Remaining service life prediction method of chemical rotating equipment

Abstract

The invention provides a method for predicting the residual service life of chemical rotating equipment, which comprises the following steps of: firstly, acquiring full-life data of the rotating equipment, performing wavelet transform denoising and Z-SCORE standardization processing on the acquired data, then segmenting the data by using a sliding window to construct a training sample, then constructing a GRU-Attention network, and inputting the sample into the network for training. Then, parameters such as vibration signals, temperature and rotating speed in the operation process are obtained according to real-time monitoring of a sensor in a workshop on rotating equipment; and finally, pre-processing the obtained real-time data, and inputting the pre-processed real-time data into the model to predict the residual service life, thereby realizing real-time and accurate prediction of the residual service life of the equipment, and being beneficial to optimizing a maintenance strategy, formulating an operation plan and improving the production efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent manufacturing workshop performance prediction, and in particular to a method for predicting the remaining service life of chemical rotating equipment. Background Art

[0002] With the development of Industry 4.0 and intelligent manufacturing, predictive maintenance has become an important means to improve equipment reliability and production efficiency. By analyzing the real-time operating status data of the equipment through intelligent algorithms, potential equipment failures can be discovered in advance, thereby reducing equipment downtime, extending equipment service life, and optimizing production resource allocation. Chemical rotating equipment plays a key role in chemical production, and its operating status directly affects the stability and safety of the entire production line. Traditional equipment maintenance strategies are usually divided into two categories: regular maintenance and post-maintenance. Regular maintenance does not take into account the actual status of the equipment, which may lead to excessive or insufficient maintenance and increase maintenance costs; post-maintenance, a passive maintenance strategy, may cause serious equipment damage, high repair costs and easily cause safety problems. Therefore, accurate prediction of the remaining service life of chemical rotating equipment has become the key to ensuring production efficiency and safety.

[0003] Predictive maintenance uses equipment operation data to predict the remaining service life of equipment in advance, realize condition-based maintenance, and avoid unnecessary maintenance operations. Traditional prediction methods include methods based on physical models and statistical analysis. The method based on physical models takes the actual operating parameters of the equipment as input and uses physical equations to predict the remaining service life of the equipment. This method relies on a lot of professional knowledge and complex mathematical models, and it is difficult to model nonlinear systems; the method based on statistical analysis relies on historical data and probability distribution to predict the probability of equipment failure. When faced with changes in equipment operating conditions or nonlinear complexity of data, the prediction accuracy is not ideal. In recent years, machine learning methods have been widely used in equipment fault diagnosis and life prediction, such as support vector machines (SVM), random forests (RF), and neural networks. These methods predict the future state of the equipment by learning the characteristic patterns in historical data. However, traditional machine learning methods may encounter bottlenecks when processing long time series data and cannot fully utilize the dynamic information in the time series. Some scholars have proposed the use of deep learning methods, especially recurrent neural networks (RNN), long short-term memory networks (LSTM), and gated recurrent units (GRU), which have significant advantages in processing time series data and can effectively capture the time dependency and nonlinear characteristics in equipment operation data.

[0004] Although the above research has a good promoting effect on the prediction of the remaining service life of equipment, breakthroughs are still needed in the following new problems: Due to the complex and changeable manufacturing resources and production environment, the sensor signals obtained by the sensors inside the workshop contain too many useless signals and noise, which will affect the accuracy of the prediction; secondly, the equipment operation data is multidimensional, and different features may contribute differently to the health status of the equipment in different time periods. The existing technology ignores the importance of different features, which will lead to reduced prediction accuracy. Summary of the invention

[0005] Purpose of the invention: In order to solve the deficiencies in the prior art, the present application proposes a method for predicting the remaining service life of chemical rotating equipment, which aims to predict the remaining service life of chemical rotating equipment and solve the problem of inaccurate prediction in the prior art, which affects normal production.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for predicting the remaining service life of chemical rotating equipment comprises the following steps:

[0008] Step 1: Obtain the full life operation status data of the rotating equipment through the monitoring equipment (sensors, etc.) in the workshop;

[0009] Step 2: Build a GRU-Attention model, input the data obtained in step 1 into the model for training, and output the full life degradation curve of the rotating equipment;

[0010] Step 3: The real-time operating status data of the rotating equipment is preprocessed and input into the trained GRU-Attention model to predict the remaining service life and provide support for maintenance personnel to make maintenance decisions.

[0011] Step 1 includes the following steps:

[0012] Step 1.1: Based on the sensors inside the workshop, collect the equipment's full life cycle operation status data, including the characteristic data monitored by the sensors and the equipment degradation status data;

[0013] Step 1.2: Use wavelet transform to remove noise from the equipment life cycle operation status data. By performing multi-scale decomposition on the original signal, the noise component is effectively separated and the key equipment degradation signal is retained.

[0014] The continuous wavelet transform of a signal f(t) is expressed as:

[0015]

[0016] Where W(a,b) is the wavelet coefficient, t is time, a is the scale parameter, b is the translation parameter, ψ(t) is the complex conjugate of the wavelet basis function, and d is the differential sign;

[0017] The signal is split into high-frequency and low-frequency parts through wavelet decomposition, and the high-frequency part is threshold processed to remove the noise; finally, the denoised signal is reconstructed using inverse wavelet transform;

[0018] Step 1.3: Perform Z-SCORE standardization on the denoised data. The specific formula is:

[0019]

[0020] Among them, Z is the standardized data, X is the original data, μ is the mean of the data, and σ is the standard deviation of the data;

[0021] Step 1.4: Calculate the Pearson correlation coefficient between the standardized characteristic data of each sensor and the equipment degradation status data. The calculation formula of the correlation coefficient r is:

[0022]

[0023] Among them, Z i is the i-th sample in the characteristic data of the sensor, is the mean of variable Z, Y i is the i-th sample in the equipment degradation state data, is the mean of variable Y;

[0024] By setting the correlation threshold τ, feature data with |r|>τ is screened out, where τ is generally set to 0.5;

[0025] Step 1.5: Use the sliding window technology to slice the equipment operation status data to construct the training samples of the model: by setting a time window of fixed length, sliding the time window on the entire time series data, and generating training samples containing continuous time steps.

[0026] In step 2, the GRU-Attention model includes a GRU layer and an attention mechanism. Step 2 specifically includes the following steps:

[0027] Step 2.1: Construct a GRU layer (Gated Recurrent Unit, GRU) to capture the time-dependent characteristics of the device degradation process. The GRU layer controls the flow of information through the update gate and reset gate. The GRU layer receives the full life data of the device as input and outputs the hidden state of each time step. Each hidden state h t Both contain the operating characteristics of the device at the current moment;

[0028] Step 2.2: Build an attention mechanism and use the Efficient Channel Attention (ECA) module to identify different patterns in the time series and assign weights, thereby dynamically focusing on the key time steps in the device degradation process;

[0029] Step 2.3: The hidden state h′ enhanced by the GRU layer and the efficient channel attention mechanism module t The data is passed into the multi-layer perceptron MLP, and nonlinear mapping is performed through the fully connected layer to generate the final prediction result.

[0030] Step 2.4: After building the GRU-Attention model, use the training samples obtained in step 1 as input for training. When training the model, the mean square error (MSE) is used as the loss function to measure the error between the model's predicted value and the actual value. The formula is:

[0031]

[0032] Where n is the total number of samples, is the predicted value, y i is the actual value;

[0033] Update the model parameters through back propagation, gradually reducing the error on the training samples;

[0034] Adam optimizer is used to optimize model parameters;

[0035] After the training, the optimal weights and the full life degradation curve of the rotating equipment are obtained.

[0036] In step 2.1, the formula for updating the gate is:

[0037] z t =σ(W z ·[x t ,h t-1 ]+b z )

[0038] Among them, z t is the output of the update gate, σ is the Sigmoid activation function, x t is the current step input, h t-1 is the hidden state of the previous time step, W z is the weight matrix of the update gate, b z is the bias of the update gate.

[0039] In step 2.1, the formula for resetting the gate is:

[0040] rt =σ(W r ·[x t ,h t-1 ]+b r )

[0041] Among them, r t is the output of the reset gate, W r is the weight matrix of the reset gate, b r To reset the gate bias.

[0042] In step 2.1, the candidate hidden state is calculated using the following formula:

[0043]

[0044] in, is the candidate hidden state, tanh is the hyperbolic tangent activation function, W h is the weight matrix of the candidate hidden state, b h is the bias of the candidate hidden state.

[0045] In step 2.1, the final hidden state is calculated using the following formula:

[0046]

[0047] Among them, h t The final hidden state.

[0048] Step 2.2 includes:

[0049] First, global information is extracted from the hidden state of the GRU layer through global pooling, and global features are directly retained:

[0050]

[0051] in, is the hidden state of the GRU layer after global pooling, h t is the hidden state of the GRU layer, and AvgPool is the global pooling operation;

[0052] Then, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels:

[0053]

[0054] Where C is the number of channels, γ is the scaling factor, and b is the offset. To round up to the nearest odd number;

[0055] Through 1D convolution operation, the local interaction relationship between channels is dynamically captured, and the attention weight w of the i-th channel is i for:

[0056]

[0057] Among them, σ is the Sigmoid activation function, α j is the jth weight of the convolution kernel, is the value of the jth adjacent channel of the i-th value after global average pooling;

[0058] The one-dimensional convolution result is mapped to [0,1] through the Sigmoid activation function, indicating the relative importance of each channel:

[0059]

[0060] Among them, W′ channel is the generated channel weight matrix, Conv1D is a one-dimensional convolution operation;

[0061] DropAttention regularization is introduced to randomly discard some attention weights to alleviate the overfitting risk introduced by the attention mechanism:

[0062] W channel = Dropout(W′ channel )

[0063] Among them, Dropout is a regularization operation; W channel represents the attention weight;

[0064] Finally, the attention weight W channel With the hidden state h t Multiply channel by channel to generate the enhanced hidden state h′ t :

[0065] h′ t =W channel ⊙h t

[0066] Among them, ⊙ is the element-by-element multiplication.

[0067] Step 3 includes the following steps:

[0068] Step 3.1: Based on the internal sensors of the workshop, the operating status data of the equipment is obtained in real time. The operating status data includes temperature, vibration and rotation speed. First, outliers are removed and missing values ​​are filled in the operating status data. Then, the noise is removed by wavelet transform and standardized by Z-SCORE.

[0069] Step 3.2: Use a sliding window to segment the data preprocessed in step 3.1 according to the time window size, generate time series blocks suitable for model input, and then input them into the trained model to achieve real-time prediction of the remaining service life of the equipment.

[0070] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0071] The present invention has the following beneficial effects: Compared with the prior art, the present invention adopts the GRU-Attention model, combined with the trend characteristics of the equipment degradation process in the time series, and can capture the time dependency and key degradation information in the equipment operation. On the one hand, the gating mechanism of the GRU model helps to deal with long-term and short-term dependencies, and the Attention mechanism focuses on important time points in the equipment operation process, and the prediction accuracy of the remaining service life of the equipment is effectively improved; at the same time, adding regularization after ECA can effectively alleviate the overfitting problem that may be caused by the attention mechanism; compared with other attention mechanisms, the computational complexity of the ECA module is lower, and it is more applicable in embedded devices or edge computing environments; on the other hand, the feature data highly related to the equipment degradation trend is screened out through the Pearson correlation coefficient, redundant features are removed, and the input data dimension is reduced, so that the model can process multiple sensor data more efficiently in a complex industrial environment. At the same time, wavelet transform denoising and Z-SCORE standardization processing ensure the stability of data quality, so that the model can still maintain good prediction performance under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0073] Figure 1 It is an overall schematic diagram of the present invention.

[0074] Figure 2 It is a schematic diagram of the GRU cycle unit structure.

[0075] Figure 3 It is a schematic diagram of the GRU-Attention structure. DETAILED DESCRIPTION

[0076] Reference Figure 1 The embodiment of the present invention provides a method for predicting the remaining service life of chemical rotating equipment, which specifically includes:

[0077] Step 1: Obtain the full life cycle data of rotating equipment, including:

[0078] Step 1.1: Based on the sensors inside the workshop, collect the status data of the equipment during its entire life cycle.

[0079] Step 1.2: Use wavelet transform to remove noise from equipment monitoring data. By performing multi-scale decomposition on the original signal, the noise component is effectively separated and the key equipment degradation signal is retained.

[0080] For a signal f(t), its continuous wavelet transform can be expressed as:

[0081]

[0082] Where W(a,b) is the wavelet coefficient, t is the time, a is the scale parameter representing the frequency component of the analysis, b is the translation parameter representing the position of the wavelet in the time domain, ψ(y) is the complex conjugate of the wavelet basis function, and d is the differential sign.

[0083] The signal is split into high-frequency and low-frequency parts through wavelet decomposition. The high-frequency part mainly contains noise information, while the low-frequency part retains the main characteristics of the signal. Then the high-frequency part is thresholded to remove the noise. Finally, the denoised signal is reconstructed using the inverse wavelet transform.

[0084] Step 1.3: Perform Z-SCORE standardization on the denoised data to eliminate the impact of differences in sensor units, ensure that different features have the same scale during model training, and prevent some feature values ​​from being too large or too small and causing an unbalanced impact on the model. The specific formula is as follows:

[0085]

[0086] Among them, Z is the standardized data, X is the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0087] Step 1.4: Calculate the Pearson correlation coefficient for the standardized characteristic data of each sensor and the equipment degradation status data. The Pearson correlation coefficient is an indicator to measure the linear correlation between two variables. Its value range is [-1,1]. Values ​​close to 1 indicate positive correlation, values ​​close to -1 indicate negative correlation, and values ​​close to 0 indicate no correlation. The calculation formula of the Pearson correlation coefficient is as follows:

[0088]

[0089] Among them, r is the correlation coefficient, Z i is the i-th sample in the characteristic data of the sensor, is the mean of variable Z, Y i is the i-th sample in the equipment degradation state data, is the mean of the variable Y.

[0090] By setting the correlation threshold τ, feature data with |r|>τ are screened out, and features with higher correlation are retained for model training.

[0091] Step 1.5: Use the sliding window technology to slice the equipment monitoring data to construct the training samples of the model. By setting a fixed-length time window l = 24, corresponding to using the first 24 time steps to predict the value of the next time step, sliding the window over the entire time series data, and generating training samples containing continuous time steps.

[0092] Step 2: Build the GRU-Attention model, such as Figure 3 As shown in the figure, the prediction model structure is input layer, GRU layer, attention mechanism layer, and fully connected layer. The GRU layer is used to extract long-term features of the input data, and the attention module focuses on key information of the current task. The data obtained in step 1 is input into the model for training, and the full life degradation curve of the rotating equipment is output, which includes:

[0093] Step 2.1: Construct the GRU layer, Figure 2 Schematic diagram of the GRU cycle unit structure, where z t The update gate determines the new information stored in the cell state and updates the old cell state; r t It is the reset gate that determines the information output by the cell. It captures the time-dependent characteristics of the device degradation process. GRU controls the flow of information through two gates (update gate and reset gate). The main formula is as follows:

[0094] Update Gate:

[0095] z t =σ(W z ·[x t ,h t-1 ]+b z )

[0096] Among them, z t is the output of the update gate, σ is the Sigmoid activation function, x t is the current step input, h t-1 is the hidden state of the previous time step, W z is the weight matrix of the update gate, b z is the bias of the update gate.

[0097] Reset the gate:

[0098] r t =σ(W r ·[x t ,h t-1 ]+b r )

[0099] Among them, r t is the output of the reset gate, σ is the Sigmoid activation function, xt is the current step input, h t-1 is the hidden state of the previous time step, W r is the weight matrix of the reset gate, b r To reset the gate bias.

[0100] Calculation of candidate hidden states:

[0101]

[0102] in, is the candidate hidden state, tanh is the hyperbolic tangent activation function, r t is the output of the reset gate, x t is the current step input, h t-1 is the hidden state of the previous time step, W h is the weight matrix of the candidate hidden state, b h is the bias of the candidate hidden state.

[0103] Final hidden state:

[0104]

[0105] Among them, g t is the final hidden state, z t is the output of the update gate, is a candidate hidden state, x t is the current step input, h t-1 is the hidden state at the previous time step.

[0106] First, the input x of the current step t and the hidden state h of the previous time step t-1 As input into the update gate, the output is mapped to [0,1] through the Sigmoid activation function to control the amount of information that needs to be transmitted between the current time step and the previous time step. The closer it is to 1, the more information needs to be transmitted. The reset gate controls which information of the previous hidden state needs to be forgotten when generating candidate hidden states. The output is mapped to [0,1] through the Sigmoid activation function. The closer it is to 0, the more historical information is forgotten. Combined with the current step input x t and the reset hidden state h t-1 Compute candidate hidden states Characterize the important features of the current moment. Finally, the hidden state h of the previous time step t-1 With candidate hidden states Weighted average to calculate the final hidden state h t .

[0107] The GRU layer receives the equipment lifecycle data as input and outputs the hidden state at each time step. Each hidden state ht Both contain the operating characteristics of the device at the current moment.

[0108] Step 2.2: Construct an attention mechanism and use the Efficient Channel Attention (ECA) module to identify different patterns in the time series and assign weights, thereby dynamically focusing on the key time steps in the device degradation process.

[0109] First, global information is extracted from the hidden state of the GRU layer through global pooling, and the global features are directly retained.

[0110]

[0111] in, is the hidden state of the GRU layer after global pooling, h t is the hidden state of the GRU layer, and AvgPool is the global pooling operation.

[0112] Then, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels:

[0113]

[0114] Where C is the number of channels, γ is the scaling factor, and b is the offset. Round up to the nearest odd number.

[0115] Through 1D convolution operation, the local interaction relationship between channels is dynamically captured. For the i-th channel, its attention weight w i for:

[0116]

[0117] Among them, σ is the Sigmoid activation function, α j is the jth weight of the convolution kernel, is the value of the jth adjacent channel of the i-th value after global average pooling.

[0118] The one-dimensional convolution result is mapped to [0,1] through the Sigmoid activation function to indicate the relative importance of each channel:

[0119]

[0120] Among them, W′ channel is the generated channel weight matrix, σ is the Sigmoid activation function, Conv1D is the one-dimensional convolution operation, is the hidden state of the GRU layer after global pooling.

[0121] After that, DropAttention regularization is introduced to randomly discard some attention weights to alleviate the overfitting risk introduced by the attention mechanism:

[0122] W channel = Dropout(W′ channel )

[0123] Among them, Dropout is a regularization operation; W channel represents the attention weight.

[0124] Finally, the attention weight W channel With the hidden state h t Multiply channel by channel to generate the enhanced hidden state h′ t :

[0125] h′ t =W channel ⊙h t

[0126] Among them, ⊙ is the element-by-element multiplication.

[0127] Step 2.3: Multiply the hidden state obtained after the data is processed by the GRU layer with the weight obtained after the Attention layer is enhanced to obtain the enhanced hidden state h′ t The data is passed into the multi-layer perceptron MLP, and nonlinear mapping is performed through the fully connected layer to generate the final prediction result.

[0128] Step 2.4: After building the GRU-Attention model, set the network training parameters, the number of units in the hidden layer is 128, the batch size is 32, and the learning rate is 0.001. Use the training samples obtained in step 1 as input for training. When training the model, the mean square error (MSE) is used as the loss function to measure the error between the model's predicted value and the actual value. The formula is as follows:

[0129]

[0130] Where n is the total number of samples, is the predicted value, y i is the actual value.

[0131] The model parameters are updated through back propagation, gradually reducing the error on the training samples.

[0132] In addition, the Adam optimizer is used to optimize the model parameters. Adam combines momentum and adaptive learning rate adjustment mechanisms to accelerate the convergence of the model while avoiding falling into local optimality. Finally, after the training is completed, the optimal weights and the degradation curve of the rotating equipment over its entire life are obtained.

[0133] Step 3: The real-time rotating equipment operating status data is preprocessed and input into the trained GRU-Attention model to predict the remaining service life and provide support for maintenance personnel to make maintenance decisions, including:

[0134] Step 3.1: Based on the sensors inside the workshop, the operating status data of the equipment is obtained in real time, including parameters such as temperature, vibration, and speed. First, the data is subjected to outlier removal and missing value filling. After that, the noise is removed using wavelet transform and standardized using Z-SCORE.

[0135] Step 3.2: Use a sliding window to segment the data preprocessed in step 3.1 according to the time window size, generate time series blocks suitable for model input, and then input them into the trained model to achieve real-time prediction of the remaining service life of the equipment.

[0136] The present invention provides a method for predicting the remaining service life of a chemical rotating equipment. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A method for predicting the remaining service life of chemical rotating equipment, characterized in that: The following steps are involved: Step 1: Obtain the full life cycle operation status data of the rotating equipment through the monitoring equipment in the workshop; Step 2: Build a GRU-Attention model, input the data obtained in step 1 into the model for training, and output the full life degradation curve of the rotating equipment; Step 3: The real-time operating status data of the rotating equipment is preprocessed and input into the trained GRU-Attention model to predict the remaining service life and provide support for maintenance personnel to make maintenance decisions.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Step 1.1: Based on the sensors inside the workshop, collect the equipment's full life cycle operation status data, including the characteristic data monitored by the sensors and the equipment degradation status data; Step 1.2: Use wavelet transform to remove noise from the equipment life cycle operation status data. By performing multi-scale decomposition on the original signal, the noise component is effectively separated and the key equipment degradation signal is retained. The continuous wavelet transform of a signal f(t) is expressed as: Where W(a, b) is the wavelet coefficient, t is time, a is the scale parameter, b is the translation parameter, ψ(t) is the complex conjugate of the wavelet basis function, and d is the differential sign; The signal is split into high-frequency and low-frequency parts through wavelet decomposition, and the high-frequency part is threshold processed to remove the noise; finally, the denoised signal is reconstructed using inverse wavelet transform; Step 1.3: Perform Z-SCORE standardization on the denoised data. The specific formula is: Among them, Z is the standardized data, X is the original data, μ is the mean of the data, and σ is the standard deviation of the data; Step 1.4: Calculate the Pearson correlation coefficient between the standardized characteristic data of each sensor and the equipment degradation status data. The calculation formula of the correlation coefficient r is: Among them, Z i is the i-th sample in the characteristic data of the sensor, is the mean of variable Z, Y i is the i-th sample in the equipment degradation state data, is the mean of variable Y; By setting the correlation threshold τ, feature data with |r|>τ is screened out, where τ is generally set to 0.5; Step 1.5: Use the sliding window technology to slice the equipment operation status data to construct the training samples of the model: by setting a time window of fixed length, sliding the time window on the entire time series data, and generating training samples containing continuous time steps.

3. The method according to claim 2, characterized in that In step 2, the GRU-Attention model includes a GRU layer and an attention mechanism. Step 2 specifically includes the following steps: Step 2.1: Construct a GRU layer to capture the time-dependent features of the equipment degradation process. The GRU layer controls the information flow through the update gate and the reset gate. The GRU layer receives the equipment life data as input and outputs the hidden state of each time step. Each hidden state h t Both contain the operating characteristics of the device at the current moment; Step 2.2: Build an attention mechanism and use an efficient channel attention mechanism module to identify different patterns in the time series and assign weights, thereby dynamically focusing on the key time steps in the device degradation process; Step 2.3: The hidden state h′ enhanced by the GRU layer and the efficient channel attention mechanism module t The data is passed into the multi-layer perceptron MLP, and nonlinear mapping is performed through the fully connected layer to generate the final prediction result. Step 2.4: After building the GRU-Attention model, use the training samples obtained in step 1 as input for training. When training the model, the mean square error (MSE) is used as the loss function to measure the error between the model's predicted value and the actual value. The formula is: Where n is the total number of samples, is the predicted value, y i is the actual value; Update the model parameters through back propagation, gradually reducing the error on the training samples; Adam optimizer is used to optimize model parameters; After the training, the optimal weights and the full life degradation curve of the rotating equipment are obtained.

4. The method according to claim 3, characterized in that In step 2.1, the formula for updating the gate is: z t =σ(W z ·[x t ,h t-1 ]+b z ) Among them, z t is the output of the update gate, σ is the Sigmoid activation function, x t is the current step input, h t-1 is the hidden state of the previous time step, W z is the weight matrix of the update gate, b z is the bias of the update gate.

5. The method according to claim 4, characterized in that In step 2.1, the formula for resetting the gate is: r t =σ(W r ·[x t ,h t-1 ]+b r ) Among them, r t is the output of the reset gate, W r is the weight matrix of the reset gate, b r To reset the gate bias.

6. The method according to claim 5, characterized in that In step 2.1, the candidate hidden state is calculated using the following formula: in, is the candidate hidden state, tanh is the hyperbolic tangent activation function, W h is the weight matrix of the candidate hidden state, b h is the bias of the candidate hidden state.

7. The method according to claim 6, characterized in that In step 2.1, the final hidden state is calculated using the following formula: Among them, h t The final hidden state.

8. The method according to claim 7, characterized in that Step 2.2 includes: First, global information is extracted from the hidden state of the GRU layer through global pooling, and global features are directly retained: in, is the hidden state of the GRU layer after global pooling, h t is the hidden state of the GRU layer, and AvgPool is the global pooling operation; Then, the kernel size k of the one-dimensional convolution is adaptively calculated according to the number of channels: Where C is the number of channels, γ is the scaling factor, and b is the offset. To round up to the nearest odd number; Through 1D convolution operation, the local interaction relationship between channels is dynamically captured, and the attention weight w of the i-th channel is i for: Among them, σ is the Sigmoid activation function, α j is the jth weight of the convolution kernel, is the value of the jth adjacent channel of the i-th value after global average pooling; The one-dimensional convolution result is mapped to [0, 1] through the Sigmoid activation function, indicating the relative importance of each channel: Among them, W′ channel is the generated channel weight matrix, Conv1D is a one-dimensional convolution operation; DropAttention regularization is introduced to randomly discard some attention weights to alleviate the overfitting risk introduced by the attention mechanism: W channel =Dropout(W′ chaannel ) Among them, Dropout is a regularization operation; W channel represents the attention weight; Finally, the attention weight W channel With the hidden state h t Multiply channel by channel to generate the enhanced hidden state h′ t : h′ t =W channel ⊙h t Among them, ⊙ is the element-by-element multiplication.

9. The method according to claim 8, characterized in that Step 3 includes the following steps: Step 3.1: Based on the internal sensors of the workshop, the operating status data of the equipment is obtained in real time. The operating status data includes temperature, vibration and rotation speed. First, outliers are removed and missing values ​​are filled in the operating status data. Then, the noise is removed by wavelet transform and standardized by Z-SCORE. Step 3.2: Use a sliding window to segment the data preprocessed in step 3.1 according to the time window size, generate time series blocks suitable for model input, and then input them into the trained model to achieve real-time prediction of the remaining service life of the equipment.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 9.

Images