Degradation equipment health index construction method fusing multi-source sensing information

Abstract

The invention provides a degraded equipment health index construction method fusing multi-source sensing information, and aims to solve the problem that a traditional health index construction method is difficult to comprehensively reflect the equipment degradation state. The method is based on a deep learning algorithm, considers time domain features, frequency domain features and spatial features of multi-source sensing information at the same time, and constructs a one-dimensional health index through feature extraction, fusion and optimization. Experimental results show that the method has remarkable advantages in the aspects of reflecting the degradation trend of the equipment and improving the time-varying trend and quality consistency, and the accuracy of predictive maintenance and life prediction of the equipment can be effectively improved. The method is suitable for health management of various industrial devices, has wide universality and adaptability, and provides powerful support for safe operation and intelligent management of the industrial devices.

Description

Technical Field

[0001] The present invention relates to the field of industrial equipment health management, and in particular to a method for constructing a health indicator of degraded equipment by integrating multi-source sensor information. Background Art

[0002] With the rapid development of industrial technology, modern industrial equipment is continuously progressing towards larger-scale, integrated, and automated systems. While these devices are improving production efficiency and expanding their application areas, they also face the challenges of an increasingly complex operating environment. Over long-term operation, equipment inevitably degrades and fails due to factors such as the environment and operating conditions. This poses a serious threat to production safety and can result in significant economic losses. Therefore, effective industrial equipment health management is crucial.

[0003] One of the core issues in equipment health management is the remaining useful life (RUL) prediction, which is to predict the remaining use time of the equipment before it is completely damaged from its current state based on the historical measurement data or mechanism model of the equipment. This process mainly includes two key steps: health indicator extraction and life prediction model construction. Traditional health indicator construction methods usually assume that the health status of the equipment is directly reflected by a single performance degradation variable and obtain monitoring data through a single sensor. However, for complex degraded equipment, the monitoring data obtained by a single sensor is often difficult to fully and comprehensively reflect the potential health status of the equipment, nor can it effectively characterize its random evolution process. Therefore, how to effectively integrate the multi-source sensor information of degraded equipment and construct a health indicator that can fully reflect the degradation trend of the equipment is the key to achieving accurate prediction of the remaining useful life of the equipment.

[0004] Currently, multi-source information feature extraction methods can be primarily categorized into time-domain and frequency-domain methods. Time-domain analysis evaluates component faults by calculating the statistical characteristics of signals in the time domain. Commonly used features include dimensional features such as maximum, variance, and root mean square (RMS), as well as dimensionless features such as kurtosis, skewness, and margin factor. While computationally simple, time-domain methods are susceptible to noise during monitoring and are primarily used for signals with obvious fault characteristics. Frequency-domain analysis, based on the Fourier transform, converts time-domain signals into frequency-domain signals and identifies fault states by analyzing the frequency distribution within the spectrum. Common methods include envelope analysis, refined spectrum, power spectrum, and holographic spectrum. However, with the increasing complexity of degraded equipment and multi-source information, statistical or machine learning-based methods struggle to extract the implicit information underlying the degradation process. Existing methods, most of which focus solely on time-domain or frequency-domain features, are unable to effectively extract degradation trends and construct high-quality health indicators, thus impacting the accuracy of subsequent fault diagnosis or lifespan prediction.

[0005] In order to solve the above problems, the applicant proposes a method for constructing a health indicator of degraded equipment by fusing multi-source sensor information. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for constructing a health indicator of a degraded device by fusing multi-source sensor information, so as to solve the problems in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a health indicator of a degraded device by fusing multi-source sensor information, comprising the following steps:

[0008] S1. Data selection and preprocessing: Obtain multi-source sensor signals from degraded equipment, perform maximum-minimum normalization on the data of each sensor, and unify the data from different sensors to the same scale;

[0009] S2. Time Domain Health Index Construction: Extract time domain statistical features from preprocessed multi-source sensor signals, and extract deep time domain features through a multi-head self-attention mechanism and neural network to generate time domain health indicators;

[0010] S3. Frequency Domain Health Index Construction: Perform continuous wavelet transform on the preprocessed multi-source sensor signals to generate a spectrum. Use convolutional neural networks to extract frequency domain features and generate frequency domain health indicators.

[0011] S4. Spatial Health Index Construction: Based on the linear correlation of sensor data in multi-source sensor signals, a graph structure is constructed. Graph convolutional neural networks are used to extract spatial features and generate spatial health indicators.

[0012] S5. Multi-domain feature fusion: Optimize and weight the time domain health index, frequency domain health index, and spatial health index to generate a composite health index;

[0013] S6. Life prediction: The composite health index is input into the life prediction model, and the model parameters are optimized by minimizing the error between the predicted value and the true value to achieve the prediction of the remaining service life of the equipment.

[0014] Optionally, the time domain statistical features in step S2 include mean, root mean square value, linear trend, root square amplitude, peak-to-peak value and standard deviation.

[0015] Optionally, the specific operation of the multi-head self-attention mechanism in step S2 is:

[0016] After embedding and position encoding the time domain statistical features, the multi-head attention weights are calculated by mapping the query matrix, key matrix and value matrix, and the time domain features are extracted by combining residual connection, layer normalization and feedforward neural network.

[0017] Optionally, after the continuous wavelet transform in step S3, the generated spectrum graph is subjected to bilinear interpolation dimensionality reduction processing, and local time-frequency block features and cross-band global features are extracted respectively through a two-layer convolutional neural network.

[0018] Optionally, the method for constructing the graph structure in step S4 is:

[0019] The Pearson correlation coefficient between multi-source sensor data is calculated. When the absolute value of the correlation coefficient is greater than the set threshold, the edge relationship between the sensor nodes is constructed to generate an adjacency matrix.

[0020] Optionally, the graph convolutional neural network in step S4 includes two GCN layers, where the first layer extracts first-order neighbor information of adjacent sensors, and the second layer extracts second-order neighbor information across sensors, and a Dropout layer is added between the layers to prevent overfitting.

[0021] Optionally, the specific method for optimizing weighted fusion in step S5 is:

[0022] The fusion weights of time domain, frequency domain and spatial health indicators are optimized through the back propagation algorithm to minimize the loss function of the life prediction model.

[0023] Optionally, the life prediction model is a random process model or a regression model based on deep learning.

[0024] Optionally, the multi-source sensing signals include vibration, temperature, pressure and current.

[0025] Optionally, the composite health indicator is used for predictive maintenance management of industrial equipment, specifically including remaining useful life prediction of aerospace engines, wind turbines or CNC machine tools.

[0026] Beneficial effects: 1. Improve the comprehensiveness and accuracy of health indicators:

[0027] This invention fuses multi-source sensor information across the spatial, temporal, and frequency domains to more comprehensively reflect the degradation status of a device. Traditional methods often rely solely on the characteristics of a single sensor or a single domain, making it difficult to capture comprehensive information about device degradation. This invention, however, fuses and processes multi-source information using deep learning algorithms to extract more accurate and comprehensive health indicators, thereby more precisely reflecting the degradation trends and status of the device.

[0028] 2. Enhance the time-varying trend and quality consistency of health indicators:

[0029] This invention takes the time-varying nature of equipment degradation into account when constructing health indicators. Time-domain analysis extracts long-term dependencies and degradation trends from time-series data, frequency-domain analysis captures the energy distribution of signals at different frequencies, and spatial analysis considers inter-sensor correlations. The resulting health indicator demonstrates strong time-varying trends and high-quality consistency. This helps more accurately predict the remaining useful life and maintenance needs of equipment.

[0030] 3. Improve the accuracy of equipment predictive maintenance and life prediction:

[0031] Using the health indicators constructed in this invention for predictive maintenance and lifespan prediction of equipment can significantly improve prediction accuracy. As a crucial input to the prediction model, the quality and comprehensiveness of health indicators directly impact the quality of the prediction results. By integrating high-quality health indicators from multiple sources, this invention provides a more reliable data foundation for the prediction model, thereby improving the accuracy and reliability of the predictions.

[0032] 4. Enhance the versatility and adaptability of the method:

[0033] The proposed method is not only applicable to specific types of industrial equipment but also possesses broad versatility and adaptability. By adjusting the parameters of the deep learning algorithm and the feature extraction method, it can adapt to the monitoring data and degradation characteristics of different types of equipment. Experimental results demonstrate that the proposed method achieves good results on various types of equipment and has broad application value.

[0034] 5. Promote intelligent and automated health management of industrial equipment:

[0035] The implementation of this invention will help advance the intelligent and automated health management of industrial equipment. By building high-precision health indicators and predictive models, real-time monitoring and early warning of equipment status can be achieved, reducing manual intervention and downtime, and improving equipment operational efficiency and safety. Furthermore, this method can provide a scientific basis for the development of equipment maintenance plans, optimizing maintenance resources and costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 Schematic diagram of six time-domain statistical characteristics of a device according to an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of a spectrum diagram of a device failure moment according to an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of original multi-source information and composite health indicators of engine 90 and engine 98 in the FD001 dataset according to an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the original multi-source information and composite health indicators of engine 83 and engine 87 in the FD003 data set according to an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of the process of an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following describes preferred embodiments of the present invention with reference to the accompanying drawings to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0042] The specific implementation of the present invention is as follows:

[0043] Data selection and preprocessing process

[0044] Considering the multi-source sensor signal U of the degraded equipment i ∈R T×Q :

[0045]

[0046] in, is the monitoring data of the jth sensor of the i-th degraded device at the k-th moment.

[0047] Perform maximum-minimum normalization on the data of each sensor to convert the data of different sensors to the same scale:

[0048]

[0049] in, is the minimum value of the jth sensor at all monitoring moments on all degraded devices, is the maximum value of the jth sensor at all monitoring moments on all degraded devices.

[0050] Time domain health indicator construction

[0051] During the operation of the equipment, the time series data collected by the sensor contains the basic information of equipment degradation. First, based on the statistical principles, some statistics describing the data distribution and trend are extracted from the original data. Figure 1 As shown in the figure, for the input multi-source signal, this paper selects six time domain statistical features: mean, root mean square value, linear trend, root square amplitude, peak-to-peak value and standard deviation:

[0052] Mean: RMS: Linear Trend: Trend = -10.a

[0053] Square root amplitude:

[0054] Peak-to-peak value:

[0055]

[0056] Standard Deviation: At this time, the information of the i-th degraded device becomes:

[0057] Taking the data at the kth moment as an example, for z i,k Perform embedding and positional encoding

[0058]

[0059] Perform multi-head self-attention calculation on the encoded sequence to further extract

[0060] Domain characteristics: Q = XW Q K=XW K V=XW V

[0061]

[0062] Head i =Attention(Q i ,K i ,V i ),i=1,…,n

[0063] MultiHead(Q,K,V)=Concat(head1,…,head n )W 0

[0064] Among them, d k Represents the dimension of the input sequence, Q, K, V represent the mapping vectors of the input sequence on the query matrix, key matrix and value matrix respectively, and the dimension is d k , n represents the number of attention heads, W 0 is the weight matrix.

[0065] Residual connections and layer normalization:

[0066] X atten =LayerNorm(X+Dropout(MultiHead(Q,K,V)))

[0067] Enter the feedforward neural network:

[0068] FFN(X atten)=RELU(X atten W1+b1)W2+b2

[0069] in

[0070] Residual connection and layer normalization again:

[0071] X FFN =LayerNorm(X atten +Dropout(FFN(X atten )))

[0072] Perform mean pooling and full connection operations in the time domain feature dimension:

[0073]

[0074] W f +b

[0075] in, b fc ∈R 1 .

[0076] The time domain health indicator of the i-th device is expressed as:

[0077]

[0078] Frequency domain health indicator construction

[0079] Frequency domain features reflect the energy distribution information of the signal in different time-frequency ranges. First, continuous wavelet transform (CWT) is used to process the original multi-source signal. Consider the input data Z i ∈R T×Q :

[0080]

[0081] Among them, s (t) is the wavelet basis function under scale S, and the spectrum matrix X CWT ∈R T×S×Q .

[0082] Perform bilinear interpolation dimensionality reduction on the spectrum graph to retain key time-frequency features:

[0083]

[0084] Among them, w i,j is the interpolation weight, X resize ∈R T×Z×Z .

[0085] In order to further extract deep features, the spectrum graph after dimension reduction is continuously subjected to two layers of convolution for frequency domain feature extraction.

[0086] First convolution layer:

[0087] Xconv1=ReLU(W*Xresize+b)

[0088] Among them, W is the convolution kernel parameter, Xconv1∈RT×C×Z×Z, and C is the number of convolution kernels.

[0089] Second convolution layer:

[0090] Among them, Xconv 2∈RT×D×Z×Z.

[0091] The first convolution layer learns the amplitude distribution characteristics of local time-frequency blocks, and the second layer integrates cross-band information through a larger receptive field to identify residual periodic patterns or energy distribution laws in the degraded signal.

[0092] Flatten layer:

[0093] X flatten =reshape(X conv2 ,(1,D×Z×Z))

[0094] Fully connected layer:

[0095]

[0096] Among them, W fc ∈R( D×Z×Z)×1 , b fc ∈R T×1 .

[0097] Construction of spatial health indicators

[0098] Graph structure construction:

[0099] In multi-source sensor data, there is often a certain correlation between sensors. Consider the input data Z i ∈R T ×Q , use the Pearson correlation coefficient to calculate the linear correlation of each dimension in the multidimensional time data composed of multi-source sensor monitoring data, and construct the graph structure. The formula is as follows:

[0100]

[0101] Among them, M i and M j are two columns of multidimensional time data, σ i and σ j Mi and M j The standard deviation of i and μ j are the average values ​​of the two columns of data, and E represents the expected value. When the absolute value of the Pearson correlation coefficient is greater than or equal to the set value, there is an edge relationship between the two nodes (sensor data), and thus the edge relationship is determined and the adjacency matrix A is obtained.

[0102] After the graph structure is constructed, the sensor data is input as node features into a two-layer graph convolutional neural network for spatial feature extraction. The first layer GCN:

[0103]

[0104] Where D is the degree matrix, W 0 ∈R Q×H is the weight matrix of the graph convolutional neural network, X GCN ∈R T×H To prevent overfitting, a Dropout layer is added between the first and second GCN layers:

[0105]

[0106] Second layer GCN network:

[0107]

[0108] Among them, W 1 ∈R H×O ,X GCN2 ∈R T×O .

[0109] The first layer of GCN captures the local correlation between adjacent sensor data (first-order neighbor information), and the second layer of GCN further integrates local information from different nodes, performs deeper information integration, and extracts global high-level features across sensors (second-order neighbor information).

[0110] Fully connected layer:

[0111]

[0112] Among them, W fc ∈R O×1 ,b fc ∈R T×1 .

[0113] Multi-domain feature fusion

[0114] After obtaining the health indicators in the spatial, time, and frequency domains, we optimize and integrate the three to obtain the final composite health indicator:

[0115]

[0116] Where w = [w1 w2 w3] represents the fusion weights of the time domain health index, frequency domain health index, and spatial health index, respectively.

[0117] The constructed health indicators are then input into the life prediction model (random process model or deep learning model) to predict the equipment life, and the following optimization objective function is constructed:

[0118]

[0119] in, is the predicted value of the life of the ith equipment, T i is the true value of the life of the i-th device. s ,W t ,W f ,W respectively represent the neural network parameters of spatial health index, time domain health index and frequency domain health index and the fusion weights of time domain health index, frequency domain health index and spatial health index.

[0120] By minimizing the objective function, the above parameters can be optimized in reverse until the objective function converges, thus obtaining the optimized weight coefficients.

[0121] The technical solution of the present invention is described in detail below with reference to the embodiments.

[0122] Example 1

[0123] Data selection and preprocessing

[0124] Consider the degradation monitoring data for a certain engine model, which includes values ​​from multiple sensors. These sensors, such as temperature, pressure, and vibration sensors, monitor different engine components and performance indicators. For each sensor's data, missing value processing and outlier detection are first performed, followed by data preprocessing using the maximum-minimum normalization method. The normalization formula is as follows:

[0125] ^x_i^norm=\frac{x_i-x_{min}}{x_{max}-x_{min}}

[0126] Among them, x i represents the raw data of the i-th sensor, x min and x max represent the minimum and maximum values ​​of the sensor at all monitoring moments on all degraded devices, and ^x_i^norm represents the normalized data.

[0127] Time domain health indicator construction

[0128] Six time-domain statistical features are extracted from the preprocessed data: mean, root mean square value, linear trend, root mean square amplitude, peak-to-peak value, and standard deviation. These features describe the distribution and trend of the data. The extracted time-domain features are then further processed using a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism captures correlations between different locations in the time series data, while the feedforward neural network extracts deeper nonlinear features. Finally, through mean pooling and fully connected operations, the engine's time-domain health indicator is obtained.

[0129] Frequency domain health indicator construction

[0130] The original multi-source signal is processed using a continuous wavelet transform to obtain a spectrum matrix. The spectrum is then subjected to bilinear interpolation dimensionality reduction to preserve key time-frequency features. A two-layer convolutional neural network is then used to extract frequency-domain features from the reduced spectrum. The first convolution layer learns the amplitude distribution characteristics of local time-frequency blocks, while the second convolution layer integrates cross-band information through a larger receptive field to identify residual periodic patterns or energy distribution patterns in the degraded signal. Finally, through the Flatten layer and the fully connected layer, the engine's frequency-domain health indicators are obtained.

[0131] Construction of spatial health indicators

[0132] The Pearson correlation coefficient is used to calculate the linear correlation of each dimension in the multidimensional temporal data composed of multi-source sensor monitoring data, and a graph structure is constructed. The sensor data is then input as node features into a two-layer graph convolutional neural network for spatial feature extraction. The first layer of the GCN captures the local correlation between adjacent sensor data (first-order neighbor information). The second layer of the GCN further integrates local information from different nodes, performing deeper information integration and extracting global high-level features across sensors (second-order neighbor information). Finally, through the fully connected layer, the engine's spatial health indicators are obtained.

[0133] Multi-domain feature fusion

[0134] After obtaining health indicators in the spatial, time, and frequency domains, an optimization fusion strategy is employed to integrate these three. During the fusion process, a corresponding fusion weight is assigned to each health indicator. These weights are optimized by constructing and minimizing an optimization objective function. This optimization objective function considers the accuracy of the equipment life prediction, specifically minimizing the error between the predicted and actual lifespan. A reverse optimization algorithm is used to iteratively update the parameters until the objective function converges, ultimately yielding the optimized weight coefficients and the final composite health indicator.

[0135] Experimental results analysis

[0136] Experimental validation was conducted using data from actual engines. The results demonstrate that the proposed method for constructing a degraded equipment health index by integrating multi-source sensor information significantly improves time-varying trends and quality consistency. Comparing the raw multi-source information and composite health indicators from different engines clearly demonstrates the superiority of the proposed method in reflecting equipment degradation trends. Furthermore, the constructed health index also achieved good results in predicting equipment lifespan.

[0137] Example 2

[0138] In order to further verify the effectiveness, versatility and adaptability of the present invention, the present invention selected different types of degraded equipment data for experiments based on Example 1.

[0139] 1. Data selection and preprocessing

[0140] Consider another type of industrial equipment, such as a wind turbine or petrochemical plant. Its degradation monitoring data also includes values ​​from multiple sensors. These sensors may monitor various performance indicators, such as temperature, pressure, vibration, and flow. For each sensor's data, missing value processing and outlier detection are performed, and data preprocessing using the maximum-minimum normalization method ensures data accuracy and consistency.

[0141] 2. Construction of time domain health indicators

[0142] Similar to Example 1, time-domain statistical features, such as mean, root mean square value, and linear trend, are extracted from the preprocessed data. These extracted time-domain features are then further processed using deep learning algorithms (such as long short-term memory (LSTM) networks or gated recurrent units (GRU)) to capture long-term dependencies and degradation trends in the time series data. Finally, through appropriate pooling operations and fully connected layers, the device's time-domain health indicators are obtained.

[0143] 3. Construction of frequency domain health indicators

[0144] The same frequency domain feature extraction method as in Example 1 was used, namely, continuous wavelet transform (CWT) was used to process the original multi-source signal to obtain a spectrum matrix. The spectrum graph was then subjected to dimensionality reduction, and frequency domain feature extraction was performed using a convolutional neural network. The number and parameters of the convolutional layers were adjusted to accommodate the frequency domain characteristics of different types of devices. Finally, the frequency domain health indicators of the device were obtained through the Flatten layer and the fully connected layer.

[0145] 4. Construction of spatial health indicators

[0146] Similar to Example 1, the Pearson correlation coefficient is used to calculate the linear correlation between multi-source sensor monitoring data and construct a graph structure. The sensor data is then input as node features into a graph convolutional neural network for spatial feature extraction. The number of GCN layers and parameters are adjusted to accommodate the spatial characteristics of different types of devices. Finally, a fully connected layer is used to obtain the device's spatial health indicator.

[0147] 5. Multi-domain feature fusion

[0148] After obtaining the health indicators in the spatial, time, and frequency domains, the three are fused using the same optimization fusion strategy as in Example 1. During the fusion process, a corresponding fusion weight is assigned to each health indicator, and these weights are optimized by constructing an optimization objective function. The optimization objective function can consider multiple aspects such as the accuracy of equipment life prediction, the time-varying trend of health indicators, and quality consistency. The parameters are iteratively updated through the reverse optimization algorithm until the objective function converges, ultimately obtaining the optimized weight coefficients and the final composite health indicator.

[0149] In summary, the method proposed in this paper provides detailed and specific implementation details and solutions for data preprocessing, time-domain feature extraction, frequency-domain feature extraction, spatial feature extraction, multi-domain feature fusion, and experimental verification. This method can comprehensively and accurately reflect the degradation status of equipment, providing strong support for predictive maintenance and lifespan prediction.

[0150] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, from all points of view, the embodiments should be regarded as illustrative and non-restrictive. The scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes that fall within the meaning and range of equivalents of the claims are included in the present invention. Any reference signs in the claims should not be construed as limiting the claim to which they relate.

[0151] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for constructing a health indicator of degraded equipment by integrating multi-source sensor information, characterized in that: The following steps are involved: S1. Data selection and preprocessing: Obtain multi-source sensor signals from degraded equipment, perform maximum-minimum normalization on the data of each sensor, and unify the data from different sensors to the same scale; S2. Time Domain Health Index Construction: Extract time domain statistical features from preprocessed multi-source sensor signals, and extract deep time domain features through a multi-head self-attention mechanism and neural network to generate time domain health indicators; S3. Frequency Domain Health Index Construction: Perform continuous wavelet transform on the preprocessed multi-source sensor signals to generate a spectrum. Use convolutional neural networks to extract frequency domain features and generate frequency domain health indicators. S4. Spatial Health Index Construction: Based on the linear correlation of sensor data in multi-source sensor signals, a graph structure is constructed. Graph convolutional neural networks are used to extract spatial features and generate spatial health indicators. S5. Multi-domain feature fusion: Optimize and weight the time domain health index, frequency domain health index, and spatial health index to generate a composite health index; S6. Life prediction: The composite health index is input into the life prediction model, and the model parameters are optimized by minimizing the error between the predicted value and the true value to achieve the prediction of the remaining service life of the equipment.

2. The method according to claim 1, characterized in that The time domain statistical features in step S2 include mean, root mean square value, linear trend, root square amplitude, peak-to-peak value and standard deviation.

3. The method according to claim 1, characterized in that The specific operation of the multi-head self-attention mechanism in step S2 is: After embedding and position encoding the time domain statistical features, the multi-head attention weights are calculated by mapping the query matrix, key matrix and value matrix, and the time domain features are extracted by combining residual connection, layer normalization and feedforward neural network.

4. The method according to claim 1, wherein After the continuous wavelet transform in step S3, the generated spectrum graph is subjected to bilinear interpolation dimensionality reduction processing, and the local time-frequency block features and the cross-band global features are extracted respectively through a two-layer convolutional neural network.

5. The method according to claim 1, wherein The method for constructing the graph structure in step S4 is: The Pearson correlation coefficient between multi-source sensor data is calculated. When the absolute value of the correlation coefficient is greater than the set threshold, the edge relationship between the sensor nodes is constructed to generate an adjacency matrix.

6. The method according to claim 1, wherein The graph convolutional neural network in step S4 includes two GCN layers, where the first layer extracts the first-order neighbor information of adjacent sensors, and the second layer extracts the second-order neighbor information across sensors, and a Dropout layer is added between the layers to prevent overfitting.

7. The method according to claim 1, characterized in that The specific method for optimizing weighted fusion in step S5 is: The fusion weights of time domain, frequency domain and spatial health indicators are optimized through the back propagation algorithm to minimize the loss function of the life prediction model.

8. The method according to claim 1, characterized in that The life prediction model is a random process model or a regression model based on deep learning.

9. The method according to claim 1, characterized in that The multi-source sensor signals include vibration, temperature, pressure and current.

10. The method according to claim 1, characterized in that The composite health indicator is used for predictive maintenance management of industrial equipment, specifically including the remaining service life prediction of aerospace engines, wind turbines or CNC machine tools.

Images