Knife jump detection and compensation method and system and storage medium

Abstract

The invention provides a knife jumping detection and compensation method and system and a storage medium, and the method comprises the steps: collecting the operation state data between a doctor blade and a plate cylinder, and carrying out the time synchronization and noise reduction processing of the operation state data, and obtaining a knife jumping feature matrix; judging whether a blade jumping event occurs or not based on the blade jumping characteristic matrix, and if not, collecting operation state data between the doctor blade and the plate cylinder; if yes, determining a knife jumping type based on the knife jumping characteristic matrix, obtaining working state data of the gravure press, determining compensation information based on the knife jumping type and the working state data, and generating a compensation instruction based on the compensation information; compensation state data obtained after the compensation instruction is responded is obtained, whether a blade jumping event exists or not is determined based on the compensation state data, and if not, operation state data between the doctor blade and the plate cylinder are collected; and if yes, generating early warning information. The operation state of the doctor blade can be monitored in real time, and the blade jumping event and the type thereof can be accurately judged in time, so that the blade jumping influence is eliminated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of knife jumping, and in particular to a method and system for detecting and compensating for knife jumping, and a storage medium. BACKGROUND

[0002] As a high-efficiency and high-quality printing process, gravure printing is widely used in packaging, publishing, and decorative materials. In the gravure printing process, the doctor blade is one of the key functional components, and its main function is to remove excess ink on the surface of the printing cylinder and only keep the ink in the pits to achieve image transfer. However, under high-speed printing conditions, the doctor blade and the printing cylinder often unexpectedly separate or vibrate abnormally.

[0003] Knife jumping can cause quality defects such as knife lines, uneven ink color, and missing images, resulting in material waste and affecting the service life of the printing cylinder and the efficiency of the equipment. In the prior art, visual inspection or regular replacement of the doctor blade is usually relied on the experience of the operator, and there is a lack of real-time and accurate knife jumping detection means. Therefore, after identifying the knife jumping in the prior art, manual intervention is often relied on for adjustment, which is slow in response and low in adjustment progress. SUMMARY

[0004] In order to monitor the running state of the doctor blade in real time, accurately determine the knife jumping event and its type in a timely manner, and eliminate the influence of knife jumping, the present application provides a method and system for detecting and compensating for knife jumping, and a storage medium.

[0005] In a first aspect, the present application provides a method for detecting and compensating for knife jumping, which comprises: Collecting running state data between the doctor blade and the printing cylinder, and performing time synchronization and noise reduction processing on the running state data to obtain a knife jumping feature matrix; Determining whether a knife jumping event occurs based on the knife jumping feature matrix, and if not, continuing to collect the running state data between the doctor blade and the printing cylinder; If yes, determining the type of knife jumping based on the knife jumping feature matrix, obtaining working state data of the gravure press, determining compensation information based on the type of knife jumping and the working state data, and generating corresponding compensation instructions based on the compensation information; Obtaining compensation state data after responding to the compensation instructions, determining whether there is a knife jumping event based on the compensation state data, and if not, continuing to collect the running state data between the doctor blade and the printing cylinder; If yes, generating a warning information.

[0006] In some embodiments, the running state data includes gap information, vibration information and contact pressure information, a laser displacement sensor, a vibration sensor and a pressure sensor are respectively arranged at preset positions near the contact area between the doctor blade and the plate cylinder, the detection direction of the laser displacement sensor is perpendicular to the surface of the plate cylinder, the vibration sensor is fixedly installed at the non-center of force of the doctor blade holder, and the pressure sensor is integrated on the pressure transmission path of the pressure adjusting mechanism of the doctor blade. The laser displacement sensor collects gap data between the blade edge of the doctor blade and the surface of the plate cylinder to obtain the gap information. The vibration sensor collects vibration frequency and amplitude data of the doctor blade during use to generate the vibration information. The pressure sensor collects contact pressure data between the doctor blade and the plate cylinder to generate the contact pressure information.

[0007] In some embodiments, the running state data includes gap information, vibration information and contact pressure information, a laser displacement sensor, a vibration sensor and a pressure sensor are respectively arranged at preset positions near the contact area between the doctor blade and the plate cylinder, the detection direction of the laser displacement sensor is perpendicular to the surface of the plate cylinder, the vibration sensor is fixedly installed at the non-center of force of the doctor blade holder, and the pressure sensor is integrated on the pressure transmission path of the pressure adjusting mechanism of the doctor blade. The laser displacement sensor collects gap data between the blade edge of the doctor blade and the surface of the plate cylinder to obtain the gap information. The vibration sensor collects vibration frequency and amplitude data of the doctor blade during use to generate the vibration information. The pressure sensor collects contact pressure data between the doctor blade and the plate cylinder to generate the contact pressure information.

[0008] In some embodiments, the determining whether a skip occurs based on the skip feature matrix includes: inputting the skip feature matrix into a preset skip identification model to output a probability of occurrence of a skip event; The skip identification model is a multi-channel time sequence classification network based on an attention mechanism, including a feature embedding layer for mapping the skip feature matrix into a high-dimensional feature vector, a multi-head self-attention layer for extracting time sequence dependency between different channel features, a time sequence convolution layer for capturing skip pattern features within a local time window, and a fully connected classification layer for outputting a skip event probability value. If the occurrence probability exceeds a first preset threshold, a skip event occurs. If the occurrence probability does not exceed the first preset threshold, no skip event occurs.

[0009] In some embodiments, the method further comprises: monitoring online decision accuracy and prediction confidence of the jump blade identification model; calculating a drift index between data distribution of the jump blade feature matrix and historical training base distribution; detecting whether a preset working condition change event occurs in the gravure press; determining that the jump blade identification model needs to be adjusted when at least one of the online decision accuracy is lower than a first threshold, the prediction confidence is lower than a second threshold, the drift index is higher than a third threshold, or the preset working condition change event occurs; wherein, during the adjustment of the jump blade identification model, a focal loss function is introduced as an optimization target, a sliding window enhancement strategy is used for data enhancement of training samples, and an early stopping mechanism and a learning decay strategy are used to optimize the training process.

[0010] In some embodiments, the method further comprises: monitoring online decision accuracy and prediction confidence of the jump blade identification model; calculating a drift index between data distribution of the jump blade feature matrix and historical training base distribution; detecting whether a preset working condition change event occurs in the gravure press; determining that the jump blade identification model needs to be adjusted when at least one of the online decision accuracy is lower than a first threshold, the prediction confidence is lower than a second threshold, the drift index is higher than a third threshold, or the preset working condition change event occurs; wherein, during the adjustment of the jump blade identification model, a focal loss function is introduced as an optimization target, a sliding window enhancement strategy is used for data enhancement of training samples, and an early stopping mechanism and a learning decay strategy are used to optimize the training process.

[0011] In some embodiments, the determining compensation information based on the jump blade type and the working state data comprises: obtaining a historical jump blade event library corresponding to the jump blade type, retrieving a matching historical jump blade event from the historical jump blade event library based on the jump blade type and the working state data, and extracting a compensation strategy corresponding to the historical jump blade event, the compensation strategy being determined as a candidate compensation strategy; based on the jump blade feature matrix and the working state data, constructing a context feature vector of a current jump blade event; performing similarity matching between the context feature vector and a historical context feature vector corresponding to the candidate compensation strategy, and screening at least one target historical event with a similarity higher than a second preset threshold. obtain a compensation effect evaluation result of the target historical event, determine a target compensation strategy from the candidate compensation strategies based on the compensation effect evaluation result, and generate compensation information containing a compensation parameter and an execution timing.

[0012] In some of the embodiments, the method further comprises: Calibrating zero points of the laser displacement sensor, the vibration sensor, and the pressure sensor before collecting the running state data between the doctor blade and the plate cylinder.

[0013] In a second aspect, the embodiments provide a system for detecting and compensating for a skip, which comprises a skip detection module, a skip analysis module, and a skip compensation module, wherein: The skip detection module is configured to collect running state data between a doctor blade and a plate cylinder. The skip analysis module is configured to perform time synchronization and noise reduction processing on the running state data to obtain a skip feature matrix, and determine whether a skip event has occurred based on the skip feature matrix. The skip detection module is further configured to continue collecting running state data between the doctor blade and the plate cylinder if no skip event has occurred. The skip compensation module is further configured to determine a skip type based on the skip feature matrix if a skip event has occurred, obtain working state data of the gravure press, determine compensation information based on the skip type and the working state data, and generate corresponding compensation instructions based on the compensation information. The skip analysis module is further configured to obtain compensation state data in response to the compensation instructions, and determine whether there is a skip event based on the compensation state data. The skip detection module is further configured to continue collecting running state data between the doctor blade and the plate cylinder if there is no skip event, and generate a warning message if there is a skip event.

[0014] In a third aspect, the embodiments provide a computer-readable storage medium having stored thereon a computer program capable of running on a processor, wherein the computer program, when executed by the processor, implements a skip detection and compensation method as described in the first aspect.

[0015] By adopting the above method, the running state data between the doctor blade and the plate cylinder is first collected, time synchronization and noise reduction processing are performed on the running state data to obtain a knife jump feature matrix. Then, whether a knife jump event occurs is determined based on the knife jump feature matrix, if not, the running state data between the doctor blade and the plate cylinder is continuously collected; if yes, the type of the knife jump is determined based on the knife jump feature matrix, the working state data of the gravure press is obtained, the compensation information is determined based on the type of the knife jump and the working state data, and the corresponding compensation instruction is generated based on the compensation information. Finally, the compensation state data after responding to the compensation instruction is obtained, whether the knife jump event exists is determined based on the compensation state data, if not, the running state data between the doctor blade and the plate cylinder is continuously collected; if yes, the warning information is generated. In this way, the running state of the doctor blade can be monitored in real time, the knife jump event and its type can be accurately determined in time, so that the influence of the knife jump is eliminated. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 It is a detection and compensation method block diagram of a knife jump provided by the present application.

[0017] Figure 2 It is a method block diagram for performing time synchronization and noise reduction processing on running state data to obtain a knife jump feature matrix provided by the present application.

[0018] Figure 3 It is a method block diagram for determining the type of a knife jump based on a knife jump feature matrix provided by the present application.

[0019] Figure 4 It is a method block diagram for determining compensation information based on the type of a knife jump and working state data provided by the present application.

[0020] Figure 5 It is a connection schematic diagram of a detection and compensation system of a knife jump provided by the present application. DETAILED DESCRIPTION

[0021] In order to more clearly understand the purpose, technical scheme and advantages of the present application, the present application is described and explained below in combination with the drawings and embodiments. However, it should be understood by those skilled in the art that the present application can be implemented without these details. It is obvious for those skilled in the art that various changes can be made to the embodiments disclosed in the present application, and the universal principles defined in the present application can be applied to other embodiments and application scenarios without deviating from the principles and scope of the present application. Therefore, the present application is not limited to the shown embodiments, but conforms to the broadest range claimed in the present application.

[0022] The embodiments of the present application are further described in detail below in combination with the drawings of the specification.

[0023] The specific application scenario of the present application is a gravure press.Figure 1 is a method for detecting and compensating for a knife jump provided by the present application. As shown in Figure 1 a method for detecting and compensating for a knife jump includes the following steps: Step S100, collecting running state data between the doctor blade and the plate cylinder, and performing time synchronization and noise reduction processing on the running state data to obtain a knife jump feature matrix.

[0024] The present application is described at the processing end. The above-mentioned running state data includes gap information, vibration information and contact pressure information. Laser displacement sensors, vibration sensors and pressure sensors are respectively arranged at preset positions near the contact area between the doctor blade and the plate cylinder. The detection direction of the laser displacement sensor is perpendicular to the surface of the plate cylinder. The vibration sensor is fixedly installed at the non-center of force point of the doctor blade holder. The pressure sensor is integrated on the pressure sensing path of the pressure adjusting mechanism of the doctor blade. Collecting the running state data between the doctor blade and the plate cylinder includes the following steps: Step S101, the laser displacement sensor collects the gap data between the doctor blade edge and the surface of the plate cylinder to obtain the gap information.

[0025] Step S102, the vibration sensor collects the vibration frequency and amplitude data of the doctor blade during the running process to generate vibration information.

[0026] Step S103, the pressure sensor collects the contact pressure data between the doctor blade and the plate cylinder to generate contact pressure information.

[0027] Specifically, for the gap information, it refers to the instantaneous and small distance change between the edge of the doctor blade and the surface of the high-speed rotating plate cylinder during the printing process. In the ideal contact state, the gap should be zero or a constant micron-level negative gap, which indicates a slight interference contact. When the knife jump occurs, the doctor blade will instantaneously separate from the surface of the cylinder, causing the gap value to change abruptly.

[0028] A laser displacement sensor can be used for collection. The laser displacement sensor is installed near the doctor blade assembly, and the laser beam emission direction thereof is perpendicular to the tangent direction of the plate cylinder, i.e., substantially perpendicular to the surface of the cylinder, and directly aims at the contact area between the doctor blade edge and the cylinder. The sensor emits laser to the surface of the cylinder and receives reflected light. Through the principles of laser triangulation or phase interference, the distance between the sensor and the surface of the cylinder is calculated in real time. Through calibration, the distance is converted into the actual gap between the edge and the cylinder. The process is carried out at a high frequency of thousands of hertz, so that the transient knife jump event in high-speed operation can be captured.

[0029] For vibration information, it refers to the abnormal mechanical vibration of the doctor blade and its holder during the printing process. Normal doctoring process will produce background vibration of a certain frequency range, while the knife jump is often accompanied by vibration signals caused by impact, resonance, etc., which are different from the normal state in amplitude and frequency.

[0030] A vibration sensor can be used for collection, which is usually an accelerometer. The vibration sensor is fixedly installed on the holder of the doctor blade, and is preferably installed at the non-center of force point. This position can more sensitively capture the bending or torsional vibration of the holder caused by the knife jump, while avoiding the damage or signal saturation of the sensor caused by the high-frequency impact of the direct force point. The piezoelectric crystal or MEMS structure inside the sensor will generate an electric signal proportional to the acceleration when subjected to vibration. After conditioning and analog-to-digital conversion, the vibration information is obtained.

[0031] For contact pressure information, it refers to the vertical pressure of the doctor blade blade on the surface of the plate cylinder. Stable contact pressure is the premise to ensure the quality of doctoring. The phenomenon of knife jump will directly cause the sharp fluctuation of the pressure, i.e. the pressure drops sharply when jumping up, and impact pressure may be generated when falling back.

[0032] A pressure sensor can be used for collection. The sensor is not directly measuring the contact point of the blade and the cylinder, but is integrated in the pressure adjusting mechanism of the doctor blade, and is installed on the pressure transmission path, such as the end of the piston rod of the pressurizing cylinder, the oil path of the hydraulic actuator, or the load point of the pressure spring loading mechanism. When the actuator applies pressure to the doctor blade, the pressure will finally act on the cylinder through the transmission path. The pressure sensor indirectly but accurately reflects the contact pressure between the doctor blade and the cylinder by detecting the force or pressure on the transmission path. This integrated installation method has small modification to the existing equipment, and does not affect the structural rigidity and movement freedom of the doctor blade assembly.

[0033] Among them, the laser displacement sensor, the vibration sensor and the pressure sensor are cooperatively arranged in space and synchronously sampled in time, so as to ensure that the collected gap, vibration and pressure data have spatio-temporal consistency, and provide reliable data basis for subsequent construction of the knife jump feature matrix. In this way, on the one hand, the knife jump phenomenon is comprehensively represented from three dimensions of gap, vibration and pressure, overcoming the possible misjudgment or missed detection problem of a single signal source. On the other hand, the sensor arrangement position is optimized and designed, which can avoid the interference of the mechanical structure to the measurement signal, and improve the data reliability. In addition, the collected multi-dimensional running state data provide rich and structured input features for subsequent knife jump identification, classification and compensation strategy generation, supporting the system to realize precise and adaptive closed-loop control.

[0034] Subsequently, the collected running state data are further subjected to the detection work of the knife jump. Figure 2is a method block diagram provided by the present application for time synchronization and noise reduction processing of running state data to obtain a jump knife feature matrix. As shown in Figure 2 Time synchronization and noise reduction processing of running state data to obtain a jump knife feature matrix includes the following steps: Step S104, time stamp alignment is performed on the gap information, vibration information and contact pressure information, and multi-channel time series data is constructed.

[0035] Step S105, wavelet noise reduction processing is performed on the multi-channel time series data, and the effective signal components of each channel are extracted to obtain the multi-channel signal after noise reduction.

[0036] Step S106, the multi-channel signal after noise reduction is sliced according to a preset time window, and the time domain features and frequency domain features in each time window are extracted, wherein the time domain features include mean, variance and peak factor, and the frequency domain features include spectral centroid, frequency standard deviation and energy entropy.

[0037] Step S107, the extracted time domain features and frequency domain features are spliced according to the channel dimension, and a jump knife feature matrix is constructed, wherein the rows of the jump knife feature matrix correspond to the time window, and the columns of the jump knife feature matrix correspond to the feature dimension.

[0038] Since the laser displacement sensor, the vibration sensor and the pressure sensor are physically independently deployed, the data collected by them has a slight sampling delay or clock drift in time. In order to realize the space-time consistency of the data, a hardware synchronization signal or a software time stamp alignment mechanism is adopted. Specifically, a unified clock source is used to label the data of each sensor with an accurate time label, and the data of other channels is interpolated or resampled based on the highest sampling rate, so as to ensure that the time gap, vibration and pressure data of each sampling point correspond to the same time. The aligned data is organized as multi-channel time series data, wherein each row represents a sampling time, and each column represents a sensor channel. In this way, the time synchronization effectively eliminates the data misalignment problem caused by the asynchronous sampling of the sensors, provides an accurate time sequence basis for subsequent multi-modal feature fusion, and avoids the jump knife misjudgment or missed detection caused by time deviation.

[0039] The printing machine field environment is complex, and various noises such as electromagnetic interference, mechanical resonance background noise and the like are inevitably mixed in the sensor signal. After the construction of the multi-channel time sequence data is completed, the noise reduction work of the multi-channel time sequence data is carried out, and the wavelet noise reduction method is adopted in the present application. Because it can analyze the signal in time and frequency domains at the same time, it is particularly suitable for processing non-stationary, sudden jump signals, so it is superior to the traditional Fourier filter. Specifically, the first step is to carry out wavelet decomposition, wherein a commonly used wavelet basis function library is built in the processing end. When calibrating initially or for a specific machine type, an optimal wavelet basis and decomposition level will be automatically selected from a preset self by evaluating the entropy values of different wavelets on a known normal vibration signal, and the time sequence signal of each channel is decomposed into subbands of different frequencies. Usually, the selected wavelet can represent the signal characteristics with the least coefficient energy. The second step is threshold processing, that is, the threshold is applied to the decomposition of each layer of detail coefficients, such as using unbiased risk estimation threshold or fixed threshold, and the coefficients less than the threshold are regarded as noise and are set to zero, while the coefficients greater than the threshold, representing the signal mutation components, are retained. The third step is wavelet reconstruction, that is, using the detail coefficients after threshold processing and the original approximation coefficients, inverse wavelet transform is performed to reconstruct the multi-channel signal after noise reduction of each channel. In this way, the background noise can be effectively filtered out, while the signal mutation edge and detail characteristics generated when the jump occurs are maximally retained, providing a clean signal source for subsequent feature extraction.

[0040] The jump event is a transient process, and its pattern needs to be captured in high-frequency data. The present application adopts a sliding time window to analyze the sliced signals after noise reduction. Specifically, for the signal in each time window, statistical features with clear physical meaning are extracted from the time domain and the frequency domain respectively. That is, the time domain features include the mean value that can reflect the signal reference level in the time period; the variance that characterizes the fluctuation intensity of the signal; and the peak factor for detecting whether there is an impact peak in the signal. Among them, the variance of the jump will usually increase significantly, and the peak factor is a key indicator for identifying instantaneous jump. The frequency domain features are calculated after the fast Fourier transform of the signal in the window. The frequency domain features include the spectral barycenter that can reflect the frequency position of the concentrated spectrum energy; the frequency standard deviation that describes the dispersion degree of the spectrum; and the energy entropy that can measure the distribution disorder of the spectrum energy. Among them, the jump may cause the spectral barycenter to shift; abnormal vibration will make the spectrum spread, affecting the frequency standard deviation; when a new vibration mode appears, the energy entropy will change.

[0041] Finally, all the time-domain and frequency-domain features extracted from all channels in the same time window are spliced according to the channel dimension to form a feature vector representing the comprehensive state of the window. Specifically, the feature vectors of all time windows are arranged in chronological order, and finally a jump knife feature matrix is constructed. The rows of the matrix correspond to different time windows, and the columns correspond to all feature dimensions. This matrix is a structured, high information density data representation, providing an ideal input for subsequent model determination. In this way, the time stamp is aligned, eliminating the feature phase confusion problem caused by the asynchronization of multi-sensor data, ensuring the spatio-temporal correlation of gap, vibration and pressure information when analyzing, and laying a foundation for accurately associating the multi-element causes of the jump knife. On the other hand, wavelet denoising can adaptively filter out various background noises while preserving the sharp features of the jump knife transient signal. This greatly reduces false positives and false negatives caused by noise interference, and still maintains high reliability in complex industrial environments. In addition, by jointly extracting time-domain and frequency-domain features, the jump knife phenomenon is fully described from different dimensions. Time-domain features directly capture the amplitude changes and impact characteristics of the signal, while frequency-domain features reveal the vibration patterns and frequency structure behind it. This multi-dimensional, multi-perspective feature representation enables the system to distinguish between different types of jump knives caused by different mechanisms and exhibiting different characteristics.

[0042] Step S200, based on the jump knife feature matrix, determine whether a jump knife event occurs, if not, continue to collect the running state data between the doctor blade and the plate cylinder.

[0043] After obtaining the jump knife feature matrix, whether a jump knife event occurs is determined based on the jump knife feature matrix. Wherein, based on the jump knife feature matrix to determine whether a jump knife event occurs includes the following steps: Step S201, input the jump knife feature matrix into a preset jump knife recognition model, and output the occurrence probability of the jump knife event.

[0044] Step S202, wherein the jump knife recognition model is a multi-channel time sequence classification network based on attention mechanism, including a feature embedding layer for mapping the jump knife feature matrix into a high-dimensional feature vector, a multi-head self-attention layer for extracting the time sequence dependency between different channel features, a time sequence convolution layer for capturing the jump knife pattern features in the local time window, and a fully connected sub-type for outputting the jump knife event probability value.

[0045] Step S203, if the occurrence probability exceeds the first preset threshold, a jump knife event occurs.

[0046] Step S204, if the occurrence probability does not exceed the first preset threshold, no jump knife event occurs.

[0047] The input data of the jump recognition model is a jump feature matrix, and the output data is a probability of occurrence. The structure of the jump feature matrix is N*M, N is the number of time windows, and M is the feature dimension. The probability of occurrence is a probability value between 0 and 1, indicating the possibility of a jump event occurring in the current time window.

[0048] The jump recognition model is a multi-channel time series classification network based on attention mechanism, which includes four core layers: feature embedding layer, multi-head self-attention layer, time series convolution layer and fully connected classification layer. The feature embedding layer maps the input jump feature matrix to a high-dimensional feature vector, enhancing the expression ability of the features. Specifically, a fully connected layer or a one-dimensional convolution can be used to perform nonlinear transformation on the features of each time window, improving the model's ability to capture complex jump patterns.

[0049] The multi-head self-attention layer extracts the time series dependency between different channel features, enhancing the model's understanding of multi-sensor data fusion. Specifically, multiple attention heads can be used to calculate in parallel, focusing on the correlation between different feature channels, and then concatenating the outputs of each head to form a comprehensive time series feature representation.

[0050] The time series convolution layer captures the jump pattern features within the local time window, enhancing the sensitivity to transient jump events. Specifically, a one-dimensional convolution kernel can be used to slide in the time dimension to extract local time series patterns, and a ReLU activation function can be used to enhance the non-linear modeling ability.

[0051] The fully connected classification layer is based on the high-dimensional features extracted by the previous layers, and outputs the probability value of the jump event. Specifically, one or more fully connected layers can be used to map the features to a scalar, and then normalized to a probability value using Sigmoid.

[0052] Finally, the probability of occurrence obtained above is compared with the first preset threshold value. If the probability of occurrence is greater than the first preset threshold value, it indicates that a jump event has occurred. If the probability of occurrence is not greater than the first preset threshold value, it indicates that no jump event has occurred.

[0053] The first preset threshold value is not a fixed empirical value, but is determined by combining historical data, model performance evaluation and process requirements. Specifically, the following steps are included.

[0054] The first step is to build a labeled data set. That is, on a specific type of gravure press, a large number of jump feature matrix samples are collected under different working conditions, i.e. different speeds, different ink viscosities and different printing materials. Each sample is accurately labeled by an expert or by comparing with a high-precision synchronous high-speed camera system, and is divided into positive samples with jump events and negative samples without jump events, forming a basic data set for model training and threshold calibration.

[0055] Second step is to carry out preliminary training and probability output of the model. That is, using the above-mentioned labeled data set to train and verify the jump detection model, to ensure that the model itself has good feature extraction and classification ability. Use a reserved validation set input to the trained model, let the model output a jump probability value for each sample in the validation set.

[0056] Third step is to generate performance evaluation curve and key indicators. That is, based on the probability and true label output by the model on the validation set, draw the precision-recall curve and the receiver operating characteristic curve. Focus on the harmonic mean of precision and recall, which is an ideal indicator to measure the overall performance of the model. Among them, the harmonic mean of precision and recall = 2*precision*recall / (precision+recall).

[0057] Fourth step is to determine the preliminary preset threshold based on multi-face optimization. That is, through the precision-recall curve or through the calculation process of the harmonic mean of precision and recall, find the probability value that maximizes the harmonic mean of precision and recall, and set the probability value corresponding to this point as the preliminary preset threshold.

[0058] Fifth step is to optimize the preliminary preset threshold to obtain the first preset threshold. That is, if the false alarm prevention requirement is very high, appropriately increase the harmonic mean of precision and recall accordingly, so as to obtain the corresponding first preset threshold. If the requirement of capturing all jump events is very high, appropriately reduce the harmonic mean of precision and recall accordingly, so as to obtain the corresponding first preset threshold.

[0059] In this way, the blindness of setting threshold by experience is overcome, and based on data-driven and model performance quantitative indicators, the threshold is set at the optimal or near-optimal operating point.

[0060] In the case of no jump event, no subsequent compensation will be performed, in order to eliminate the influence of jump in time, it is necessary to continue to collect the running state data between the doctor blade and the plate cylinder in the above-mentioned manner, which will not be repeated here.

[0061] Preferably, the processing end also monitors the online judgment accuracy and prediction confidence of the knife jump identification type. The shift index between the data distribution of the knife jump feature matrix and the historical training base distribution is calculated. It is detected whether a preset working condition change event occurs in the gravure press. When at least one of the online judgment accuracy is lower than a first threshold, the prediction confidence is lower than a second threshold, the shift index is higher than a third threshold, or a preset working condition change event occurs, it is determined that the knife jump identification model needs to be adjusted. During the adjustment of the knife jump identification model, the focal loss function is introduced as the optimization target, the sliding window enhancement strategy is used for data enhancement of the training samples, and the early stop mechanism and the learning decay strategy are used to optimize the training process.

[0062] Specifically, first, the online judgment accuracy and prediction confidence of the knife jump identification model are monitored. That is, the first step of online judgment accuracy monitoring. The processing end periodically extracts samples from the current running data, compares them with the true labels recorded synchronously by artificial labeling or high-speed cameras, and calculates the online accuracy of the model. If the accuracy of multiple consecutive periods is lower than the first threshold, the model adjustment mechanism is triggered. The second step of prediction execution degree monitoring. The knife jump identification model outputs a confidence score for each knife jump event, which reflects its certainty of judgment. The average of all predicted confidence scores in the current window is calculated, and if it is lower than the second threshold, it is determined that the judgment is unstable and needs to be adjusted.

[0063] Then, the shift index between the data distribution of the knife jump feature matrix and the historical training base distribution is calculated. That is, the processing end periodically extracts the statistical features of the current knife jump feature matrix, such as mean, variance, and covariance, and compares them with the distribution of the historical data set used in the training stage. The KL divergence, maximum mean difference, or Mahalanobis distance is used to function the distribution shift degree. If the shift index exceeds the third threshold, it means that the current working condition is significantly different from the training data, and the model performance may decrease.

[0064] Next, it is detected whether a preset working condition change event occurs in the gravure press. That is, the processing end detects the working condition change events in real time, such as changes in printing speed, ink type or viscosity, doctor blade type or wear state, and substrate type. Any event is considered as a working condition change, triggering the model adjustment process.

[0065] Next, it is determined whether the model needs to be adjusted. That is, when any of the following conditions is met, the processing end determines that the knife jump identification model needs to be adjusted. That is, the online judgment accuracy is less than the first threshold, the prediction confidence is less than the second threshold, the shift index is greater than the third threshold, and a preset working condition change event occurs.

[0066] Finally, the model is adjusted. The specific means for adjusting the jump detection model are as follows. First, a focal loss function is introduced as an optimization objective. That is, in the model retraining stage, the focal loss function is used to replace the traditional cross-entropy loss to alleviate the imbalance problem of jump samples. The focal loss adjusts the weights of difficult and easy samples, so that the jump detection model pays more attention to the jump samples that are difficult to classify, and improves the recognition ability of small probability jump types.

[0067] Second, a sliding window enhancement strategy is used to enhance the training samples. That is, multiple overlapping time period samples are extracted from the current running data in a sliding window manner to construct an enhanced training set. By performing time stretching, noise injection, amplitude scaling and other operations on the window samples, the jump characteristics under different working conditions are simulated, and the generalization ability of the model is improved.

[0068] Third, the early stopping mechanism and learning rate decay strategy are used to optimize the training process. That is, in the model retraining process, the validation set loss is monitored. If it decreases for several consecutive periods, the early stopping mechanism is started to prevent overfitting. The cosine annealing or step learning rate decay strategy is used to gradually reduce the learning rate to ensure that the model converges to a better solution. In this way, on the one hand, the jump detection model can automatically adjust according to the change of working conditions to avoid the decline of recognition performance caused by the change of printing conditions. On the other hand, through multi-dimensional monitoring and dynamic adjustment, the system can still maintain high precision and stability in complex industrial environments. In addition, the focal loss and data enhancement strategy effectively utilize limited samples to improve the recognition ability of the model for rare jump types, and the early stopping and learning rate decay mechanism ensures that the model converges quickly under limited computing resources to avoid overfitting.

[0069] Step S300, if yes, determining the jump type based on the jump feature matrix, obtaining the working state data of the gravure press, determining the compensation information based on the jump type and the working state data, and generating corresponding compensation instructions based on the compensation information.

[0070] After determining that a jump event occurs, the jump type needs to be further determined. Figure 3 The method for determining the jump type based on the jump feature matrix provided in the present application is shown in the method block diagram. Figure 3 As shown in the method block diagram, the method for determining the jump type based on the jump feature matrix comprises the following steps: Step S301, inputting the jump feature matrix into a pre-trained multi-class jump classification model to output the probability distribution of the jump type.

[0071] Step S302, the multi-class jump classification model is a hybrid model based on a time sequence attention mechanism and a convolutional neural network, wherein the multi-class jump classification model comprises a feature selection layer, a time sequence feature extraction layer and a type discrimination layer.

[0072] Step S303, the feature selection layer is used to filter out the key features related to the jump type from the jump feature matrix based on the basic attention weight.

[0073] Step S304, the time sequence feature extraction layer is used to extract the dynamic evolution features of the jump event in the time dimension by using a bidirectional long short-term memory network.

[0074] Step S305, the type discrimination layer is used to output the probability of each jump type based on the extracted dynamic evolution features through a fully connected neural network and an activation function.

[0075] Step S306, the type corresponding to the maximum probability in the probability is determined as the current jump type.

[0076] The application realizes the determination of the jump type by using a pre-trained multi-class jump classification model which combines the advantages of the time sequence attention mechanism and the convolutional neural network to accurately identify the jump type from the jump features in multiple dimensions and time sequences.

[0077] Firstly, the jump feature matrix is input into the pre-trained multi-class jump classification model to output the probability distribution of the jump type. That is, the input data of the multi-class jump classification model is the jump feature matrix, and the structure of the jump feature matrix is N*M, where N is the number of time windows and M is the feature dimension. The output data of the multi-class jump classification model is a probability distribution vector.

[0078] Then, the feature selection layer is used to filter out the key features related to the jump type from the jump feature matrix based on the attention weight. That is, the feature selection layer adopts the channel attention mechanism to dynamically enhance the feature dimension that contributes most to the jump classification by learning the weight of each feature channel. Specifically, the global information of each channel is obtained by global average pooling, and then the attention weight vector is generated by a fully connected layer and a Sigmoid activation function. Finally, the original feature matrix is weighted by the weight channel by channel, so as to effectively suppress the noise feature interference, highlight the key information related to the jump type, and improve the sensitivity and robustness of the multi-class jump classification model to the jump features.

[0079] Then, the time sequence feature extraction layer adopts a bidirectional long short-term memory network to extract the dynamic evolution features of the jump event in the time dimension. That is, a bidirectional LSTM structure is adopted to capture the dependence relationship of the jump event in the time sequence from the forward and backward directions respectively. Each LSTM unit receives the feature vector weighted by the feature selection layer and outputs the hidden state. Finally, the final hidden states of the forward and backward directions are spliced to form a comprehensive time sequence feature representation. Thus, the complete dynamic process of the jump event from occurrence, development to subsidence can be captured, which is especially suitable for identifying the jump type with time sequence evolution characteristics.

[0080] Next, the type discrimination layer outputs the probability of each type of jump-cut based on the extracted dynamic evolution features through a fully connected neural network and an activation function. That is, the comprehensive timing features output by the bidirectional LSTM are input into a classification network composed of multiple fully connected layers, and finally the probability distribution of each category is output through the Softmax activation function. Thus, the multi-classification discrimination of jump-cut types is realized, and the output result has clear probability interpretation, which is convenient for subsequent decision-making and compensation strategy matching.

[0081] Finally, the type corresponding to the maximum probability in the probability is determined as the current jump-cut type. That is, the class label corresponding to the maximum value in the output probability distribution is selected as the final classification result of the current jump-cut event. Thus, the certainty and operability of the classification result are ensured, and a clear basis is provided for the generation of subsequent compensation strategies. In this way, by filtering key features through the feature selection layer and capturing timing dynamics through the bidirectional LSTM, the model can accurately distinguish different types of jump-cuts, and the classification accuracy is significantly improved. Moreover, the attention mechanism can effectively suppress noise and non-key feature interference, so that the model still maintains stable performance in complex industrial environments.

[0082] The processing end determines the working state data by real-time reading at least one of the printing speed, the ink viscosity, the doctor blade angle, the blade wear state, and the environmental temperature and humidity. The printing speed can be determined by the real-time rotating speed of the plate cylinder fed back by an encoder or a servo driver. The ink viscosity can be obtained by real-time detection through an online viscometer or conversion from the ink formula and temperature. The doctor blade angle can be obtained from the position feedback of an angle sensor or a servo motor. The blade wear state can be determined based on the cumulative running time or the wear grade output by a visual detection system. The environmental temperature and humidity can be obtained from the temperature and humidity sensor for compensating for the change in ink fluidity. The pressure setting value can be determined by the set pressure of the doctor blade pressure adjusting mechanism. The blade holder position can be determined by the installation position information of the doctor blade holder relative to the plate cylinder. In this way, the compensation information can be determined more accurately.

[0083] Figure 4 is a method block diagram for determining compensation information based on the jump-cut type and the working state data provided by the present application. As shown in Figure 4 determining compensation information based on the jump-cut type and the working state data includes the following steps: Step S307, a historical jump-cut event library corresponding to the jump-cut type is obtained, a matching historical jump-cut event is retrieved from the historical jump-cut event library based on the jump-cut type and the working state data, a compensation strategy corresponding to the historical jump-cut event is extracted, and the compensation strategy is determined as a candidate compensation strategy.

[0084] Step S308, based on the jump-cut feature matrix and the working state data, a context feature vector of the current jump-cut event is constructed.

[0085] Step S309, similarity matching is performed between the context feature vector and the historical context feature vectors corresponding to the candidate compensation strategies, and at least one target historical event with a similarity higher than a second preset threshold is screened out.

[0086] Step S310, a compensation effect evaluation result of the target historical event is obtained, a target compensation strategy is determined from the candidate compensation strategies based on the compensation effect evaluation result, and compensation information containing a compensation parameter and an execution timing is generated.

[0087] First, the candidate compensation strategy is obtained. That is, a historical skip event library is built in the processing end, which stores records of past skip events, each record including a skip type, such as instantaneous skip, periodic skip, and persistent skip; working state data, such as printing speed, ink viscosity, and doctor blade angle; a compensation strategy adopted; and a compensation effect evaluation result, such as successful elimination of skip, partial improvement, and invalidity. When it is determined that a current skip event occurs and the type of the skip is determined, the event library is searched according to the current skip type and the working state data, historical events similar to the current situation are screened out, and the compensation strategies corresponding to the historical events are used as candidate compensation strategies. In the searching process, a multi-condition matching algorithm can be used, and if the skip type is matched, the range of the working state data is matched. The candidate strategies include pressure adjustment, angle fine-tuning, ink overfeed, and blade feed compensation.

[0088] Then, the context feature vector is constructed. That is, the skip feature matrix of the current skip event is fused with the working state data to construct a context feature vector. The vector includes statistical quantities of the skip feature matrix, such as mean, variance, and peak value of each channel; working state data, such as printing speed, ink viscosity, and doctor blade angle; and an encoding representation of the skip type. The context feature vector reflects the multidimensional state of the current skip event and provides a structured input for subsequent similarity matching.

[0089] Next, the target historical event is screened out by similarity matching. That is, similarity calculation is performed between the current context feature vector and the historical context feature vectors corresponding to the candidate compensation strategies. The similarity calculation can use methods such as cosine similarity, Euclidean distance, or Mahalanobis distance. Historical events with a similarity higher than a second preset threshold are screened out as target historical events. The second preset threshold can be adjusted according to the actual application scenario. If multiple target events are matched, they can be sorted in descending order of similarity for subsequent strategy optimization.

[0090] Finally, the target compensation strategy is determined and compensation information is generated. That is, the compensation effect evaluation result of the target historical event is obtained. According to the evaluation result, the optimal target compensation strategy is determined from the candidate compensation strategies. Finally, the compensation information is generated. If multiple target events are all evaluated as successful, the strategy with the closest occurrence or the closest working condition can be selected preferentially. Manual intervention confirmation is supported, and the compensation parameters can be audited and fine-tuned on the human-machine interface. In this way, on the one hand, based on the similarity matching of historical data and the current state, the compensation strategy that has been proven effective in practice can be automatically selected, avoiding blind adjustment and improving the success rate of compensation. On the other hand, the context vector is constructed by fusing the skip blade feature and the working state data, ensuring that the compensation strategy is highly adaptable to the current printing working condition and enhancing the working condition adaptability.

[0091] Subsequently, the compensation information is packaged as an action execution, i.e., a compensation instruction, and sent to the corresponding components, so that these components can respond to the compensation instruction in a timely and accurate manner, take corresponding actions, and achieve compensation.

[0092] In step S400, the compensation state data after responding to the compensation instruction is obtained, and it is determined whether there is a skip blade event based on the compensation state data. If there is not, the running state data between the doctor blade and the plate cylinder is continuously collected.

[0093] In step S500, if there is, warning information is generated.

[0094] In addition, the processing end also obtains the compensation state data after the component responds to the compensation instruction. The compensation state data refers to the running state data between the doctor blade and the plate cylinder collected after compensation. The compensation state data is obtained in the same way as the running state data, which will not be described here. Then, the above method of determining whether a skip blade event occurs based on the skip blade feature matrix is used to determine whether there is a skip blade event based on the compensation state data. When it is determined based on the compensation state data that there is no skip blade event, it means that accurate compensation has been achieved, and the running state data between the doctor blade and the plate cylinder can be continuously collected to indirectly determine the skip blade event and its type in a timely manner. When it is determined based on the compensation state data that there is a skip blade event, it means that accurate compensation has not been achieved, and warning information is generated at this time, so that manual compensation can be performed in a timely manner.

[0095] Preferably, the zero points of the laser displacement sensor, the vibration sensor, and the pressure sensor are calibrated before the running state data between the doctor blade and the plate cylinder is collected. In this way, the calibrated sensors can more accurately capture micron-level gap changes, weak vibration signals, and pressure fluctuations, significantly improving the ability to identify early or slight skip blade events, thereby indirectly achieving timely and accurate determination of the skip blade event and its type, and eliminating the impact of the skip blade.

[0096] Figure 5is a connection diagram of a jump detection and compensation system provided by the present application. As shown in Figure 5 The jump detection and compensation system comprises a jump detection module, a jump analysis module and a jump compensation module.

[0097] The jump detection module is configured to collect running state data between the doctor blade and the plate cylinder. The jump analysis module is configured to perform time synchronization and noise reduction processing on the running state data to obtain a jump feature matrix, and determine whether a jump event occurs based on the jump feature matrix. The jump detection module is further configured to continue collecting the running state data between the doctor blade and the plate cylinder if no jump event occurs. The jump compensation module is further configured to determine a jump type based on the jump feature matrix, obtain working state data of the gravure press, determine compensation information based on the jump type and the working state data, and generate corresponding compensation instructions based on the compensation information if a jump event occurs. The jump analysis module is further configured to obtain compensation state data after responding to the compensation instructions, and determine whether a jump event exists based on the compensation state data. The jump detection module is further configured to continue collecting the running state data between the doctor blade and the plate cylinder if no jump event exists, and generate a warning information if a jump event exists.

[0098] The jump detection module, the jump analysis module and the jump compensation module perform other functions and the technical details of each function are the same as or similar to the corresponding features in the jump detection and compensation method described above, and thus are not described again here.

[0099] The embodiments of the present application further provide a computer storage medium having a computer program stored thereon, which, when executed on a computer, enables the computer to perform the steps of the jump detection and compensation method described above.

[0100] It should be understood that, although each step in the flowchart of the accompanying drawings is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise explicitly stated herein, the execution of these steps is not strictly limited in sequence, and they can be executed in other orders.

[0101] The above only describes some embodiments of the present application, and it should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should also be considered as falling within the scope of protection of the present application.

Claims

1. A method for detecting and compensating for switch blades, characterized in that, The method includes: Collect the operating status data between the doctor blade and the printing cylinder, and perform time synchronization and noise reduction processing on the operating status data to obtain the doctor blade feature matrix; Based on the aforementioned tool skipping feature matrix, it is determined whether a tool skipping event has occurred. If not, the operating status data between the doctor blade and the printing cylinder continues to be collected. If so, determine the blade type based on the blade feature matrix, obtain the gravure printing machine's working status data, determine compensation information based on the blade type and the working status data, and generate a corresponding compensation instruction based on the compensation information; Acquire compensation status data after responding to the compensation command, determine whether a doctor blade skipping event exists based on the compensation status data, and if not, continue to collect the operating status data between the doctor blade and the printing cylinder; If it exists, generate an early warning message.

2. The method according to claim 1, characterized in that, The operational status data includes gap information, vibration information, and contact pressure information. A laser displacement sensor, a vibration sensor, and a pressure sensor are deployed at preset positions near the contact area between the doctor blade and the printing plate cylinder. The detection direction of the laser displacement sensor is perpendicular to the surface of the printing plate cylinder. The vibration sensor is fixedly installed at the non-stress center point of the doctor blade holder. The pressure sensor is integrated into the pressure transmission path of the doctor blade pressure adjustment mechanism. The collected operational status data between the doctor blade and the printing plate cylinder includes: The laser displacement sensor collects gap data between the doctor blade and the surface of the printing cylinder to obtain gap information; The vibration sensor collects the vibration frequency and amplitude data of the doctor blade during use to generate vibration information; The pressure sensor collects contact pressure data between the doctor blade and the printing cylinder to generate contact pressure information.

3. The method according to claim 2, characterized in that, The step of performing time synchronization and noise reduction processing on the running status data to obtain the jump cutter feature matrix includes: The gap information, vibration information, and contact pressure information are timestamped to construct multi-channel time series data; Wavelet denoising is performed on the multi-channel time series data to extract the effective signal components of each channel and obtain the denoised multi-channel signal. The denoised multi-channel signal is sliced ​​according to a preset time window, and the time domain features and frequency domain features within each time window are extracted. The time domain features include mean, variance and peak factor, and the frequency domain features include spectral centroid, frequency standard deviation and energy entropy. The extracted time-domain features and frequency-domain features are concatenated along the channel dimension to construct a jump knife feature matrix, wherein the rows of the jump knife feature matrix correspond to the time window and the columns of the jump knife feature matrix correspond to the feature dimension.

4. The method according to claim 1, characterized in that, The step of determining whether a jump tool event has occurred based on the jump tool feature matrix includes: Input the switch tool feature matrix into a preset switch tool recognition model and output the probability of switch tool events occurring; The jump knife recognition model is a multi-channel temporal classification network based on an attention mechanism, including a feature embedding layer for mapping the jump knife feature matrix into a high-dimensional feature vector, a multi-head self-attention layer for extracting the temporal dependencies between features of different channels, a temporal convolutional layer for capturing jump knife pattern features within a local time window, and a fully connected classification layer for outputting the probability value of jump knife events. If the probability of occurrence exceeds a first preset threshold, a jump cut event occurs; If the probability of occurrence does not exceed the first preset threshold, no tool skipping event occurs.

5. The method according to claim 4, characterized in that, The method further includes: Monitor the online judgment accuracy and prediction confidence of the switch knife recognition model; Calculate the offset index between the data distribution of the jump knife feature matrix and the distribution of the historical training base; Detect whether a preset operating condition change event occurs on the gravure printing machine; When at least one of the following occurs: the online judgment accuracy is lower than the first threshold, the prediction confidence is lower than the second threshold, the drift index is higher than the third threshold, or a preset working condition change event occurs, it is determined that the switch blade recognition model needs to be adjusted. In the process of adjusting the knife recognition model, a focus loss function is introduced as the optimization objective, a sliding window enhancement strategy is used to augment the training samples, and an early stopping mechanism and a learning decay strategy are used to optimize the training process.

6. The method according to claim 1, characterized in that, Determining the jump knife type based on the jump knife feature matrix includes: The jump knife feature matrix is ​​input into a pre-trained multi-class jump knife classification model, which outputs the probability distribution of jump knife types. The multi-class knife classification model is a hybrid model based on temporal attention mechanism and convolutional neural network, wherein the multi-class knife classification model includes a feature selection layer, a temporal feature extraction layer and a type discrimination layer; The feature selection layer is used to filter out key features related to the Blink Dagger type from the Blink Dagger feature matrix based on attention weights; The temporal feature extraction layer is used to extract the dynamic evolution features of the jump knife event in the time dimension using a bidirectional long short-term memory network; The type discrimination layer is used to output the probability of each jump knife type based on the extracted dynamic evolution features through a fully connected neural network and activation function; The type corresponding to the highest probability among the probabilities is determined as the current Blink Dagger type.

7. The method according to claim 1, characterized in that, The determination of compensation information based on the switch blade type and the working status data includes: Obtain the historical jump event library corresponding to the jump tool type, retrieve matching historical jump events from the historical jump event library based on the jump tool type and the working status data, extract the compensation strategy corresponding to the historical jump event, and determine the compensation strategy as a candidate compensation strategy; Based on the tool skipping feature matrix and the working status data, construct the context feature vector of the current tool skipping event; The context feature vector is matched with the historical context feature vector corresponding to the candidate compensation strategy to filter out at least one target historical event with a similarity higher than the second preset threshold. Obtain the compensation effect evaluation result of the target historical event, determine the target compensation strategy from the candidate compensation strategies based on the compensation effect evaluation result, and generate compensation information containing compensation parameters and execution sequence.

8. The method according to claim 2, characterized in that, The method further includes: Before collecting operational status data between the doctor blade and the printing cylinder, calibrate the zero points of the laser displacement sensor, vibration sensor, and pressure sensor.

9. A system for detecting and compensating for switch blades, characterized in that, The system includes: a tool switch detection module, a tool switch analysis module, and a tool switch compensation module; wherein... The doctor blade detection module is used to collect the operating status data between the doctor blade and the printing plate cylinder; The tool skipping analysis module is used to perform time synchronization and noise reduction processing on the running status data to obtain a tool skipping feature matrix, and to determine whether a tool skipping event has occurred based on the tool skipping feature matrix; The doctor blade detection module is also used to continue collecting the operating status data between the doctor blade and the printing plate cylinder if no doctor blade event occurs. The tool skipping compensation module is also used to determine the tool skipping type based on the tool skipping feature matrix, obtain the working status data of the gravure printing machine, determine compensation information based on the tool skipping type and the working status data, and generate a corresponding compensation instruction based on the compensation information if a tool skipping event occurs. The tool skipping analysis module is also used to obtain compensation status data after responding to the compensation command, and to determine whether a tool skipping event exists based on the compensation status data; The doctor blade detection module is also used to continue collecting operating status data between the doctor blade and the printing plate cylinder if the doctor blade is not present, and to generate an early warning message if the doctor blade is present.

10. A computer-readable storage medium having a computer program stored thereon that can run on a processor, characterized in that, When the computer program is executed by the processor, it implements a method for detecting and compensating for a switch blade as described in any one of claims 1 to 8.

Images