MCAT residual life prediction method and system for mechanical equipment

Abstract

The invention relates to an MCAT residual life prediction method and system for mechanical equipment, and the method comprises the steps: constructing a historical operation full-life-cycle data set, carrying out the preprocessing of the historical operation full-life-cycle data set, and carrying out the division to obtain a training set; constructing a residual life prediction model; training a residual life prediction model by using the training set, constructing a loss function in each round of training, and adjusting parameters of the residual life prediction model according to the loss function to obtain a trained residual life prediction model; and receiving monitoring data streams output by the mechanical equipment to be detected detected by the multi-dimensional sensor in real time, preprocessing the monitoring data streams, inputting the preprocessed monitoring data streams into the trained residual life prediction model to obtain a residual life prediction value of the mechanical equipment to be detected at the current moment, and then performing graded early warning according to the residual life prediction value. The method is used for predicting the residual life of the mechanical equipment, and the prediction precision and robustness are high.

Description

Technical Field

[0001] This invention relates to the field of remaining service life prediction technology for mechanical equipment, and particularly to a method and system for predicting the remaining service life of mechanical equipment using the MCAT (Mechanical, Technical, and Availability) method. Background Technology

[0002] With the rapid advancement of modern industry and infrastructure construction, large-scale mechanical equipment such as engineering machinery and heavy mining machinery operate under harsh and complex conditions for extended periods. The health status of their key components directly impacts production safety and economic efficiency. Therefore, accurate remaining life prediction is crucial for achieving condition-based maintenance and reducing operation and maintenance costs. Although traditional models based on physical failure mechanisms are widely used in simple systems, they often struggle to establish accurate degradation models for complex and highly coupled mechanical equipment. In recent years, data-driven methods, represented by Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), have become mainstream due to their powerful nonlinear feature extraction capabilities. However, they still have significant limitations in practical applications: most existing models employ single-scale feature extraction methods, making it difficult to simultaneously consider both short-term fault impacts and long-term degradation trends; they also lack effective adaptive key feature selection mechanisms, making them highly susceptible to environmental noise interference from multi-sensor data; furthermore, traditional networks are insufficient in capturing long-distance temporal dependencies and dynamically aggregating information across the entire lifecycle, resulting in poor stability and robustness of prediction results. Summary of the Invention

[0003] This invention provides a method and system for predicting the remaining useful life (MCAT) of mechanical equipment, in order to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the technical solution of the present invention is implemented as follows: This invention provides a method for predicting the remaining useful life (MCAT) of mechanical equipment, comprising the following steps: S1. Collect multi-dimensional sensor monitoring data during the operation of mechanical equipment, and then construct a historical full life cycle dataset based on this data. The historical full life cycle dataset is then preprocessed and divided into training sets. S2. Construct a remaining lifetime prediction model, which includes a multi-scale local feature extraction module, a feature adaptive enhancement module, a global temporal coding module, and a temporal attention prediction module connected in sequence. S3. Train the remaining lifetime prediction model using the training set, and construct a loss function in each round of training. Adjust the parameters of the remaining lifetime prediction model according to the loss function until the remaining lifetime prediction model converges, and obtain the trained remaining lifetime prediction model. S4. Receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, preprocess it, and input it into the trained remaining life prediction model to obtain the remaining life prediction value of the mechanical equipment under test at the current time. Then, perform graded early warning based on the remaining life prediction value.

[0005] Furthermore, step S1 specifically includes the following steps: S11. Collect the operating status data of the mechanical equipment throughout its entire life cycle using multi-dimensional sensors installed on the mechanical equipment, and then construct a historical operating life cycle dataset based on this data. The historical operating life cycle dataset includes the original multi-dimensional monitoring sequence matrix of multiple equipment units. Among them, the The original multidimensional monitoring sequence matrix of each equipment unit The expression is as follows: ; In the formula, Indicates the first The total number of time steps before a single equipment unit completely fails; Indicates the first Sensor feature vectors at each time step, and ; This represents the total dimension of the features monitored by the sensor; Indicates the first The time step The physical measurement values ​​of each sensor; T Indicates transpose; S12. Perform Z-Score standardization on the data in the original multidimensional monitoring sequence matrix of each equipment unit to generate multiple standard normal distribution sequences; S13. For each standard normal distribution sequence, a preset window length is used. and sliding step size The standard normal distribution sequence is segmented to generate multiple time-series samples of fixed length. ; S14. For each time series sample Construct each time series sample True remaining life label ; S15, Collect all time series samples and the corresponding actual remaining life tag by The data is summarized in the form of a preprocessed historical runtime dataset. S16. Then, according to the set ratio, the preprocessed historical running lifecycle dataset is divided into training set and test set.

[0006] Furthermore, the expression for Z-Score normalization in S12 is as follows: ; in, Represents the standardized physical measurement values ; This represents the first time step in the historical running full lifecycle dataset. The average of the physical measurements from each sensor; This represents the first time step in the historical running full lifecycle dataset. The standard deviation of the physical measurements of each sensor; The expression for the standard normal distribution sequence in S12 is: ; in, This represents the standardized original multidimensional monitoring sequence matrix, i.e., the standard normal distribution sequence; Represents the standardized sensor feature vector ; The time-series samples in S13 The expression is as follows: ; in, Representing time series samples The start time index; This represents the time step index in the entire lifecycle sequence; Represents the standardized sensor feature vector ; Represents the standardized sensor feature vector ; The actual remaining life tag in S14 The expression is as follows: ; in, This is the moment when the mechanical equipment fails. This indicates the last time step of the sample window; This is the preset degradation saturation threshold.

[0007] Furthermore, the multi-scale local feature extraction module includes three parallel convolutional neural networks connected in sequence and a concatenation module, wherein each parallel convolutional neural network includes a convolutional kernel. A one-dimensional convolutional layer and a Gaussian error linear unit activation function; The feature adaptive enhancement module includes a global average pooling layer, a fully connected layer, a linear rectified unit, a fully connected layer, a Sigmoid activation function, and a fusion module connected in sequence. The global timing coding module includes sequentially connected... Convolutional layers, positional encoding modules, and Transformer encoders; the Transformer encoder consists of N stacked Transformer encoding layers; The time attention prediction module includes an attention score calculation module, a summation module, and a regression network connected in sequence.

[0008] Furthermore, each parallel convolutional neural network in the multi-scale local feature extraction module is used to extract multi-scale feature maps. The expression is as follows: ; in, Represents the input feature tensor; Indicates that the convolution kernel is Convolution operations in a one-dimensional convolutional layer; Represents the activation function of the Gaussian error linear unit; Represents the set of real numbers; The number of convolution channels; The stitching module is used to fuse feature maps at multiple scales to obtain a multi-scale fused feature tensor. The expression is as follows: ; in, These represent feature maps at three different scales; Indicates a splicing operation; The global average pooling layer in the feature adaptive enhancement module is used to fuse multi-scale feature tensors. The spatial dimension is compressed to generate a channel statistics vector. The expression is as follows: ; in, Represents the multi-scale fusion feature tensor In the Channel, First The numerical value at any given moment; The fully connected layer is used to learn the nonlinear interactions between channels and generate channel weight vectors. The expression is as follows: ; in, and Here are the weight matrices for the two fully connected layers; It is the ReLU activation function; Use the Sigmoid activation function; The fusion module is used to convert the channel weight vector With multi-scale fusion feature tensor Element-wise multiplication yields the enhanced feature tensor. The expression is as follows: ; in, This represents element-wise multiplication based on the channel dimension.

[0009] Furthermore, in the global timing coding module Convolutional layers are used to enhance the feature tensors The channel dimension from Projected onto model dimension , to obtain projection features ; The position encoding module is used for projection features Positional encoding is introduced to preserve the sequential information of the time series, thereby obtaining the feature sequence. The expression is as follows: ; in, The position encoding matrix. No. Line 1 Column and number Line 1 The elements of the column are calculated as follows: ; ; in, Indicates the position encoding matrix of the first... Line 1 Column elements; The Transformer encoder is used to mine temporal dependencies using a multi-head self-attention mechanism, resulting in a deep feature matrix containing global temporal context information. , of which The operation process of the Transformer encoding layer is as follows: ; ; ; in, , , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the first l -1 layer feature sequence; This represents multi-head attention computation; This represents a feedforward neural network; Representation layer normalization; Indicates the first The feature sequence of the layer; Indicates the first The intermediate feature sequence after layer normalization in the Transformer encoding layer; The attention score calculation module is used to calculate the score at each time step. Attention score The calculation formula is as follows: ; ; in, Deep feature matrix The Middle Feature vectors at each time step; , , All of these are learnable attention network parameters; Indicates the first Attention intermediate hidden layer vectors at each time step; The summation module is used to perform a weighted summation of the feature vectors at all time steps to obtain the global context vector. The expression is as follows: ; in, This represents the time step after attention weighting. eigenvectors; The regression network is used to base on the global context vector. Predicted remaining life value The expression is as follows: ; in, and The weights and biases of the regression network.

[0010] Furthermore, step S3 specifically includes the following steps: S31. Input a batch of training samples from the training set into the remaining lifetime prediction model, perform forward propagation calculations, and obtain the remaining lifetime prediction value for the current batch. ; S32, then the remaining lifetime prediction value Combined with the actual remaining life label of the current batch Calculate the loss value of the current loss function. ; S33. Calculate the loss function based on the backpropagation algorithm. Parameters of each layer in the remaining life prediction model The gradient vector is then processed by norm clipping. S34. The AdamW optimization algorithm is used, and the parameters of the remaining lifetime prediction model are adjusted based on the pruned gradient vector. Perform an update, determine whether the predetermined iteration stopping condition has been met, and if so, return the trained remaining lifetime prediction model; otherwise, return to S31. S35. Use the test set to verify and evaluate the remaining life prediction model during the training process, record the prediction error of each training cycle, and retain the model parameters with the smallest error on the test set as the final model parameters; otherwise, return to S31 to continue training.

[0011] Furthermore, the loss function in S3 is: ; in, B This refers to the number of training samples within the same batch. i Indicates the first i One training sample; Indicates smoothness L 1. Loss function, calculated as follows: ; in, This indicates the prediction error.

[0012] Furthermore, step S4 specifically includes the following steps: S41. Deploy the trained remaining lifetime prediction model to the MCAT remaining lifetime prediction system. S42. Utilize a data acquisition card to receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, and load the average value. Standard deviation As a preset parameter, whenever a new sensor feature vector is received at a time step... Immediately apply preset parameters to the new sensor feature vector. Perform online normalization mapping to calculate new normalized vectors. ; S43, The length of the real-time window built based on the FIFO queue is L A first-in, first-out (FIFO) data cache queue, whenever a new normalized vector... The MCAT remaining lifetime prediction system will generate a new standardized vector. Push the data to the end of the queue and remove the oldest data from the head of the queue. This allows for dynamic updates to the current real-time time series sample matrix. ; S44. The current real-time time series sample matrix... The input is fed into the trained remaining life prediction model, which outputs the predicted remaining life of the machine under test at the current time. ; S45, MCAT remaining life prediction system based on remaining life prediction values Implement tiered early warning systems.

[0013] In another aspect, the present invention provides an MCAT remaining lifetime prediction system, configured to execute the above-described MCAT remaining lifetime prediction method, the MCAT remaining lifetime prediction system comprising: The data processing module is used to acquire multidimensional sensor data and construct time-series samples; The model storage module contains a built-in trained remaining lifespan prediction model. The lifespan monitoring and alarm module is used to provide graded early warnings based on the remaining lifespan prediction values ​​output by the trained remaining lifespan prediction model.

[0014] The beneficial effects of this invention are: This invention discloses an MCAT (Mean Time Acquisition) remaining useful life prediction method for mechanical equipment. The method incorporates a remaining useful life prediction model, which includes a multi-scale local feature extraction module, a feature adaptive enhancement module, a global temporal coding module, and a temporal attention prediction module connected in sequence. Compared with the existing single-scale feature extraction module, the multi-scale local feature extraction module can simultaneously take into account both short-term fault impact and long-term degradation trend, thereby improving the feature extraction capability of this invention. The feature adaptive enhancement module in this invention has a built-in key feature adaptive filtering mechanism, which can reconstruct the feature channels with weights, thereby improving the environmental noise resistance of the remaining lifetime prediction model.

[0015] In addition, the global temporal coding module has a built-in Transformer encoder, which can use a multi-head self-attention mechanism to mine long-distance temporal dependencies. Meanwhile, the temporal attention prediction module can dynamically aggregate deep feature matrices containing global temporal context information, which improves the prediction stability and robustness of the remaining lifetime prediction model. Attached Figure Description

[0016] Figure 1 This is a structural block diagram of the remaining lifetime prediction model in this invention; Figure 2 This is a schematic diagram of the remaining lifetime prediction results in an embodiment of the present invention. Detailed Implementation

[0017] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many other different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0018] Reference Figure 1 This application provides a method for predicting the remaining life (MCAT) of mechanical equipment, comprising the following steps: S1. Install multi-dimensional sensors in key parts of mechanical equipment (such as drive end bearings, fan end bearings, etc.) to collect monitoring data from multi-dimensional sensors (such as vibration sensors, pressure sensors, temperature sensors, etc.) during the operation of the mechanical equipment. Then, construct a historical full life cycle dataset based on this data. Then, preprocess the historical full life cycle dataset to eliminate the differences in the dimensions of different sensors, and then divide it into training sets. S2. Construct a remaining lifetime prediction model (Multi-scale Convolution AttentionTransformer, MCAT), which includes a multi-scale local feature extraction module, a feature adaptive enhancement module, a global temporal coding module, and a temporal attention prediction module connected in sequence. S3. Train the remaining lifetime prediction model using the training set, and construct a loss function in each round of training. Adjust the parameters of the remaining lifetime prediction model according to the loss function until the remaining lifetime prediction model converges, and obtain the trained remaining lifetime prediction model. S4. Receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, preprocess it, and input it into the trained remaining life prediction model to obtain the remaining life prediction value of the mechanical equipment under test at the current time. The remaining life prediction value can provide a basis for the condition-based maintenance decision of the mechanical equipment, and then perform graded early warning based on the remaining life prediction value.

[0019] In some embodiments, S1 specifically includes the following steps: S11. Install multi-dimensional sensors on key parts of mechanical equipment, collect operating status data of the mechanical equipment throughout its entire life cycle using the multi-dimensional sensors installed on the mechanical equipment, convert the continuous analog signals output by the multi-dimensional sensors into discrete time series, and then construct a historical operating life cycle dataset based on this. The historical operating life cycle dataset includes the original multi-dimensional monitoring sequence matrix of multiple equipment units. Among them, the The original multidimensional monitoring sequence matrix of each equipment unit The expression is as follows: ; In the formula, Indicates the first The total number of time steps before a single equipment unit completely fails; Indicates the first Sensor feature vectors at each time step, and ; The total dimension representing the sensor-monitored features (i.e., the number of channels in the input data, corresponding to the number of sensors). Indicates the first The time step The physical measurement values ​​of each sensor; T Indicates transpose; S12. To eliminate the differences in physical dimensions of different sensors, the data in the original multidimensional monitoring sequence matrix of each equipment unit are subjected to Z-Score standardization to generate multiple standard normal distribution sequences. S13. For each standard normal distribution sequence, a preset window length is used. and sliding step size The standard normal distribution sequence is segmented to generate multiple time-series samples of fixed length. ; S14. For each time series sample Each time series sample is constructed using a piecewise linear function. True remaining life label ; S15, Collect all time series samples and the corresponding actual remaining life tag by The data is summarized in the form of a preprocessed historical runtime dataset. S16. Then, according to the set ratio, the preprocessed historical running lifecycle dataset is divided into training set and test set.

[0020] In some embodiments, the expression for Z-Score normalization in S12 is as follows: ; in, Represents the standardized physical measurement values ; This represents the first time step in the historical running full lifecycle dataset. The average of the physical measurements from each sensor; This represents the first time step in the historical running full lifecycle dataset. The standard deviation of the physical measurements of each sensor; The expression for the standard normal distribution sequence in S12 is: ; in, This represents the standardized original multidimensional monitoring sequence matrix, i.e., the standard normal distribution sequence; Represents the standardized sensor feature vector ; The time-series samples in S13 The expression is as follows: ; in, Representing time series samples The start time index; This represents the time step index in the entire lifecycle sequence; Represents the standardized sensor feature vector ; Represents the standardized sensor feature vector ; The actual remaining life tag in S14 The expression is as follows: ; in, This is the moment when the mechanical equipment fails. This indicates the last time step of the sample window; This is the preset degradation saturation threshold.

[0021] In some embodiments, the multi-scale local feature extraction module includes three parallel convolutional neural networks connected in sequence and a concatenation module, wherein each parallel convolutional neural network includes a convolutional kernel. The system consists of a one-dimensional convolutional layer and a Gaussian error linear unit activation function; a three-way parallel convolutional neural network is used to perform parallel convolution operations on the input to capture local degradation features at different time spans; the kernel sizes of the three-way parallel convolutional neural network are respectively... ; The feature adaptive enhancement module includes a global average pooling layer, a fully connected layer, a linear rectified unit, a fully connected layer, a sigmoid activation function, and a fusion module connected in sequence; the feature adaptive enhancement module is used to reconstruct the feature channels with weights based on the built-in key feature adaptive screening mechanism. The global timing coding module includes sequentially connected... Convolutional layers, positional encoding modules, and Transformer encoders; the Transformer encoder consists of N stacked Transformer encoding layers; The time attention prediction module includes an attention score calculation module, a summation module, and a regression network connected in sequence.

[0022] In some embodiments, each parallel convolutional neural network in the multi-scale local feature extraction module is used to extract multi-scale feature maps. The expression is as follows: ; in, Represents the input feature tensor; Indicates that the convolution kernel is Convolution operations in a one-dimensional convolutional layer; Represents the activation function of the Gaussian error linear unit; Represents the set of real numbers; The number of convolution channels; The stitching module is used to fuse feature maps at multiple scales to obtain a multi-scale fused feature tensor. The expression is as follows: ; in, These represent feature maps at three different scales; Indicates a splicing operation; The global average pooling layer in the feature adaptive enhancement module is used to fuse multi-scale feature tensors. The spatial dimension is compressed to generate a channel statistics vector. The expression is as follows: ; in, Represents the multi-scale fusion feature tensor In the Channel, First The numerical value at any given moment; The fully connected layer is used to learn the nonlinear interactions between channels and generate channel weight vectors. The expression is as follows: ; in, and Here are the weight matrices for the two fully connected layers; It is the ReLU activation function; Use the Sigmoid activation function; The fusion module is used to convert the channel weight vector With multi-scale fusion feature tensor Element-wise multiplication yields the enhanced feature tensor. The expression is as follows: ; in, This represents channel-wise multiplication, a step that enhances channels containing critical degradation information while effectively suppressing noisy channels.

[0023] In some embodiments, the global timing coding module Convolutional layers are used to enhance the feature tensors The channel dimension from Projected onto model dimension , to obtain projection features ; The position encoding module is used for projection features Positional encoding is introduced to preserve the sequential information of the time series, thereby obtaining the feature sequence. The expression is as follows: ; in, The position encoding matrix. No. Line 1 Column and number Line 1 The elements of the column are calculated as follows: ; ; in, Indicates the position encoding matrix of the first... Line 1 Column elements; The Transformer Encoder is used to mine long-distance temporal dependencies using a multi-head self-attention mechanism, obtaining a deep feature matrix containing global temporal context information. , of which The operation process of the Transformer encoding layer is as follows: ; ; ; in, , , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the first l -1 layer feature sequence; This represents multi-head attention computation; This represents a feedforward neural network; Representation layer normalization; Indicates the first The feature sequence of the layer; Indicates the first The intermediate feature sequence after layer normalization in the Transformer encoding layer; The attention score calculation module is used to calculate the score at each time step. Attention score The calculation formula is as follows: ; ; in, Deep feature matrix The Middle Feature vectors at each time step; , , All of these are learnable attention network parameters; Indicates the first Attention intermediate hidden layer vectors at each time step; The summation module is used to perform a weighted summation of the feature vectors at all time steps to obtain the global context vector. The expression is as follows: ; in, This represents the time step after attention weighting. eigenvectors; The regression network is used to base on the global context vector. Predicted remaining life value The expression is as follows: ; in, and The weights and biases of the regression network.

[0024] In some embodiments, S3 specifically includes the following steps: S31. Input a batch of training samples from the training set into the remaining lifetime prediction model, perform forward propagation calculations, and obtain the remaining lifetime prediction value for the current batch. ; S32, then the remaining lifetime prediction value Combined with the actual remaining life label of the current batch Calculate the loss value of the current loss function. ; S33. Calculate the loss function based on the backpropagation algorithm. Parameters of each layer in the remaining life prediction model The gradient vector is then processed by norm clipping; in this embodiment, the loss function... Select smooth L 1. Loss function (Smooth L1 Loss), compared to mean squared error (MSE), smoothing L 1. The loss function behaves as follows when the error is small: L 2-norm, convergent and smooth; exhibits characteristics when the error is large. L The 1 norm is insensitive to outliers and can effectively prevent gradient explosion. S34. The AdamW (Adaptive Moment Estimation with Weight Decay) optimization algorithm is used to optimize the parameters of the remaining lifetime prediction model based on the pruned gradient vector. Update; Specifically, in each round of training iteration, the batch of training samples is first input into the remaining lifetime prediction model for forward propagation calculation to obtain the remaining lifetime prediction value of the current batch, and the current loss function is calculated in combination with the true labels. The loss value is then calculated; subsequently, the loss function is calculated based on the backpropagation algorithm. Regarding the parameters of each layer of the model gradient vector To effectively prevent gradient explosion during deep network training, this embodiment performs norm clipping on the calculated gradient vector before parameter updates, i.e., the gradient's norm is reduced. L The 2-norm is limited to a preset threshold. Within; finally, the AdamW optimization algorithm is used to optimize the model parameters based on the clipped gradient vector. The model is updated, and a weight decay term is introduced during the update process to suppress model overfitting, thereby improving the generalization ability of the remaining life prediction model. After training is complete, determine whether the predetermined iteration stopping condition has been met. If so, return the trained remaining lifetime prediction model; otherwise, return to S31. S35. Use the test set to verify and evaluate the remaining life prediction model during the training process, record the prediction error of each training cycle, and retain the model parameters with the smallest error on the test set as the final model parameters; otherwise, return to S31 to continue training.

[0025] In some embodiments, the loss function in S3 is: ; in, B This refers to the number of training samples within the same batch. i Indicates the first i One training sample; Indicates smoothness L 1. Loss function, calculated as follows: ; in, This indicates the prediction error.

[0026] In some embodiments, S4 specifically includes the following steps: S41. Deploy the trained remaining lifetime prediction model to the MCAT remaining lifetime prediction system. S42. Utilize a data acquisition card to receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, and load the average value. Standard deviation As a preset parameter, whenever a new sensor feature vector is received at a time step... Immediately apply preset parameters to the new sensor feature vector. Perform online normalization mapping to calculate new normalized vectors. This step ensures that the distribution characteristics of the real-time online monitoring data stream are consistent with the distribution space of the model training data, avoiding prediction bias caused by data drift. S43. The real-time window construction length based on a FIFO (First-In-First-Out) queue is... L A first-in, first-out (FIFO) data cache queue, whenever a new normalized vector... The MCAT remaining lifetime prediction system will generate a new standardized vector. Push the data to the end of the queue and remove the oldest data from the head of the queue. This allows for dynamic updates to the current real-time time series sample matrix. This matrix always includes the current time and past times. Continuous state information at each moment; S44. The current real-time time series sample matrix... The input is fed into the trained remaining life prediction model, which outputs the predicted remaining life of the machine under test at the current time. This process requires no human intervention and can complete end-to-end inference from raw data input to lifetime value output with a response speed of milliseconds. S45, MCAT remaining life prediction system based on remaining life prediction values Implement tiered early warning systems.

[0027] Specifically, a first-level warning threshold is set. and Level 2 shutdown threshold The MCAT remaining lifetime prediction system compares the current predicted remaining lifetime in real time. Compared with the above threshold: when When the MCAT remaining life prediction system triggers a yellow alert, it prompts maintenance personnel to monitor the equipment status and develop a maintenance plan; when When this occurs, the MCAT remaining life prediction system triggers a red alert and recommends immediate shutdown for maintenance. This mechanism enables a shift from traditional scheduled maintenance to a data-driven, condition-based maintenance model, maximizing equipment lifespan and ensuring operational safety.

[0028] This invention incorporates a remaining lifetime prediction model, which includes a multi-scale local feature extraction module, a feature adaptive enhancement module, a global temporal coding module, and a temporal attention prediction module connected in sequence. Compared with the existing single-scale feature extraction module, the multi-scale local feature extraction module can simultaneously take into account both short-term fault impacts and long-term degradation trends, thereby improving the feature extraction capability of this invention. The feature adaptive enhancement module in this invention has a built-in key feature adaptive filtering mechanism, which can reconstruct the feature channels with weights, thereby improving the environmental noise resistance of the remaining lifetime prediction model.

[0029] In addition, the global temporal coding module has a built-in Transformer encoder, which can use a multi-head self-attention mechanism to mine long-distance temporal dependencies. Meanwhile, the temporal attention prediction module can dynamically aggregate deep feature matrices containing global temporal context information, which improves the prediction stability and robustness of the remaining lifetime prediction model.

[0030] To verify the effectiveness of the MCAT remaining life prediction method proposed in this invention, this embodiment uses a historical full life cycle dataset of mechanical equipment for experimental verification. The trained remaining life prediction model is inferred on the test set. To clearly demonstrate the prediction effect, this embodiment sorts the test set samples in descending order of actual remaining life and plots them as shown below. Figure 2 The comparison curves shown.

[0031] like Figure 2 As shown, the horizontal axis represents the sorted sequence of test set samples, and the vertical axis represents the remaining lifetime (RUL) value. The blue solid line in the figure represents the true remaining lifetime label of the test samples, and the orange broken line represents the predicted remaining lifetime value output by the remaining lifetime prediction model. Figure 2 The experimental results show that: a. High overall fit: The orange prediction curve and the blue actual curve maintain a high degree of consistency in the overall trend, indicating that the remaining life prediction model can effectively capture the degradation law of mechanical equipment.

[0032] b. Accurate prediction of critical stages: In the latter half of the sample sequence (i.e., the interval between 80 and 100 on the horizontal axis), corresponding to the "late degradation stage" (the area with a small RUL value) when mechanical equipment is about to fail, the predicted curve almost coincides with the actual curve. This shows that the remaining life prediction model has extremely high prediction accuracy when equipment is close to failure, and can provide accurate downtime maintenance warnings for operation and maintenance personnel, which has extremely high engineering practical value.

[0033] c. Strong anti-interference capability: Although there is noise in the original sensor data, and in the early degradation stage ( Figure 2 The area with larger RUL values ​​on the left side has greater uncertainty. Although the model's predicted values ​​fluctuate, they always revolve around the true values, which verifies the effectiveness of multi-scale convolution and attention mechanisms in noise resistance.

[0034] In another aspect, the present invention provides an MCAT remaining lifetime prediction system, configured to execute the above-described MCAT remaining lifetime prediction method, the MCAT remaining lifetime prediction system comprising: The data processing module is used to acquire multidimensional sensor data and construct time-series samples; The model storage module contains a built-in trained remaining lifespan prediction model. The lifespan monitoring and alarm module is used to provide graded early warnings based on the remaining lifespan prediction values ​​output by the trained remaining lifespan prediction model.

[0035] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining useful life (MCAT) of mechanical equipment, characterized in that, Includes the following steps: S1. Collect multi-dimensional sensor monitoring data during the operation of mechanical equipment, and then construct a historical full life cycle dataset based on this data. The historical full life cycle dataset is then preprocessed and divided into training sets. S2. Construct a remaining lifetime prediction model, which includes a multi-scale local feature extraction module, a feature adaptive enhancement module, a global temporal coding module, and a temporal attention prediction module connected in sequence. S3. Train the remaining lifetime prediction model using the training set, and construct a loss function in each round of training. Adjust the parameters of the remaining lifetime prediction model according to the loss function until the remaining lifetime prediction model converges, and obtain the trained remaining lifetime prediction model. S4. Receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, preprocess it, and input it into the trained remaining life prediction model to obtain the remaining life prediction value of the mechanical equipment under test at the current time. Then, perform graded early warning based on the remaining life prediction value.

2. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 1, characterized in that, S1 specifically includes the following steps: S11. Collect the operating status data of the mechanical equipment throughout its entire life cycle using multi-dimensional sensors installed on the mechanical equipment, and then construct a historical operating life cycle dataset based on this data. The historical operating life cycle dataset includes the original multi-dimensional monitoring sequence matrix of multiple equipment units. Among them, the The original multidimensional monitoring sequence matrix of each equipment unit The expression is as follows: ； In the formula, Indicates the first The total number of time steps before a single equipment unit completely fails; Indicates the first Sensor feature vectors at each time step, and ; This represents the total dimension of the features monitored by the sensor; Indicates the first The time step The physical measurement values ​​of each sensor; T Indicates transpose; S12. Perform Z-Score standardization on the data in the original multidimensional monitoring sequence matrix of each equipment unit to generate multiple standard normal distribution sequences; S13. For each standard normal distribution sequence, a preset window length is used. and sliding step size The standard normal distribution sequence is segmented to generate multiple time-series samples of fixed length. ; S14. For each time series sample Construct each time series sample True remaining life label ; S15, Collect all time series samples and the corresponding actual remaining life tag by The data is summarized in the form of a preprocessed historical runtime dataset. S16. Then, according to the set ratio, the preprocessed historical running lifecycle dataset is divided into training set and test set.

3. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 2, characterized in that, The expression for Z-Score normalization in S12 is as follows: ； in, Represents the standardized physical measurement values ; This represents the first time step in the historical running full lifecycle dataset. The average of the physical measurements from each sensor; This represents the first time step in the historical running full lifecycle dataset. The standard deviation of the physical measurements of each sensor; The expression for the standard normal distribution sequence in S12 is: ； in, This represents the standardized original multidimensional monitoring sequence matrix, i.e., the standard normal distribution sequence; Represents the standardized sensor feature vector ; The time-series samples in S13 The expression is as follows: ； in, Representing time series samples The start time index; This represents the time step index in the entire lifecycle sequence; Represents the standardized sensor feature vector ; Represents the standardized sensor feature vector ; The actual remaining life tag in S14 The expression is as follows: ； in, This is the moment when the mechanical equipment fails. This indicates the last time step of the sample window; This is the preset degradation saturation threshold.

4. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 3, characterized in that, The multi-scale local feature extraction module includes three parallel convolutional neural networks connected in sequence and a concatenation module, wherein each parallel convolutional neural network includes a convolutional kernel. A one-dimensional convolutional layer and a Gaussian error linear unit activation function; The feature adaptive enhancement module includes a global average pooling layer, a fully connected layer, a linear rectified unit, a fully connected layer, a Sigmoid activation function, and a fusion module connected in sequence. The global timing coding module includes sequentially connected... Convolutional layers, positional encoding modules, and Transformer encoders; the Transformer encoder consists of N stacked Transformer encoding layers; The time attention prediction module includes an attention score calculation module, a summation module, and a regression network connected in sequence.

5. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 4, characterized in that, Each parallel convolutional neural network in the multi-scale local feature extraction module is used to extract multi-scale feature maps. The expression is as follows: ； in, Represents the input feature tensor; Indicates that the convolution kernel is Convolution operations in a one-dimensional convolutional layer; Represents the activation function of the Gaussian error linear unit; Represents the set of real numbers; The number of convolution channels; The stitching module is used to fuse feature maps at multiple scales to obtain a multi-scale fused feature tensor. The expression is as follows: ； in, These represent feature maps at three different scales; Indicates a splicing operation; The global average pooling layer in the feature adaptive enhancement module is used to fuse multi-scale feature tensors. The spatial dimension is compressed to generate a channel statistics vector. The expression is as follows: ； in, Represents the multi-scale fusion feature tensor In the Channel, First The numerical value at any given moment; The fully connected layer is used to learn the nonlinear interactions between channels and generate channel weight vectors. The expression is as follows: ； in, and Here are the weight matrices for the two fully connected layers; It is the ReLU activation function; Use the Sigmoid activation function; The fusion module is used to convert the channel weight vector With multi-scale fusion feature tensor Element-wise multiplication yields the enhanced feature tensor. The expression is as follows: ； in, This represents element-wise multiplication based on the channel dimension.

6. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 5, characterized in that, The global timing coding module Convolutional layers are used to enhance the feature tensors The channel dimension from Projected onto model dimension , to obtain projection features ; The position encoding module is used for projection features Positional encoding is introduced to preserve the sequential information of the time series, thereby obtaining the feature sequence. The expression is as follows: ； in, The position encoding matrix. No. Line 1 Column and number Line 1 The elements of the column are calculated as follows: ； ； in, Indicates the position encoding matrix of the first... Line 1 Column elements; The Transformer encoder is used to mine temporal dependencies using a multi-head self-attention mechanism, resulting in a deep feature matrix containing global temporal context information. , of which The operation process of the Transformer encoding layer is as follows: ； ； ； in, , , These represent the query matrix, key matrix, and value matrix, respectively. Indicates the first l -1 layer feature sequence; This represents multi-head attention computation; This represents a feedforward neural network; Representation layer normalization; Indicates the first The feature sequence of the layer; Indicates the first The intermediate feature sequence after layer normalization in the Transformer encoding layer; The attention score calculation module is used to calculate the score at each time step. Attention score The calculation formula is as follows: ； ； in, Deep feature matrix The Middle Feature vectors at each time step; , , All of these are learnable attention network parameters; Indicates the first Attention intermediate hidden layer vectors at each time step; The summation module is used to perform a weighted summation of the feature vectors at all time steps to obtain the global context vector. The expression is as follows: ； in, This represents the time step after attention weighting. eigenvectors; The regression network is used to base on the global context vector. Predicted remaining life value The expression is as follows: ； in, and The weights and biases of the regression network.

7. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 6, characterized in that, S3 specifically includes the following steps: S31. Input a batch of training samples from the training set into the remaining lifetime prediction model, perform forward propagation calculations, and obtain the remaining lifetime prediction value for the current batch. ; S32, then the remaining lifetime prediction value Combined with the actual remaining life label of the current batch Calculate the loss value of the current loss function. ; S33. Calculate the loss function based on the backpropagation algorithm. Parameters of each layer in the remaining life prediction model The gradient vector is then processed by norm clipping. S34. The AdamW optimization algorithm is used, and the parameters of the remaining lifetime prediction model are adjusted based on the pruned gradient vector. Perform an update, determine whether the predetermined iteration stopping condition has been met, and if so, return the trained remaining lifetime prediction model; otherwise, return to S31. S35. Use the test set to verify and evaluate the remaining life prediction model during the training process, record the prediction error of each training cycle, and retain the model parameters with the smallest error on the test set as the final model parameters; otherwise, return to S31 to continue training.

8. The MCAT remaining life prediction method for mechanical equipment according to claim 7, characterized in that, The loss function in S3 is: ； in, B This refers to the number of training samples within the same batch. i Indicates the first i One training sample; Indicates smoothness L 1. Loss function, calculated as follows: ； in, This indicates the prediction error.

9. The MCAT (Mechanical Equipment Terminal) remaining life prediction method for mechanical equipment according to claim 1, characterized in that, S4 specifically includes the following steps: S41. Deploy the trained remaining lifetime prediction model to the MCAT remaining lifetime prediction system. S42. Utilize a data acquisition card to receive the monitoring data stream output by the multi-dimensional sensor detecting the mechanical equipment under test in real time, and load the average value. Standard deviation As a preset parameter, whenever a new sensor feature vector is received at a time step... Immediately apply preset parameters to the new sensor feature vector. Perform online normalization mapping to calculate new normalized vectors. ; S43, The length of the real-time window built based on the FIFO queue is L A first-in, first-out (FIFO) data cache queue, whenever a new normalized vector... The MCAT remaining lifetime prediction system will generate a new standardized vector. Push the data to the end of the queue and remove the oldest data from the head of the queue. This allows for dynamic updates to the current real-time time series sample matrix. ; S44. The current real-time time series sample matrix... The input is fed into the trained remaining life prediction model, which outputs the predicted remaining life of the machine under test at the current time. ; S45, MCAT remaining life prediction system based on remaining life prediction values Implement tiered early warning systems.

10. An MCAT remaining lifetime prediction system, characterized in that, The MCAT remaining lifetime prediction system is configured to perform the MCAT remaining lifetime prediction method according to any one of claims 1 to 9, and includes: The data processing module is used to acquire multidimensional sensor data and construct time-series samples; The model storage module contains a built-in trained remaining lifespan prediction model. The lifespan monitoring and alarm module is used to provide graded early warnings based on the remaining lifespan prediction values ​​output by the trained remaining lifespan prediction model.

Images