A method and system for monitoring chain slippage of tobacco leaf cutting machine
By constructing a chain slip monitoring model based on CNN and DNN, combining motor current and speed data, real-time monitoring of chain slip slip of tobacco leaf cutters has been solved, and the problem of insufficient chain slip monitoring in the existing technology has been improved, and the quality of tobacco and equipment intelligence level has been improved.
Patent Information
- Application Number
- CN202310950843.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-07-31
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2043-07-31
AI Technical Summary
The existing technology lacks effective monitoring methods to detect the slipping of the tobacco leaf shredder as early as possible, which leads to fluctuations in the width of the tobacco wire and affects the quality of the shred and yield.
Using a monitoring model based on CNN and DNN, combining motor current, chain speed and tobacco width data, through data preprocessing, feature extraction and classifier training, the chain slip situation is monitored in real time, and a chain slip monitoring model is constructed.
It improves the yield rate of tobacco, enhances the adaptability and accuracy of the monitoring model, and can effectively identify the synchronous slippage of upper and lower sequential chains, realizing intelligent upgrade of the equipment.
Smart Images
Figure CN116831309B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of cigarette manufacturing, and more particularly, relates to a method and system for monitoring chain slippage of a tobacco leaf cutting machine. Background Art
[0002] The copper chain is a key component of the tobacco cutter, responsible for compacting and conveying tobacco leaves from the feed trolley to the cutter rollers. The chain's conveying speed and the roller's rotational speed jointly determine the width of the tobacco. To ensure consistent tobacco width, the chain's conveying speed and the roller's rotational speed are mechanically or electronically locked in the cutter. During the conveying and extrusion process, moisture and organic matter from the tobacco leaves are retained on the chain and the side walls of the feed channel. This can cause the copper chain to slip, altering the transmission relationship between the chain's conveying speed and the roller's rotational speed. This can lead to fluctuations in tobacco leaf width during the cutting process, affecting the quality of the cut tobacco.
[0003] As tobacco companies increasingly demand digital transformation and upgrades, they are increasingly focusing on production process monitoring, hoping to detect manufacturing anomalies early and minimize losses. However, there are no solutions or methods for monitoring copper busbar slippage. Summary of the Invention
[0004] In response to the above-mentioned defects or improvement needs of the prior art, the present invention provides a method and system for monitoring the slippage of the tobacco leaf cutter chain, in order to monitor the slippage of the tobacco leaf cutter chain, reduce the problems of tobacco width fluctuation and yield reduction caused by copper chain slippage, thereby improving the yield of tobacco.
[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:
[0006] The method for monitoring chain slippage of a tobacco leaf cutter according to the present invention is characterized in that it comprises the following steps:
[0007] Step 1: Collect historical operating data of the tobacco cutter, including process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width; wherein the process parameters include tobacco blend, feed height, and blade door pressure;
[0008] Step 2: extracting operation data of an effective processing period from the historical operation data, and then aligning the intercepted rotational speeds of the upper and lower chains, currents of the upper and lower chain motors, process parameters, and tobacco width in chronological order within the effective processing period to obtain an aligned original data packet;
[0009] Step 3: performing dimensionality reduction processing on the currents of the upper and lower chain motors in the original data packet, and then filtering the current eigenvalues after dimensionality reduction to obtain filtered current eigenvalues;
[0010] Step 4: Obtain the rotation speeds and tobacco widths of the upper and lower chain under the same process parameters from the original data packet, and match them with the filtered current characteristic values to form an analysis data packet under the same process parameters. Together with the corresponding chain slippage fault mark, this data packet constitutes a training sample, thereby forming a training data set consisting of N training samples under different process parameters.
[0011] The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as p n,i , the chain slip fault of the nth analysis data packet is recorded as the real label flag n , and flag n ∈{1,2...,k,...,K}, K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as n k ;
[0012] The current of the upper and lower chain motors in the nth original data packet is recorded as c n ;
[0013] Step 5: Build a chain slippage monitoring network, including: CNN-based feature extraction layer, digital processing module, DNN-based feature extraction layer, feature fusion layer and classifier, and train the network:
[0014] Step 5.1: Construct a CNN-based feature extraction layer, which is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence;
[0015] The current c in the nth original data packet n Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained;
[0016] Step 5.2: The digital processing module processes the current c in the nth original data packet. n and the i-th current characteristic value p n,i After performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained;
[0017] Step 5.3: Construct a DNN-based feature extraction layer, which is composed of q2 DNN unit blocks, and each DNN unit block is composed of a normalized BN, an activation function ReLU, and a DNN layer in sequence;
[0018] The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain a feature processing result dnnResult_n;
[0019] Step 5.4: The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and inputs it into the fully connected layer for processing, outputting the nth fusion feature vector;
[0020] Step 5.5: The nth fused feature vector is input into the classifier, and after being processed by the Softmax function, the chain slip fault prediction label corresponding to the nth fused feature vector is output;
[0021] Step 5.6: Based on the real label flag n The cross entropy loss function L of the chain slippage monitoring network is constructed with the predicted label n , based on the training data set, the chain slip monitoring network is trained using the gradient descent algorithm, and the cross entropy loss function is calculated to update the network parameters until the cross entropy loss function converges or reaches the maximum number of iterations, thereby obtaining the trained chain slip monitoring model.
[0022] Step 6: Use the trained chain slippage monitoring model to predict the real-time collected cutter process parameters, upper and lower chain speeds, and upper and lower chain motor currents to obtain chain slippage fault prediction results to determine whether the tobacco cutter chain is slipping.
[0023] The method for establishing a chain slip monitoring model according to the present invention is also characterized in that the screening process in step 3 includes:
[0024] If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value.
[0025] If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
[0026] The present invention provides a tobacco leaf cutter chain slippage monitoring system that includes:
[0027] The data acquisition module is used to collect historical operating data of the tobacco leaf cutter, including process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width;
[0028] A data preprocessing module is used to preprocess the historical operation data to obtain training samples under different process parameters and form a training data set;
[0029] The data input module is used to input the real-time collected process parameters of the wire cutting machine, the upper and lower chain speeds, and the upper and lower chain motor currents into the diagnosis output module;
[0030] The diagnostic output module is used to build a chain slip monitoring network and train the chain slip monitoring network with a training data set to obtain a trained chain slip monitoring model. The trained chain slip monitoring model is then used to predict the real-time collected cutter process parameters, upper and lower chain speeds, and upper and lower chain motor currents to obtain prediction labels to determine whether the chain of the tobacco leaf cutter is slipping.
[0031] The system of the present invention is also characterized in that the data preprocessing module includes the following sub-parts:
[0032] an alignment submodule, configured to extract the operation data of an effective processing period from the historical operation data, and then align the extracted rotational speeds of the upper and lower chain arranging machines, the currents of the upper and lower chain arranging machines, the process parameters, and the width of the tobacco in chronological order within the effective processing period to obtain an aligned original data packet;
[0033] a feature extraction submodule for performing noise reduction filtering on the currents of the upper and lower chain motors in the original data packet, extracting current characteristic values from the currents after noise reduction filtering, and filtering the extracted current characteristic values to obtain filtered current characteristic values; wherein the current characteristic values include energy, peak value, and frequency;
[0034] The data marking submodule obtains the rotation speeds and tobacco widths of the upper and lower chain strips under the same process parameters from the original data packet, and matches them with the filtered current characteristic values to form an analysis data packet under the same process parameters. The data packet and the corresponding chain strip slippage fault together constitute a training sample, thereby forming a training data set composed of N training samples under different process parameters;
[0035] The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as p n,i , the chain slip fault of the nth analysis data packet is recorded as the real label flag n , and flag n ∈{1,2...,k,...,K}, K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as n k ;
[0036] The current of the upper and lower chain motors in the nth original data packet is recorded as c n .
[0037] The screening process in the feature extraction submodule includes:
[0038] If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value.
[0039] If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
[0040] The chain slip monitoring network in the diagnostic output module includes: a CNN-based feature extraction layer, a digital processing module, a DNN-based feature extraction layer, a feature fusion layer and a classifier;
[0041] The CNN-based feature extraction layer is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence;
[0042] The current c in the nth original data packet n Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained;
[0043] The digital processing module processes the current c in the nth original data packet. n and the i-th current characteristic value p n,i After performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained;
[0044] The DNN-based feature extraction layer is composed of q2 DNN unit blocks, and each DNN unit block is composed of a normalized BN, an activation function ReLU and a DNN layer in sequence;
[0045] The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain a feature processing result dnnResult_n;
[0046] The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and inputs it into the fully connected layer for processing, and outputs the nth fusion feature vector;
[0047] The nth fused feature vector is input into the classifier, and after being processed by the Softmax function, the predicted label corresponding to the nth fused feature vector is output;
[0048] Based on the true label flag n The cross entropy loss function L of the chain slippage monitoring network is constructed with the predicted label n, based on the training data set, the chain slip monitoring network is trained using the gradient descent algorithm, and the cross entropy loss function is calculated to update the network parameters until the cross entropy loss function converges or reaches the maximum number of iterations, thereby obtaining the trained chain slip monitoring model.
[0049] Compared with the prior art, the present invention has the following beneficial effects:
[0050] 1. Based on technical parameters such as motor current, upper and lower chain speeds, and tobacco width, the present invention uses correlation analysis, machine learning, and other methods to establish a slippage monitoring model. This model is then used to monitor copper chain slippage in real time, preventing it from affecting tobacco width and thus improving tobacco yield.
[0051] 2. The present invention matches the current characteristic values, upper and lower chain speed data, and tobacco width data under the same process parameters to form an analysis data package. The analysis data package is used to construct a monitoring model, making the monitoring model adaptable under different process parameters, thereby effectively improving the model monitoring accuracy and model generalization ability;
[0052] 3. When the upper and lower chains of the tobacco leaf cutter slip synchronously, the tobacco width will not fluctuate. Therefore, when constructing the chain slip monitoring model, the present invention introduces tobacco width data and eliminates the current data when the upper and lower chains slip synchronously, avoiding abnormal data contamination of the data model. Therefore, the present invention can effectively monitor the situation where the upper and lower chains slip synchronously, and avoid the situation where conventional monitoring methods cannot identify the situation where the upper and lower chains slip synchronously.
[0053] 4. The technical solution of the present invention can be implemented with a modular approach and can be directly installed on existing tobacco leaf cutters, easily realizing the intelligent upgrade of existing equipment. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 is a flow chart of the method of the present invention;
[0055] Figure 2 Schematic diagram of the chain slip monitoring model of the present invention. DETAILED DESCRIPTION
[0056] In this embodiment, Figure 1 As shown, a method for monitoring chain slippage of a tobacco leaf cutting machine comprises the following steps:
[0057] Step 1: Install the data acquisition system:
[0058] Install the communication module of the wire cutter control system to collect and transmit relevant parameters when the wire cutter is working; install the current data acquisition module of the upper and lower chain motors in the wire cutter control cabinet to collect and transmit the current of the upper and lower chain motors;
[0059] Check and adjust the time of the data acquisition system to align it with the time of the shredder control system to ensure the consistency of the collected data time;
[0060] Step 2: Collect data from the cutting machine:
[0061] Collect data such as the power on / off, feeding height, knife door pressure, and upper and lower chain speeds of the wire cutting machine control system; and also collect current data of the upper and lower chain motors of the wire cutting machine;
[0062] Sampling and testing of shred width is carried out in accordance with the shredding process requirements. The time interval between the time when the sample tobacco leaves are put into the feeding trolley and the time when the shredded tobacco leaves are cut is monitored and recorded, as well as the time when the sample tobacco leaves are put into the feeding trolley.
[0063] The resulting historical operating data of the tobacco cutter includes process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width. The process parameters include tobacco blend, feed height, and blade pressure.
[0064] Step 3: Data preprocessing of the shredder:
[0065] The operation data of the effective processing period is intercepted from the historical operation data, and then the rotation speeds of the upper and lower chains, the currents of the upper and lower chain motors, the process parameters and the width of the tobacco are aligned in chronological order within the effective processing period to obtain the aligned original data packet;
[0066] During the specific implementation process, data preprocessing performs the following operations:
[0067] By recording the start and stop time of the wire cutter, we can obtain the system start and stop signals and the time points of the current and sensor signals, and align the upper and lower chain speeds with the current data time.
[0068] Record the time interval data from the time the sample tobacco leaves are put into the feeding trolley to the time they are cut, and the time interval data from the tobacco leaves being transferred from the cutter to the blanking frame at the visual inspection area. Calculate the effective processing period data for cutting the sample tobacco leaves, record the time period for cutting the sample tobacco leaves, and align the tobacco sampling width with the cutting process data.
[0069] The aligned process parameters (tobacco leaf formula, feeding height and knife gate pressure), the rotation speed of the upper and lower chain, the current data of the upper and lower chain motors and the tobacco width data are packaged to obtain the aligned original data packet;
[0070] Step 4: Screening of chain current characteristic values:
[0071] The currents of the upper and lower chain motors in the original data packets are subjected to dimensionality reduction processing, including noise reduction filtering. Time-frequency domain eigenvalues, including energy, peak value, and frequency, are then extracted from the noise-reduction filtered currents. The reduced current eigenvalues are then screened to select those that can be used to initially identify chain slippage faults. This reduces the risk of chain slippage marker contamination and the workload of subsequent model training, ultimately yielding the screened current eigenvalues.
[0072] Specifically, the screening process includes:
[0073] If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value.
[0074] If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
[0075] Step 5: Mark the shredding process data:
[0076] The rotation speeds and tobacco widths of the upper and lower chains under the same process parameters are obtained from the original data packet and matched with the filtered current characteristic values to form an analysis data packet under the same process parameters. Together with the corresponding chain slippage fault mark, this constitutes a training sample. Thus, a training data set is formed by training samples under N different process parameters.
[0077] The specific operations are as follows:
[0078] The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as p n,i , the chain slip fault of the nth analysis data packet is recorded as the real label flag n , and flag n ∈{1,2...,k,...,K}, K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as n k ;
[0079] The current of the upper and lower chain motors in the nth original data packet is recorded as c n ;
[0080] Step 6: Monitor model training:
[0081] Construct a chain slippage monitoring network, including: CNN-based feature extraction layer, digital processing module, DNN-based feature extraction layer, feature fusion layer and classifier, such as Figure 2 As shown, and train the network:
[0082] In this embodiment, the model has two different inputs. One input has the basic structure of a 7-layer CNN, but it uses large convolution kernels to extract features. The other input has the basic structure of a 3-layer DNN, which processes traditional features in a fully connected manner. The two dimensions of features are then fused to blend the information. This method allows the features to interfere with each other, forming a mutually reinforcing mechanism, further enabling the learned features to have a deeper expression.
[0083] Step 6.1: Construct a CNN-based feature extraction layer, which is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence;
[0084] The current c in the nth original data packet n Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained;
[0085] Step 6.2: The digital processing module processes the current c in the nth original data packet. n and the i-th current characteristic value p n,i After performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained;
[0086] Step 6.3: Construct a DNN-based feature extraction layer, which is composed of q2 DNN unit blocks, and each DNN unit block is composed of a normalized BN, an activation function ReLU, and a DNN layer in sequence;
[0087] The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain the feature processing result dnnResult_n;
[0088] Step 6.4: The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and inputs it into the fully connected layer for processing, outputting the nth fusion feature vector.
[0089] Step 6.5: The nth fused feature vector is input into the classifier and processed by the Softmax function, and the chain slip fault prediction label corresponding to the nth fused feature vector is output;
[0090] Step 6.6: Based on the real label flag n The cross entropy loss function L of the chain slippage monitoring network is constructed with the predicted label nBased on the training data set, the chain slip monitoring network is trained using the gradient descent algorithm, and the cross entropy loss function is calculated to update the network parameters until the cross entropy loss function converges or reaches the maximum number of iterations, thereby obtaining the trained chain slip monitoring model.
[0091] Step 7: Application of chain slip monitoring model:
[0092] The trained chain slip monitoring model is used to predict the real-time collected process parameters of the tobacco cutter, the upper and lower chain speeds, and the upper and lower chain motor currents to obtain the chain slip fault prediction results to determine whether the tobacco cutter chain is slipping.
[0093] In this embodiment, a tobacco cutter chain slippage monitoring system includes:
[0094] The data acquisition module is used to collect historical operating data of the tobacco leaf cutter, including process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width;
[0095] The data preprocessing module is used to preprocess the historical operation data to obtain training samples under different process parameters and form a training data set;
[0096] The data preprocessing module includes the following sub-parts:
[0097] The alignment submodule is used to intercept the operation data of the effective processing period from the historical operation data, and then align the intercepted upper and lower chain speeds, upper and lower chain machine currents, process parameters and tobacco width in chronological order within the effective processing period to obtain the aligned original data packet;
[0098] The feature extraction submodule is used to perform noise reduction filtering on the currents of the upper and lower chain motors in the original data packet, extract current characteristic values from the noise reduction filtered currents, and filter the extracted current characteristic values to obtain filtered current characteristic values; the current characteristic values include: energy, peak value, and frequency;
[0099] Among them, the screening process in the feature extraction submodule includes:
[0100] If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value.
[0101] If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
[0102] The data labeling submodule obtains the rotation speed and tobacco width of the upper and lower chain under the same process parameters from the original data packet, and matches them with the filtered current characteristic values to form an analysis data packet under the same process parameters. This data packet, together with the corresponding chain slippage fault, constitutes a training sample. Thus, the training data set is composed of N training samples under different process parameters.
[0103] The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as p n,i , the chain slip fault of the nth analysis data packet is recorded as the real label flag n , and flag n ∈{1,2...,k,...,K}, K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as n k ;
[0104] The current of the upper and lower chain motors in the nth original data packet is recorded as c n .
[0105] The data input module is used to input the real-time collected process parameters of the wire cutting machine, the upper and lower chain speeds, and the upper and lower chain motor currents into the diagnosis output module;
[0106] The diagnostic output module is used to build a chain slip monitoring network and train the chain slip monitoring network with a training data set to obtain a trained chain slip monitoring model. The trained chain slip monitoring model is then used to predict the real-time collected cutter process parameters, upper and lower chain speeds, and upper and lower chain motor currents to obtain prediction labels to determine whether the chain of the tobacco leaf cutter is slipping.
[0107] In this embodiment, Figure 2 As shown in the figure, the chain slip monitoring network in the diagnostic output module includes: a CNN-based feature extraction layer, a digital processing module, a DNN-based feature extraction layer, a feature fusion layer, and a classifier;
[0108] The CNN-based feature extraction layer is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence;
[0109] The current c in the nth original data packet n Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained;
[0110] The digital processing module processes the current c in the nth original data packet. n and the i-th current characteristic value p n,iAfter performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained;
[0111] The DNN-based feature extraction layer is composed of q2 DNN unit blocks, and each DNN unit block is composed of normalized BN, activation function ReLU and DNN layer in sequence;
[0112] The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain the feature processing result dnnResult_n;
[0113] The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and input it into the fully connected layer for processing, outputting the nth fusion feature vector;
[0114] The nth fused feature vector is input into the classifier and processed by the Softmax function, and the predicted label corresponding to the nth fused feature vector is output;
[0115] Based on the true label flag n The cross entropy loss function L of the chain slippage monitoring network is constructed with the predicted label n Based on the training data set, the chain slip monitoring network is trained using the gradient descent algorithm, and the cross entropy loss function is calculated to update the network parameters until the cross entropy loss function converges or reaches the maximum number of iterations, thereby obtaining the trained chain slip monitoring model.
Claims
1. A method for monitoring chain slippage of a tobacco leaf cutter, characterized in that: The following steps are involved: Step 1: Collect historical operating data of the tobacco cutter, including process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width; wherein the process parameters include tobacco blend, feed height, and blade door pressure; Step 2: extracting operation data of an effective processing period from the historical operation data, and then aligning the intercepted rotational speeds of the upper and lower chains, currents of the upper and lower chain motors, process parameters, and tobacco width in chronological order within the effective processing period to obtain an aligned original data packet; Step 3: performing dimensionality reduction processing on the currents of the upper and lower chain motors in the original data packet, and then filtering the current eigenvalues after dimensionality reduction to obtain filtered current eigenvalues; Step 4: Obtain the rotation speeds and tobacco widths of the upper and lower chain under the same process parameters from the original data packet, and match them with the filtered current characteristic values to form an analysis data packet under the same process parameters. Together with the corresponding chain slippage fault mark, this data packet constitutes a training sample, thereby forming a training data set consisting of N training samples under different process parameters. The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as , the chain slip fault of the nth analysis data packet is recorded as the true label ,and , K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as ; The current of the upper and lower chain motors in the nth original data packet is recorded as ; Step 5: Build a chain slippage monitoring network, including: CNN-based feature extraction layer, digital processing module, DNN-based feature extraction layer, feature fusion layer and classifier, and train the network: Step 5.1: Construct a CNN-based feature extraction layer, which is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence; The current in the nth original data packet Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained; Step 5.2: The digital processing module processes the current in the nth original data packet. and the i-th current eigenvalue After performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained; Step 5.3: Construct a DNN-based feature extraction layer, which is composed of q2 DNN unit blocks, and each DNN unit block is composed of a normalized BN, an activation function ReLU, and a DNN layer in sequence; The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain a feature processing result dnnResult_n; Step 5.4: The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and inputs it into the fully connected layer for processing, outputting the nth fusion feature vector; Step 5.5: The nth fused feature vector is input into the classifier, and after being processed by the Softmax function, the chain slip fault prediction label corresponding to the nth fused feature vector is output; Step 5.6: Based on the real label The cross entropy loss function of the chain slippage monitoring network is constructed by using the predicted labels , thereby training the chain slippage monitoring network based on the training data set using a gradient descent algorithm, and calculating the cross entropy loss function to update the network parameters until the cross entropy loss function converges or reaches a maximum number of iterations, thereby obtaining a trained chain slippage monitoring model; Step 6: Use the trained chain slippage monitoring model to predict the real-time collected cutter process parameters, upper and lower chain speeds, and upper and lower chain motor currents to obtain chain slippage fault prediction results to determine whether the tobacco cutter chain is slipping.
2. The method for monitoring chain slippage of a tobacco leaf cutter according to claim 1, characterized in that: The screening process in step 3 includes: If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value. If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
3. A tobacco leaf cutter chain slip monitoring system, characterized in that: include: The data acquisition module is used to collect historical operating data of the tobacco leaf cutter, including process parameters, upper and lower chain speeds, upper and lower chain motor currents, and tobacco width; A data preprocessing module is used to preprocess the historical operation data to obtain training samples under different process parameters and form a training data set; The data preprocessing module includes the following sub-sections: an alignment submodule, configured to extract the operation data of an effective processing period from the historical operation data, and then align the extracted rotational speeds of the upper and lower chain arranging machines, the currents of the upper and lower chain arranging machines, the process parameters, and the width of the tobacco in chronological order within the effective processing period to obtain an aligned original data packet; a feature extraction submodule for performing noise reduction filtering on the currents of the upper and lower chain motors in the original data packet, extracting current characteristic values from the currents after noise reduction filtering, and filtering the extracted current characteristic values to obtain filtered current characteristic values; wherein the current characteristic values include energy, peak value, and frequency; The data marking submodule obtains the rotation speeds and tobacco widths of the upper and lower chain strips under the same process parameters from the original data packet, and matches them with the filtered current characteristic values to form an analysis data packet under the same process parameters. The data packet and the corresponding chain strip slippage fault together constitute a training sample, thereby forming a training data set composed of N training samples under different process parameters; The i-th current characteristic value of the n-th analysis data packet in any n-th training sample is recorded as , the chain slip fault of the nth analysis data packet is recorded as the true label ,and , K represents the number of real labels, and the number of analysis data packets of the k-th type of chain slip fault is recorded as ; The current of the upper and lower chain motors in the nth original data packet is recorded as ; The data input module is used to input the real-time collected process parameters of the wire cutting machine, the upper and lower chain speeds, and the upper and lower chain motor currents into the diagnosis output module; The diagnostic output module is used to build a chain slip monitoring network and train the chain slip monitoring network with a training data set to obtain a trained chain slip monitoring model. The trained chain slip monitoring model is then used to predict the real-time collected cutter process parameters, upper and lower chain speeds, and upper and lower chain motor currents to obtain prediction labels to determine whether the chain of the tobacco leaf cutter is slipping.
4. The system according to claim 3, characterized in that The screening process in the feature extraction submodule includes: If the tobacco width is normal, determine whether the fluctuation between the current characteristic value after dimension reduction and the historical current characteristic value under the same process parameters exceeds the threshold. If so, discard the current characteristic value after dimension reduction; otherwise, retain the current characteristic value after dimension reduction and use it as the effective current characteristic value. If the tobacco width is abnormal, the current characteristic value after dimensionality reduction is recorded and used as the alarm threshold.
5. The system according to claim 4, characterized in that The chain slip monitoring network in the diagnostic output module includes: a CNN-based feature extraction layer, a digital processing module, a DNN-based feature extraction layer, a feature fusion layer and a classifier; The CNN-based feature extraction layer is composed of q1 CNN unit blocks, and each CNN unit block is composed of normalized BN, activation function ReLU and one-dimensional convolution in sequence; The current in the nth original data packet Input into the CNN-based feature extraction layer, and after being processed by q1 CNN unit blocks in sequence, the nth feature processing result cnnResult_n is obtained; The digital processing module processes the current in the nth original data packet. and the i-th current eigenvalue After performing time domain feature extraction, frequency feature extraction and time-frequency feature extraction respectively, the nth current time-frequency feature matrix is obtained; The DNN-based feature extraction layer is composed of q2 DNN unit blocks, and each DNN unit block is composed of a normalized BN, an activation function ReLU and a DNN layer in sequence; The nth current time-frequency feature matrix is input into the DNN-based feature and processed by q2 DNN unit blocks in sequence to obtain a feature processing result dnnResult_n; The feature fusion layer fuses cnnResult_n and dnnResult_n to obtain the nth fusion result and inputs it into the fully connected layer for processing, and outputs the nth fusion feature vector; The nth fused feature vector is input into the classifier, and after being processed by the Softmax function, the predicted label corresponding to the nth fused feature vector is output; Based on the real label The cross entropy loss function of the chain slippage monitoring network is constructed by using the predicted labels , based on the training data set, the chain slip monitoring network is trained using the gradient descent algorithm, and the cross entropy loss function is calculated to update the network parameters until the cross entropy loss function converges or reaches the maximum number of iterations, thereby obtaining the trained chain slip monitoring model.