A method and system for intelligently predicting the remaining amount of printer consumables based on deep learning

Abstract

The present invention discloses a method and system for intelligently predicting the remaining amount of printer consumables based on deep learning, including: S1: collecting printing consumables data and obtaining the printing consumables data to be processed; S2: performing integrity verification and abnormal value detection on the printing consumables data to be processed, and outputting the printing consumables data after cleaning; S3: extracting short-time features, long-time features, and long-short-time fusion features based on the printing consumables data after cleaning; S4: extracting steady-state features and transient features based on the long-short-time fusion features, and then calculating the inkjet consumption features and the inkjet consumption prediction value; S5: calculating the remaining ink volume and the ink consumption rate prediction value based on the inkjet consumption prediction value, and then constructing an ink volume warning classifier, and outputting the ink volume alarm level according to the probability distribution. The present invention can solve the problem of insufficient consumables remaining amount prediction accuracy of the traditional printer consumables prediction method in the multi-material mixed printing scenario.

Description

Technical Field

[0001] The present invention relates to the technical field of printer consumables prediction, and in particular to a method and system for intelligently predicting the remaining amount of printer consumables based on deep learning. Background Art

[0002] In the field of printer consumables prediction technology, the lack of accuracy in predicting the remaining consumables in multi-material mixed printing scenarios is a challenge currently faced by the industry. Due to the significant differences in ink absorption characteristics, surface roughness and thermal conductivity of different printing media, such as photo paper, film, and special fabrics, the consumables consumption rate will fluctuate nonlinearly with the switching of media.

[0003] Traditional prediction methods based on fixed page count estimation, single sensor detection or traditional machine learning algorithms cannot effectively cope with such dynamic changes. Estimation based on fixed page count ignores the correlation between physical parameters of the medium and consumables consumption, which will produce cumulative errors when the material is switched; optical or pressure sensors are easily affected by differences in the reflectivity of the surface of the medium and mechanical vibration, resulting in distortion of the detection signal; although existing machine learning algorithms can process time series data, they do not include key parameters such as media type and printing resolution into the feature system, making it difficult to capture the sudden change of consumables caused by material switching.

[0004] In addition, the fixed threshold warning mechanism cannot adapt to the dynamic consumption characteristics in multi-material scenarios, and often leads to delayed warnings or false alarms. The above technical defects lead to insufficient prediction accuracy of the existing system in mixed printing environments, and there is an urgent need for a consumables consumption prediction solution that can integrate multi-dimensional media properties. Summary of the invention

[0005] In view of this, the present invention provides a method and system for intelligently predicting the remaining consumables of printers based on deep learning, aiming to solve the problem of insufficient prediction accuracy of the remaining consumables of traditional printer consumables prediction methods in multi-material mixed printing scenarios.

[0006] A method for intelligently predicting the remaining amount of printer consumables based on deep learning, comprising:

[0007] S1: Collecting printing consumables data, wherein the printing consumables data includes a timestamp, duration, medium type, number of printed pages, and inkjet consumption of each printing task; preprocessing the printing consumables data except the timestamp and medium type, and then merging the timestamp, medium type, and the preprocessed printing consumables data to obtain the printing consumables data to be processed;

[0008] S2: Perform integrity verification and outlier detection on the printing consumables data to be processed, and output the printing consumables data after cleaning;

[0009] S3: Sort the printing consumables data after cleaning by timestamp to construct a time series feature sequence; splice each printing consumables time series data to obtain a multimodal input vector; map the medium type to a medium type embedding vector, and calculate the convolution kernel offset matrix and the convolution kernel offset weight matrix;

[0010] According to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, the short-term features and long-term features of the printing task are obtained; based on the short-term features, long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task;

[0011] S4: extracting steady-state features and transient features according to the long- and short-time fusion features of the printing task; calculating inkjet consumption features based on the steady-state features and transient features, and calculating the predicted value of inkjet consumption;

[0012] S5: According to the predicted value of inkjet consumption, the remaining ink volume and the predicted value of ink consumption rate are calculated, and then an ink volume warning classifier is constructed, and the ink volume alarm level is output according to the probability distribution.

[0013] Furthermore, in the step S1, the printing consumables data also includes the printing resolution, page size, medium ink absorption rate, medium surface roughness, and color mode of each printing task;

[0014] The printing consumables data other than the timestamp and the media type are preprocessed, including: the timestamp and the media type are not processed; the duration, the printing resolution, the page size, the number of printed pages, the ink absorption rate of the media, the surface roughness of the media and the inkjet consumption are normalized by the maximum and minimum value method; and the color mode is uniquely encoded.

[0015] Furthermore, the step S2 includes:

[0016] S21: verify the existence of the fields of media type, number of printed pages, and inkjet consumption for the printing consumables data to be processed, delete the data entries with missing fields, and obtain data that passes the integrity verification;

[0017] S22: Based on the data that has passed the integrity verification, the Z-score method is used to detect global outliers, and the data is grouped by media type, and the abnormal data of printing consumables is identified by the IQR method;

[0018] S23: For the abnormal data of printing consumables, the ink jetting volume deviation is calculated by using a theoretical ink jetting volume formula, and data entries with ink jetting volume deviation exceeding 20% ​​are eliminated to obtain the printing consumables data after cleaning.

[0019] Furthermore, in the step S3, the short-term features and long-term features of the printing task are obtained according to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, and the calculation method is:

[0020] ;

[0021] in, is the short-term feature of the t-th printing task, t is the printing task index, is the ReLU function, is batch normalization, is a one-dimensional convolution, For Hadamard, is the long-term feature of the t-th printing task, is the GeLU function, is channel normalization, is a dilated convolution, Trilinear interpolation, is the fully connected layer, is the base convolution kernel, is the adaptive long-term convolution kernel for the t-th printing task, is the level expansion coefficient of the t-th printing task, is the multimodal input vector of the t-th printing task, is the convolution kernel offset matrix of the t-th printing task, It is the convolution kernel offset weight matrix of the t-th printing task.

[0022] Furthermore, in the step S3, based on the short-term features, the long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task, and the calculation method is:

[0023] ;

[0024] in, is the fusion coefficient of the t-th printing task, is the hyperbolic tangent function, is the fusion coefficient weight matrix of the t-th printing task, For feature splicing, is the fusion coefficient bias vector of the t-th printing task, is the long-short time fusion feature of the t-th printing task; M is the medium type switching flag. When the medium type of the t-th printing task is the same as that of the t-1-th printing task, ,otherwise .

[0025] Furthermore, in the step S3, the medium type is mapped to a medium type embedding vector, and a convolution kernel offset matrix and a convolution kernel offset weight matrix are calculated, including:

[0026] The medium type is mapped to a medium type embedding vector, and the convolution kernel offset matrix and the convolution kernel offset weight matrix are calculated according to the medium type embedding vector. The calculation method is:

[0027] ;

[0028] in, is the media type embedding vector of the t-th printing task, m is the media type, For the embedding layer operation, is one-hot encoding, is the media type of the tth print task, is the sigmoid function.

[0029] In step S3 of the present invention, the dynamic convolution and medium perception fusion mechanism are used to significantly improve the accuracy and scene adaptability of printing consumables prediction; in the feature extraction stage, the reference convolution kernel is dynamically adjusted according to the convolution kernel offset matrix, and the adaptive long-term convolution kernel is generated by synchronously combining trilinear interpolation to achieve accurate separation and extraction of short-term local fluctuations and long-term trends; through the constraint design of sharing the reference convolution kernel parameters, the cross-media generalization ability is enhanced while reducing the model complexity;

[0030] In the feature fusion stage, a media switching trigger mechanism is innovatively introduced. When a change in the media type of an adjacent printing task is detected, a dynamic weighted fusion module based on a hyperbolic tangent function is automatically activated. The nonlinear fusion of long- and short-time features is achieved through the joint calculation of feature splicing and gating coefficients. Compared with the traditional static fusion method, this design can effectively suppress the prediction bias caused by sudden changes in feature distribution in the media switching scenario, while retaining the ability to perceive subtle changes in the same media task.

[0031] The entire solution addresses the technical difficulties of multi-media alternation and time-varying operating parameters in printing tasks. Through the collaborative optimization of parameter sharing and dynamic gating, it achieves highly robust prediction while ensuring computational efficiency. It is particularly suitable for consumables management scenarios of high-frequency, multi-media mixed industrial printing equipment.

[0032] Furthermore, in the step S4, the predicted value of inkjet consumption dynamically adjusts the contribution ratio of steady-state characteristics and transient characteristics through the inkjet consumption adaptive weight; the inkjet consumption adaptive weight is calculated based on the multimodal input vector of the current printing task, combined with the fully connected layer and the sigmoid function.

[0033] Furthermore, the step S4 includes:

[0034] S41: According to the long- and short-time fusion features of the printing task, the steady-state features are calculated by low-pass filtering. The calculation method is:

[0035] ;

[0036] in, is the steady-state characteristic of the t-th printing task, is one-dimensional average pooling, is the steady-state gate of the t-th printing task, is the sigmoid function;

[0037] S42: According to the long-short time fusion feature of the current printing task and the long-short time fusion feature of the previous printing task, the transient feature is obtained by high-frequency difference calculation. The calculation method is:

[0038] ;

[0039] in, is the transient characteristic of the t-th printing task, is a convolutional neural network, is the long- and short-time fusion feature of the t-1th printing task, is the high-pass convolution kernel, is the transient gate of the t-th printing task;

[0040] S43: Calculate the inkjet consumption characteristics according to the steady-state characteristics and the transient characteristics, and calculate the predicted value of the inkjet consumption, the calculation method is:

[0041] ;

[0042] in, is the inkjet consumption characteristic of the t-th printing task, is the adaptive weight of inkjet consumption for the t-th printing task, is the predicted value of inkjet consumption for the tth printing task.

[0043] Step S4 of the present invention realizes the working condition adaptive adjustment of inkjet consumption prediction through the steady-state-transient dual-modal fusion mechanism; in view of the coexistence of steady-state baseline fluctuations and transient interference in printing tasks, the present invention innovatively adopts low-pass filtering and high-frequency differentiation, and comprehensively utilizes dual-path features. In the steady-state feature extraction, the long-term trend baseline of consumables consumption is effectively captured through the combination of average pooling and gated weighting; in the transient feature extraction, convolution calculation is performed based on the difference in long- and short-time fusion features of adjacent tasks to accurately locate the transient fluctuations caused by sudden parameter changes; in addition, the present invention further designs the architecture of steady-state gates and transient gates, uses the sigmoid function to constrain the smooth transition of steady-state features, and enhances the dynamic response sensitivity of transient features through the hyperbolic tangent function; finally, the adaptive weight of inkjet consumption of the printing task is calculated, and the contribution ratio of steady-state and transient features is dynamically balanced, which not only avoids the prediction inaccuracy of traditional single features when the working conditions suddenly change, but also overcomes the defect of fixed weight fusion being insensitive to subtle changes.

[0044] Furthermore, the step S5 includes:

[0045] S51: Calculate the remaining ink volume according to the predicted ink consumption value, and the calculation method is:

[0046] ;

[0047] in, is the predicted value of the remaining ink volume after the tth printing task, is the initial ink volume, is the inkjet consumption of the i-th printing task, is the predicted value of inkjet consumption for the t-th printing task;

[0048] S52: Calculate the predicted value of ink consumption rate according to the predicted value of ink consumption, and the calculation method is:

[0049] ;

[0050] in, is the predicted value of ink consumption rate, is the duration of the tth printing task;

[0051] S53: construct an ink volume warning classifier and output the ink volume warning level according to the probability distribution. The calculation method is:

[0052] ;

[0053] in, is the probability distribution of ink volume alarm level, is the softmax function, Ink level warning level. is the fully connected layer, For feature splicing, is the argmax function.

[0054] The present invention also discloses a deep learning-based intelligent prediction system for remaining consumables of a printer, comprising:

[0055] Printing consumables data collection and preprocessing module: collects printing consumables data; preprocesses the printing consumables data except the timestamp and the media type, and combines the timestamp, the media type and the preprocessed printing consumables data to obtain the printing consumables data to be processed;

[0056] Printing consumables data cleaning module: performs integrity verification and outlier detection on the printing consumables data to be processed, and outputs the cleaned printing consumables data;

[0057] Long- and short-time fusion feature extraction module: sort the printing consumables data after cleaning by timestamp to construct a time series feature sequence; splice each printing consumables time series data to obtain a multimodal input vector; map the medium type to a medium type embedding vector, and calculate the convolution kernel offset matrix and the convolution kernel offset weight matrix; obtain the short-time features and long-time features of the printing task based on the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix; perform feature fusion based on the short-time features, long-time features, and the difference in media type between the current printing task and the previous printing task to obtain the long- and short-time fusion features of the printing task;

[0058] Inkjet consumption prediction module: extract steady-state features and transient features based on the long- and short-time fusion features of the printing task; calculate the inkjet consumption features based on the steady-state features and transient features, and calculate the inkjet consumption prediction value;

[0059] Ink volume alarm module: Based on the predicted value of inkjet consumption, calculate the remaining ink volume and ink consumption rate predicted value, then build an ink volume warning classifier, and output the ink volume alarm level according to the probability distribution.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) To address the problem of insufficient accuracy of consumables remaining prediction in traditional inkjet consumption prediction methods in multi-material mixed printing scenarios, the present invention designs a steady-state-transient dual-modal fusion mechanism and a media type switching flag to achieve accurate prediction under complex working conditions, significantly improving the consumables remaining prediction accuracy in multi-material mixed printing scenarios.

[0062] (2) To address the problem of insufficient adaptability of traditional convolution models in feature extraction in multi-media scenarios, the present invention generates an adaptive long-term convolution kernel through dynamic offset of the baseline convolution kernel and trilinear interpolation, accurately separating short-term fluctuations and long-term trend features. At the same time, a medium switching trigger mechanism is introduced to automatically activate the hyperbolic tangent gated fusion module when a change in medium type is detected. By using feature splicing and dynamic weighting to suppress mutation interference, the model generalization capability in multi-material mixed printing scenarios is improved.

[0063] (3) The present invention proposes a steady-state-transient dual-modal fusion method, which extracts the long-term consumption baseline through low-pass filtering and captures sudden fluctuations using high-frequency differential convolution. It also designs a differentiated gating mechanism to dynamically balance the weights of steady-state characteristics and transient characteristics, so that the model can maintain long-term trend stability and quickly respond to transient disturbances under complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of the process of the intelligent prediction method for remaining consumables of a printer based on deep learning provided by the present invention;

[0065] Figure 2 A schematic diagram of the flow of calculating the convolution kernel offset matrix and the convolution kernel offset weight matrix provided by the present invention;

[0066] Figure 3 A schematic diagram of the algorithm flow for long- and short-time fusion feature extraction provided by the present invention. DETAILED DESCRIPTION

[0067] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.

[0068] Embodiment 1:

[0069] A method for intelligently predicting the remaining amount of printer consumables based on deep learning, such as Figure 1 As shown, the following steps are included:

[0070] S1: Collect printing consumables data, wherein the printing consumables data includes the timestamp, duration, media type, number of printed pages, and inkjet consumption of each printing task; pre-process the printing consumables data except the timestamp and media type, and then merge the timestamp, media type and the pre-processed printing consumables data to obtain the printing consumables data to be processed.

[0071] Furthermore, the printing consumables data also includes the printing resolution, page size, medium ink absorption rate, medium surface roughness, and color mode of each printing task;

[0072] The printing consumables data other than the timestamp and the media type are preprocessed, including: the timestamp and the media type are not processed; the duration, the printing resolution, the page size, the number of printed pages, the ink absorption rate of the media, the surface roughness of the media and the inkjet consumption are normalized by the maximum and minimum value method; and the color mode is uniquely encoded.

[0073] S2: Perform integrity verification and abnormal value detection on the printing consumables data to be processed, and output the printing consumables data after cleaning, including:

[0074] S21: verify the existence of the fields of media type, number of printed pages, and inkjet consumption for the printing consumables data to be processed, delete the data entries with missing fields, and obtain data that passes the integrity verification;

[0075] S22: Based on the data that has passed the integrity verification, the Z-score method is used to detect global outliers, and the data is grouped by media type, and the abnormal data of printing consumables is identified by the IQR method;

[0076] S23: For the abnormal data of printing consumables, the ink jetting volume deviation is calculated by using a theoretical ink jetting volume formula, and data entries with ink jetting volume deviation exceeding 20% ​​are eliminated to obtain the printing consumables data after cleaning.

[0077] S3: Sort the printing consumables data after cleaning by timestamp to construct a time series feature sequence; splice each printing consumables time series data to obtain a multimodal input vector; map the medium type to a medium type embedding vector, and calculate the convolution kernel offset matrix and the convolution kernel offset weight matrix;

[0078] According to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, the short-term features and long-term features of the printing task are obtained; based on the short-term features, long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task, including:

[0079] S31: Sort the cleaning consumables data by timestamp to construct the time series data of the consumables, and then concatenate each time series data of the consumables to obtain a multi-modal input vector ;

[0080] S32: Map the medium type to a medium type embedding vector, and calculate the convolution kernel offset matrix and the convolution kernel offset weight matrix according to the medium type embedding vector, such as Figure 2 As shown, the calculation method is:

[0081] ;

[0082] in, is the media type embedding vector of the t-th printing task, t is the printing task index, m is the media type, For the embedding layer operation, is one-hot encoding, is the media type of the tth print task, is the convolution kernel offset matrix of the t-th printing task, is the fully connected layer, is the convolution kernel offset weight matrix for the t-th printing task, is the sigmoid function;

[0083] S33: According to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, the short-term features and long-term features of the printing task are obtained by combining the one-dimensional convolution and the dilated convolution. The calculation method is:

[0084] ;

[0085] in, is the short-term feature of the t-th printing task, is the ReLU function, is batch normalization, is a one-dimensional convolution, For Hadamard, is the long-term feature of the t-th printing task, is the GeLU function, is channel normalization, is a dilated convolution, is the base convolution kernel, is the adaptive long-term convolution kernel for the t-th printing task, is trilinear interpolation, is the level expansion coefficient of the t-th printing task, is the multimodal input vector of the t-th printing task;

[0086] The reference convolution kernel is initialized by Xavier normal distribution, the convolution kernel sizes include 3, 5, and 7, and the reference convolution kernel parameters are shared for all media types during training;

[0087] S34: Based on the short-term features and the long-term features, combined with the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task, such as Figure 3 As shown, the calculation method is:

[0088] ;

[0089] in, is the fusion coefficient of the t-th printing task, is the hyperbolic tangent function, is the fusion coefficient weight matrix of the t-th printing task, For feature splicing, is the fusion coefficient bias vector of the t-th printing task, is the long-short time fusion feature of the t-th printing task; M is the medium type switching flag. When the medium type of the t-th printing task is the same as that of the t-1-th printing task, ,otherwise .

[0090] It should be further explained that steps S33-S34 of the present invention significantly improve the accuracy and scene adaptability of printing consumables prediction through dynamic convolution and medium perception fusion mechanism; in the feature extraction stage (S33), the reference convolution kernel is dynamically adjusted according to the convolution kernel offset matrix, and the adaptive long-term convolution kernel is generated by combining trilinear interpolation to achieve accurate separation and extraction of short-term local fluctuations and long-term trends; through the constraint design of sharing the reference convolution kernel parameters, the cross-media generalization ability is enhanced while reducing the model complexity;

[0091] In the feature fusion stage (S34), a media switching trigger mechanism is innovatively introduced. When a change in the media type of an adjacent printing task is detected, a dynamic weighted fusion module based on a hyperbolic tangent function is automatically activated. The nonlinear fusion of long- and short-time features is achieved through the joint calculation of feature splicing and gating coefficients. Compared with the traditional static fusion method, this design can effectively suppress the prediction bias caused by sudden changes in feature distribution in the media switching scenario, while retaining the ability to perceive subtle changes in the same media task.

[0092] The entire solution addresses the technical difficulties of multi-media alternation and time-varying operating parameters in printing tasks. Through the collaborative optimization of parameter sharing and dynamic gating, it achieves highly robust prediction while ensuring computational efficiency. It is particularly suitable for consumables management scenarios of high-frequency, multi-media mixed industrial printing equipment.

[0093] S4: Extract steady-state features and transient features based on the long- and short-time fusion features of the printing task; calculate inkjet consumption features based on the steady-state features and transient features, and calculate the predicted value of inkjet consumption, including:

[0094] S41: According to the long- and short-time fusion features of the printing task, the steady-state features are calculated by low-pass filtering. The calculation method is:

[0095] ;

[0096] in, is the steady-state characteristic of the t-th printing task, is one-dimensional average pooling, is the steady-state gate of the t-th printing task;

[0097] S42: According to the long-short time fusion feature of the current printing task and the long-short time fusion feature of the previous printing task, the transient feature is obtained by high-frequency difference calculation. The calculation method is:

[0098] ;

[0099] in, is the transient characteristic of the t-th printing task, is a convolutional neural network, is the long- and short-time fusion feature of the t-1th printing task, is the high-pass convolution kernel, is the transient gate of the t-th printing task;

[0100] S43: Calculate the inkjet consumption characteristics according to the steady-state characteristics and the transient characteristics, and calculate the predicted value of the inkjet consumption, the calculation method is:

[0101] ;

[0102] in, is the inkjet consumption characteristic of the t-th printing task, is the adaptive weight of inkjet consumption for the t-th printing task, is the predicted value of inkjet consumption for the tth printing task.

[0103] Step S4 of the present invention realizes the working condition adaptive adjustment of inkjet consumption prediction through the steady-state-transient dual-modal fusion mechanism; in view of the coexistence of steady-state baseline fluctuations and transient interference in printing tasks, the present invention innovatively adopts low-pass filtering and high-frequency differentiation, and comprehensively utilizes dual-path features. In the steady-state feature extraction, the long-term trend baseline of consumables consumption is effectively captured through the combination of average pooling and gated weighting; in the transient feature extraction, convolution calculation is performed based on the difference in long- and short-time fusion features of adjacent tasks, and the transient fluctuations caused by sudden parameter changes are accurately located; in addition, the present invention further designs the architecture of steady-state gates and transient gates, uses the sigmoid function to constrain the smooth transition of steady-state features, and enhances the dynamic response sensitivity of transient features through the hyperbolic tangent function; finally, the adaptive weight of inkjet consumption of the printing task is calculated, and the contribution ratio of steady-state and transient features is dynamically balanced, which not only avoids the prediction inaccuracy of traditional single features when the working conditions suddenly change, but also overcomes the defect of fixed weight fusion being insensitive to subtle changes.

[0104] S5: Calculate the remaining ink volume and ink consumption rate prediction value according to the inkjet consumption prediction value, then construct an ink volume warning classifier, and output the ink volume warning level according to the probability distribution;

[0105] S51: Calculate the remaining ink volume according to the predicted ink consumption value, and the calculation method is:

[0106] ;

[0107] in, is the predicted value of the remaining ink volume after the tth printing task, is the initial ink volume, is the inkjet consumption of the i-th printing task;

[0108] S52: Calculate the predicted value of ink consumption rate according to the predicted value of ink consumption, and the calculation method is:

[0109] ;

[0110] in, is the predicted value of ink consumption rate, is the duration of the tth printing task;

[0111] S53: construct an ink volume warning classifier and output the ink volume warning level according to the probability distribution. The calculation method is:

[0112] ;

[0113] in, is the probability distribution of ink volume alarm level, is the softmax function, Ink level warning level. is the argmax function.

[0114] For example, the initial ink volume of an industrial inkjet printer is 5000ml, and it has completed three high-load printing tasks. The parameters of each task are as follows:

[0115] Task 1: Inkjet consumption y¹=420ml, lasting 1.0 hour;

[0116] Task 2: Inkjet consumption y²=580ml, lasting 1.2 hours;

[0117] Task 3: Inkjet consumption y³=650ml, lasting 1.5 hours;

[0118] Task 4 prediction: inkjet consumption prediction =720ml, task duration = 2 hours; then the predicted value of the remaining ink volume is:

[0119] ;

[0120] Ink consumption rate prediction ;

[0121] The probability distribution is obtained through the softmax function, and the probability distribution of the ink volume alarm level is calculated , representing the probabilities of the four alarm levels respectively , and finally issue a no-alarm signal (L0);

[0122] The alarm levels are defined as follows:

[0123] Low-level alarm (L1): 20%≤remaining ink volume<30%, and 450ml / h<consumption rate≤550ml / h;

[0124] Intermediate alarm (L2): 10%≤Remaining ink volume<20%, or consumption rate≤650ml / h;

[0125] Advanced alarm (L3): The remaining ink volume is less than 10%, or the consumption rate is greater than 650 ml / h;

[0126] No alarm (L0): other situations;

[0127] If multiple alarm levels are met at the same time, an alarm will be issued at the higher alarm level;

[0128] In particular, if there is a scenario in the printing task that is sensitive to the ink consumption rate, the following formula can be used to replace the calculation formula of the ink warning classifier in step S53, specifically:

[0129] ;

[0130] in, is the ink consumption rate characteristic, are the ink consumption rates of the first, second, ..., t-1th times, respectively, It is the probability distribution of ink level warning level.

[0131] Embodiment 2:

[0132] A printer consumables remaining intelligent prediction system based on deep learning, comprising:

[0133] Printing consumables data collection and preprocessing module: collects printing consumables data; preprocesses the printing consumables data except the timestamp and the media type, and combines the timestamp, the media type and the preprocessed printing consumables data to obtain the printing consumables data to be processed;

[0134] Printing consumables data cleaning module: performs integrity verification and outlier detection on the printing consumables data to be processed, and outputs the cleaned printing consumables data;

[0135] Long- and short-time fusion feature extraction module: sort the printing consumables data after cleaning by timestamp to construct a time series feature sequence; splice each printing consumables time series data to obtain a multimodal input vector; map the medium type to a medium type embedding vector, and calculate the convolution kernel offset matrix and the convolution kernel offset weight matrix; obtain the short-time features and long-time features of the printing task based on the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix; perform feature fusion based on the short-time features, long-time features, and the difference in media type between the current printing task and the previous printing task to obtain the long- and short-time fusion features of the printing task;

[0136] Inkjet consumption prediction module: extract steady-state features and transient features based on the long- and short-time fusion features of the printing task; calculate the inkjet consumption features based on the steady-state features and transient features, and calculate the inkjet consumption prediction value;

[0137] Ink volume alarm module: Based on the predicted value of inkjet consumption, calculate the remaining ink volume and ink consumption rate predicted value, then build an ink volume warning classifier, and output the ink volume alarm level according to the probability distribution.

[0138] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0139] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in multiple embodiments of the present invention.

[0140] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for intelligently predicting the remaining amount of printer consumables based on deep learning, characterized in that: The following steps are involved: S1: Collecting printing consumables data, wherein the printing consumables data includes a timestamp, duration, media type, number of printed pages, and inkjet consumption of each printing task; Preprocessing the printing consumables data except the timestamp and the medium type, and then merging the timestamp, the medium type and the preprocessed printing consumables data to obtain the printing consumables data to be processed; S2: Perform integrity verification and outlier detection on the printing consumables data to be processed, and output the printing consumables data after cleaning; S3: sort the printing consumables data after cleaning by timestamp to construct a time series feature sequence; Each piece of printing consumables time series data is spliced ​​to obtain a multimodal input vector; the media type is mapped to a media type embedding vector, and the convolution kernel offset matrix and the convolution kernel offset weight matrix are calculated; According to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, the short-term features and long-term features of the printing task are obtained; Based on the short-term features, long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task; S4: extracting steady-state features and transient features according to the long- and short-time fusion features of the printing task; calculating inkjet consumption features based on the steady-state features and transient features, and calculating the predicted value of inkjet consumption; S5: According to the predicted value of inkjet consumption, the remaining ink volume and the predicted value of ink consumption rate are calculated, and then an ink volume warning classifier is constructed, and the ink volume alarm level is output according to the probability distribution.

2. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 1, characterized in that: In the step S1, the printing consumables data also includes the printing resolution, page size, medium ink absorption rate, medium surface roughness, and color mode of each printing task; The printing consumables data other than the timestamp and the media type are preprocessed, including: the timestamp and the media type are not processed; the duration, the printing resolution, the page size, the number of printed pages, the ink absorption rate of the media, the surface roughness of the media and the inkjet consumption are normalized by the maximum and minimum value method; and the color mode is uniquely encoded.

3. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 1, characterized in that: The step S2 comprises: S21: verify the existence of the fields of media type, number of printed pages, and inkjet consumption for the printing consumables data to be processed, delete the data entries with missing fields, and obtain data that passes the integrity verification; S22: Based on the data that has passed the integrity verification, the Z-score method is used to detect global outliers, and the data is grouped by media type, and the abnormal data of printing consumables is identified by the IQR method; S23: For the abnormal data of printing consumables, the ink jetting volume deviation is calculated by using a theoretical ink jetting volume formula, and data entries with ink jetting volume deviation exceeding 20% ​​are eliminated to obtain the printing consumables data after cleaning.

4. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 1, characterized in that: In the step S3, the short-term features and long-term features of the printing task are obtained according to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix. The calculation method is: ； in, is the short-term feature of the t-th printing task, t is the printing task index, is the ReLU function, is batch normalization, is a one-dimensional convolution, For Hadamard, is the long-term feature of the t-th printing task, is the GeLU function, is channel normalization, is a dilated convolution, is trilinear interpolation, is the fully connected layer, is the base convolution kernel, is the adaptive long-term convolution kernel for the t-th printing task, is the level expansion coefficient of the t-th printing task, is the multimodal input vector of the t-th printing task, is the convolution kernel offset matrix of the t-th printing task, It is the convolution kernel offset weight matrix of the t-th printing task.

5. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 4, characterized in that: In the step S3, based on the short-term features, the long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task. The calculation method is: ； in, is the fusion coefficient of the t-th printing task, is the hyperbolic tangent function, is the fusion coefficient weight matrix of the t-th printing task, For feature splicing, is the fusion coefficient bias vector of the t-th printing task, is the long-short time fusion feature of the t-th printing task; M is the medium type switching flag. When the medium type of the t-th printing task is the same as that of the t-1-th printing task, ,otherwise .

6. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 5, characterized in that: In the step S3, the medium type is mapped to a medium type embedding vector, and the convolution kernel offset matrix and the convolution kernel offset weight matrix are calculated, including: The medium type is mapped to a medium type embedding vector, and the convolution kernel offset matrix and the convolution kernel offset weight matrix are calculated according to the medium type embedding vector. The calculation method is: ； in, is the media type embedding vector of the t-th printing task, m is the media type, For the embedding layer operation, is one-hot encoding, is the media type of the tth print task, is the sigmoid function.

7. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 5, characterized in that: In the step S4, the predicted value of inkjet consumption dynamically adjusts the contribution ratio of steady-state characteristics and transient characteristics through the inkjet consumption adaptive weight; the inkjet consumption adaptive weight is calculated based on the multimodal input vector of the current printing task, combined with the fully connected layer and the sigmoid function.

8. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 7, characterized in that: The S4 step comprises: S41: According to the long- and short-time fusion features of the printing task, the steady-state features are calculated by low-pass filtering. The calculation method is: ； in, is the steady-state characteristic of the t-th printing task, is one-dimensional average pooling, is the steady-state gate of the t-th printing task, is the sigmoid function; S42: According to the long-short time fusion feature of the current printing task and the long-short time fusion feature of the previous printing task, the transient feature is obtained by high-frequency difference calculation. The calculation method is: ； in, is the transient characteristic of the t-th printing task, is a convolutional neural network, is the long- and short-time fusion feature of the t-1th printing task, is the high-pass convolution kernel, is the transient gate of the t-th printing task; S43: Calculate the inkjet consumption characteristics according to the steady-state characteristics and the transient characteristics, and calculate the predicted value of the inkjet consumption, the calculation method is: ； in, is the inkjet consumption characteristic of the t-th printing task, is the adaptive weight of inkjet consumption for the t-th printing task, is the predicted value of inkjet consumption for the tth printing task.

9. The method for intelligently predicting the remaining amount of printer consumables based on deep learning according to claim 1, characterized in that: The step S5 comprises: S51: Calculate the remaining ink volume according to the predicted ink consumption value, and the calculation method is: ； in, is the predicted value of the remaining ink volume after the tth printing task, is the initial ink volume, is the inkjet consumption of the i-th printing task, is the predicted value of inkjet consumption for the tth printing task; S52: Calculate the predicted value of ink consumption rate according to the predicted value of ink consumption, and the calculation method is: ； in, is the predicted value of ink consumption rate, is the duration of the tth printing task; S53: construct an ink volume warning classifier and output the ink volume warning level according to the probability distribution. The calculation method is: ； in, is the probability distribution of ink volume alarm level, is the softmax function, Ink level warning level. is the fully connected layer, For feature splicing, is the argmax function.

10. A printer consumables remaining intelligent prediction system based on deep learning, characterized in that: include: Printing consumables data collection and preprocessing module: collect printing consumables data; Preprocessing the printing consumables data except the timestamp and the medium type, and merging the timestamp, the medium type and the preprocessed printing consumables data to obtain the printing consumables data to be processed; Printing consumables data cleaning module: performs integrity verification and outlier detection on the printing consumables data to be processed, and outputs the cleaned printing consumables data; Long-short time fusion feature extraction module: sort the printing consumables data after cleaning by timestamp and construct a time series feature sequence; Each piece of printing consumables time series data is spliced ​​to obtain a multimodal input vector; the medium type is mapped to a medium type embedding vector, and the convolution kernel offset matrix and the convolution kernel offset weight matrix are calculated; according to the multimodal input vector, the convolution kernel offset matrix, and the convolution kernel offset weight matrix, the short-term features and long-term features of the printing task are obtained; Based on the short-term features, long-term features, and the difference in media type between the current printing task and the previous printing task, feature fusion is performed to obtain the long-term and short-term fusion features of the printing task; Inkjet consumption prediction module: extracts steady-state features and transient features based on the long- and short-time fusion features of the printing task; and calculating the inkjet consumption characteristics based on the steady-state characteristics and the transient characteristics, and calculating the inkjet consumption prediction value; Ink volume alarm module: Calculate the remaining ink volume and ink consumption rate prediction value based on the inkjet consumption prediction value, then build an ink volume warning classifier, and output the ink volume alarm level based on the probability distribution; To realize a method for intelligently predicting remaining printer consumables based on deep learning as described in any one of claims 1-9.

Images