Intelligent maintenance method for small piece sorting machine in logistics transfer yard based on digital twinning

Abstract

The invention discloses a digital twinning-based intelligent maintenance method for a small piece sorting machine in a logistics transfer yard, which comprises the following steps of: S1, deploying a plurality of sensors at key positions of a cross-belt sorting machine, acquiring in real time, transmitting to a preprocessing module for preprocessing to obtain a preprocessed data vector, and constructing a data matrix in a time window; s2, constructing a digital twin model of the cross-belt sorting machine; s3, based on historical data, accumulated health indexes and residual life of the cross-belt sorting machine are calculated, and then a training set based on the fault state, the accumulated health indexes and the residual life is constructed; and S4, constructing a CNN model to predict a fault state, a cumulative health index and residual life, performing optimization design on the CNN model, and calculating an early warning index to specify a maintenance strategy. According to the invention, on the basis of considering real-time data acquisition, digital twinning and deep learning intelligent maintenance, dynamic health monitoring, fault prediction and maintenance early warning of the small piece sorting machine in the logistics transfer yard are realized.

Description

Technical Field

[0001] The present invention relates to the field of logistics sorting, and particularly to an intelligent maintenance method for small-piece sorting machines in logistics transfer yards based on digital twins. Background Art

[0002] In the current logistics industry, as an important device in the logistics transfer yard, the efficient and stable operation of the small-piece sorting machine is directly related to logistics efficiency and equipment maintenance costs. The traditional maintenance mode of small-piece sorting machines mainly relies on periodic inspections or after-the-fact maintenance, and has the following deficiencies: 1. Insufficient real-time monitoring: Key parts of the equipment (such as the drive system, control unit, etc.) lack real-time monitoring, which is likely to cause the accumulation of potential hazards; 2. Data silos: There is a lack of effective integration among various types of data obtained by multiple sensors, PLC controllers, and cameras, making it difficult to form health indicators for the entire life cycle of the equipment; 3. Imperfect prediction model: Traditional experience-based maintenance methods are difficult to predict faults based on the operating status of the equipment, historical fault data, and environmental factors, resulting in inaccurate equipment maintenance warnings. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an intelligent maintenance method for small-piece sorting machines in logistics transfer yards based on digital twins. On the basis of taking into account real-time data collection, digital twins, and deep learning intelligent maintenance, dynamic health monitoring, fault prediction, and maintenance warning of small-piece sorting machines in logistics transfer yards are realized.

[0004] The purpose of the present invention is achieved through the following technical solutions: An intelligent maintenance method for small-piece sorting machines in logistics transfer yards based on digital twins, including the following steps: S1. Deploy a variety of sensors at key positions of the cross-belt sorting machine, collect data in real time, and transmit it to the preprocessing module for preprocessing to obtain a preprocessed data vector, and construct a data matrix within a time window; S2. Construct a digital twin model of the cross-belt sorting machine; S3. Based on historical data, calculate the cumulative health index and remaining life of the cross-belt sorting machine, and then construct a training set based on the fault status, cumulative health index, and remaining life; S4. Construct a CNN model for predicting the fault status, cumulative health index, and remaining life, optimize the design of the CNN model, and calculate the warning index to specify the maintenance strategy.

[0005] The beneficial effects of the present invention are as follows: By deploying various sensors such as vibration, temperature, operation duration, current, and noise, a PLC controller, a camera, and key monitoring devices, the present invention realizes the real-time acquisition and preprocessing of the operating state of the device. Specifically, the present invention establishes a digital twin model driven by multi-source real-time data, realizes two-way data synchronization between the physical device and the virtual model, dynamically restores the mechanical structure, control logic, and environmental impact of the device, and quantitatively describes the health state of the device through a full-life-cycle health assessment model, revealing the internal coupling relationship between the operating state, health index, and potential fault hazards of the key parts of the device. To further achieve intelligent maintenance, the present invention introduces a fault diagnosis and health state assessment method based on a convolutional neural network (CNN). This method automatically learns features from the preprocessed multi-dimensional sensor data by constructing a deep network structure containing spatio-temporal feature extraction and attention mechanisms, and establishes a multi-task model for fault state, cumulative health index, and remaining life prediction. Thus, the system can accurately diagnose the device state, accurately predict potential fault risks, and quantitatively evaluate the remaining life. In summary, on the basis of taking into account real-time data acquisition, digital twin, and deep learning intelligent maintenance, the present invention realizes dynamic health monitoring, fault prediction, and maintenance warning of the small-piece sorting machine in the logistics transfer yard, breaking through the technical bottlenecks of the traditional periodic maintenance mode, such as insufficient perception of the device health state and lagging fault prediction, and providing a new theoretical basis and technical support for the field of intelligent logistics equipment maintenance. Description of the Drawings

[0006] Figure 1 It is a flowchart of the method of the present invention. Detailed Embodiments

[0007] The technical solution of the present invention will be further described in detail below with reference to the drawings, but the protection scope of the present invention is not limited to the following.

[0008] With the development of the Internet of Things, digital twin, and deep learning technologies, by establishing a two-way data synchronization mechanism between physical devices and virtual models, not only can the state of the device be restored in real time, but also multi-source data can be fused and analyzed with the help of deep learning algorithms to achieve fault prediction, remaining life estimation, and intelligent maintenance scheduling. Specifically, the entire method of the present invention is divided into three levels: Physical layer: including actual sorting equipment, sensor network, PLC controller, etc., for data acquisition; Virtual modeling layer: Restoring the device state through functions to form a virtual device model; Intelligent analysis layer, namely CNN: nested in the digital twin system, used to automatically learn device state features from multi-dimensional data to achieve prediction and decision-making support. Such as Figure 1As shown in the figure, an intelligent maintenance method for small-piece sorting machines in logistics transfer yards based on digital twins includes the following steps: S1. Deploy a variety of sensors at key positions of the cross-belt sorting machine, collect data in real time, and transmit it to the preprocessing module for preprocessing to obtain a preprocessed data vector, and construct a data matrix within a time window; Deploy a variety of sensors at key positions such as the conveyor belt, sorting compartments, and scanners of the cross-belt sorting machine to collect information such as vibration, temperature, running duration, current, and noise in real time. Set the value sampled by each sensor at time t as Then, the sensor data collected is formed into a preprocessed data vector: where m is the number of sensors. We use the PLC controller and camera to transmit the data to the data preprocessing module through industrial Ethernet or other field buses, and perform denoising and normalization processing on the original data to form a data matrix suitable for subsequent modeling.

[0009] Based on the preprocessed data vector, construct a data matrix within a time window: where T is the length of the time window.

[0010] S2. Construct a digital twin model of the cross-belt sorting machine; According to the CAD model, kinematic and dynamic parameters of the actual cross-belt sorting machine, establish an accurate three-dimensional model of the cross-belt sorting machine, and import the preprocessed data vector into the digital twin model to achieve the synchronization of the sorting machine digital twin model and the actual data.

[0011] In some embodiments, feedback control can also be performed based on the digital twin model: Define the function F to express the combination of real-time data collected by sensors with physical, control, and environmental models to achieve the dynamic restoration of equipment status.

[0012] where represents the virtual model state at time t (including mechanical structure, control logic, and environmental state), is a set of physical parameters, control parameters, and environmental impact parameters.

[0013] To ensure the consistency between the digital twin model and the physical equipment status, the system establishes a two-way data synchronization mechanism: On the one hand, it is to preprocess the data vector Import the digital twin model to synchronize the digital twin model of the sorting machine with the actual data. On the other hand, observe whether its actual operating state is consistent with the expected state through the digital twin model. If not, the control parameters need to be adjusted to achieve feedback control. to achieve feedback control; After the digital twin model is constructed, it can be used to display the operation of the sorting machine for real-time monitoring by staff.

[0014] S3. Calculate the cumulative health index and remaining life of the cross-belt sorting machine based on historical data, and then construct a training set based on the fault state, cumulative health index, and remaining life; S301. Obtain the data within multiple time windows according to step S1, and the data within each time window is recorded as: S302. For the data within any time window, calculate the cumulative health index and estimate the remaining life, including: Set the instantaneous health index function of the device and obtain the cumulative health index by integrating over time : where is the mapping function: where represents the input variable of the mapping function, is the adjustment parameter based on manual experience to control the response sensitivity between the health value and the actual device degradation degree through historical sorting performance; is used to reflect the non-linear relationship between the cumulative health index and the overall state of the device is expressed as: where is the weight of the i th sensor feature; ; Estimate the remaining life according to the cumulative health index and the preset health threshold : : where is the device health degradation rate, obtained by fitting historical data and regarded as a known value; S303. Take the data of any time window as sample features, and add the cumulative health index , remaining life and As a label; is a fault status label, : , indicating that the cross-belt sorter has had a fault within the time window; , indicating that the cross-belt sorter has not had a fault within the time window; The training samples of the time window data are denoted as ; S304. For each time window data, repeat steps S302 - S303 to obtain the training samples corresponding to each time window data, add them to the same set, and form a training set.

[0015] In some embodiments, the samples can also be divided into a training set and a test set. After training the CNN model using the training set, use the test set to test the accuracy of the model; Traditional health assessment function methods have problems such as difficult modeling and complex parameter adjustment when facing complex working conditions and multi-source heterogeneous data. In response, the present invention introduces a data-driven method based on CNN to directly learn the fault status, cumulative health index, and remaining life prediction model of the device from sensor data. Specifically, CNN can be regarded as a function approximate fitter, and the cumulative health index output by it can replace the traditional for real-time monitoring; at the same time, the output by CNN can be used as the remaining life estimate value, avoiding complex integration and mapping processes. The above two model paths can be flexibly selected according to actual application requirements: in the model training stage, use the traditional health status function and as the source of pseudo-labels for CNN; in the operation stage, the output result of CNN can be used as the main indicator, while the traditional analysis path can be used as an auxiliary analysis tool or model interpretation means, thus realizing a fusion health assessment architecture of "data-driven + knowledge modeling".

[0016] S4. Construct a CNN model for fault status, cumulative health index, and remaining life prediction, optimize the design of the CNN model, and calculate the warning index to specify the maintenance strategy.

[0017] S401. Input data tensor construction: Through the data preprocessing module, the sensor data collected in real time is normalized, denoised, and windowed to form the following data matrix: Reconstruct the data into a three-dimensional tensor format, set it in the form of "height - width - channel", where the height corresponds to the number of deployed sensors, and the width Corresponding time step, number of channels According to the specific fusion method, for each sensor data, a single channel is formed, then represents the feature channel corresponding to each sensor; the final input tensor is denoted as ; S402. Randomly crop the input tensor on the time axis, inject noise to simulate sensor errors, and randomly scale the input tensor to expand the dataset, which can also enable the model to better adapt to data variations in actual work; S403. The model gradually captures features from local to global through multiple stacked two-dimensional convolutional layers; in the multiple stacked two-dimensional convolutional layers, the output feature map of the previous two-dimensional convolutional layer serves as the input feature map of the next two-dimensional convolutional layer, and the input feature map of the first two-dimensional convolutional layer is the input tensor; the output feature map of the last two-dimensional convolutional layer is the final output of the multiple stacked two-dimensional convolutional layers; For the th layer, the eigenvalue of the th output channel at position is defined as follows: Among them, is the weight of the th convolutional kernel in the th layer, where is the index of the convolutional kernel in the height and width directions, is the index of the input channel, is the index of the output channel, and the kernel size is ; the output channel represents the th channel of the feature map generated by the convolutional kernel; is the value of the th channel of the input feature map in the th layer at position ( ); is the bias value; is the ReLU activation function; the number of channels of the input tensor is , and the number of output tensor channels is determined by the set number of convolutional kernels; In this process, a pooling layer needs to be added after each two-dimensional convolutional layer for downsampling and then output; S404. Process the input tensor through multiple stacked one-dimensional convolutional layers; the input feature map of the first one-dimensional convolutional layer is the input tensor, the output feature map of the previous one-dimensional convolutional layer serves as the input feature map of the next one-dimensional convolutional layer, and the output of the last one-dimensional convolutional layer is the final output of the multiple stacked one-dimensional convolutional layers; The multiple stacked one-dimensional convolutional layers and the multiple stacked two-dimensional convolutional layers are parallel different branch structures; During this process, it is necessary to add a pooling layer after each one-dimensional convolutional layer for downsampling and then output the result; S405. By merging the outputs of different branch structures, that is, splicing the outputs of the multi-layer stacked two-dimensional convolutional layer and the multi-layer stacked one-dimensional convolutional layer, the spliced feature map is obtained. Then, the feature map is flattened, and the extracted features are integrated through a fully connected network, and finally the fault state, cumulative health index, and remaining life prediction are output. The output is a multi-dimensional vector: .

[0018] Model training and optimization: Design of loss function The optimization design of the CNN model includes: A1. Assume that the CNN model is updated by the mini-batch gradient descent method; the samples in the training set are divided into multiple batches, and each batch contains N samples. The CNN model is updated using the samples in each batch: For any batch of samples, based on the constructed CNN network, for multi-task output, the loss function is designed as a weighted sum of multiple sub-losses, including: For each sample in the current batch, the sample features are used as the input of the CNN model, and the output of the CNN model is used as the predicted value; For the fault state, cumulative health index, and remaining life prediction, the total loss is expressed as: , is the importance weight for adjusting the losses of each part; where, represents the loss function of the fault state, and the cross-entropy loss function is adopted: where, in the current batch of samples, the health state label and predicted value of the nth sample are respectively denoted as , ; n = 1, 2, …, N; represents the cumulative health index loss function: where, in the current batch of samples, the cumulative health index label and predicted value of the nth sample are respectively denoted as , ; n = 1, 2, …, N; Denote the remaining useful life loss function, and adopt the mean square error loss: Among them, in the current batch of samples, the cumulative remaining useful life label and the predicted value of the nth sample are respectively denoted as 、 ; n = 1, 2, …, N; A2. Based on each batch of samples in the training set, repeat step A1, and then train the CNN model based on the total loss function until the model converges or all batches of samples are trained, and a mature CNN model is obtained for prediction.

[0019] In the embodiments of the present application, regularization and prevention of overfitting can also be performed To avoid overfitting of the model, Dropout and L2 regularization can be added during training: Dropout layer: Insert a Dropout layer in the fully connected layer, so that some nodes are "discarded" with a certain probability during training, thereby enhancing the generalization ability of the model. Assume that the Dropout probability is , then the output is: Among them, is the input of the Dropout layer, is the output of the Dropout layer, represents the Hadamard product, is a Bernoulli distribution random vector. For the jth element in the vector , it follows ; ; that is The probability of taking 1 is , The probability of taking 0 is .

[0020] L2 regularization: Add a weight decay term to the loss function, and add the L2 norm of the weight parameter to the loss function: Among them, is the regularization parameter.

[0021] For the label definition methods corresponding to the three tasks: Fault status label: For the label : Indicates that the cross-belt sorter has failed within the time window; , indicating that the cross-belt sorter did not malfunction within the time window. The construction method is to trace back the fault log. Within the given time window , it is judged whether a fault occurred in the future. If a fault occurred, it is 1, otherwise it is 0. The loss function is the one defined above ; Cumulative health index label: a continuous value label obtained through the defined analytical model (pseudo-label), and the loss function is the one defined above . Although the CNN model uses the output of the traditional health function as the pseudo-supervised label for optimization during the training phase, its final output is not limited to 's expression form, but learns a stronger state expression ability through a data-driven approach. During the actual operation phase, it can completely replace to achieve real-time prediction of the health score (cumulative health index), thus avoiding the complexity and limitations of traditional models in aspects such as parameter tuning and function modeling; Remaining useful life label: a continuous value label obtained through the defined analytical model R (pseudo-label), and the loss function is the one defined above . Although the CNN model uses the output of the traditional remaining useful life estimation function R as the pseudo-supervised label for optimization during the training phase, its final output is not limited to R 's analytical expression, but automatically learns the implicit mapping relationship between the device state and the remaining useful life through a data-driven approach. During the actual operation phase, the output by the CNN can be directly used to replace the traditional R model for remaining useful life prediction, thus avoiding the decline in accuracy and stability caused by problems such as difficult degradation rate modeling and health score error transmission, and improving the real-time performance and generalization ability of remaining useful life prediction.

[0022] The cumulative health index output in the CNN model can be regarded as a quantitative indicator of the current state of the device. Combining historical data and real-time data, a dynamic monitoring curve of the health state can be formulated. Using time series analysis methods (such as moving average or Kalman filter), the health score data is smoothed in real time, and at the same time, the warning level is constructed by combining the change trends of the fault state and the remaining useful life prediction, specifically defined as: Among them, , is a weight parameter, is a preset life threshold. Different maintenance strategies can be set according to the value of the warning index.

[0023] In the embodiments of the present application: when the warning index is between 0.0 and 0.3, it indicates that the device is in a healthy state and can continue to operate normally. Only status monitoring needs to be carried out according to the regular cycle; when the index is in the range of 0.3 to 0.6, it belongs to medium risk. It is recommended to arrange on-site engineers for inspection and strengthen data collection and analysis of key parts; when the index reaches 0.6 to 0.8, it indicates that there are relatively large potential risks in the device, and partial shutdown inspection should be carried out to prepare corresponding maintenance or replacement plans; if the warning index exceeds 0.8, it is in a high-risk or critical state, and the device must be shut down immediately, and emergency repair or component replacement should be arranged first to prevent systematic shutdown caused by equipment failure.

[0024] This grading strategy can be further optimized by combining historical failure data and actual operation experience to achieve more accurate and efficient intelligent maintenance scheduling.

[0025] In the embodiments of the present application, in order to meet the real-time requirements in the industrial field, the CNN model can be lightweight processed through techniques such as pruning, quantization, and knowledge distillation to ensure high-speed and efficient operation on low-power embedded devices. The deployed model can perform online prediction while collecting data in the established digital twin system of the logistics transfer yard, providing real-time decision support information for equipment maintenance personnel.

Claims

1. An intelligent maintenance method for small-parts sorting machines in logistics transfer yards based on digital twins, characterized in that: The following steps are involved: S1. Deploy multiple sensors at key locations of the cross-belt sorter to collect data in real time, transmit the data to the preprocessing module for preprocessing to obtain preprocessed data vectors, and construct a data matrix within the time window; S2. Build a digital twin model of the cross-belt sorter; S3. Based on historical data, the cumulative health index and remaining life of the cross-belt sorter are calculated, and then a training set based on fault status, cumulative health index and remaining life is constructed; S4. Construct a CNN model to predict fault status, cumulative health index, and remaining life, optimize the design of the CNN model, and calculate the early warning index to specify the maintenance strategy.

2. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The key positions of the cross-belt sorter include the conveyor belt, the sorting grid and the scanner; The various sensors are used to collect vibration, temperature, operation time, current and noise information.

3. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The data of the various sensors are collected by a camera and then transmitted to a PLC controller, which transmits the data to a data preprocessing module via an industrial Ethernet or a field bus. The data preprocessing module processes the collected data including denoising and normalization.

4. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: Suppose the value of the data preprocessed by the i-th sensor sampled at time t is , then the preprocessed data of each sensor constitutes the preprocessed data vector: Where m represents the number of sensors; Based on the preprocessed data vector, a data matrix is ​​constructed within the time window, which is recorded as: Where T is the length of the time window.

5. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The construction of the cross-belt sorter digital twin model includes: Based on the CAD model, kinematic and dynamic parameters of the actual cross-belt sorter, an accurate three-dimensional model of the cross-belt sorter is established, and the preprocessed data vectors are imported into the digital twin model to achieve synchronization between the sorter digital twin model and the actual data.

6. According to claim 4, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The step S3 comprises: S301. Obtain data in multiple time windows according to step S1, and the data in each time window is recorded as: S302. For data in any time window, calculate the cumulative health index and estimate the remaining life span, including: Set the device instantaneous health index function , the cumulative health index is obtained by integrating the time : in, For the mapping function: in, represents the mapping function input variable, To adjust the parameters; Used to reflect the nonlinear relationship between the cumulative health index and the overall status of the equipment It is expressed as: in For the i The weight of each sensor feature; ; Based on the accumulated health status and the preset health threshold , estimated remaining life : in, is the equipment health degradation rate; S303. Take any time window data as sample feature and add cumulative health index , Remaining life and as a label; is the fault status label, : , indicating that the cross belt sorter has failed within the time window; , indicating that the cross-belt sorter has not failed within the time window; is the cumulative health index label; is the remaining life label; The training sample of time window data is recorded as ; S304. For each time window data, repeat steps S302 to S303 to obtain the training samples corresponding to each time window data, add them to the same set to form a training set.

7. According to claim 6, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: In step S4, a CNN model is constructed to perform fault diagnosis, cumulative health index and remaining life prediction: S401. Input data tensor construction: After the data preprocessing module, the real-time collected sensor data is normalized, denoised and windowed to form the following data matrix: Reconstruct the data into a three-dimensional tensor format, set in the form of "height-width-channel", where the height Corresponding to the number of sensors deployed, the width Corresponding time step, number of channels According to the specific fusion method, each sensor data constitutes a channel separately. Represents the feature channel corresponding to each sensor; the final input tensor is recorded as ; S402. Randomly crop the input tensor on the time axis, inject noise to simulate sensor errors, and randomly scale the input tensor to expand the data set, so that the model can better adapt to data variations in actual work; S403. By stacking multiple layers of two-dimensional convolutional layers, the model gradually captures features from local to global; In a multi-layer stacked 2D convolutional layer, the output feature map of the previous 2D convolutional layer is used as the input feature map of the next 2D convolutional layer, and the input feature map of the first 2D convolutional layer is the input tensor; the output feature map of the last 2D convolutional layer is the final output of the multi-layer stacked 2D convolutional layer; For In the layer, Output channels at position The eigenvalues ​​of are defined as follows: in, For the Tier convolution kernel weights, where is the index of the convolution kernel in the height and width directions, is the index of the input channel, is the index of the output channel, and the kernel size is ; Output channel Represents the feature map generated by the convolution kernel channels; The input feature map Tier Channels at position ( ) value; is the bias value; is the ReLU activation function; the number of channels of the input tensor is , the number of output tensor channels is determined by the number of set convolution kernels; In this process, a pooling layer needs to be added after each two-dimensional convolutional layer to downsample and output; S404. Process the input tensor through multiple stacked one-dimensional convolutional layers; the input feature map of the first one-dimensional convolutional layer is the input tensor, the output feature map of the previous one-dimensional convolutional layer is used as the input feature map of the next one-dimensional convolutional layer, and the output of the last one-dimensional convolutional layer is used as the final output of the multiple stacked one-dimensional convolutional layers; The multi-layer stacked one-dimensional convolutional layers and the multi-layer stacked two-dimensional convolutional layers are parallel different branch structures; In this process, a pooling layer needs to be added after each one-dimensional convolution layer to downsample and output; S405. By merging the outputs of different branch structures, that is, splicing the outputs of multi-layer stacked two-dimensional convolutional layers and multi-layer stacked one-dimensional convolutional layers, a spliced ​​feature map is obtained, and then the feature map is flattened, and the extracted features are integrated through a fully connected network, and finally the fault status, cumulative health index, and remaining life prediction are output. As a multidimensional vector: ； They represent the predicted values ​​of fault status, cumulative health index and remaining life respectively.

8. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The optimization design of the CNN model includes: A1. Assume that the CNN model is updated by mini-batch gradient descent. Divide the samples in the training set into multiple batches. Assume that each batch contains N samples. Use the samples in each batch to update the CNN model: For any batch of samples, based on the constructed CNN network, for multi-task output, the loss function is designed as the weighted sum of multiple sub-losses, including: For each sample in the current batch, the sample features are used as the input of the CNN model, and the output of the CNN model is used as the predicted value; For the fault status, cumulative health index and remaining life prediction, the total loss It is expressed as: , To adjust the importance weight of each part of the loss; in, The loss function representing the fault state adopts the cross entropy loss function: Among them, in the current batch of samples, the health status label and predicted value of the nth sample are recorded as , ; n = 1, 2, ..., N; Represents the cumulative health index loss function: Among them, in the current batch of samples, the cumulative health index label and predicted value of the nth sample are recorded as , ; n = 1, 2, ..., N; Represents the remaining life loss function, using mean square error loss: Among them, in the current batch of samples, the cumulative remaining life label and predicted value of the nth sample are recorded as , ; n = 1, 2, ..., N; A2. Repeat step A1 for each batch of samples in the training set, and then train the CNN model based on the total loss function until the model converges or all batches of samples are trained, and then a mature CNN model is obtained for prediction.

9. According to claim 1, a digital twin-based intelligent maintenance method for small-parts sorting machines in logistics transfer yards is characterized by: The calculation of the early warning index and the designation of the maintenance strategy include: Cumulative health index output in CNN model It is regarded as a quantitative indicator of the current status of the equipment, and the warning level is constructed by combining the changing trend of the fault status and the remaining life prediction. The specific definition is: in, , is the weight parameter, is the preset lifespan threshold; Different maintenance strategies are set according to the value of the warning index. Specifically, the warning index is divided into intervals, and different maintenance strategies are formulated for each interval warning index.

Images