A rotating machinery residual service life prediction method based on GBIT-GAN

Abstract

The application provides a rotating machinery residual service life prediction method based on GBIT-GAN. The method combines a graph neural network GNN and a bidirectional Transformer generative adversarial network BIT-GAN, can fully utilize the structure information of the rotating machinery to construct a graph structure, and the GNN layer can accurately capture the topological relationship between components. The bidirectional Transformer performs well when processing time series data, can well mine the long-term dependence relationship in the data, and the adversarial training mechanism of the generator and the discriminator further improves the accuracy and reliability of the model. The method has obvious advantages in the residual service life prediction of the rotating machinery, can provide more accurate and reliable decision basis for the maintenance and management of the rotating machinery in industrial production, greatly reduces the equipment failure risk, and has important significance for improving the industrial production efficiency and economic benefits.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery condition monitoring and life prediction, and in particular relates to a method for predicting the remaining service life of rotating machinery based on GBIT-GAN. Background Art

[0002] In industrial production, the stable operation of rotating machinery, such as bearings and aircraft engines, is crucial. For example, bearings are used in large numbers in various mechanical transmissions on automobile production lines. A bearing failure can disrupt the operation of machinery on the production line, ultimately halting the entire production process. This not only means that parts being processed may be scrapped, resulting in direct material loss, but also significantly reduces production efficiency, delays product delivery, and triggers a series of chain reactions, such as contract damages and other financial losses. Aircraft engines are a key power source for low-altitude flight operations. For example, in low-altitude mapping and geographic exploration, specialized surveying aircraft equipped with high-precision mapping instruments rely on the stable operation of aircraft engines to collect data over large areas at low altitudes. An in-flight engine failure not only disrupts ongoing mapping and exploration missions, wasting the manpower, material, and time invested initially, but also creates incomplete data that can affect the subsequent development and analysis of geographic information systems, hindering the smooth progress of projects such as urban planning and resource exploration. In addition, in some scientific research projects, such as long-term low-altitude monitoring of the ecological environment of a specific area, engine failure may cause the loss of data during critical periods, affecting the accuracy and reliability of scientific research results, and thus delaying research progress in related fields.

[0003] The purpose of predicting the remaining useful life of rotating machinery is to proactively assess the health of the equipment, enabling the rational planning of maintenance and replacement. This is crucial because it effectively prevents production downtime, safety incidents, and economic losses caused by sudden equipment failures, ensuring the continuity and stability of industrial production, improving production efficiency, reducing operating costs, and ensuring personnel safety and sustainable business development. Accurate predictions enable companies to take timely measures before equipment failures occur, such as stockpiling spare parts and arranging professional maintenance personnel, minimizing the impact of equipment failures.

[0004] In the field of rotating machinery remaining useful life prediction, early deep learning models, such as the multi-layer perceptron (MLP), were relatively simple. Composed of an input layer, hidden layers, and an output layer, the MLP processes and transforms input data through the connection weights between neurons. However, it suffers from significant drawbacks when processing rotating machinery data. Due to its structural limitations, it struggles to capture the time series characteristics and spatial structure of the data. Faced with the complex operational state changes of rotating machinery, it is unable to effectively mine potential dependencies from long-term time series data, and its utilization of equipment structural information is almost non-existent.

[0005] While technological advancements have led to the emergence of more complex models, existing models still fall short when it comes to predicting rotating machinery. For example, recurrent neural networks (RNNs) and their variants, while capable of processing time series data to a certain extent, suffer from vanishing or exploding gradients, making them less than ideal for capturing long-term dependencies. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art, such as insufficient feature extraction of rotating machinery data, insufficient model generalization ability, and ineffective utilization of component topological relationships, so as to achieve more accurate and reliable prediction of the remaining useful life of rotating machinery. A method for predicting the remaining useful life of rotating machinery based on GBIT-GAN is proposed.

[0007] The present invention is implemented through the following technical solutions. The present invention proposes a method for predicting the remaining service life of rotating machinery based on GBIT-GAN. The method is specifically as follows:

[0008] Data acquisition and preprocessing: Utilizing acceleration, temperature, and pressure sensors installed on rotating machinery, we collect multi-channel vibration, temperature, and pressure data, covering the entire equipment lifecycle from normal operation to failure. We filter to remove outliers and noise, then normalize the data to fit within the [0, 1] range. The data is then segmented into specific time windows and labeled with remaining useful life.

[0009] Constructing a GBIT-GAN model: The generator is built based on a bidirectional Transformer. The Transformer's encoder and decoder are both composed of multiple Transformer layers, each of which contains a multi-head self-attention mechanism, a feedforward neural network, residual connections, and layer normalization. The generator can effectively extract data features, input historical data and equipment information, output predicted values, and optimize through adversarial and reconstruction losses. The discriminator uses a structure that processes sequence data to distinguish the authenticity of data. A GNN layer is introduced after the generator's bidirectional Transformer encoder to capture the topological relationship and information propagation between rotating machinery components. The output of the GNN layer is concatenated or weighted summed with the output of the bidirectional Transformer encoder to perform information fusion. The fused features are then input into the bidirectional Transformer decoder to generate predicted future sensor data sequences and remaining service life prediction values.

[0010] Model training: divide the training data set and use the training data set to iteratively train the GBIT-GAN model;

[0011] Prediction of the remaining useful life of rotating machinery: Newly collected data is preprocessed and input into the trained GBIT-GAN model. The predicted results are compared with a preset threshold. If the predicted remaining useful life is lower than the threshold, an early warning signal is issued, prompting maintenance personnel to inspect and maintain the rotating machinery to avoid failures.

[0012] Furthermore, the pre-processed data is segmented according to a certain time window to form multiple data samples; each data sample contains multi-channel sensor data within a period of time and the corresponding remaining service life label; the size of the time window is selected according to the operating characteristics of the rotating machinery and the sampling frequency of the data, and the remaining service life label y i Calculation is based on the actual operating time and failure time of the rotating machinery.

[0013] Furthermore, in the Transformer layer, for the input sequence X = [x1, x2, ..., x n ], where x i The i-th element in the sequence is the preprocessed data sample, which is mapped to query Q, key K and value V through linear transformation, that is, Q = XW Q ,K=XW K ,V=XW V , where W Q ,W K ,W V is a learnable weight matrix; then calculate the attention weight where d kis the dimension of the key, and the softmax function is used to normalize the attention weights to [0,1] so that the sum of the weights is 1; finally, the output of the multi-head self-attention is MultiHead(X)=Concat(head1,head2,…,head h ,)W O , where head i =Attention(QW i Q ,KW i K ,VW i V ), h is the number of heads, W i Q ,W i K ,W i V and W O It is a learnable parameter matrix; the output of the multi-head self-attention is further transformed by the feedforward neural network. The feedforward neural network consists of two fully connected layers, with the ReLU activation function in the middle. The formula is FFN(x)=ReLU(xW1+b1)W2+b2, where W1 and W2 are weight matrices, and b1 and b2 are bias vectors.

[0014] Furthermore, the input of the generator includes historical sensor data sequences and basic information of rotating machinery. This basic information is converted into vector representation through the embedding layer, and then spliced ​​with the historical sensor data sequence and input into the encoder of the bidirectional Transformer; the output of the generator is the predicted sensor data sequence of the next n time windows and the corresponding remaining service life prediction value.

[0015] Furthermore, adversarial loss: The goal of the generator is to make the discriminator unable to distinguish between the generated sequence and the real sequence. The adversarial loss of the generator is expressed as where p g is the data distribution generated by the generator, and D(x) is the discriminator's judgment result on the input x;

[0016] Reconstruction loss: Using mean square error MSE as reconstruction loss, then where p data is the real data distribution, x is the real sequence, is the sequence generated by the generator;

[0017] Total loss: The total loss of the generator is the weighted sum of the adversarial loss and the reconstruction loss, i.e. Where α is a weight parameter used to balance the importance of adversarial loss and reconstruction loss.

[0018] Furthermore, the loss function in the discriminator is: The goal of the discriminator is to accurately distinguish between real sequences and generated sequences, and its loss function is expressed as

[0019] Furthermore, the GNN layer adopts the graph convolutional network GCN. For node i, its feature update formula at the l+1 layer is: in is the feature vector of node i in layer l, is the feature vector of node j in layer l, N(i) is the set of neighbor nodes of node i, c ij is the normalization constant, W l is the learnable weight matrix of layer l, b l is the bias vector, and σ is the activation function. By stacking multiple GNN layers, the features of the nodes can be continuously updated, thereby fusing the topological information and sensor data information between components.

[0020] Furthermore, in each training iteration, the discriminator parameters are first fixed, and a batch of data is sampled from the training set and input into the generator to generate a batch of predicted future sensor data sequences and remaining service life prediction values; the generated sequences and the real sequences are input into the discriminator, and the loss of the generator and the loss of the discriminator are calculated; the backpropagation algorithm is used to calculate the gradients of the generator, discriminator and GNN layer parameters respectively, and the model parameters are updated according to the gradients to minimize the loss of the generator and maximize the loss of the discriminator; after each training cycle, the model is evaluated using the validation set, and the loss function value on the validation set and the error index between the predicted remaining service life and the actual remaining service life are calculated; according to the evaluation results of the validation set, the hyperparameters of the model are adjusted. If the performance on the validation set no longer improves, the training can be stopped early to prevent overfitting.

[0021] The present invention also proposes an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting the remaining useful life of rotating machinery based on GBIT-GAN are implemented.

[0022] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the method for predicting the remaining useful life of rotating machinery based on GBIT-GAN.

[0023] Beneficial effects of the present invention:

[0024] (1) By combining the bidirectional Transformer and GNN layers, the present invention can fully extract the temporal features and topological relationship features between components in the rotating machinery sensor data, improving the model's data representation and feature extraction capabilities, thereby more accurately predicting the remaining service life of the rotating machinery;

[0025] (2) The adversarial training mechanism in the improved BITGAN model, namely the GBIT-GAN model, enables the generator to generate more realistic future sensor data sequences, which helps to capture the potential distribution and change trends of the data, further improving the accuracy and reliability of the remaining useful life prediction;

[0026] (3) The present invention can monitor the operating status of rotating machinery in real time and issue early warnings in a timely manner, providing strong support for the maintenance and management of equipment, helping to reduce losses caused by equipment failures and improve production efficiency and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0028] Figure 1 This is the network architecture diagram of GBIT-GAN.

[0029] Figure 2 Schematic diagram of the GNN graph structure.

[0030] Figure 3 Schematic diagram of the Bi-Transformer structure. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] Combine Figure 1-Figure 3 The present invention proposes a method for predicting the remaining service life of rotating machinery based on GBIT-GAN, and the method is specifically as follows:

[0033] Data acquisition and preprocessing: Utilize acceleration, temperature, and pressure sensors installed on rotating machinery to collect multi-channel vibration, temperature, and pressure data. The collected data should include a sufficiently long time series to cover the entire process of the rotating machinery from normal operation to failure. Filtering is used to remove outliers and noise, and the data is normalized to the [0, 1] range. The data is then segmented into specific time windows and labeled with the remaining useful life.

[0034] Data preprocessing:

[0035] Data cleaning: Remove outliers and noise from the collected data. For example, use methods such as mean filtering and median filtering to smooth the vibration signal and remove abnormal data points caused by factors such as sensor failure or environmental interference.

[0036] Data normalization: Normalize the data collected by different sensors so that their value range is between [0,1] to facilitate model training and processing. Common normalization methods include minimum-maximum normalization, the formula is: norm =(xx min ) / (x max -x min ), where x is the original data, x min and x max are the minimum and maximum values ​​in the data sequence, respectively, norm The data are normalized.

[0037] Data segmentation: The pre-processed data is segmented according to a certain time window to form multiple data samples; each data sample contains multi-channel sensor data within a period of time and the corresponding remaining service life label; the size of the time window is selected according to the operating characteristics of the rotating machinery and the sampling frequency of the data. For example, the window size can be selected as W and the step size can be selected as S, then the i-th data sample can be expressed as X i =[x i,1 ,x i,2 ,…,x i,W ], where X i Indicates the xth i,j The sensor data vector of the i-th time step in samples, j = 1, 2, ..., W. The remaining useful life label y i The calculation is based on the actual operating time and failure time of the rotating machinery. For example, if the rotating machinery fails after the operating time T, and the time range corresponding to the current data sample is from t1 to t2, then the remaining service life label y of the sample is i =T-(t2-t1).

[0038] Constructing a GBIT-GAN model: The generator is built based on a bidirectional Transformer. The encoder and decoder of the Transformer are composed of multiple Transformer layers. Each Transformer layer contains a multi-head self-attention mechanism, a feed-forward neural network, a residual connection, and layer normalization. The generator can effectively extract data features, input historical data and equipment information, output predicted values, and optimize through adversarial and reconstruction losses. The discriminator uses a structure that processes sequence data to distinguish the authenticity of data. A GNN layer is introduced after the bidirectional Transformer encoder of the generator to capture the topological relationship and information propagation between rotating machinery components. The output of the GNN layer is concatenated or weighted summed with the output of the bidirectional Transformer encoder to perform information fusion. The fused features are then input into the decoder of the bidirectional Transformer to generate predicted future sensor data sequences and remaining service life prediction values.

[0039] In the Transformer layer, for the input sequence X = [x1, x2, ..., x n ], where x i The i-th element in the sequence is the preprocessed data sample, which is mapped to query Q, key K and value V through linear transformation, that is, Q = XW Q ,K=XW K ,V=XW V , where W Q ,W K ,W V is a learnable weight matrix; then calculate the attention weight where d k is the dimension of the key, and the softmax function is used to normalize the attention weights to [0,1] so that the sum of the weights is 1; finally, the output of the multi-head self-attention is MultiHead(X)=Concat(head1,head2,…,head h ,)W O , where head i =Attention(QW i Q ,KW i K ,VW i V ), h is the number of heads, W iQ ,W i K ,W i V and W O is a learnable parameter matrix. The output of the multi-head self-attention network undergoes further feature transformation through a feedforward neural network consisting of two fully connected layers with a ReLU activation function in between. The formula is FFN(x) = ReLU(xW1+b1)W2+b2, where W1 and W2 are weight matrices and b1 and b2 are bias vectors. To avoid vanishing and exploding gradients, residual connections and layer normalization are used in each Transformer layer, namely LayerNorm(x+Sublayer(x)), where Sublayer(x) represents the output of the multi-head self-attention or feedforward neural network, and LayerNorm represents the layer normalization operation. The input to the generator consists of a historical sensor data sequence (e.g., data from the past n time windows) and basic information about the rotating machinery (such as model, speed, and load). This basic information is converted into a vector representation through an embedding layer, then concatenated with the historical sensor data sequence and input into the encoder of the bidirectional Transformer. The output of the generator is the predicted sensor data sequence for the next n time windows and the corresponding predicted remaining useful life.

[0040] Adversarial Loss: The goal of the generator is to make the discriminator unable to distinguish between the generated sequence and the real sequence. The adversarial loss of the generator is expressed as where p g is the data distribution generated by the generator, and D(x) is the discriminator's judgment result on the input x (the probability of being true);

[0041] Reconstruction Loss: In order to make the generated sequence closer to the real sequence in some features, reconstruction loss can also be added. Using mean square error MSE as reconstruction loss, then where p data is the real data distribution, x is the real sequence, is the sequence generated by the generator;

[0042] Total loss: The total loss of the generator is the weighted sum of the adversarial loss and the reconstruction loss, i.e. Where α is a weight parameter used to balance the importance of adversarial loss and reconstruction loss.

[0043] Discriminator: The discriminator can use a convolutional neural network (CNN) or other architecture suitable for processing sequence data. It receives an input sequence (real or generated) and outputs a discriminant score, which represents the probability that the input sequence is real. The discriminator's network structure can include convolutional layers, pooling layers, and fully connected layers. The specific number of layers and parameters can be adjusted according to actual conditions.

[0044] The loss function in the discriminator is: The goal of the discriminator is to accurately distinguish between real sequences and generated sequences, and its loss function is expressed as

[0045] Introducing the GNN layer:

[0046] Graph construction: Construct a graph based on the physical structure of the rotating machinery and the connection relationship between the components. The nodes in the constructed graph represent the various components of the rotating machinery (such as bearings, gears, shafts, etc.). The characteristics of the nodes may include information such as the type, size, material, and the corresponding sensor data of the component (such as the vibration signal and temperature signal collected by the sensor installed near the component). The edge represents the connection relationship between the components, and the weight of the edge can be set according to factors such as the physical distance between the components and the connection strength. For example, the adjacency matrix can be used to represent the structure of the graph, where A ij Indicates whether there is an edge connection between node i and node j. If there is an edge connection, then A ij =1, otherwise A ij =0.

[0047] GNN layer: A GNN layer is introduced after the bidirectional Transformer encoder of the generator to capture the topological relationships and information propagation between rotating machinery components. The GNN layer can use a graph convolutional network (GCN) or other types of GNN models. Taking GCN as an example, for node i, its feature update formula at layer l+1 is: in is the feature vector of node i in layer l, is the feature vector of node j in layer l, N(i) is the set of neighbor nodes of node i, c ij is a normalization constant (e.g. where d i and d j are the node and the degree of the node respectively), W l is the learnable weight matrix of layer l, b l is the bias vector, σ is the activation function (such as ReLU); by stacking multiple layers of GNN layers, the features of the nodes can be continuously updated, thereby fusing the topological information and sensor data information between components.

[0048] Information fusion: The output of the GNN layer is fused with the output of the bidirectional Transformer encoder by concatenating or weighted summing the information. The fused features are then input into the decoder of the bidirectional Transformer to generate the predicted future sensor data sequence and the remaining service life prediction value.

[0049] Model training: The training dataset is divided and the GBIT-GAN model is iteratively trained using the training dataset. This includes:

[0050] Training data set partitioning: Divide the preprocessed data samples into training set, validation set, and test set according to a certain ratio. For example, 70% of the data can be used as the training set, 15% as the validation set, and 15% as the test set.

[0051] Initialize model parameters: Randomly initialize the parameters of the generator, discriminator, and GNN layers.

[0052] Iterative training: In each training iteration, the discriminator parameters are first fixed, and a batch of data is sampled from the training set and input into the generator to generate a batch of predicted future sensor data sequences and remaining service life prediction values; the generated sequences and the actual sequences are input into the discriminator, and the generator loss and the discriminator loss are calculated; the backpropagation algorithm is used to calculate the gradients of the generator, discriminator and GNN layer parameters respectively, and the model parameters are updated according to the gradients to minimize the generator loss and maximize the discriminator loss; after each training cycle, the model is evaluated using the validation set, and the loss function value on the validation set (such as the total loss of the generator, the loss of the discriminator, etc.) and the error indicators between the predicted remaining service life and the actual remaining service life (such as the root mean square error RMSE, the mean absolute error MAE, etc.) are calculated; according to the evaluation results of the validation set, the model hyperparameters (such as the learning rate, the weight parameter α of the generator and discriminator losses, the number of layers and parameters of the GNN layer, etc.) are adjusted. If the performance on the validation set no longer improves, the training can be stopped early to prevent overfitting.

[0053] Model optimization: Some optimization techniques can be used to improve the training efficiency and performance of the model, such as using a learning rate decay strategy (such as gradually reducing the learning rate as training progresses), gradient clipping (to prevent gradient explosion), batch normalization (to accelerate model convergence), etc.

[0054] Prediction of the remaining useful life of rotating machinery: Newly collected data is preprocessed and input into the trained GBIT-GAN model. The predicted results are compared with a preset threshold. If the predicted remaining useful life is lower than the threshold, an early warning signal is issued, prompting maintenance personnel to inspect and maintain the rotating machinery to avoid failures.

[0055] Online prediction: When the trained model is deployed in an actual rotating machinery monitoring system, the newly collected real-time sensor data is first cleaned, normalized, and windowed according to the data preprocessing steps to obtain the input data samples at the current moment.

[0056] Model inference: The input data sample at the current moment is input into the trained model, and after being processed by the generator's bidirectional Transformer and GNN layers, the predicted future sensor data sequence and remaining service life prediction value are obtained.

[0057] Result output and early warning: The predicted remaining service life is compared with the preset threshold. If the predicted remaining service life is lower than the threshold, an early warning signal is issued to prompt maintenance personnel to inspect and maintain the rotating machinery to avoid failures.

[0058] The GBIT-GAN-based remaining service life prediction method for rotating machinery proposed in the present invention effectively overcomes the problems in the prior art. By combining the graph neural network (GNN) with the bidirectional Transformer generative adversarial network (BIT-GAN), the structural information of the rotating machinery can be fully utilized to construct a graph structure, and the GNN layer can accurately capture the topological relationship between components. The bidirectional Transformer performs well in processing time series data and can well mine the long-term dependencies in the data. The adversarial training mechanism of the generator and the discriminator further improves the accuracy and reliability of the model. This innovative fusion architecture makes the method described in the present invention have significant advantages in the prediction of the remaining service life of rotating machinery, and can provide a more accurate and reliable decision-making basis for the maintenance and management of rotating machinery in industrial production, greatly reduce the risk of equipment failure, and ensure the safe and stable operation of production, which is of great significance to improving industrial production efficiency and economic benefits.

[0059] Example: Prediction of Remaining Bearing Service Life

[0060] To monitor bearings on large rotating equipment at a factory, vibration, temperature, and pressure sensors were installed at key locations on the bearings. Four months of data were collected, and 1,200 valid samples were obtained after screening and sorting. During data preprocessing, median filtering was used to remove noise. The data was mapped to the [0, 1] interval using min-max normalization. The data was segmented using a 30-minute time window and a 15-minute step size. Remaining useful life labels were calculated based on equipment operating and failure times. An improved BITGAN model was constructed, with a generator consisting of a three-layer bidirectional Transformer encoder and decoder, a four-head multi-head self-attention mechanism, and a feedforward neural network with a specific architecture. The input consisted of data from the past 12 hours and information such as equipment speed. The output consisted of data from the next 24 hours and a predicted remaining useful life. The loss function was calculated using a given formula, with appropriate weights determined. The discriminator employed a simple multilayer perceptron architecture with two hidden layers and was trained using its own loss function. When introducing the GNN layer, a graph was constructed based on the bearing structure. Nodes represented each bearing component and associated sensor data, while edges represented component connections and were appropriately weighted. A three-layer GCN layer is connected after the Transformer encoder of the generator. After updating the node features according to the formula, it is weightedly fused with the Transformer output (the weight is set to 0.5) and then input into the decoder. During model training, the training, validation, and test sets are divided into 70%, 15%, and 15% respectively. The Adam optimizer is used with an initial learning rate of 0.002. During training, the discriminator is fixed to train the generator, and then the parameters are updated through backpropagation. The hyperparameters are adjusted according to the performance of the validation set. If there is no improvement after 8 consecutive cycles, the training is stopped. At the same time, learning rate decay and gradient clipping optimization are used. During online prediction, the newly collected data is preprocessed and input into the trained model to obtain the prediction result. If the predicted remaining service life is less than 400 hours, an early warning is triggered to notify maintenance personnel to check the bearing lubrication, wear and other conditions to ensure the normal operation of the equipment.

[0061] The present invention relates to the technical field of rotating machinery condition monitoring and life prediction, and specifically to a method for predicting the remaining useful life of rotating machinery using a fusion technology (GBIT-GAN) of a graph neural network (GNN) and a bidirectional Transformer generative adversarial network (BIT-GAN). The GBIT-GAN-based rotating machinery remaining useful life prediction method of the present invention first performs data acquisition and preprocessing. Using acceleration, temperature, and pressure sensors installed on the rotating machinery, multi-channel vibration, temperature, and pressure data is collected, covering the entire process from normal operation to failure. Outliers and noise are removed through filtering, and the data is normalized to the [0, 1] interval. The data is then segmented into specific time windows and labeled with remaining useful life labels. In terms of model construction, the generator is built based on a bidirectional Transformer. Its unique structure effectively extracts data features, inputs historical data and equipment information, outputs predicted values, and optimizes through adversarial and reconstruction losses. The discriminator uses a suitable structure to distinguish between authentic and fake data. A GNN layer is also introduced to construct a graph based on the mechanical structure and component connections, integrating component and sensor data into nodes. The GNN layer is then integrated into the generator to integrate topological information. During model training, the data set is divided, training is iteratively performed, and hyperparameters are adjusted based on the validation set. Techniques such as learning rate decay are employed to improve performance. In practice, newly collected data is preprocessed and fed into the model. The predicted results are compared with a threshold, and if they fall below the threshold, an alert is issued, ensuring stable operation of rotating machinery and effectively improving prediction accuracy and reliability.

[0062] The present invention also proposes an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for predicting the remaining useful life of rotating machinery based on GBIT-GAN are implemented.

[0063] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the method for predicting the remaining useful life of rotating machinery based on GBIT-GAN.

[0064] The memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus RAM (DRRAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0065] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as high-density digital video disc (digital video disc, DVD)), or semiconductor media (such as solid state disc (solid state disc, SSD)) and the like.

[0066] In the implementation process, each step of the above method can be completed by integrated logic circuit of hardware in the processor or instruction in the form of software. The steps of the method disclosed in the embodiments of the present application can be directly embodied as hardware processor execution completion, or executed by combination of hardware and software modules in the processor. The software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, register or other mature storage medium in the art. The storage medium is located in the memory, and the processor reads the information in the memory and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0067] It should be noted that the processor in the embodiments of the present application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiment can be completed by an integrated logic circuit of the hardware in the processor or by instructions in the form of software. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, or a discrete hardware component. The various methods, steps, and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0068] The above is a detailed introduction to the GBIT-GAN-based remaining service life prediction method for rotating machinery proposed in the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting the remaining useful life of rotating machinery based on GBIT-GAN, characterized in that: The method is specifically as follows: Data acquisition and preprocessing: Utilize acceleration, temperature, and pressure sensors installed on rotating machinery to collect multi-channel vibration, temperature, and pressure data, covering the entire process from normal operation to failure. Remove outliers and noise through filtering, normalize the data to the range [0,1], and then segment the data into specific time windows and label the remaining useful life. Building the GBIT-GAN model: The generator is based on a bidirectional Transformer. Both the encoder and decoder consist of multiple Transformer layers, each of which incorporates a multi-head self-attention mechanism, a feedforward neural network, residual connections, and layer normalization. The generator effectively extracts data features, inputs historical data and device information, outputs predicted values, and optimizes through adversarial and reconstruction losses. The discriminator uses a structure that processes sequential data to distinguish between authentic and fake data. A GNN layer is introduced after the generator's bidirectional Transformer encoder to capture the topological relationships and information propagation between rotating machinery components. The output of the GNN layer is concatenated or weighted summed with the output of the bidirectional Transformer encoder to fuse the information. The fused features are then input into the bidirectional Transformer decoder to generate the predicted future sensor data sequence and the remaining service life prediction value. Model training: divide the training data set and use the training data set to iteratively train the GBIT-GAN model; Prediction of the remaining useful life of rotating machinery: Newly collected data is preprocessed and input into the trained GBIT-GAN model. The predicted results are compared with a preset threshold. If the predicted remaining useful life is lower than the threshold, an early warning signal is issued, prompting maintenance personnel to inspect and maintain the rotating machinery to avoid failures.

2. The method according to claim 1, characterized in that The pre-processed data is divided into a certain time window to form multiple data samples; each data sample contains multi-channel sensor data within a period of time and the corresponding remaining service life label; the size of the time window is selected according to the operating characteristics of the rotating machinery and the sampling frequency of the data, and the remaining service life label y i Calculation is based on the actual operating time and failure time of the rotating machinery.

3. The method according to claim 1, characterized in that In the Transformer layer, for the input sequence X = [x1, x2, ..., x n ], where x i The i-th element in the sequence is the preprocessed data sample, which is mapped to query Q, key K and value V through linear transformation, that is, Q = XW Q ,K=XW K ,V=XW V , where W Q ,W K ,W V is a learnable weight matrix; then calculate the attention weight where d k is the dimension of the key, and the softmax function is used to normalize the attention weights to [0,1] so that the sum of the weights is 1; finally, the output of the multi-head self-attention is MultiHead(X)=Concat(head1,head2,…,head h )W O , where head i =Attention(QW i Q ,KW i K ,VW i V ), h is the number of heads, W i Q ,W i K ,W i V and W O It is a learnable parameter matrix; the output of the multi-head self-attention is further transformed by the feedforward neural network. The feedforward neural network consists of two fully connected layers, with the ReLU activation function in the middle. The formula is FFN(x)=ReLU(xW1+b1)W2+b2, where W1 and W2 are weight matrices, and b1 and b2 are bias vectors.

4. The method according to claim 3, characterized in that The input of the generator includes historical sensor data sequences and basic information about rotating machinery. This basic information is converted into vector representation through the embedding layer, and then concatenated with the historical sensor data sequence and input into the encoder of the bidirectional Transformer. The output of the generator is the predicted sensor data sequence of the next n time windows and the corresponding remaining service life prediction value.

5. The method according to claim 1, wherein Adversarial loss: The goal of the generator is to make the discriminator unable to distinguish between the generated sequence and the real sequence. The adversarial loss of the generator is expressed as where p g is the data distribution generated by the generator, and D(x) is the discriminator's judgment result on the input x; Reconstruction loss: Using mean square error MSE as reconstruction loss, then where p data is the real data distribution, x is the real sequence, is the sequence generated by the generator; Total loss: The total loss of the generator is the weighted sum of the adversarial loss and the reconstruction loss, i.e. Where α is a weight parameter used to balance the importance of adversarial loss and reconstruction loss.

6. The method according to claim 5, characterized in that The loss function in the discriminator is: The goal of the discriminator is to accurately distinguish between real sequences and generated sequences, and its loss function is expressed as 7. The method according to claim 1, characterized in that The GNN layer uses the graph convolutional network GCN. For node i, its feature update formula in the l+1 layer is: in is the feature vector of node i in layer l, is the feature vector of node j in layer l, N(i) is the set of neighbor nodes of node i, c ij is the normalization constant, W l is the learnable weight matrix of layer l, b l is the bias vector, σ is the activation function; By stacking multiple GNN layers, the features of the nodes can be continuously updated, thereby fusing the topological information and sensor data information between components.

8. The method according to claim 1, characterized in that In each training iteration, the discriminator parameters are first fixed, and a batch of data is sampled from the training set and input into the generator to generate a batch of predicted future sensor data sequences and remaining service life prediction values; the generated sequences and the actual sequences are input into the discriminator, and the loss of the generator and the loss of the discriminator are calculated; the backpropagation algorithm is used to calculate the gradients of the generator, discriminator and GNN layer parameters respectively, and the model parameters are updated according to the gradients to minimize the loss of the generator and maximize the loss of the discriminator; after each training cycle, the model is evaluated using the validation set, and the loss function value on the validation set and the error index between the predicted remaining service life and the actual remaining service life are calculated; according to the evaluation results of the validation set, the hyperparameters of the model are adjusted. If the performance on the validation set no longer improves, the training can be stopped early to prevent overfitting.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium for storing computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images