Data-based complex equipment life prediction method

Abstract

The invention relates to the technical field of equipment prediction and health management, in particular to a data-based complex equipment life prediction method. The method comprises the following steps: carrying out sliding filtering and self-attention calculation on time sequence data of a sensor so as to denoise and strengthen key features; degeneration features of short, medium and long scales are extracted through a multi-scale time convolution network; constructing two prediction branches of basic feature fusion and gating enhancement feature fusion; based on the gating enhancement feature, adaptively selecting a branch and outputting a residual life prediction result; during training, a weighted mean square error loss function is adopted, and a higher weight is given to the prediction error at the end of the life. According to the method, self-adaptive fusion and selection of multi-scale features are realized, the life prediction precision is improved, and the method has higher reliability especially in a key stage close to failure, and is suitable for predictive maintenance of complex equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment prediction and health management technology, and more specifically to a data-based method for predicting the lifespan of complex equipment. Background Technology

[0002] Complex equipment faces challenges such as material aging, component fatigue, and performance degradation caused by energy issues during long-term operation, which adversely affect the safe and efficient operation of the equipment. There is an urgent need to accurately and in real-time predict the remaining service life of complex equipment in the field of equipment prediction and health management. Based on equipment status information, combined with historical data and relevant models, potential failures of the equipment can be monitored in real time and maintenance can be implemented to extend service life, reduce failure probability, reduce unplanned downtime, and minimize major accidents caused by catastrophic failures.

[0003] Machine life prediction technologies can be mainly divided into two categories: The first category is prediction based on physical models. This involves introducing reasonable and necessary assumptions based on the working principles, failure mechanisms, and domain knowledge of mechanical components to construct a physical model. This model captures the physical laws governing the degradation process of mechanical components and predicts future performance degradation trends. However, for complex mechanical equipment, accurately establishing a physical model is often impractical. Its complex structural forms, numerous components, and intricate interactions between them all make accurate modeling extremely difficult.

[0004] The second category is data-driven prediction, which uses collected equipment status data as key support. Based on the feature elements extracted from the status data, it can predict the operating status of the equipment over a future period. This approach eliminates the reliance on specific physical models of faults or degradation processes, and can handle complex degradation processes. It is particularly suitable for life prediction tasks of complex equipment with intricate structures, diverse operating conditions, and where accurate physical models are difficult to establish. However, existing data-driven methods typically rely on large amounts of high-quality data and lack interpretability, limiting their reliability in practical applications. Furthermore, equipment sensor data contains rich multi-scale information, but most methods extract features only from a single scale or ignore cross-scale temporal dependencies and feature correlations, thus affecting prediction performance.

[0005] Complex equipment often has complex structures, many operating conditions, and is difficult to model accurately. Therefore, data-driven prediction technology is more suitable for the life prediction task of complex equipment. Summary of the Invention

[0006] In view of this, the present invention provides a data-based method for predicting the lifespan of complex equipment, which aims to solve the problems of insufficient accuracy and limited generalization ability in lifespan prediction caused by the coupling of degradation features, significant multi-scale characteristics of time-series data, diverse operating conditions, and noisy data in complex equipment during operation, thereby improving the accuracy and reliability of lifespan prediction to support predictive maintenance of equipment.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a data-based method for predicting the lifespan of complex equipment, comprising: Trainable sliding filtering and self-attention calculation are performed on sensor time-series data to obtain denoised and enhanced feature sequences; The feature sequence is input into a multi-scale temporal convolutional network to extract short-scale, medium-scale, and long-scale features. The short-scale, medium-scale, and long-scale features are concatenated to obtain multi-scale features; The multi-scale features are respectively input into the basic multi-scale feature fusion prediction branch and the gated enhanced multi-scale feature fusion prediction branch to obtain the basic branch prediction result and the gated enhanced branch prediction result. Based on the multi-scale features, a selection vector is generated through an adaptive prediction branch selection module; The prediction results of the basic branch and the prediction results of the gated enhancement branch are weighted and fused according to the selection vector to output the final remaining lifetime prediction value. The model training employs a weighted mean squared error loss function, which assigns a higher weight to the end-of-life prediction error.

[0008] In one specific implementation, the multi-scale temporal convolutional network includes TCN1, TCN2, and TCN3, which are used to extract the short-scale features, the medium-scale features, and the long-scale features, respectively.

[0009] In one specific implementation, the basic multi-scale feature fusion prediction branch performs the following operations: Based on the multi-scale features, time aggregate features are obtained by calculating time weights and weighting them. Based on the aforementioned temporal aggregation features, scale weights are generated using MLP and the Sigmoid function. The time-aggregated features are weighted and fused based on the scale weights to obtain global features; Based on the global features, scale weights are generated through MLP and the multi-scale features are weighted and fused to obtain fused features; The fused features are input into the regressor to obtain the basic branch prediction results.

[0010] In one specific implementation, the gated enhanced multi-scale feature fusion prediction branch performs the following operations: For the multi-scale features, scale-gated features, time-gated features, and feature-gated features are generated respectively through scale gating, time gating, and feature gating; The scale-gated feature, the time-gated feature, and the feature-gated feature are multiplied element-wise to obtain the gated enhanced multi-scale feature. Based on the gated enhanced multi-scale features, a gated enhanced global feature is constructed; Based on the gating augmentation global features, the gating augmentation branch prediction results are obtained.

[0011] In one specific implementation, the scale gating, the time gating, and the feature gating are all implemented by an MLP.

[0012] In one specific implementation, the adaptive prediction branch selection module performs the following operations: Global features constructed based on the gated and enhanced multi-scale features are used to generate selection scores via MLP; Apply Gumbel-Softmax to the selection score to generate an approximate one-hot selection vector, which is then used as the selection vector.

[0013] In a specific feasible implementation, the weighted mean square error loss function is: , Where B is the batch size and T is the sequence length. and Predicted and actual remaining lifetime values, As weight.

[0014] In one specific feasible implementation, the weight The remaining lifetime is dynamically calculated based on the actual remaining lifetime value, and higher weights are assigned to moments with shorter remaining lifetimes.

[0015] In one specific implementation, the parameters of the trainable sliding filter are updated via backpropagation and initialized as a mean filter.

[0016] In one specific implementation, the self-attention computation is a self-attention computation with residuals.

[0017] Compared with existing technologies, the data-based method for predicting the lifespan of complex equipment described in this invention has the following advantages: This invention effectively suppresses noise interference and highlights key features crucial for lifetime prediction by performing signal sliding denoising on the sensor sequence during the input stage and introducing a self-attention mechanism. Based on a multi-scale temporal convolutional network, it simultaneously extracts both the original multi-scale features and those enhanced by scale, time, and feature gating mechanisms, enabling dynamic filtering and adaptive enhancement of information at different time scales and feature dimensions, avoiding information loss caused by single-scale modeling. Furthermore, it constructs global features separately and achieves multi-scale information fusion through a learnable weighting method, improving the effectiveness and stability of feature fusion. Finally, an adaptive branch selection mechanism dynamically generates the final remaining lifetime prediction result, allowing the model to flexibly adjust prediction strategies for different degradation stages. Simultaneously, a weighted mean squared error loss function is introduced, assigning higher weight to prediction errors in the later stages of the sequence, thereby improving prediction accuracy near the critical failure stage.

[0018] Experimental results show that the present invention achieves better prediction performance than existing mainstream methods on both the stratospheric airship dataset and the single-fault mode dataset of the aero-engine dataset (C-MAPSS). This verifies its comprehensive advantages in multi-scale feature extraction, information fusion, adaptive prediction, and late-life prediction accuracy, and can provide an efficient and reliable technical solution for health management and predictive maintenance of complex industrial equipment. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the architecture of a data-based method for predicting the lifespan of complex equipment as described in this invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, the data-based method for predicting the lifespan of complex equipment described in this invention is applied to accurately and in real-time predict the remaining lifespan of complex equipment, thereby improving the reliability of the equipment. The method includes: (I) Multi-scale feature modeling module The multi-scale feature modeling module is used to extract degradation information at short, medium, and long scales from time series data, capturing rapid changes, trends, and periodic patterns in the sequence. This module mainly consists of three parts: trainable sliding filter, self-attention computation, and multi-scale temporal convolution. The original time series is represented as follows: , Indicates the number of variables in the sequence. Indicates the size of the sliding window. express The values ​​of each variable at each time point. First, a trainable sliding filter is used to process the time series data to smooth high-frequency fluctuations in complex equipment time series, while also handling potential spikes or outliers. For each position of the sliding window... The filtered output can be expressed as: , in, Indicates feature index, Indicates a time index. , representing the radius of the area covered by the convolution kernel. Indicates to Round down to the nearest integer. This is the kernel size (11 in the example). Representation of features The convolution weights, Representation of features The bias, and These are trainable parameters, updated via backpropagation. The filter is initialized as a mean filter.

[0023] Output after sliding filter processing Represented as: , Considering that the degradation characteristics of complex equipment are often coupled and different features have different impacts on equipment lifespan, a self-attention calculation with residuals is performed on the filtered sequence to highlight the features more important to the prediction task. The output of the self-attention can be expressed as: , in, , , , , respectively, represent the query vector, key vector, and value vector mapped from the input sequence. , and It is a learnable weight matrix, initialized using the Xavier Uniform method. For the embedding dimension, this embodiment takes .

[0024] Time-series data of complex equipment typically exhibits characteristics such as long-term dependencies and short-term fluctuations. To fully mine and utilize the degradation information in time-series data, three-scale Temporal Convolutional Networks (TCNs)—short, medium, and long—are used to process the data after self-attention computation. Specifically, in this embodiment, the short-scale TCN uses 2 base convolutional kernels, the medium-scale TCN uses 3 base convolutional kernels, and the long-scale TCN uses 5 base convolutional kernels. The processed output can be represented as: , in, This represents a series of transformation operations within the TCN module, namely dilated causal convolution, weight normalization, ReLU activation, and Dropout.

[0025] By setting different sized base convolution kernels to achieve short, medium, and long-scale temporal convolution operations, it is possible to simultaneously cover the extraction and modeling of short-term state changes and long-term degradation trends, which helps to enrich the degradation information required for life prediction of complex equipment. To better integrate the information at the three scales, the temporal convolution outputs at the three scales are concatenated into a three-dimensional vector along the channel dimension. ,Right now: .

[0026] (II) Dual-branch multi-scale feature fusion prediction module The dual-branch multi-scale feature fusion prediction module includes two parallel prediction branches: basic multi-scale feature fusion prediction and gated enhanced multi-scale feature fusion prediction. These branches are used to model and predict the global evolution trend and local fine-grained change features of multi-scale temporal features, respectively.

[0027] The basic multi-scale feature fusion prediction module first utilizes the results output by the multi-scale temporal convolutional network. To construct global features, specifically, the time weights are obtained by averaging the feature dimensions and then applying a softmax function. ,Right now: , Then, the time-weighted features of the three scales are weighted to obtain the time-aggregated features. : , This operation allows us to capture overall trends at various scales while suppressing short-term noise.

[0028] Subsequently, the aggregated features are processed using a two-layer feedforward fully connected neural network, and the output is mapped to the [0,1] interval using the Sigmoid function to obtain the final scale weights. : , in, These are the learnable weights.

[0029] The global feature is obtained by weighting and fusing the three scale features according to their scale weights. : .

[0030] After the global features are constructed, they are used to effectively fuse multi-scale convolutional features to fully utilize feature information at each scale. First, a two-layer feedforward fully connected neural network is used to process the global features to obtain the weights for each scale. : , in, These are the learnable weights.

[0031] Then, the multi-scale features are weighted and summed using this weight to obtain the fused features. : , After completing the multi-scale feature fusion, the fused features are input into a regressor consisting of two fully connected neural networks for regression prediction, thus obtaining the basic multi-scale feature fusion prediction result. : .

[0032] The gated-enhanced multi-scale feature fusion prediction module adds a gated enhancement mechanism to the basic multi-scale feature fusion prediction module. This mechanism enhances the output of the multi-scale temporal convolutional network from the dimensions of scale, time, and features. Feature selection and reconstruction are performed to highlight locally sensitive features. Specifically, scale gating unfolds the features at each scale into one dimension and inputs them into a two-layer feedforward fully connected neural network to obtain the gating tensor for each scale. Temporal gating maps the temporal dimension of each scale feature to a two-layer feedforward network, generating a gating tensor in the temporal dimension. Feature gating processes the feature dimensions at each scale, obtaining the feature-gating tensor through a two-layer feedforward network. Then, the three gate tensors are multiplied element-wise to obtain the gated enhanced multi-scale features. : .

[0033] This multi-dimensional gating selection adaptively emphasizes locally sensitive features across scale, time, and characteristics, improving responsiveness to short-term abrupt changes, local anomalies, and early fault signals while preserving the overall degradation trend, thus enhancing lifetime prediction accuracy. The resulting gated enhanced features are then sequentially processed through the same global feature construction as the basic multi-scale feature fusion prediction module. Multi-scale feature fusion And regression prediction, ultimately yielding the prediction results of this module. .

[0034] (III) Adaptive Prediction Branch Selection Module The adaptive prediction branch selection module, based on global features constructed from gated enhanced features, introduces a learnable branch selection mechanism that can automatically select the appropriate prediction path according to the overall feature distribution of the sample.

[0035] First, the global features constructed based on gated enhanced multi-scale features are... Input a two-layer feedforward fully connected network and obtain the selection scores of the two prediction branches. To achieve hard selection, where each sample activates only one prediction branch, this invention employs Gumbel-Softmax to generate an approximately differentiable one-hot selection vector. Each sample's selection vector contains two elements. And satisfy Final remaining lifetime prediction results It is obtained by weighting the outputs of the two branches: .

[0036] (iv) Weighted mean square error loss function In remaining useful life prediction tasks, traditional methods using mean squared error as the loss function treat prediction errors at all time steps equally, failing to reflect the varying importance of different time steps. To increase attention to prediction errors in later stages of the sequence (i.e., when the actual remaining useful life is relatively small and nearing failure), this invention designs the following weighted mean squared error loss function: , Where B is the batch size and T is the sequence length. and Let RUL be the predicted value and the actual value of the i-th time window at time step t, respectively. These are the corresponding weights. The allocation is dynamically based on the actual RUL value for each time window, calculated using the following formula: , in, The base weight is set to 1.1 in this embodiment. The weight control factor is set to 0.9 in this embodiment. Average value normalization is used to ensure the overall weight mean is 1, making the loss magnitude consistent with the original mean squared error magnitude, which is beneficial for model training. Through the above loss function design, the focus on the low remaining lifetime interval can be adaptively increased during training, making the final prediction results more accurate and sensitive when the lifetime approaches the critical point.

[0037] On the stratospheric airship dataset, the hyperparameter settings of each module of this invention are shown in Table 1, and the evaluation metrics are shown in Table 2. On the C-MAPSS dataset, the hyperparameter settings of each module of this invention are shown in Table 3, and the comparison of evaluation metrics is shown in Table 4. Among the comparison methods listed in Table 4, BiLSTM is a bidirectional long short-term memory network, SCTA-LSTM is a long short-term memory framework based on spatial correlation and temporal attention mechanisms, ADFG is an adaptive dynamic fusion graph neural network, MSIDSN is a multi-scale integrated deep self-attention network, MDFA is a multi-scale dilated fusion attention mechanism, AGATT is a method based on attention-graph network-Transformer, and ARR-Transformer is an adaptive weighting technique based on remaining lifetime.

[0038] The root mean square error (RMSE) is a commonly used measure of the difference between measured values. The formula for calculating it is as follows: , Scoring functions are used for quantitative evaluation. The formula for the cumulative deviation between a predicted value and the actual value is shown below: , , Mean Absolute Percentage Error (MAPE) is a metric used to measure the difference between predicted and observed values. It is commonly used in time series analysis, regression analysis, and other fields to evaluate the predictive performance of models. The calculation formula is as follows: , Accuracy measures the reliability or pass rate of model predictions by calculating the proportion of samples that fall within a specific allowable error range. The formula is as follows: , , in, This represents the difference between the predicted and actual lifespan. In accuracy calculations, the discrimination interval for stratospheric airships is... The engine discrimination interval is .

[0039] Table 2. Lifespan Prediction and Assessment Indicators for Stratospheric Airships

[0040] Table 4. Lifetime Prediction Evaluation Metrics for the C-MAPSS Dataset

[0041] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A data-driven method for predicting the lifespan of complex equipment, characterized in that, include: Trainable sliding filtering and self-attention calculation are performed on sensor time-series data to obtain denoised and enhanced feature sequences; The feature sequence is input into a multi-scale temporal convolutional network to extract short-scale, medium-scale, and long-scale features. The short-scale, medium-scale, and long-scale features are concatenated to obtain multi-scale features; The multi-scale features are respectively input into the basic multi-scale feature fusion prediction branch and the gated enhanced multi-scale feature fusion prediction branch to obtain the basic branch prediction result and the gated enhanced branch prediction result. Based on the multi-scale features, a selection vector is generated through an adaptive prediction branch selection module; The prediction results of the basic branch and the prediction results of the gated enhancement branch are weighted and fused according to the selection vector to output the final remaining lifetime prediction value. The model training employs a weighted mean squared error loss function, which assigns a higher weight to the end-of-life prediction error.

2. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The multi-scale temporal convolutional network includes TCN1, TCN2, and TCN3, which are used to extract the short-scale features, the medium-scale features, and the long-scale features, respectively.

3. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The basic multi-scale feature fusion prediction branch performs the following operations: Based on the multi-scale features, time aggregate features are obtained by calculating time weights and weighting them. Based on the aforementioned temporal aggregation features, scale weights are generated using MLP and the Sigmoid function. The time-aggregated features are weighted and fused based on the scale weights to obtain global features; Based on the global features, scale weights are generated through MLP and the multi-scale features are weighted and fused to obtain fused features; The fused features are input into the regressor to obtain the basic branch prediction results.

4. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The gated enhanced multi-scale feature fusion prediction branch performs the following operations: For the multi-scale features, scale-gated features, time-gated features, and feature-gated features are generated respectively through scale gating, time gating, and feature gating; The scale-gated feature, the time-gated feature, and the feature-gated feature are multiplied element-wise to obtain the gated enhanced multi-scale feature. Based on the gated enhanced multi-scale features, a gated enhanced global feature is constructed; Based on the gating augmentation global features, the gating augmentation branch prediction results are obtained.

5. The data-based method for predicting the lifespan of complex equipment according to claim 4, characterized in that, The scale gating, the time gating, and the feature gating are all implemented by MLP.

6. The data-based method for predicting the lifespan of complex equipment according to claim 4, characterized in that, The adaptive prediction branch selection module performs the following operations: Global features constructed based on the gated and enhanced multi-scale features are used to generate selection scores via MLP; Apply Gumbel-Softmax to the selection score to generate an approximate one-hot selection vector, which is then used as the selection vector.

7. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The weighted mean square error loss function is: ； Where B is the batch size and T is the sequence length. and Predicted and actual remaining lifetime values, For weights.

8. The data-based method for predicting the lifespan of complex equipment according to claim 7, characterized in that, The weight The remaining lifetime is dynamically calculated based on the actual remaining lifetime value, and higher weights are assigned to moments with shorter remaining lifetimes.

9. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The parameters of the trainable sliding filter are updated through backpropagation and initialized as a mean filter.

10. The data-based method for predicting the lifespan of complex equipment according to claim 1, characterized in that, The self-attention calculation is a self-attention calculation with residuals.

Images