data is calculated according to the following formula:

[0013] in, Indicates the selected number i One piece of raw RTF data, This indicates the weight corresponding to the data. Indicates by the first i The original RTF data is used as the initial sequence to generate virtual RTF data.

[0014] More preferably, in step S1, the preprocessing includes data screening and normalization, wherein the data screening is to select experimental data that are strongly correlated with the degradation process according to preset rules, and the normalization is to normalize the data according to different working conditions.

[0015] More preferably, in step S1, when marking the remaining lifespan of each segment of data, the remaining lifespan at the last moment of each segment of data is used as the label.

[0016] According to another aspect of the present invention, a system employing a TWDBA-DCRN-based method for predicting the remaining useful life of mechanical equipment is provided, comprising a processor for executing the TWDBA-DCRN-based method for predicting the remaining useful life of mechanical equipment described above.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN.

[0018] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: 1. This invention first uses the proposed TWDBA algorithm to augment the original finite RTF data, generating virtual RTF data with similar degradation characteristics. When generating virtual RTF data, the similarity between different RTF data is determined in the time dimension rather than the time step dimension. This allows data with high similarity to be used for the generation of virtual RTF data, which can preserve the inherent degradation trend of the original data and avoid noise in the generated virtual data as much as possible, thereby improving the accuracy of RUL prediction. 2. The improved DenseNet, or DCRN, designed in this invention promotes the interaction of degradation information between different layers. By comprehensively considering information from multiple levels (shallow detailed degradation information and deep global degradation information), it can reduce the negative impact of noisy data on RUL prediction performance to a certain extent. In addition, when designing the model, we comprehensively considered the balance between model complexity and the information it captures, which reduced the number of model parameters and computational burden to a certain extent. Finally, DCRN emphasizes extracting degradation information from the temporal dimension of the data and focuses on capturing contextual degradation information at each level, which enhances the model's ability to capture degradation information. 3. The data augmentation method and prediction model proposed in this invention both revolve around the time dimension of the data and consider the degradation trend of the data. The data augmentation part can generate high-quality and diverse data, which helps to improve the generalization performance of the prognostic model. At the same time, the prediction model promotes the interaction and reuse of degradation information between layers. This ensures that the model effectively learns the inherent degradation trend of the data (including the original data and the generated virtual data) during the training process. Therefore, it has been shown to be more suitable for the remaining lifetime estimation problem in scenarios with insufficient samples and noisy data. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN, provided according to a preferred embodiment of the present invention. Figure 2 The above are visualization results of the original RTF data and the generated virtual RTF data applied to a specific instance of the present invention. Among them, (a) is the monitoring data of the second sensor in the FD001 dataset (the first sub-dataset in the C-MAPSS dataset) (only 5 original RTF data are shown), and (b) is the virtual RTF data generated by applying the TWDBA algorithm. Figure 3 This is a schematic diagram of the structure of the dense convolutional neural network designed in a preferred embodiment of the present invention; Figure 4This is a visualization of the RUL prediction results of the present invention on a specific example (predicting the RUL of all engines in the last recorded data points of the C-MAPSS dataset test data). Among them, (a) is the RUL prediction result diagram of the present invention on the FD001 test data, (b) is the RUL prediction result diagram of the present invention on the FD002 test data, (c) is the RUL prediction result diagram of the present invention on the FD003 test data, and (d) is the RUL prediction result diagram of the present invention on the FD004 test data. Figure 5 This is a visualization of the RUL prediction results of the present invention in a specific example (predicting data of the entire life cycle of four randomly selected engines). Among them, (a) is the RUL prediction result diagram of the present invention on the 34th engine randomly selected in FD001, (b) is the RUL prediction result diagram of the present invention on the 231st engine randomly selected in FD002, (c) is the RUL prediction result diagram of the present invention on the 99th engine randomly selected in FD003, and (d) is the RUL prediction result diagram of the present invention on the 135th engine randomly selected in FD004. Figure 6 This is a visualization of the RUL prediction results and errors of the present invention applied to a specific instance (FD001 and FD002 sub-datasets in the C-MAPSS dataset) and compared with those without data augmentation models. Among them, (a) is the RUL prediction result diagram of the augmented model (TWDBA-DCRN) and the baseline model (DCRN) on the FD001 test data, (b) is the RUL estimation error distribution diagram of the augmented model and the baseline model on the FD001 test data, (c) is the RUL prediction result diagram of the augmented model and the baseline model on the FD002 test data, and (d) is the RUL estimation error distribution diagram of the augmented model and the baseline model on the FD002 test data. Figure 7 The present invention is applied to specific instances (FD003 and FD004 sub-datasets in the C-MAPSS dataset) and compared with the data augmentation model without it, and the RUL prediction results and error visualizations are shown. Among them, (a) is the RUL prediction result diagram of the augmented model and the baseline model on the FD003 test data, (b) is the RUL estimation error distribution diagram of the augmented model and the baseline model on the FD003 test data, (c) is the RUL prediction result diagram of the augmented model and the baseline model on the FD004 test data, and (d) is the RUL estimation error distribution diagram of the augmented model and the baseline model on the FD004 test data. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] Figure 1 The flowchart below illustrates a TWDBA-DCRN-based method for predicting the RUL (Run-to-Failure) data of mechanical equipment, constructed according to a preferred embodiment of the present invention. This method first preprocesses the run-to-failure (RTF) data from multiple sensors monitoring the mechanical equipment. Then, a proposed Time-dimension-weighted dynamic time warping barycentric averaging (TWDBA) algorithm is used to augment the preprocessed RTF data, fully utilizing the correlation between the time series of the monitoring data to expand the scale of available data. Next, the original RTF data and the generated virtual RTF data are integrated into a designed Dense Convolution Regression Network (DCRN) to extract the degradation features of the RTF data. DCRN utilizes the close connections between layers to improve the information flow in the network, promoting the interaction of degradation information between layers, thereby reducing the possibility of gradient vanishing. Simultaneously, DCRN comprehensively considers the degradation features of multiple levels of RTF data (deep global degradation features and shallow detailed degradation features), mitigating the negative impact of noisy data on model estimation performance to some extent. Finally, the fully connected layers in the DCRN use nonlinear mapping to transform the multi-level degradation features extracted from the data by the model into specific values ​​for RUL estimation. This method specifically includes the following steps: S1. Data Acquisition and Processing (1) Collect RTF data monitored by multiple sensors of mechanical equipment; depending on the research purpose, RTF data of one or more mechanical equipment can be collected as training dataset and test dataset; the RTF data of the training set is used for data augmentation and model training, and the test data is used to evaluate model performance; In this embodiment, the mechanical equipment's full lifecycle signals are collected from multiple sensors. The present invention uses multiple sensors to collect the degradation time signals of the mechanical equipment.

[0022] (2) The collected RTF data were preprocessed. RTF data with strong correlation to the degradation process were selected as experimental data and normalized according to the working conditions of the RTF data to eliminate the differences between the dimensions of different monitoring signals.

[0023] Assume that the degradation data of mechanical equipment includes T At that moment, the input complete degradation process N RTF data monitored by individual sensors , No. j RTF data monitored by individual sensors Sensor data that is not strongly correlated with the degradation process or whose values ​​remain largely unchanged during degradation are deleted, and finally, only the following data are retained. RTF data monitored by individual sensors ,in Less than N .

[0024] Furthermore, in order to eliminate the significant differences in RTF data sizes monitored by different sensors and to eliminate sensor data variations caused by different operating conditions, [the following measures were taken]. The sensor data is normalized according to the formula. ,in and They represent the first j The sensor at the first k Calculate the normalized data by finding the minimum and maximum values ​​of RTF data under each working condition; (3) Data augmentation is performed on the preprocessed RTF data. Starting from the time dimension of the original data, RTF data monitored by a sensor is randomly selected as the initial data. The RTF data with a small distance to the initial data is searched by DTW distance and adaptively weighted and combined. Using multiple RTF data with small distances to generate virtual RTF data can fully preserve the inherent degradation trend of the original RTF data, thereby ensuring the quality of the generated data. In addition, the adaptive weighted combination ensures the diversity of the generated RTF data; Data augmentation is performed using different RTF data collected from the same sensor. Since RUL prediction is a regression problem, unlike classification problems, it predicts a continuous variable with no clear boundaries between them. Furthermore, if each RUL value is treated as a category, the inherent degradation trend of the original RTF data cannot be preserved during data augmentation. The generated virtual RTF data may contain significant noise, meaning there is an incorrect correspondence between the generated virtual RTF data and its RUL label, thus affecting the final RUL prediction accuracy. Therefore, for RUL prediction problems, data augmentation should not be performed at the class dimension, i.e., at each time step.

[0025] The proposed TWDBA algorithm is used to perform data augmentation on the original RTF data in the time series dimension. For multiple RTF data points collected by a single sensor, the similarity between the RTF data points is measured. Different RTF data points are stretched or compressed, and the RTF data points with high similarity are adaptively weighted to generate high-quality RTF data. Specific details are as follows: The dimensions of the original data used for data augmentation can be represented as follows: ,in, , , These represent the number of RTF data points, the time dimension (maximum sequence length), and the number of selected features, respectively. The original data, according to the selected features and time dimension, can be represented as follows: and Then with the first i The dataset corresponding to each feature and the i The datasets corresponding to each time step can be represented as follows:

[0026]

[0027] in, and Each column represents the first i One RTF data point recorded by each sensor and all sensors at the [number]th ... i Data recorded at each time step. First, from the dataset... Randomly select one RTF data Instead of choosing All sensor data recorded at a specific time step. RTF data. The characteristics of this feature ensure that data augmentation is performed along the time series dimension, rather than the dimension of each time step. To improve the quality of the generated virtual RTF data, we use DTW (Dynamic Time Warping) distance as a metric to search for the five most similar nearest neighbors of the RTF data. For any two distinct RTF data sets, DTW calculates the similarity distance of all subsequences, obtaining the first... i RTF data and the j RTF data The similarity matrix between them, where It is a sequence ]and ] The squared similarity distance between them.

[0028]

[0029] but express and The square root of the similarity distance between them. This refers to the similarity distance between RTF data points. RTF data with a smaller similarity distance is more suitable for data augmentation, as it does not disrupt the inherent degradation trend of the original RTF data.

[0030] Furthermore, to create diversity, weights are adaptively assigned to the selected RTF data. The randomly selected initial sequence is assigned a weight of 0.5. This higher weight ensures that the degradation trend of the generated virtual RTF data is similar to that of the original RTF data to a certain extent, thus guaranteeing that the generated data retains the inherent degradation information of the original data. Next, instead of all five, two RTF data points are randomly selected from these five, and each RTF data point is assigned a weight of 0.15. The total weights assigned here are 0.5 + 2 × 0.15 = 0.8. The remaining unweighted RTF data points share the remaining 0.2 weight to ensure that the sum of the weights is 1. Finally, all weighted RTF data points are used to synthesize the virtual RTF data. i A synthesized virtual RTF data can be represented as:

[0031] By selecting different initial RTF data and repeating the above process, a sufficient amount of virtual RTF data can be generated to form a virtual dataset. , can be represented as:

[0032] (4) Perform sliding window processing on the original RTF data and the generated virtual RTF data. The length of the sample is the size of the sliding window, and the width of the sample is the number of selected sensors. The sample can be represented as ,in ,in The remaining lifetime of each sample at the last moment is the label; S2. Based on the original RTF data and the generated virtual RTF data, a prediction model based on a Dense Convolution Regression Network (DCRN) is established to achieve accurate prediction of RUL. The input layer, Dense Block, and Transition Block in the DCRN enable feature reuse, effectively mitigating the possibility of gradient vanishing. Furthermore, the extracted shallow and deep features of the input data comprehensively consider multi-level degradation information, helping to reduce the negative impact of noisy data on the model's prediction accuracy. Finally, the fully connected layer in the DCRN fuses all features to output the predicted RUL value.

[0033] The basic DenseNet model is improved and applied to RUL prediction. The DCRN model proposed in this invention mainly consists of an input layer, a Dense Block, a Transition Block, and a fully connected layer, as detailed below: (1) Input layer: two-dimensional form of mixed data The input convolutional layer extracts shallow time-series features from the RTF data. The input layer contains a single convolutional layer. This convolutional layer contains five 7×1 filters, which perform convolution operations on the time-series dimension of the original input using one-dimensional filters.

[0034] (2) Dense Block: The Dense Block contains five stacked convolutional layers to extract deep features from the time series. Each convolutional layer contains a one-dimensional convolution plus a Tanh activation function. All convolutional layers are zero-padding to maintain the dimensionality of the feature maps. Batch normalization is omitted in the Dense Block to avoid destroying the statistical features of the original RTF data and to preserve the inherent temporal model in the data, thereby enhancing the model's sensitivity to subtle signals. The convolutional operation focuses on extracting the spatial features of the RTF data, while the Tanh function can accelerate the convergence speed of the model. This relatively short single Dense Block configuration can facilitate the transfer of degradation information and extract shallow information. By extracting shallow and deep features of the time series with the fewest Dense Blocks, richer degradation information can be obtained.

[0035] (3) Transition Block: The output of the Dense Block is of dimension 1. The Transition Block combines the outputs of the Dense Block into a single, unique feature map. The Transition Block consists of a convolutional layer and a 3×1 filter. After the convolution operation in the Transition Block, the spatial size of the feature map is small enough that pooling is unnecessary, thus discarding the pooling operation from the original Transition Block.

[0036] (4) Fully connected layer: By establishing a flatten layer and two fully connected layers to fuse the local features extracted by all convolutional layers, the degradation process of mechanical equipment can be more comprehensively characterized, and finally the predicted value of remaining service life (RUL) is output.

[0037] The present invention will be further described below with reference to specific embodiments.

[0038] The following section uses an aero-engine as a specific example to further illustrate the TWDBA-DCRN-based method for predicting the remaining useful life of mechanical equipment according to this invention. The specific steps are as follows: (1) The C-MAPSS dataset was used for training and testing. The C-MAPSS dataset consists of four subsets (FD001, FD002, FD003, and FD004) with different failure modes and operating conditions, as shown in Table 1. Each subset contains engine degradation information recorded by 21 sensors, such as temperature and pressure. The training dataset contains sensor data from multiple engines from start-up to failure, used for offline training of the TWDBA-DCRN model. The test set contains only sensor data before the failure occurs, used for online RUL estimation. This invention comprehensively evaluates all four subsets.

[0039]

[0040] (2) The C-MAPSS dataset contains RTF data collected by 21 sensors. However, some sensor data remain constant or undergo slight changes throughout the engine's lifecycle. Therefore, we retained the valid sensor data and discarded useless data. Specifically, we retained data from 14 sensors, numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21. Next, to eliminate significant differences in the size of RTF data monitored by different sensors and to eliminate sensor data variations caused by different operating conditions, the retained sensor data was normalized.

[0041] (3) In the data augmentation stage, we used the proposed TWDBA algorithm to augment the training data of each subset, generating degradation data for 100, 260, 100, and 249 engines respectively. In the C-MAPSS dataset, engines are typically in the degradation stage during the last 125 cycles. Therefore, to ensure that the data label range after the sliding window operation covers 0–125, the last 125 cycles of the original data for each engine are augmented. ( (Number) periods are sampled for generating virtual RTF data. Therefore, the input to the TWDBA algorithm is a normalized data matrix, with input sizes of 100×155×14, 260×145×14, 100×155×14, and 249×140×14 for each subset. Finally, the generated data is mixed with the original data to form a new training dataset for training the RUL estimation model. Figure 2 The original data and the virtual RTF time series generated using TWDBA are displayed. The original data comes from the monitoring data of the second sensor in the FD001 dataset. It can be seen that the generated RTF data has similar degradation characteristics to the original RTF data, and their degradation curve trends are also basically consistent.

[0042] (4) Apply sliding window processing to the original RTF data and the generated virtual RTF data. The length of the sample is the size of the sliding window. The sample width is equal to the number of sensors selected. =14; FD001, FD002, FD003, and FD004 subsets The values ​​were set to 30, 20, 30 and 15 respectively.

[0043] (5) Establish a prediction model based on DCRN, the structure of which is shown in the figure below. Figure 3 As shown in Table 2, the specific settings for DCRN are as follows. A variable learning rate strategy was employed during model training. In the first stage, the learning rate was set to 0.001 to quickly update the trainable parameters. In the second stage, the learning rate was set to 0.0001 to fine-tune the trainable parameters and obtain stable convergence results. All hyperparameters were determined through grid search, as shown in Table 3. During backpropagation, the loss function was the mean absolute error, and the optimization algorithm was RMSProp. After model training was completed, a test dataset was input, and the RUL prediction value for the test dataset was output. All experiments were run 10 times, and the average value was taken as the final experimental result to reduce randomness.

[0044]

[0045]

[0046] (6) Visualize the RUL prediction results. Taking the test data of FD001, FD002, FD003 and FD004 as examples, the RUL prediction is performed using a DCRN-based prediction model. The prediction results are as follows: Figure 4 and Figure 5 As shown. From Figure 4 and Figure 5 As can be seen, this patent performs excellently in solving the RUL prediction problem of aero-engines, verifying the effectiveness of the invention.

[0047] To further verify the effectiveness of this patent in augmenting RTF data and designing the DCRN prediction model, three baseline RUL prediction models were selected for comparison, including two representative deep learning methods: Deep Convolutional Neural Network (DCNN), Long Short-Term Memory (LSTM), and the DCRN designed in this patent. To verify the superiority of data augmentation, two scenarios were discussed for each baseline model: (1) a baseline model using only the original data and (2) an augmented model using both the original and generated data. The experimental results were analyzed using root mean square error (RMSE) and a score function, the calculation methods of which are explained in detail below:

[0048]

[0049]

[0050] in, and The values ​​represent the predicted and actual RUL values, respectively. Table 4 summarizes all estimation results based on the root mean square error and score, where the improvement (Imp) is calculated as the reduction rate of the reference index value. The last row of Table 4 shows the overall results of different methods on all subsets. Compared with DCRN, TWDBA-DCRN improves the overall estimation performance by 2.26% and 17.33% using RMSE and Score, respectively.

[0051] Figure 6 and Figure 7The table shows the RUL estimation errors of DCRN for the engines and their distributions in four subsets, serving as auxiliary evaluation metrics for RMSE and the scoring function. The blue and red curves represent the estimation results of DCRN and TWDBA-DCRN, respectively. The red curve is closer to the true values ​​than the blue curve, indicating that more accurate estimation results can be obtained by adding generated data. Furthermore, error values ​​less than 0 in the histogram indicate early predictions. TWDBA-DCRN has a higher proportion of early predictions and low-value ranges, further validating the superiority of the proposed data augmentation framework and the reliability and effectiveness of the generated data.

[0052]

[0053] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN, characterized in that, The method includes the following steps: S1 collects data on the equipment to be predicted to fail, preprocesses and enhances the data to obtain processed data, uses a sliding window to extract the processed data and marks the remaining lifespan of each segment of the data to form a database in which each window data corresponds one-to-one with the remaining lifespan. The data enhancement is achieved by adaptively weighting data with similarity that meets a preset threshold to generate new data. S2 constructs a prediction model, trains the prediction model using data from the database to obtain a converged prediction model, and uses the converged prediction model to predict the remaining service life of the device to be predicted. In step S2, the prediction model adopts the improved DenseNet model, which includes an input layer, a Dense Block layer, a Transition Block layer and a fully connected layer connected in sequence. In step S1, the similarity is measured using DTW distance; The similarity is calculated according to the following formula: in, It is a sequence ]and The squared similarity distance between [1:Tc], m,n∈[1:Tc] and They represent the first i Article and Section j One piece of raw RTF data, m and n They represent their lengths respectively. This indicates the maximum length in the original RTF data; The new data is calculated according to the following formula: in, Indicates the selected number i One piece of raw RTF data, This indicates the weight corresponding to the data. Indicates by the first i The original RTF data is used as the initial sequence to generate virtual RTF data.

2. The method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN as described in claim 1, characterized in that, The input layer uses a one-dimensional filter, the Dense Block layer includes five stacked convolutional layers, each containing a one-dimensional convolution and a Tanh activation function, the Transition Block layer includes a convolutional layer and a filter, and the output layer outputs the remaining lifetime.

3. The method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN as described in claim 1, characterized in that, In step S1, the preprocessing includes data screening and normalization. The data screening involves selecting experimental data that are strongly correlated with the degradation process according to preset rules, and the normalization involves normalizing the data according to different working conditions.

4. The method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN as described in claim 1, characterized in that, In step S1, when marking the remaining lifespan of each segment of data, the remaining lifespan at the last moment of each segment of data is used as the label.

5. A system employing a method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN, characterized in that, Includes a processor for executing the TWDBA-DCRN-based method for predicting the remaining useful life of mechanical equipment as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for predicting the remaining useful life of mechanical equipment based on TWDBA-DCRN as described in any one of claims 1-4.

Images