Rotary machine health state evaluation method based on LSTM and Transform fusion network

Abstract

The invention discloses a rotating machine health state evaluation method based on an LSTM and Transform fusion network, and relates to the field of intelligent monitoring, and the method comprises the following steps: collecting and preprocessing fault data of a rotating machine, and dividing the preprocessed fault data into a training set and a test set; a fusion network model based on LSTM and Transform is constructed, and the fusion network model is trained; using the trained fusion network model to predict the health state of the rotating machine, and outputting the health degree score of the rotating machine; by comparing and analyzing the health degree score prediction value of the rotating machine and the actually collected health state data of the rotating machine, the fusion network model is retrained and subjected to parameter adjustment, and the prediction accuracy and stability are improved. According to the method, the advantages of the LSTM and Transform networks are combined, the comprehensiveness of feature extraction and the reliability of prediction results are ensured, the high prediction precision and robustness of equipment health assessment under complex working conditions are ensured, and the efficiency and stability of operation monitoring of industrial equipment can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring, and in particular to a rotating machinery health status assessment method based on an LSTM and Transformer fusion network. Background Art

[0002] The operating status of rotating machinery directly impacts the stability and efficiency of industrial production. Existing health monitoring technologies struggle to simultaneously account for both local features and global dependencies when processing long-term data series, resulting in low assessment accuracy. While LSTMs excel at processing time series data, their ability to capture long-term dependencies is limited. Transformers, on the other hand, excel at modeling long-term dependencies but struggle with extracting short-term features. Therefore, a model that combines the strengths of both is needed to improve rotating machinery health assessment.

[0003] Therefore, technicians in this field are committed to developing a rotating machinery health status assessment method based on LSTM and Transformer fusion network. Summary of the Invention

[0004] In view of the above-mentioned defects in the prior art, the technical problem to be solved by the present invention is to improve the accuracy and reliability of equipment health status prediction.

[0005] To achieve the above objectives, the present invention provides a rotating machinery health status assessment method based on an LSTM and Transformer fusion network. The method constructs an LTFusion module, comprehensively utilizes the advantages of both to process time-frequency domain data features, and adaptively adjusts attention weights through an ACC module to achieve high-precision health assessment. The method comprises the following steps: S101: collecting fault data of the rotating machinery and preprocessing the data, and dividing the preprocessed fault data into a training set and a test set; S103: Constructing a fusion network model based on LSTM and Transformer, and training the fusion network model; S105: using the trained fusion network model to predict the health status of the rotating machinery and output a health score of the rotating machinery; S107: By comparing and analyzing the predicted health score value of the rotating machinery with the actually collected health status data of the rotating machinery, the fusion network model is retrained and the parameters are adjusted, and the health status of the rotating machinery is evaluated based on the fusion network model after retraining and parameter adjustment.

[0006] Furthermore, in step S101, the preprocessing includes the following sub-steps: S1011: cutting the one-dimensional time domain signal in the fault data according to a predetermined window size, and normalizing the one-dimensional time domain signal in each window; S1012: Performing continuous wavelet transform on the normalized one-dimensional time domain signal to generate a time-frequency signal and obtain two-dimensional spectrum graph data.

[0007] Furthermore, the continuous wavelet transform is performed using the following formula:

[0008] in, is the translation parameter, is the scaling parameter, For my mother Xiaobo, is the input signal, For time.

[0009] Furthermore, in step S103, the fusion network model includes an LTFusion module, an ACC module, an output classification module and a health assessment module, wherein: The LTFusion module sequentially extracts time series features from the two-dimensional spectrogram data, and includes a Transformer submodule and an LSTM submodule. The Transformer submodule and the LSTM submodule perform feature extraction in parallel and perform a splicing operation on the output data. The ACC module is configured as an attention-based channel convolution module, which generates channel attention weights for the output of the LTFusion module to enhance attention to key channel features; The output classification module includes an MLP classifier, which is composed of a linear layer with Softmax activation, and the MLP classifier predicts the fault information of the rotating machinery based on the features output by the LTFusion module; The health assessment module scores the health of the equipment according to the fault type and severity.

[0010] Furthermore, the Transformer submodule includes the following steps: S1031: Input embedding and position encoding, adding position information to the input data through the position encoding. The original position encoding uses sine and cosine functions to calculate the absolute position:

[0011]

[0012] S1032: Calculate attention weights using multi-head self-attention processing and complete multi-head output; S1033: Residual connection and layer normalization processing, the output of the multi-head attention is processed by residual connection and layer normalization; S1034: Perform feature transformation and nonlinear mapping on each position in the sequence data using a feedforward neural network. The feedforward neural network is composed of two layers of linear transformation and a nonlinear activation function. The nonlinear activation function uses a ReLU activation function. The feedforward neural network uses the following calculation formula for processing:

[0013] S1035: Perform the second residual connection and normalization process, and output the final encoder features; in, 、 is the position code, is the position index, is the dimension index, is the total dimension of the feature vector, is the input vector, is the weight matrix of the first layer, is the weight matrix of the second layer, is the bias vector of the first layer, is the bias vector of the second layer.

[0014] Furthermore, in step S1032, the attention weight is calculated using the following method:

[0015]

[0016] Multi-head output uses the following method:

[0017] in, are query, key, and value matrices, is the input vector, is the linearly changing weight matrix, The dimension of the key, used for scaling to avoid excessive values. represents the output linear transformation matrix, For splicing operation, is an independent attention calculation unit, Count the number of attention units.

[0018] Furthermore, the LSTM submodule is set to a bidirectional LSTM, including a forward LSTM and a reverse LSTM. The forward LSTM processes the input sequence from left to right, and the reverse LSTM processes the input sequence from right to left. Finally, the outputs of the forward LSTM and the reverse LSTM are concatenated or weightedly combined as the output representation of each time step, wherein,

[0019]

[0020]

[0021] in, is the forward output, For reverse output, is the fusion output, is the time step, and is the weighting coefficient.

[0022] Furthermore, the ACC module includes the following steps: Preliminary channel feature extraction: The input data first passes through the first one-dimensional convolution operation to extract preliminary channel information and generate a new feature map; Channel feature splicing and global description: The convolution feature map is merged along the channel to obtain a two-dimensional data feature map. The global average pooling operation is performed on the spliced ​​feature map to calculate the average value of each channel and generate a global channel description. Attention weight generation: The global channel description passes through the second one-dimensional convolutional layer to generate the attention weight of each channel; Attention weighting and feature enhancement: By element-by-element multiplication, the attention weight is multiplied by the concatenated feature map to achieve weighted features of each channel.

[0023] Furthermore, when predicting the fault information of the rotating machinery, the output classification module includes the following steps: Processing the output of the LTFusion module through a global max pooling operation along the channel dimension, converting the output of the LTFusion module into a one-dimensional vector; The one-dimensional vector is input into an MLP classifier to predict equipment fault conditions of the rotating machinery.

[0024] Furthermore, the health assessment module uses the following formula to calculate the device health score: .

[0025] In a preferred embodiment of the present invention, compared with the prior art, the present invention has the following beneficial effects: 1. The accuracy and robustness of rotating machinery health assessment have been significantly improved. It can effectively capture the local characteristics and global dependency characteristics of the equipment's operating status and adapt to the needs of multi-dimensional time series data analysis under complex working conditions.

[0026] 2. By combining the LTFusion module and the ACC module, the model can simultaneously extract short-term and long-term features of the time series, and dynamically weight key channels through the attention mechanism, further optimizing feature selection capabilities and improving the accuracy of health assessment.

[0027] 3. The network structure that combines the multi-head self-attention mechanism and the bidirectional LSTM not only improves the model's ability to model long time series data, but also significantly reduces the computational complexity and improves efficiency when processing high-dimensional time series features.

[0028] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of a method for evaluating the health status of rotating machinery based on an LSTM and Transformer fusion network according to a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the overall model network structure of a preferred embodiment of the present invention; Figure 3 Schematic diagram of a Transformer submodule of a preferred embodiment of the present invention; Figure 4 This is a bidirectional LSTM structure diagram in the LSTM submodule of a preferred embodiment of the present invention; Figure 5 1 is a schematic diagram of the LSTM structure of a preferred embodiment of the present invention; Figure 6 It is a schematic diagram of the ACC module structure of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0030] The following describes several 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 of embodiments, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0031] In the drawings, components with identical structures are denoted by the same reference numerals, and components with similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrary and are not limited by the present invention. For clarity, the thickness of components in some places in the drawings is appropriately exaggerated.

[0032] This embodiment of the present invention provides a rotating machinery health assessment method based on a fusion network of LSTM and Transformer networks. Combining the advantages of LSTM and Transformer networks, this method uses the LTFusion module to capture local and global features of time series data. The ACC module, using a channel-attention mechanism, assigns and optimizes feature weights. Ultimately, the health assessment module achieves high-precision prediction of the machinery's health status. During the data processing phase, a continuous wavelet transform (CWT) is used to convert one-dimensional vibration signals into two-dimensional spectrograms, from which key frequency features are extracted. The model architecture based on the fusion network of LSTM and Transformer provided by this embodiment of the present invention includes the LTFusion module, the ACC module, and the health assessment module. These modules work closely together to ensure comprehensive feature extraction and reliable prediction results. Experimental results demonstrate that this method demonstrates high prediction accuracy and robustness for equipment health assessment under complex operating conditions, effectively improving the efficiency and stability of industrial equipment operation monitoring.

[0033] like Figure 1 As shown, an embodiment of the present invention provides a rotating machinery health status assessment method based on an LSTM and Transformer fusion network, comprising the following steps: S101: Collecting and preprocessing fault data of rotating machinery, and dividing the preprocessed fault data into a training set and a test set; In this embodiment, bearing data with a sampling frequency of 12 kHz is used, and preprocessing is performed based on the high sampling rate and non-stationary characteristics of the signal.

[0034] During preprocessing, the collected data was segmented into windows of 512, with appropriate overlap between each window (set to 30%) to preserve temporal continuity and ensure that subsequent analysis could capture the dynamic evolution of subtle fault characteristics. To enable the model to better handle data differences, the signal in each window was normalized. The one-dimensional time-domain signal in each window was converted into a time-frequency signal using a continuous wavelet transform, and the input two-dimensional spectrogram was divided along the frequency direction, dividing the original two-dimensional data into one-dimensional data.

[0035] The time-frequency signal is generated by performing a continuous wavelet transform on the input one-dimensional time-domain signal (length 512). After the CWT, the one-dimensional data (1×512) is converted into a two-dimensional spectrogram (64×512). The processed data is then divided into training and test sets for subsequent model training and validation.

[0036] In this embodiment, Morlet wavelet is selected as the mother wavelet of CWT. After CWT, the one-dimensional data becomes a two-dimensional spectrum graph; In this embodiment, preprocessing the collected data includes the following sub-steps: S1011: cutting the one-dimensional time domain signal in the fault data according to a predetermined window size, and normalizing the one-dimensional time domain signal in each window; S1012: Performing continuous wavelet transform on the normalized one-dimensional time domain signal to generate a time-frequency signal and obtain two-dimensional spectrum graph data.

[0037] The continuous wavelet transform is performed using the following formula:

[0038] in, is the translation parameter, is the scaling parameter, For my mother Xiaobo, is the input signal, For time.

[0039] S103: Build a fusion network model based on LSTM and Transformer, and train the fusion network model.

[0040] In this embodiment, the fusion network model includes an LTFusion module, an ACC module, an output classification module, and a health assessment module, wherein: The LTFusion module sequentially extracts time series features from two-dimensional spectrogram data. It consists of a Transformer submodule and an LSTM submodule. The Transformer and LSTM submodules perform feature extraction in parallel and concatenate the output data. The ACC module is set as an attention-based channel convolution module, which generates channel attention weights for the output of the LTFusion module to enhance the focus on key channel features; The output classification module includes an MLP classifier, which consists of a linear layer with Softmax activation. The MLP classifier predicts the fault information of the rotating machinery based on the features output by the LTFusion module; The health assessment module scores the health of the equipment based on the fault type and severity.

[0041] In this embodiment, data is transferred in the above modules through the following process: LTFusion module: This module consists of a Transformer submodule and an LSTM submodule. The two submodules extract features from the input data in parallel, ensuring that the output data dimensions of the two submodules are consistent with the input dimensions. The output data of the two submodules are then concatenated. ACC module: This module generates channel attention weights for the output of the LTFusion module to reflect the importance of each channel in the final feature; Feature extraction module: It consists of 3 LTFusion modules and 2 ACC modules. The ACC module is immediately followed by the LTFusion module, and the output of the last LTFusion module is subjected to global maximum pooling.

[0042] In this embodiment, the Transformer submodule includes the following steps: S1031: Input embedding and position encoding.

[0043] The input data is first added to the positional encoding (positional embedding). This is done to add positional information to the input data. Since the Transformer submodule itself does not have the ability to perceive sequence position, positional encoding is used to add positional information to the input data so that the model can recognize the order of the input.

[0044] The original position encoding uses sine and cosine functions to calculate the absolute position:

[0045]

[0046] in, 、 is the position code, is the position index, is the dimension index, is the total dimension of the feature vector.

[0047] S1032: Use the multi-head self-attention processing system to calculate the attention weight and complete the multi-head output.

[0048] The position-encoded input data is passed through a multi-head self-attention module to calculate the attention weight. The multi-head self-attention module uses a multi-head self-attention mechanism (Multi-Head Attention), which consists of the following parts: Q, K, V generation: The input data passes through three linear layers to obtain query (Q), key (K), and value (V) matrices:

[0049] in, is the linearly varying weight matrix.

[0050] Calculate the attention weight: Q and K are matrix multiplied and scaled, and then the attention weight is calculated by the Softmax function:

[0051] in, Represents query, key, and value matrices; The dimension of the key, used for scaling to avoid excessive values.

[0052] Multi-head output:

[0053] in, represents the output linear transformation matrix, For splicing operation, is an independent attention calculation unit, totaling indivual.

[0054] S1033: Residual connection and layer normalization processing, the output of the multi-head attention is processed through residual connection and layer normalization.

[0055] The output of the multi-head attention module is connected to the residual (added to the input) and normalized (Norm) layer. The purpose of this step is to stabilize the training, avoid gradient disappearance, and improve the model optimization effect.

[0056] The specific processing adopts the following formula:

[0057] in, represents the input matrix; is the current sublayer (such as an attention layer or a feedforward network).

[0058] S1034: Use a feedforward neural network to perform feature transformation and nonlinear mapping on each position in the sequence data.

[0059] A feedforward neural network (FFN) is a core component of the Transformer architecture, primarily used to perform feature transformation and nonlinear mapping at each position in sequence data. The FFN is applied to each position in the sequence data within each Transformer sublayer, processing each data point at each time step independently without sharing weights. It consists of two layers of linear transformations and a nonlinear activation function, typically using the ReLU activation function. Its formula is as follows:

[0060] in, is the input vector, is the weight matrix of the first layer; is the bias vector of the first layer; is the weight matrix of the second layer; is the bias vector of the second layer.

[0061] S1035: Perform the second residual connection and normalization process, and output the final encoder features.

[0062] The output of the MLP is again connected to the residual network (added to the input of the MLP) and normalized. This step further stabilizes the training of the model and outputs the final encoder features.

[0063] In this embodiment, the LSTM submodule is set up using an LSTM variant, bidirectional Long Short-Term Memory (Bidirectional Long Short-Term Memory). The basic structure of the bidirectional LSTM consists of two independent LSTM networks: a forward LSTM and a backward LSTM. The forward LSTM processes the input sequence from left to right (forward), and the backward LSTM processes the input sequence from right to left (backward). Finally, the outputs of the forward LSTM and backward LSTM are concatenated or weightedly combined as the output representation of each time step.

[0064] In this embodiment, the LSTM submodule input and output process is as follows: The forward LSTM passes the input sequence in chronological order and outputs , the formula is as follows:

[0065] The reverse LSTM passes the input sequence in reverse time order and outputs , the formula is as follows:

[0066] Final forward and backward hidden state outputs and At each time step The fusion method uses weighted summation, and the formula is as follows:

[0067] in, is the forward output, For reverse output, is the fusion output, is the time step, and is the weighting coefficient.

[0068] In this embodiment, the basic structure of the LSTM in the bidirectional LSTM includes: Memory Cell: Used to store long-term information, similar to a "memory bank." Dynamically updated and maintained through a gating mechanism.

[0069] Gate Mechanisms: This mechanism controls the "addition" or "deletion" of information, ensuring that the network can dynamically adjust the memory content. It includes the following three types of gates: Forget Gate: Determines how much past information to discard.

[0070]

[0071] in, is the output of the forget gate, and are weights and biases, is the Sigmoid activation function.

[0072] Input Gate: Determines what new information is added to the memory cell.

[0073]

[0074]

[0075] in, is the output of the input gate, It is the candidate memory content.

[0076] Output Gate: Determines the output content of the current time step.

[0077]

[0078]

[0079] in, is the activation value of the output gate, ranging from [0,1], is the output.

[0080] Memory unit state update: The information of the forget gate and input gate is integrated to update the memory unit state.

[0081] .

[0082] In this embodiment, the ACC module includes the following steps: Initial channel feature extraction: The input data first passes through a one-dimensional convolution operation to extract preliminary channel information and generate a new feature map. The purpose of this step is to focus on the information interaction between local channels and prepare for subsequent attention weighting.

[0083] Channel Feature Concatenation and Global Description: The convolved feature maps are merged along the channels to produce a two-dimensional data feature map. Subsequently, a global average pooling operation is performed on the concatenated feature maps to calculate the average value for each channel and generate a global channel description. This process effectively extracts global features across channels and provides a basis for generating attention weights.

[0084] Attention weight generation: The global channel description passes through another one-dimensional convolutional layer to generate attention weights for each channel. These weights reflect the importance of each channel in the final feature.

[0085] Attention weighting and feature enhancement: Through element-by-element multiplication (Hadamard product), the attention weight is multiplied by the concatenated feature map to achieve weighted features of each channel. Channels with higher weights are enhanced in the final output, while channels with lower weights are suppressed accordingly.

[0086] The input and output relationship of each part in the ACC module can be expressed as: (1) (2) (3) (4) (5) in, Represents the combined features obtained by one-dimensional convolution and channel fusion, Express The global channel description obtained by average pooling, represents the channel attention weight generated by two layers of one-dimensional convolution, Represents the feature map after fusion of attention weights, Represents input data, represents the first convolution operation, represents the average pooling operation, represents the second convolution operation, represents element-wise multiplication, Represents a split operation, Indicates output data.

[0087] The output classification module includes the following steps when predicting fault information of rotating machinery: The output of the LTFusion module is processed through a global maximum pooling operation along the channel dimension, and the output of the LTFusion module is converted into a one-dimensional vector; the one-dimensional vector is input into the MLP classifier to predict the equipment fault of the rotating machinery.

[0088] The MLP classifier consists of linear layers with softmax activation. The input to the MLP is the features extracted from the last LTFusion module. Specifically, the output of the last LTFusion module is processed through a global max pooling operation along the channel dimension to convert it into a one-dimensional vector. This vector is then input to the MLP classifier to predict the specific equipment failure condition of the machine.

[0089] The health assessment module scores the health of the equipment based on the fault type and severity. The health assessment module uses the following formula to calculate the health of the equipment: .

[0090] When training the fusion network model, the preprocessed data is input into the designed hybrid model, and the cross entropy is used as the loss function, which is defined as:

[0091] in, is the true label, is the predicted probability.

[0092] Model parameters were optimized using the Adam optimizer. The initial learning rate was set to 0.001, dynamically adjusted using a learning rate decay strategy, and the batch size was set to 32. During model training, the loss function was reduced through multiple iterations until convergence, ensuring the model's ability to stably and accurately diagnose rolling bearing faults. The data was divided into a training set (70%), a validation set (15%), and a test set (15%). Training was performed for 100 epochs, with accuracy verified each time and the best model saved.

[0093] S105: Use the trained fusion network model to predict the health status of the rotating machinery and output a health score of the rotating machinery.

[0094] The trained fusion network model is used to predict the health data of rotating machinery (rolling bearings in this example). By extracting spatial and temporal features from the bearing vibration signals, the rolling bearing fault category is accurately predicted. The prediction results are validated using various evaluation metrics, such as average loss (AvgLoss) and average recognition accuracy (AvgAcc), to ensure the stability and accuracy of the model performance.

[0095] S107: By comparing and analyzing the predicted health score value of the rotating machinery with the actually collected health status data of the rotating machinery, the fusion network model is retrained and the parameters are adjusted, and the health status of the rotating machinery is evaluated based on the fusion network model after retraining and parameter adjustment.

[0096] Compared with the existing technology, the rotating machinery health status assessment method based on the LSTM and Transformer fusion network provided by the embodiment of the present invention has the following advantages: 1. The accuracy and robustness of rotating machinery health assessment have been significantly improved. It can effectively capture the local characteristics and global dependency characteristics of the equipment's operating status and adapt to the needs of multi-dimensional time series data analysis under complex working conditions.

[0097] 2. By combining the LTFusion module and the ACC module, the model can simultaneously extract short-term and long-term features of the time series, and dynamically weight key channels through the attention mechanism, further optimizing feature selection capabilities and improving the accuracy of health assessment.

[0098] 3. The network structure that combines the multi-head self-attention mechanism and the bidirectional LSTM not only improves the model's ability to model long time series data, but also significantly reduces the computational complexity and improves efficiency when processing high-dimensional time series features.

[0099] 4. Experimental results show that the present invention has strong generalization ability on data sets of various working conditions and different equipment failure types, can effectively avoid the overfitting problem, and shows good adaptability and stability in industrial equipment health monitoring and predictive maintenance.

[0100] The present invention will be described in detail below in conjunction with the preferred embodiments of the present invention.

[0101] like Figures 1-6 As shown, the preferred embodiment of the present invention provides a fusion network model based on LSTM and Transformer, including the following steps: Step 1: Data collection and preprocessing Step 1.1: Data Collection In this example, rolling bearings are used in rotating machinery to predict rolling bearing fault types. During the prediction process, a rolling bearing fault dataset from Case Western Reserve University (CWRU) was obtained. The one-dimensional vibration signal was segmented into windows of 512, with a moderate overlap (30%) between each window to preserve temporal continuity and ensure that subsequent analysis captures the dynamic evolution of subtle fault characteristics. To enable the model to better handle data variability, the signal in each window was normalized.

[0102] Step 1.2: Data Preprocessing A continuous wavelet transform (CWT) is performed on the input one-dimensional time-domain signal (length 512) to generate a time-frequency signal. After the CWT, the one-dimensional data (1×512) is converted into a two-dimensional spectrogram (64×512). The processed data is then divided into training and test sets for subsequent model training and validation.

[0103] Step 2: Feature extraction In this embodiment, feature extraction is implemented through a feature extraction module.

[0104] The feature extraction module includes the following modules: LTFusion module, ACC module and output classification module. Each sub-module is described in detail below.

[0105] 2.1 LTFusion module: This module contains a Transformer submodule and an LSTM submodule, such as Figure 2 To ensure the correct execution of the Concat operation within the module, parameters are designed to ensure that the output data dimensions of the two submodules are the same. For example, if the first LTFusion module inputs 64 1×512 one-dimensional data, the outputs of the Transformer submodule and the LSTM submodule are both 64 1×512 one-dimensional data. The Concat operation then generates 64 1×1024 one-dimensional data.

[0106] The Transformer submodule captures global dependencies through a multi-head self-attention mechanism, and combines residual connections and layer normalization operations to enhance the model's training stability and feature extraction capabilities. Figure 3 Each Transformer submodule uses position encoding to add sequence position information to the input data and uses a feedforward neural network (FFN) to perform nonlinear mapping on the features, further improving the ability to express complex sequence features.

[0107] The LSTM submodule in the LTFusion module adopts a bidirectional LSTM structure, which captures the forward and backward dependencies of the time series simultaneously through forward and backward calculations. Figure 4 、 Figure 5 As shown in Figure 1, each LSTM unit is controlled by a forget gate, an input gate, and an output gate to dynamically update information, thereby effectively handling long-term dependency issues and improving sequence modeling capabilities. In the LSTM structure, the three gating mechanisms work together to achieve dynamic modeling of time series information and effective capture of long-term dependencies. Among them, the forget gate receives the input vector of the current time step and the hidden state of the previous moment, and passes a The activation function generates a gating vector to control the degree of information retention in the previous memory unit state element by element, thereby selectively forgetting historical information that is no longer relevant; the input gate also generates a gating vector based on the current input and the previous hidden state, and multiplies it with the candidate memory content to determine how much new information extracted at the current time step should be written into the memory unit. This step is the key to updating the memory state; and the output gate combines the current input, the previous hidden state and the updated memory state, and adjusts the content extracted from the memory state and passed to the current hidden state through the gating mechanism as the output at the current moment, and passes it to the next time step, thereby realizing the orderly flow and control of information in time.

[0108] 2.2 ACC module: Figure 6 As shown in Figure 1, the main function of the ACC module (attention-based channel convolution module) is to optimize the channel weight distribution in the feature extraction process to enhance the attention to key channel features, thereby improving the prediction accuracy and robustness of the model.

[0109] The ACC module uses an attention-based channel convolution mechanism, which consists of two stages: preliminary channel feature extraction and attention weighting. The preliminary extraction captures local information through one-dimensional convolution and performs global average pooling to generate a global description of channel features. The attention weighting stage dynamically adjusts features by generating channel weights, highlighting key features and thus enhancing the model's feature selection ability for time series data.

[0110] The ACC module execution steps may include the following steps: Step 2.1: Preliminary channel feature extraction.

[0111] The input data first passes through a one-dimensional convolution operation to extract preliminary channel information and generate a new feature map. The purpose of this step is to focus on the information interaction between local channels and prepare for subsequent attention weighting.

[0112] Step 2.2: Channel feature concatenation and global description The convolved feature maps are merged along the channels to produce a two-dimensional data feature map. Subsequently, a global average pooling operation is performed on the concatenated feature maps to calculate the average value for each channel and generate a global channel description. This process effectively extracts global features between channels and provides a basis for the generation of attention weights.

[0113] Step 2.3: Attention weight generation The global description passes through another one-dimensional convolutional layer to generate attention weights for each channel. These weights reflect the importance of each channel in the final feature.

[0114] Step 2.4: Attention Weighting and Feature Enhancement The attention weights are multiplied by the concatenated feature maps through element-wise multiplication (Hadamard product) to weight the features of each channel. Channels with higher weights are enhanced in the final output, while channels with lower weights are suppressed accordingly.

[0115] The input and output relationship of each part in the ACC module can be expressed as: (1) (2) (3) (4) (5) in, Represents input data, represents the first convolution operation, represents the average pooling operation, represents the second convolution operation, represents element-wise multiplication, Represents a split operation, Indicates output data.

[0116] 2.3 Output Classification Module The MLP classifier in this module consists of linear layers with softmax activation. The input to the MLP is the features extracted from the last LTFusion module. Specifically, the output of the last LTFusion module is processed through a global max pooling operation along the channel dimension to convert it into a one-dimensional vector. This vector is then input into the MLP classifier to predict the specific equipment failure condition of the machine.

[0117] Step 3: Model training and optimization The preprocessed data is input into the hybrid network model for training. Cross-Entropy Loss is used as the loss function, and the parameters are optimized in combination with the gradient descent optimizer to ensure the stability of the training process and improve the predictive accuracy of health assessment and fault diagnosis.

[0118] The formula of the loss function is as follows:

[0119] in, is the true label, is the predicted probability.

[0120] The model is optimized using the Adam optimizer, with the learning rate set to 0.001 and a learning rate decay strategy used to dynamically adjust the learning rate.

[0121] In this embodiment, the trained hybrid network model is applied to actual work to predict the rolling bearing fault category.

[0122] During the operation of the equipment, the real-time collected data is input into the trained hybrid network model through the preprocessing module for prediction.

[0123] The hybrid network model combines the LTFusion module, ACC module and output classification module to analyze the equipment operation data and output the fault category of the equipment.

[0124] Step 4: Model evaluation and optimization By comparing and analyzing the predicted values ​​with the actual collected data, the performance of the model is evaluated using indicators such as average loss (AvgLoss) and average recognition accuracy (AvgAcc) to further optimize the model.

[0125] Step 4.1: Model Performance Evaluation The performance of the model is evaluated using indicators such as average loss (AvgLoss) and average recognition accuracy (AvgAcc). The performance indicator formulas of AvgLoss and AvgAcc are as follows:

[0126]

[0127] Where N is the total number of samples, is the true label of the i-th sample, is the predicted value of the i-th sample, is the loss function of the i-th sample, is the indicator function, , the value is 1, otherwise it is 0.

[0128] These evaluation results are used to retrain the model and adjust its parameters to improve the accuracy and stability of its predictions.

[0129] Step 5: Health Assessment The health assessment module is used to score the health of the equipment based on the fault type and severity. The fault location weight and fault severity weight are shown in Table 1.

[0130] Table 1 Fault location weight and fault severity weight

[0131] The score is calculated using the following formula:

[0132] In this embodiment, the complete process overview of rolling bearing fault classification diagnosis includes: 1) Startup phase: Initialize the device and model, and start data collection.

[0133] 2) Data collection and preprocessing: Real-time collection of equipment operation data and resampling.

[0134] 3) Model prediction: The preprocessed data passes through the LTFusion module, ACC module, output classification module and health assessment module in sequence to finally obtain the predicted bearing health score.

[0135] 4) Real-time monitoring and adjustment: During the operation of the equipment, the bearing health score is predicted in real time, the equipment status is monitored, and the accuracy of the prediction is ensured.

[0136] 5) Shutdown and evaluation: After the equipment is shut down, evaluate the model performance and make optimization adjustments.

[0137] This paper proposes a deep learning-based method for assessing the health of rotating machinery. By combining the advantages of LSTM and Transformer networks, a LTFusion module is designed that integrates these two approaches. Furthermore, an attention-based ACC module is introduced to optimize the weight distribution of channel features. The LTFusion module allows the present invention to simultaneously capture local features and global dependencies in time series data. The ACC module's adaptive attention mechanism further enhances the model's focus on key channel features, enabling highly accurate prediction of equipment health and fault classification.

[0138] This paper uses a standard cross-entropy loss as the loss function and optimizes model parameters via a gradient descent optimizer to ensure the stability and accuracy of model training. By combining the advantages of LSTM and Transformer, this paper effectively addresses the challenges of non-stationarity and complex feature extraction in time series data for rotating machinery operating status monitoring, significantly improving the accuracy and robustness of machinery health assessment.

[0139] This invention is suitable for industrial equipment condition monitoring and predictive maintenance based on sensor data. It can quickly and accurately diagnose equipment operating status, supporting the stability and safety of industrial production. Through this multi-module deep learning model, the invention demonstrates excellent performance when handling complex industrial scenarios and large-scale data, significantly improving the efficiency and reliability of equipment fault diagnosis and health assessment.

[0140] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. A rotating machinery health status assessment method based on LSTM and Transformer fusion network, characterized in that: The method comprises the following steps: S101: Collecting and preprocessing fault data of rotating machinery, and dividing the preprocessed fault data into a training set and a test set; S103: Constructing a fusion network model based on LSTM and Transformer, and training the fusion network model; S105: using the trained fusion network model to predict the health status of the rotating machinery and output a health score of the rotating machinery; S107: By comparing and analyzing the predicted health score value of the rotating machinery with the actually collected health status data of the rotating machinery, the fusion network model is retrained and the parameters are adjusted, and the health status of the rotating machinery is evaluated based on the fusion network model after retraining and parameter adjustment.

2. The method according to claim 1, wherein In step S101, the preprocessing includes the following sub-steps: S1011: Segmenting the one-dimensional time domain signal in the fault data according to a predetermined window size, and performing normalization processing on the one-dimensional time domain signal in each window; S1012: Performing continuous wavelet transform on the normalized one-dimensional time domain signal to generate a time-frequency signal and obtain two-dimensional spectrum graph data.

3. The method according to claim 2, wherein The continuous wavelet transform is performed using the following formula: in, is the wavelet transform operator, is the result of wavelet transform operation, is the translation parameter, is the scaling parameter, For my mother Xiaobo, is the input signal, For time.

4. The method according to claim 3, wherein In step S103, the fusion network model includes an LTFusion module, an ACC module, an output classification module and a health assessment module, wherein: The LTFusion module is used to sequentially extract time series features from the two-dimensional spectrogram data, and includes a Transformer submodule and an LSTM submodule. The Transformer submodule and the LSTM submodule perform feature extraction in parallel and perform a splicing operation on the output data; The ACC module is configured as an attention-based channel convolution module to generate channel attention weights for the output of the LTFusion module; The output classification module includes an MLP classifier, which is composed of a linear layer with Softmax activation, and the MLP classifier predicts the fault information of the rotating machinery based on the features output by the LTFusion module; The health assessment module is used to score the health of the equipment according to the fault type and severity.

5. The method according to claim 4, wherein The execution steps of the Transformer submodule include: S1031: Input embedding and position encoding, adding position information to the input data through the position encoding. The original position encoding uses sine and cosine functions to calculate the absolute position: S1032: Calculate attention weights using multi-head self-attention processing and complete multi-head output; S1033: Residual connection and layer normalization processing, the output of the multi-head attention is processed by residual connection and layer normalization; S1034: Perform feature transformation and nonlinear mapping on each position in the sequence data using a feedforward neural network. The feedforward neural network is composed of two layers of linear transformation and a nonlinear activation function. The nonlinear activation function uses a ReLU activation function. The feedforward neural network uses the following calculation formula for processing: S1035: Perform the second residual connection and normalization process, and output the final encoder features; in, 、 is the position code, is the position index, is the dimension index, is the total dimension of the feature vector, represents the output of the two-layer feedforward network, is the input vector, is the weight matrix of the first layer, is the bias vector of the first layer, is the weight matrix of the second layer, is the bias vector of the second layer.

6. The method according to claim 5, wherein In step S1032, the attention weight is calculated using the following method: Multi-head output uses the following method: in, represents the attention coefficient of the feature, is the normalized exponential function, are query, key, and value matrices respectively, represents transpose, is the input vector, is the linearly changing weight matrix, is the dimension of the key, represents the output linear transformation matrix, For splicing operations, is the attention calculation unit, Count the number of attention units.

7. The method according to claim 6, wherein The LSTM submodule is set to a bidirectional LSTM, including a forward LSTM and a reverse LSTM. The forward LSTM processes the input sequence from left to right, and the reverse LSTM processes the input sequence from right to left. Finally, the outputs of the forward LSTM and the reverse LSTM are concatenated or weighted together as the output representation of each time step, wherein, in, is the forward output, For reverse output, represents the forward LSTM unit, represents the reverse LSTM unit, express The input vector at time t, Represents the previous moment output of the forward hidden state, Represents the output of the next moment of the reverse hidden state, is the fusion output, is the time step, and is the weighting coefficient.

8. The method according to claim 7, wherein The execution steps of the ACC module include: Preliminary channel feature extraction: The input data first passes through the first one-dimensional convolution operation to extract preliminary channel information and generate a new feature map; Channel feature splicing and global description: The convolution feature map is merged along the channel to obtain a two-dimensional data feature map. The global average pooling operation is performed on the spliced ​​feature map to calculate the average value of each channel and generate a global channel description. Attention weight generation: The global channel description passes through the second one-dimensional convolutional layer to generate the attention weight of each channel; Attention weighting and feature enhancement: By element-by-element multiplication, the attention weight is multiplied by the concatenated feature map to achieve weighted features of each channel.

9. The method according to claim 8, wherein When predicting the fault information of the rotating machinery, the output classification module includes the following steps: Processing the output of the LTFusion module through a global max pooling operation along the channel dimension, converting the output of the LTFusion module into a one-dimensional vector; The one-dimensional vector is input into an MLP classifier to predict equipment fault conditions of the rotating machinery.

10. The method according to claim 9, wherein The health assessment module uses the following formula to score the health of the device: 。

Images