A device residual life prediction method and system based on a lightweight time sequence convolution network

Abstract

This application relates to the field of intelligent operation and maintenance and predictive maintenance technology for industrial equipment, specifically a method and system for predicting the remaining lifespan of equipment based on a lightweight temporal convolutional network. It employs resource-aware scheduling triggered by CPU utilization thresholds, reducing inference latency by approximately 36% under high load and RMSE by approximately 15% under low load, achieving a dynamic balance between accuracy and efficiency. Compared to EdgeMLBalancer (multi-model switching), this invention uses single-model multi-mode pre-loading switching, with a switching latency of <10ms and lower resource overhead. The lightweight temporal convolutional model used has 0.42M parameters (65% less than the standard TCN), and an inference latency of 28ms (normal mode). It requires no NPU / GPU and can run on general-purpose CPUs such as ARM Cortex-A57. Compared to existing technologies, it has greater adaptability.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance and predictive maintenance technology for industrial equipment, specifically a method and system for predicting the remaining lifespan of equipment based on lightweight temporal convolutional networks. Background Technology

[0002] In industrial automated production lines, the health status of critical components such as bearings, motors, and cutting tools directly impacts production efficiency and product quality. Accurately predicting remaining useful life (RUL) is a core technology for achieving predictive maintenance and avoiding unplanned downtime.

[0003] The current mainstream methods include: physical model-based methods (relying on accurate mechanistic models, with poor generalization), traditional machine learning methods (relying on manual feature extraction, which is labor-intensive), and deep learning methods (LSTM, TCN, etc., which automatically learn temporal features but have a large number of parameters, are computationally complex, and are difficult to deploy at the edge).

[0004] In recent years, lightweight lifetime prediction models have made progress. For example, the dual-channel lightweight convolutional Transformer model achieves lightweighting by reducing attention computation; MSSN-LST employs multi-scale separable convolutions and lightweight spatiotemporal attention; GT-MRNet reduces the number of parameters by 74.7% through Group Transformer. Furthermore, some studies have explored dynamic network tuning techniques: AdaNet dynamically adjusts the convolutional kernel size and channel size based on input data; EdgeMLBalancer dynamically switches between different models based on CPU utilization; and DynaSplit dynamically configures parameters through hardware and software collaboration.

[0005] Existing technologies using fixed network structures cannot adapt to dynamic changes in edge computing power. Existing lightweight models (including dual-channel Transformer, MSSN-LST, GT-MRNet, etc.) all use fixed network structures. When the edge gateway CPU load fluctuates (such as when simultaneously performing data acquisition and video stream processing), the fixed structure struggles to balance real-time performance and accuracy—inference latency exceeds limits when the load is too high, and hardware improvements cannot be made to enhance accuracy when the load is sufficient. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a method and system for predicting the remaining lifetime of a device based on a lightweight temporal convolutional network, so as to solve the problems in the background art.

[0007] To achieve the above objectives, this application adopts the following technical solution:

[0008] This application discloses a device remaining lifetime prediction method based on lightweight temporal convolutional networks, comprising the following steps:

[0009] The system acquires time-series monitoring data of various monitoring parameters of the target equipment on the production line, and acquires the CPU utilization rate of the edge smart gateway at the current moment. The time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration prior to the current moment.

[0010] The time-series monitoring data is preprocessed to obtain preprocessed data, wherein the preprocessing includes filling missing values, removing outliers, and standardization.

[0011] The structure of the pre-built lifetime prediction model is adaptively scheduled based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model with adjusted structure.

[0012] The preprocessed data is input into the structurally adjusted lifetime prediction model to obtain lifetime prediction results, and early warnings are issued based on the lifetime prediction results.

[0013] In one embodiment of this application, the lifetime prediction model employs a lightweight temporal convolutional network, the structure of which includes:

[0014] The input layer is used to receive preprocessed data;

[0015] The first convolution module is used to perform one-dimensional convolution, batch normalization and ReLU activation on the preprocessed data to obtain the first feature sequence;

[0016] The channel attention module is used to perform global average pooling, fully connected compression, and Sigmoid activation on the first feature sequence to obtain channel weights; and to perform channel-wise weighting on the first feature sequence based on the channel weights to obtain the enhanced first feature sequence.

[0017] The second convolution module is used to perform one-dimensional convolution, batch normalization and ReLU activation on the enhanced first feature sequence in low-load mode or normal mode to obtain the second feature sequence.

[0018] The feature fusion layer is used to fuse the first feature sequence or the second feature sequence in low-load mode to obtain fused features;

[0019] A global average pooling layer is used to perform global average pooling on the first feature sequence or fused features to obtain a feature vector.

[0020] The Dropout layer is used to randomly discard some neurons in the feature vector at a set dropout rate to obtain a regularized feature vector.

[0021] A fully connected output layer is used to perform a linear transformation on the regularized feature vector to obtain the lifetime prediction result.

[0022] In one embodiment of this application, the time-series monitoring data is preprocessed to obtain preprocessed data, including:

[0023] The time-series monitoring data is subjected to linear interpolation of missing values ​​and removal of outliers to obtain an interpolation sequence;

[0024] The mean and standard deviation of the interpolation sequence are calculated, and the interpolation sequence is standardized based on the mean and standard deviation to obtain preprocessed data.

[0025] In one embodiment of this application, the structure of a pre-built lifetime prediction model is adaptively scheduled based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure, including:

[0026] When the CPU utilization rate is less than a preset first utilization rate threshold, the edge smart gateway is determined to be in low load mode, the second convolution module and the feature fusion layer are enabled, and the life prediction model after structural adjustment is obtained.

[0027] When the CPU utilization rate is greater than or equal to a preset first utilization rate threshold and less than or equal to a preset second utilization rate threshold, the edge smart gateway is determined to be in normal mode, the second convolution module is enabled, and the feature fusion layer is disabled to obtain a life prediction model with structural adjustment, wherein the second utilization rate threshold is greater than the first utilization rate threshold.

[0028] When the CPU utilization rate exceeds a preset second utilization rate threshold, the feature fusion layer and the second convolutional module are disabled, and the number of convolutional kernels in the first convolutional module is reduced to obtain the life prediction model after structural adjustment.

[0029] In one embodiment of this application, the training method for the lifespan prediction model includes:

[0030] S1, acquire time-series monitoring data samples of various monitoring parameters throughout the entire lifecycle of the target equipment on the production line;

[0031] S2, preprocess the time-series monitoring data samples for each monitoring parameter to obtain preprocessed data samples, wherein the preprocessing includes filling missing values, removing outliers, and standardization.

[0032] S3, based on a pre-built sliding window, extract sample segments of multiple time windows from the pre-processed data sample, and perform lifetime labeling and data augmentation on the sample segments based on the end time of the time window to obtain training data with multiple labels, wherein the data augmentation includes Gaussian noise addition, time axis stretching / compression, channel dropping and time axis reversal.

[0033] S4, input the training data into a lightweight temporal convolutional network to obtain the prediction result;

[0034] S5, calculate the loss of the prediction result and the label based on the pre-built loss function, and backpropagate based on the loss, and adjust the internal parameters of the lightweight temporal convolutional network by combining gradient descent method;

[0035] S6. Repeat steps S4-S5 until training is complete and a lifespan prediction model is obtained.

[0036] In one embodiment of this application, lifetime labeling of sample segments based on the end of a time window includes:

[0037] Extracting the last moments of multiple time windows The beginning of the entire life cycle and end time ,in, Index for time windows;

[0038] End time based on time window The beginning of the entire life cycle and end time Calculate the remaining lifespan percentage label and remaining life label The remaining lifespan percentage label and remaining life label The mathematical expressions are as follows:

[0039]

[0040]

[0041] Based on the remaining lifespan percentage label and remaining life label Lifetime labels were applied to the sample fragments.

[0042] In one embodiment of this application, the mathematical expression of the loss function is:

[0043]

[0044] In the formula, Indicates the total loss. Represents the loss balance parameter. Indicates the mean absolute error. This represents the root mean square error.

[0045] In one embodiment of this application, early warning based on the lifetime prediction result includes:

[0046] The remaining lifespan in the lifespan prediction result is compared with a preset warning threshold, and a warning is triggered when the remaining lifespan is less than the preset warning threshold.

[0047] In one embodiment of this application, it further includes:

[0048] When an alert is triggered, the data collection frequency is increased, and the time-series monitoring data of various monitoring parameters of the target equipment on the production line is retrieved, along with the CPU utilization rate of the edge smart gateway at the current moment.

[0049] This application also provides a device remaining lifetime prediction system based on a lightweight temporal convolutional network, comprising:

[0050] The acquisition module is used to acquire time-series monitoring data of various monitoring parameters of the target equipment on the production line, and to acquire the CPU utilization rate of the edge smart gateway at the current moment. The time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration before the current moment.

[0051] A preprocessing module is used to preprocess the time-series monitoring data to obtain preprocessed data, wherein the preprocessing includes forward filling of missing values ​​and standardization.

[0052] An adaptive scheduling module is used to adaptively schedule the structure of the pre-built lifetime prediction model based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model after adjustment.

[0053] The lifespan prediction module is used to input the preprocessed data into the structurally adjusted lifespan prediction model to obtain lifespan prediction results, and to issue early warnings based on the lifespan prediction results.

[0054] The beneficial effects of this application are as follows: This application provides a device remaining lifetime prediction method and system based on a lightweight temporal convolutional network. It employs resource-aware scheduling triggered by CPU utilization thresholds, reducing inference latency by approximately 36% under high load and RMSE by approximately 15% under low load, achieving a dynamic balance between accuracy and efficiency. Compared to EdgeMLBalancer (multi-model switching), this invention uses single-model multi-mode preloading switching, with a switching latency of <10ms and lower resource overhead. The lightweight temporal convolutional model used has 0.42M parameters (65% less than the standard TCN), and an inference latency of 28ms (normal mode). It requires no NPU / GPU and can run on general-purpose CPUs such as ARM Cortex-A57. Compared to existing technologies, it has greater adaptability. Attached Figure Description

[0055] The present application will be further described below with reference to the accompanying drawings and embodiments:

[0056] Figure 1 This is a schematic diagram of the structure of the Lightweight Temporal Convolutional Network Light-TCN in one embodiment of this application;

[0057] Figure 2 This is a schematic diagram of the scheduler workflow in one embodiment of this application;

[0058] Figure 3 This is a flowchart of offline training in one embodiment of this application;

[0059] Figure 4 This is a flowchart illustrating a device remaining lifetime prediction method based on a lightweight temporal convolutional network in one embodiment of this application;

[0060] Figure 5 This is an example of performance comparison under different CPU loads in one embodiment of this application. Detailed Implementation

[0061] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0062] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the layers related to this application and are not drawn according to the actual number, shape and size of the layers in the actual implementation. In the actual implementation, the form, number and proportion of each layer can be arbitrarily changed, and the layer layout may also be more complex.

[0063] Numerous details are explored in the following description to provide a more thorough explanation of embodiments of this application; however, it will be apparent to those skilled in the art that embodiments of this application may be practiced without these specific details.

[0064] This application aims to overcome the shortcomings of existing technologies by providing a device remaining lifetime prediction method based on lightweight temporal convolutional networks. The core objectives are as follows:

[0065] 1) By simplifying the 1D-CNN structure and introducing channel attention, the number of parameters and computation are reduced, making it suitable for real-time inference at the edge.

[0066] 2) Introducing an adaptive scheduler: Dynamically adjusting network depth and width based on edge device CPU utilization. Unlike existing dynamic adjustment technologies, this invention employs CPU utilization threshold triggering (resource awareness), pre-loading switching (switching latency <10ms), and adjusting network depth (convolutional block level) and width (convolutional kernel level) to specifically address the real-time prediction problem under dynamically changing edge computing power.

[0067] 3) A hybrid loss function (MAE+RMSE weighted combination) is used as the training loss function, along with multi-dimensional data augmentation strategies, to improve the robustness and prediction accuracy of the model.

[0068] 4) Supports general-purpose CPU deployment, eliminating the need for dedicated hardware such as NPU / GPU.

[0069] I. Overall Architecture

[0070] This invention comprises two phases: offline training and online prediction. The offline phase trains the Light-TCN model in the cloud; the online phase deploys the model on an edge smart gateway, dynamically adjusting the network structure through a structure-adaptive scheduler, and outputs RUL prediction values ​​in real time.

[0071] II. Overall Model Architecture (Light-TCN)

[0072] Figure 1 This is a schematic diagram of the Lightweight Temporal Convolutional Network Light-TCN in one embodiment of this application, as shown below. Figure 1 As shown, the lightweight temporal convolutional network Light-TCN in this application includes:

[0073] The input layer is used to receive preprocessed data; specifically, it receives multi-channel time-series data X∈R. B×C×L C represents the number of channels (vibration, current, temperature, etc.), and L represents the sequence length.

[0074] The first convolutional module performs one-dimensional convolution, batch normalization, and ReLU activation on the preprocessed data to obtain the first feature sequence; specifically, it employs 1D convolution (kernel size 3, stride 1, same padding, 64 kernels) + batch normalization + ReLU. The output F1∈R B×64×L .

[0075] The channel attention (SE) module performs global average pooling, fully connected compression, and sigmoid activation on the first feature sequence to obtain channel weights. Based on these channel weights, it then performs channel-by-channel weighting on the first feature sequence to obtain an enhanced first feature sequence. Specifically, the process involves global average pooling → fully connected (compression ratio r=8) → sigmoid activation → generating channel weights, and then weighting the F1 score channel-by-channel. This module emphasizes key feature frequency bands.

[0076] A second convolutional module (optional) is used to perform one-dimensional convolution, batch normalization, and ReLU activation on the enhanced first feature sequence in low-load or normal mode to obtain a second feature sequence. The structure of the second convolutional module is similar to that of the first convolutional block, but with 32 kernels (50% fewer than the first block). It is enabled only in normal and low-load modes.

[0077] The feature fusion layer (only in low-load mode) is used to fuse the first feature sequence or the second feature sequence in low-load mode to obtain fused features; specifically, it uses a 1×1 convolution with 96 input channels (64+32) and 64 output channels to fuse the features of the two convolutional blocks.

[0078] A global average pooling layer is used to perform global average pooling on the first feature sequence or fused features to obtain a feature vector; the output feature vector v∈R is then processed. B×64 .

[0079] The Dropout layer is used to randomly discard some neurons in the feature vector at a set dropout rate (dropout rate 0.3) to obtain a regularized feature vector;

[0080] A fully connected output layer is used to perform a linear transformation on the regularized feature vector to obtain the lifetime prediction result.

[0081] In the above structure, the total number of convolutional kernels is reduced by more than 30% compared to the conventional TCN, and the SE module only increases by about 1-2% of FLOPs.

[0082] III. Structural Adaptive Scheduler

[0083] The structure-adaptive scheduler is deployed on the edge intelligent gateway to specifically address the real-time prediction problem under dynamic changes in edge computing power. Its essential differences from existing technologies (AdaNet, EdgeMLBalancer, etc.) are shown in the table below:

[0084] Table 1. Examples of comparison with existing technologies

[0085]

[0086] (1) Scheduler Workflow

[0087] Figure 2 This is a schematic diagram of the scheduler workflow in one embodiment of this application, illustrating the adaptive scheduler workflow, as follows: Figure 2 As shown, it includes:

[0088] Step S201: Monitor CPU utilization in real time with a sampling period of 100ms, calculate the moving average value uavg of the most recent 10 sampling points, and eliminate instantaneous fluctuations.

[0089] Step S202: Preset scheduling thresholds: High load threshold θhigh=70%, low load threshold θlow=30%. The thresholds can be configured remotely to adapt to different hardware performance.

[0090] Step S203: Scheduling Decision

[0091] 1) High-load mode (uavg>70%): ① Skip the second convolutional block (SE output is directly fed into global pooling); ② Reduce the number of convolutional kernels in the first convolutional block from 64 to 50. Reduce computational load and ensure real-time performance.

[0092] 2) Low load mode (uavg<30%): ① Restore the second convolutional block (32 convolutional kernels); ② Add a 1×1 convolutional feature fusion layer to improve prediction accuracy.

[0093] 3) Normal mode (30%≤uavg≤70%): Full structure (first convolutional block + SE + second convolutional block), no feature fusion layer.

[0094] Step S204: Perform the switch

[0095] The model parameters for the three modes are pre-loaded in memory (sharing the parameters of the first convolutional block and the SE module). Switching between modes only changes the computation graph path, with a latency of <10ms.

[0096] (2) Differences from existing dynamic adjustment technologies

[0097] 1) Differences from AdaNet: AdaNet dynamically adjusts the kernel size based on input data, making it data-driven; this invention dynamically adjusts the network depth / width based on system load (CPU utilization), making it resource-aware. The problems they solve, their triggering mechanisms, and the objects of adjustment are all different. AdaNet aims to improve the model's adaptability to diverse data, while this invention aims to ensure real-time performance under fluctuating edge computing power.

[0098] 2) Difference from EdgeMLBalancer: EdgeMLBalancer switches between multiple independent models, requiring different model files to be loaded each time, resulting in high resource overhead; this invention switches between three predefined modes of a single model, achieving low-latency switching through preloading. This invention is a "structure-level" dynamic adjustment, rather than a "model-level" adjustment.

[0099] IV. Offline Training

[0100] Figure 3 This is a flowchart of offline training in one embodiment of this application, such as... Figure 3 As shown, based on the above network structure, the process of training a preset device remaining lifetime in this application is as follows:

[0101] S1, acquire time-series monitoring data samples of various monitoring parameters throughout the entire lifecycle of the target equipment on the production line;

[0102] S2, preprocess the time-series monitoring data samples for each monitoring parameter to obtain preprocessed data samples, wherein the preprocessing includes filling missing values, removing outliers, and standardization.

[0103] In steps S1-S2, multi-channel time-series data of the entire lifecycle of the device are collected (vibration ≥1kHz, current ≥500Hz, temperature ≥100Hz). Missing values ​​are linearly interpolated, and outliers are 3σ removed.

[0104] The standardization process uses the following calculation formula:

[0105]

[0106] In the formula, For standardized data points, For the original data points, The mean, The standard deviation is denoted as .

[0107] S3, based on a pre-built sliding window, extract sample segments of multiple time windows from the pre-processed data sample, and perform lifetime labeling and data augmentation on the sample segments based on the end time of the time window to obtain training data with multiple labels, wherein the data augmentation includes Gaussian noise addition, time axis stretching / compression, channel dropping and time axis reversal.

[0108] The data augmentation strategies in this application include:

[0109] 1) Gaussian noise addition (SNR=20dB, probability 0.5);

[0110] 2) Time axis stretching / compression (scaling factor 0.8~1.2, probability 0.3);

[0111] 3) Channel drop (randomly drop one channel with a probability of 0.1);

[0112] 4) Time axis reversal (probability 0.05);

[0113] The methods for labeling samples include:

[0114] S31, extract the last moments of multiple time windows. The beginning of the entire life cycle and end time ,in, Index for time windows;

[0115] Multiple sample segments are extracted from the full lifecycle time-series data using a sliding window, with each segment corresponding to a time window. Record the last moment of this window. (i.e., the physical time of the last sampling point in the window), and simultaneously acquire the start and end times of the device's entire lifecycle from startup to complete failure. and These three time parameters form the basis for subsequent calculations of remaining lifetime.

[0116] S32, the last moment based on the time window The beginning of the entire life cycle and end time Calculate the remaining lifespan percentage label and remaining life label The remaining lifespan percentage label and remaining life label The mathematical expressions are as follows:

[0117]

[0118]

[0119] Among them, absolute remaining life label This indicates the actual time remaining (seconds, minutes, or hours) from the end of the window until device failure. Remaining life percentage label. It represents the percentage of remaining lifespan out of the total lifespan, and is dimensionless.

[0120] S33, based on the remaining lifespan percentage label and remaining life label Lifetime labels were applied to the sample fragments.

[0121] Combine the multi-channel data samples for each time window with the data calculated in the previous step. and The associations are used as target values ​​for supervised learning and stored in the training dataset. Depending on the task requirements, absolute labels, percentage labels, or a combination of both can be used for training (e.g., multi-task learning).

[0122] S4, input the training data into a lightweight temporal convolutional network to obtain the prediction result;

[0123] S5, calculate the loss of the prediction result and the label based on the pre-built loss function, and backpropagate based on the loss, and adjust the internal parameters of the lightweight temporal convolutional network by combining gradient descent method;

[0124] This application uses a hybrid loss function, expressed as follows:

[0125]

[0126] In the formula, Indicates the total loss. Represents the loss balance parameter. Indicates the mean absolute error. This represents the root mean square error.

[0127] To ensure that the overall deviation is controllable, Suppressing extreme large biases. One of the innovations of this application is using weighted combination as the training loss function (rather than just the evaluation metric).

[0128] During the training and conditioning process, the Adam optimizer was used with an initial learning rate of 0.001, a decay of 0.5 every 50 rounds, a batch size of 64, an early stop patience of 10 rounds, Dropout=0.3, and L2 regularization of 1e-5.

[0129] S6. Repeat steps S4-S5 until training is complete and a lifespan prediction model is obtained.

[0130] After training, save the model parameters (shared and independent parameters for the three modes), standardized parameters, and configuration files. Export as ONNX or TensorFlow Lite format.

[0131] V. Online Prediction Phase

[0132] Figure 4 This is a flowchart illustrating a device remaining lifetime prediction method based on lightweight temporal convolutional networks in one embodiment of this application, as shown below. Figure 4 As shown, a device remaining lifetime prediction method based on lightweight temporal convolutional networks in this application mainly includes the following steps:

[0133] S410, acquire time-series monitoring data of various monitoring parameters of the target equipment on the production line, and acquire the CPU utilization rate of the edge smart gateway at the current moment, wherein the time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration before the current moment;

[0134] Specifically, this application receives multi-channel data in real time (i.e., multiple monitoring parameters, each monitoring parameter corresponding to one channel), and uses a FIFO buffer (covering a 5-second window).

[0135] S420, the time-series monitoring data is preprocessed to obtain preprocessed data, wherein the preprocessing includes filling missing values, removing outliers, and standardization.

[0136] The preprocessing process includes:

[0137] S421, perform linear interpolation of missing values ​​and outlier removal on the time-series monitoring data to obtain an interpolation sequence;

[0138] S422, calculate the mean and standard deviation of the interpolation sequence, and standardize the interpolation sequence based on the mean and standard deviation to obtain preprocessed data.

[0139] Specifically, missing values ​​are padded forward, and a window is extracted every 2 seconds, then normalized using the saved μ and σ values. The detailed preprocessing procedure can be found above.

[0140] S430, based on the CPU utilization rate, the structure of the pre-built lifetime prediction model is adaptively scheduled to obtain the lifetime prediction model after structural adjustment. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model after structural adjustment.

[0141] Specifically, it includes:

[0142] (1) When the CPU utilization rate is less than the preset first utilization rate threshold, the edge smart gateway is determined to be in low load mode, the second convolution module and the feature fusion layer are enabled, and the life prediction model after structural adjustment is obtained.

[0143] (2) When the CPU utilization rate is greater than or equal to the preset first utilization rate threshold and less than or equal to the preset second utilization rate threshold, the edge smart gateway is determined to be in normal mode, the second convolution module is enabled, and the feature fusion layer is disabled to obtain the life prediction model after structural adjustment, wherein the second utilization rate threshold is greater than the first utilization rate threshold.

[0144] (3) When the CPU utilization rate is greater than the preset second utilization rate threshold, the feature fusion layer and the second convolution module are disabled, and the number of convolution kernels of the first convolution module is reduced to obtain the life prediction model after structural adjustment.

[0145] S440, the preprocessed data is input into the structurally adjusted lifetime prediction model to obtain lifetime prediction results, and an early warning is issued based on the lifetime prediction results.

[0146] S450, output to a visualization interface; if predicting lifetime... (e.g., 10 hours) triggers an alert and activates the dynamic data collection strategy optimizer to increase the data collection frequency.

[0147] The S460 cloud platform allows for remote updates of model parameters or scheduler thresholds.

[0148] The following are specific embodiments of this application, including:

[0149] Example 1: Offline Training (Bearing)

[0150] 1. Data Acquisition: CNC machine tool spindle bearing, accelerometer 2kHz, current transformer 1kHz, thermocouple 100Hz. Collect 200 hours of full lifecycle data. Failure criterion: vibration amplitude exceeds 5 times the initial value or temperature exceeds 85℃.

[0151] 2. Preprocessing: Linear imputation of missing values ​​(0.3%), 3σ removal of outliers (0.1%).

[0152] 3. Sliding window sampling: 5s window, 2s step. 10,000 vibration points, 5,000 current points, and 500 temperature points. This generates approximately 6,000 samples, which are then divided into training and validation sets in an 8:2 ratio.

[0153] 4. Standardization: Calculate μ and σ (vibration 0.12g / 0.03g, current 5.2A / 0.4A, temperature 32.5℃ / 1.2℃) for the first 100 health windows.

[0154] 5. Data augmentation: Apply Gaussian noise, time axis stretching, and channel dropping with probability.

[0155] 6. Model Building: By Figure 1 parameter.

[0156] 7. Training: Adam, lr=0.001, decay of 0.5 every 50 rounds, batch=64, stop 10 rounds early. The optimal RMSE was verified in round 120 at 6.48 hours.

[0157] Example 2: Online Prediction and Adaptive Scheduling

[0158] Deployed on an ARM Cortex-A57 edge gateway.

[0159] 1. High-load scenario (CPU 75%): Trigger high-load mode, prune the second convolutional block, and reduce the first convolutional kernel size from 64 to 50. Inference latency decreased from 28ms to 18ms, and RMSE decreased from 6.5% to 7.2%, meeting the <30ms requirement.

[0160] 2. Low-load scenario (CPU 25%): Trigger low-load mode, resume the second convolutional block, and add a fusion layer. Inference latency 32ms, RMSE 7.2%→6.1%.

[0161] 3. Normal scenario (CPU 50%): Normal mode, inference 28ms, RMSE 6.5%.

[0162] 4. Warning: The predicted RUL=9.8h<10h at the 150th hour is triggered. The bearing is replaced at the 158th hour. The actual remaining time is 8 hours, with an error of 1.8h.

[0163] Example 3: Comparison Experiment of Different Methods

[0164] The comparison results of different testing methods are shown in the table below:

[0165] Table 2. Examples of comparison results for multiple test methods

[0166]

[0167] This invention achieves higher accuracy than a fixed-weight lightweight CNN while being lightweight, and also has the ability to be dynamically adjusted.

[0168] Example 4: Adaptive Effect under Different CPU Loads

[0169] The adaptive performance under different CPU loads is shown in the table below.

[0170] Table 3. Examples of effects under different CPU loads

[0171]

[0172] The scheduler automatically selects the optimal mode: high load prioritizes real-time performance, while low load prioritizes precision. Figure 5 This is an example of performance comparison under different CPU loads in one embodiment of this application. The comparison results are as follows: Figure 5 As shown.

[0173] Example 5: Comparison with EdgeMLBalancer class methods

[0174] Deploy dynamic pruning methods based on activation sparsity (similar to the multi-model switching concept of EdgeMLBalancer) in the same environment:

[0175] Table 4. Comparison examples with EdgeMLBalancer class methods

[0176]

[0177] This invention features lower switching latency, faster inference, and higher accuracy.

[0178] Example 6: Hybrid Loss Function Ablation Experiment

[0179] The results of the hybrid loss function ablation experiment are shown in the table below.

[0180] Table 5. Results of ablation experiments using hybrid loss functions

[0181]

[0182] The mixed loss achieves the best balance between MAE and RMSE, with the lowest frequency of extreme errors.

[0183] Example 7: Threshold Configuration for Different Hardware Platforms

[0184] The threshold configurations for different hardware platforms are shown in the table below.

[0185] Table 6. Examples of threshold configurations for different hardware platforms

[0186]

[0187] The threshold can be configured remotely to adapt to different hardware.

[0188] Example 8: Application of Tool Life Prediction

[0189] The spindle load current, vibration, and acoustic emission of the CNC machine tool milling cutter were collected. The window size was 2 seconds, and the step size was 1 second. Results: The RMSE of this invention was 8.3 min, compared to 10.1 min for the fixed lightweight CNN, representing an improvement of 17.8%.

[0190] The beneficial effects of the above scheme are as follows:

[0191] 1. Dynamic Network Adjustment Adaptive to Computing Power: Unlike AdaNet (data-driven), this invention employs resource-aware scheduling triggered by CPU utilization thresholds. This reduces inference latency by approximately 36% under high load and RMSE by approximately 15% under low load, achieving a dynamic balance between accuracy and efficiency. Compared to EdgeMLBalancer (multi-model switching), this invention uses single-model multi-mode preloading switching, resulting in switching latency <10ms and lower resource overhead.

[0192] 2. Highly lightweight, supporting general-purpose CPU deployment: 0.42M parameters (65% less than standard TCN), inference latency 28ms (normal mode). No NPU / GPU required, can run on general-purpose CPUs such as ARM Cortex-A57.

[0193] 3. High prediction accuracy and strong robustness: The hybrid loss function (MAE+RMSE weighted) is used as the training loss function, which reduces the frequency of extreme errors by 54% (from 5.2% to 2.4%); multidimensional data augmentation improves noise resistance.

[0194] 4. Easy to deploy and maintain: Thresholds are remotely configurable and adaptable to different hardware.

[0195] 5. Edge closed-loop capability: It works in conjunction with the dynamic acquisition strategy optimizer to form a closed loop of "prediction → early warning → strategy adjustment".

[0196] This application also provides a device remaining lifetime prediction system based on a lightweight temporal convolutional network, comprising:

[0197] The acquisition module is used to acquire time-series monitoring data of various monitoring parameters of the target equipment on the production line, and to acquire the CPU utilization rate of the edge smart gateway at the current moment. The time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration before the current moment.

[0198] A preprocessing module is used to preprocess the time-series monitoring data to obtain preprocessed data, wherein the preprocessing includes forward filling of missing values ​​and standardization.

[0199] An adaptive scheduling module is used to adaptively schedule the structure of the pre-built lifetime prediction model based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model after adjustment.

[0200] The lifespan prediction module is used to input the preprocessed data into the structurally adjusted lifespan prediction model to obtain lifespan prediction results, and to issue early warnings based on the lifespan prediction results.

[0201] This application presents a device remaining lifetime prediction method and system based on a lightweight temporal convolutional network. It employs resource-aware scheduling triggered by a CPU utilization threshold, reducing inference latency by approximately 36% under high load and RMSE by approximately 15% under low load, achieving a dynamic balance between accuracy and efficiency. Compared to EdgeMLBalancer (multi-model switching), this invention uses single-model multi-mode preloading switching, with a switching latency of <10ms and lower resource overhead. The lightweight temporal convolutional model used has 0.42M parameters (65% less than standard TCN), and an inference latency of 28ms (normal mode). It requires no NPU / GPU and can run on general-purpose CPUs such as ARM Cortex-A57. Compared to existing technologies, it has greater adaptability.

[0202] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0203] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory so that the terminal performs any of the methods in this embodiment.

[0204] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0205] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0206] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0207] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0208] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. The embodiments of the present application are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0209] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for predicting the remaining lifetime of a device based on a lightweight temporal convolutional network, characterized in that, include: The system acquires time-series monitoring data of various monitoring parameters of the target equipment on the production line, and acquires the CPU utilization rate of the edge smart gateway at the current moment. The time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration prior to the current moment. The time-series monitoring data is preprocessed to obtain preprocessed data, wherein the preprocessing includes filling missing values, removing outliers, and standardization. The structure of the pre-built lifetime prediction model is adaptively scheduled based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model with adjusted structure. The preprocessed data is input into the structurally adjusted lifetime prediction model to obtain lifetime prediction results, and early warnings are issued based on the lifetime prediction results.

2. The method for predicting device remaining lifetime based on lightweight temporal convolutional networks according to claim 1, characterized in that, The lifetime prediction model employs a lightweight temporal convolutional network, the structure of which includes: The input layer is used to receive preprocessed data; The first convolution module is used to perform one-dimensional convolution, batch normalization and ReLU activation on the preprocessed data to obtain the first feature sequence; The channel attention module is used to perform global average pooling, fully connected compression, and Sigmoid activation on the first feature sequence to obtain channel weights; and to perform channel-wise weighting on the first feature sequence based on the channel weights to obtain the enhanced first feature sequence. The second convolution module is used to perform one-dimensional convolution, batch normalization and ReLU activation on the enhanced first feature sequence in low-load mode or normal mode to obtain the second feature sequence. The feature fusion layer is used to fuse the first feature sequence or the second feature sequence in low-load mode to obtain fused features; A global average pooling layer is used to perform global average pooling on the first feature sequence or fused features to obtain a feature vector. The Dropout layer is used to randomly discard some neurons in the feature vector at a set dropout rate to obtain a regularized feature vector. A fully connected output layer is used to perform a linear transformation on the regularized feature vector to obtain the lifetime prediction result.

3. The method for predicting device remaining lifetime based on lightweight temporal convolutional networks according to claim 1, characterized in that, The time-series monitoring data is preprocessed to obtain preprocessed data, including: The time-series monitoring data is subjected to linear interpolation of missing values ​​and removal of outliers to obtain an interpolation sequence; The mean and standard deviation of the interpolation sequence are calculated, and the interpolation sequence is standardized based on the mean and standard deviation to obtain preprocessed data.

4. The method for predicting device remaining lifetime based on lightweight temporal convolutional networks according to claim 2, characterized in that, Based on the CPU utilization rate, the structure of the pre-built lifetime prediction model is adaptively scheduled to obtain a structurally adjusted lifetime prediction model, including: When the CPU utilization rate is less than a preset first utilization rate threshold, the edge smart gateway is determined to be in low load mode, the second convolution module and the feature fusion layer are enabled, and the life prediction model after structural adjustment is obtained. When the CPU utilization rate is greater than or equal to a preset first utilization rate threshold and less than or equal to a preset second utilization rate threshold, the edge smart gateway is determined to be in normal mode, the second convolution module is enabled, and the feature fusion layer is disabled to obtain a life prediction model with structural adjustment, wherein the second utilization rate threshold is greater than the first utilization rate threshold. When the CPU utilization rate exceeds a preset second utilization rate threshold, the feature fusion layer and the second convolutional module are disabled, and the number of convolutional kernels in the first convolutional module is reduced to obtain the life prediction model after structural adjustment.

5. The method for predicting device remaining lifetime based on lightweight temporal convolutional networks according to claim 1, characterized in that, The training method for the lifespan prediction model includes: S1, acquire time-series monitoring data samples of various monitoring parameters throughout the entire lifecycle of the target equipment on the production line; S2, preprocess the time-series monitoring data samples for each monitoring parameter to obtain preprocessed data samples, wherein the preprocessing includes filling missing values, removing outliers, and standardization. S3, based on a pre-built sliding window, extract sample segments of multiple time windows from the pre-processed data sample, and perform lifetime labeling and data augmentation on the sample segments based on the end time of the time window to obtain training data with multiple labels, wherein the data augmentation includes Gaussian noise addition, time axis stretching / compression, channel dropping and time axis reversal. S4, input the training data into a lightweight temporal convolutional network to obtain the prediction result; S5, calculate the loss of the prediction result and the label based on the pre-built loss function, and backpropagate based on the loss, and adjust the internal parameters of the lightweight temporal convolutional network by combining gradient descent method; S6. Repeat steps S4-S5 until training is complete and a lifespan prediction model is obtained.

6. The method for predicting device remaining lifetime based on a lightweight temporal convolutional network according to claim 5, characterized in that, Lifetime labels are applied to sample segments based on the last moment of the time window, including: Extracting the last moments of multiple time windows The beginning of the entire life cycle and end time ,in, Index for time windows; End time based on time window The beginning of the entire life cycle and end time Calculate the remaining lifespan percentage label and remaining life label The remaining lifespan percentage label and remaining life label The mathematical expressions are as follows: Based on the remaining lifespan percentage label and remaining life label Lifetime labels were applied to the sample fragments.

7. The method for predicting device remaining lifetime based on a lightweight temporal convolutional network according to claim 5, characterized in that, The mathematical expression for the loss function is: In the formula, Indicates the total loss. Represents the loss balance parameter. Indicates the mean absolute error. This represents the root mean square error.

8. The method for predicting device remaining lifetime based on lightweight temporal convolutional networks according to claim 1, characterized in that, Early warnings based on the lifespan prediction results include: The remaining lifespan in the lifespan prediction result is compared with a preset warning threshold, and a warning is triggered when the remaining lifespan is less than the preset warning threshold.

9. A method for predicting the remaining lifetime of a device based on a lightweight temporal convolutional network according to claim 8, characterized in that, Also includes: When an alert is triggered, the data collection frequency is increased, and the time-series monitoring data of various monitoring parameters of the target equipment on the production line is retrieved, along with the CPU utilization rate of the edge smart gateway at the current moment.

10. A device remaining lifetime prediction system based on lightweight temporal convolutional networks, characterized in that, include: The acquisition module is used to acquire time-series monitoring data of various monitoring parameters of the target equipment on the production line, and to acquire the CPU utilization rate of the edge smart gateway at the current moment. The time-series monitoring data includes the values ​​of the detection parameters at multiple time points within the target duration before the current moment. A preprocessing module is used to preprocess the time-series monitoring data to obtain preprocessed data, wherein the preprocessing includes forward filling of missing values ​​and standardization. An adaptive scheduling module is used to adaptively schedule the structure of the pre-built lifetime prediction model based on the CPU utilization rate to obtain a lifetime prediction model with adjusted structure. The higher the CPU utilization rate, the simpler the structure of the lifetime prediction model after adjustment. The lifespan prediction module is used to input the preprocessed data into the structurally adjusted lifespan prediction model to obtain lifespan prediction results, and to issue early warnings based on the lifespan prediction results.

Images