CN102779287B - Ink key opening forecasting method having increment type learning capacity - Google Patents

Abstract

The invention is a digital printing ink preset method, and proposes an ink key opening value prediction method based on a Fuzzy ART-BP mixed neural network algorithm based on a Fuzzy ART neural network and a BP neural network. The network makes full use of the self-learning, self-organization and information fuzzy processing capabilities of the Fuzzy ART neural network to generate stable classifications of the input vectors, and uses the BP neural network to nonlinearly map the input and output vectors of the training samples for each classification. That is, the temperature, humidity, printing machine speed and dot area ratio corresponding to the ink zone are used as the input vector, and the opening value of the ink key is used as the output vector to establish the mapping relationship between the graphic and digital information of the training sample and the control parameters of the ink key , using the converged network to predict ink key opening values for new samples. The network learning is more targeted, reduces the number of iterations of the BP network, and at the same time enables the network to have the ability of incremental learning and improves the generalization of the network.

Description

A Prediction Method of Ink Key Opening with Incremental Learning Ability

technical field

The invention belongs to the field of digital printing, and in particular relates to an ink key opening prediction method with incremental learning ability.

Background technique

With the increasing use of digital files in the prepress field, digital processes are increasingly used in printing processes, and the role of digital processes in CTP (computer-to-plate) technology is becoming more and more important. At the same time, printing companies are facing more and more short boards, complex and fast printing production activities, which also put forward higher requirements for printing companies. For printing companies, an effective way to shorten the makeready time is to pre-set the ink. Pre-estimating the optimal ink output of the ink tank and presetting the ink keys can save a lot of printing machine start-up preparation time, reduce production costs, improve printing quality and efficiency, and significantly reduce paper waste.

The way that the printing press controls the amount of ink is actually to divide the printable part of the printing plate into many narrow and long areas in the direction of the vertical long side - the ink area or the ink channel. The amount of ink in each ink area can be determined according to the printing plate. The area percentage of the graphic part in the area of the ink area - the area ratio of the dot is adjusted accurately. The higher the area percentage of the graphic part, the more ink is needed. Ink preset is to obtain information such as the dot area ratio in each ink area according to the film, printing plate or other carriers before printing, and establish the functional relationship between the dot area ratio and the opening of the ink key, and then initially set the ink on the printing machine. The amount of ink applied to each ink zone. Ink preset technology is a representative technology for digital technology to enter the printing production process. It is one of the important key technologies in the digital printing workflow and plays a decisive role in printing quality and printing efficiency.

The traditional ink preset technology based on BP neural network algorithm does not have the incremental learning ability (online learning) of training sample data, and the generalization ability is weak. When training and learning new sample data, it will destroy the model that the network has memorized, resulting in The network's ink key opening prediction results are not accurate enough. In order to solve this problem, this paper uses Fuzzy ART (fuzzy adaptive resonance neural network) and BP (Back Propagation) neural network, and comprehensively applies the two neural networks to form a Fuzzy ART-BP hybrid neural network with incremental learning ability. However, the ink preset technology based on the FuzzyART-BP hybrid neural network algorithm can effectively save the adjustment time of printing machine start-up, reduce the waste of paper and ink for start-up preparation, and reduce the labor intensity of printing operators at the same time. The disadvantages caused by the personnel controlling the amount of ink based on experience alone and the disadvantages that the BP neural network cannot realize incremental learning when learning and training new sample data improve the prediction accuracy of the ink preset technology.

Contents of the invention

The invention relates to an ink key opening prediction method with incremental learning ability. In order to further improve the printing quality and production efficiency in the domestic printing industry, some enterprises have successively introduced various foreign ink preset systems to pre-adjust the ink keys of the printing machine before starting the machine, but the actual application of the ink preset system is not satisfactory. The mutual influence between ink keys and the influence of printing conditions are not taken into account, and at the same time, incremental learning cannot be realized for the learning and training of new sample data, resulting in failure to achieve the expected use effect.

The method described in the present invention is to use the printed sheet with the solid density (the surface color density printed evenly and without blanks) that meets the national standard printing standard as a training sample, and uses the Fuzzy ART-BP hybrid neural network to train the training sample with a tutor and to train the training sample. The untrained and trained samples are used to predict the ink key opening, and the BP neural network uses 3 layers (input layer, hidden layer and output layer). First digitize the customer's original manuscript, that is, to obtain complete layout data information, and then generate dot matrix information after rasterization through RIP (raster image processor), and convert the generated dot matrix information through software to generate low-resolution layout information , that is, the dot area ratio. The dot area ratio corresponding to the printing site conditions (including site temperature, site humidity, and printing machine speed) and the ink area is used as the input raw data of the Fuzzy ART-BP hybrid neural network, and the input raw data is normalized to [0, 1] After being converted, it is sent to the input layer of the Fuzzy ART-BP hybrid neural network, and the ink key opening corresponding to the training sample is used as the original data of the output layer of the Fuzzy ART-BP hybrid neural network, and the output raw data is also [0, 1] After normalization processing, it is sent to the output layer of the Fuzzy ART-BP hybrid neural network. The number of hidden layer nodes of the BP neural network is set at 21-35, and the optimal adjustment is made according to the training results, and finally the number of hidden layer nodes is selected. twenty three.

Invoke the Fuzzy ART-BP algorithm program to train and learn the qualified actual printed sheets, thus establishing the nonlinear mapping relationship between the graphic and digital information of printed sheets and printing conditions (on-site temperature, on-site humidity, printing machine speed) and ink key opening, Fuzzy ART nerve The network first performs self-adaptive clustering operations on the graphic digital information and printing conditions of printed sheets, and performs BP neural network training and learning on the classified data. When the convergence error of the BP neural network is less than 10e-4, the BP neural network finally converges and saves The weights and thresholds of the linear mapping, and the weights of the Fuzzy ART neural network to the database. If the convergence error of the BP neural network is not less than 10e-4, continue to iteratively calculate the training sample pairs until the convergence error is less than 10e-4. Use the trained FuzzyART-BP hybrid neural network to predict untrained samples, and the predicted value of ink key opening is transmitted to the real-time database of the printing machine through the network or storage media, and then the printing machine is controlled by the console to print. The method can effectively shorten the preparation time for startup, improve printing efficiency and quality, and realize incremental learning of new sample data.

The Fuzzy ART-BP hybrid neural network algorithm adopts a tutor learning method for training. Its main idea is: input the sample vector for learning, first perform adaptive competitive clustering through the Fuzzy ART neural network, and modify the weight of the category to which the input vector belongs. value, and then use the backpropagation algorithm BP to repeatedly adjust and train the weights and thresholds of the BP neural network on the category of the input vector, so that the actual output vector is as close as possible to the expected output vector. When the square error of the network output layer When the sum is less than the specified error, the training is completed, and the weight and threshold of the network at this time are saved, and the Fuzzy ART-BP hybrid neural network training is over.

In order to better use the digital ink preset technology based on the Fuzzy ART-BP hybrid neural network algorithm, this article gives the detailed steps of the Fuzzy ART-BP hybrid neural network algorithm, as follows:

(1) Parameter introduction

The input vector P=(a ₁ ,a ₂ ...a _n ) ^T of the Fuzzy ART-BP hybrid neural network, 20≤n≤30; the dot area ratio, on-site temperature, on-site humidity and printing machine speed in Jimo District are normalized digitized data.

Fuzzy ART-BP hybrid neural network expected output vector T=(s ₁ ,s ₂ ,...s _q ) ^T , q=n-3; Jimo key opening normalized data;

Hidden layer unit input vector S=(s ₁ ,s ₂ ,...s _p ) ^T , p takes 21-35; output vector B=(b ₁ ,b ₂ ,...b _p ) ^T , p takes 21-35.

Output layer unit input vector L=(l ₁ ,l ₂ ,...l _q ) ^T , q=n-3; actual output vector C=(c ₁ ,c ₂ ,c _q ) ^T , q=n-3 .

The connection weight W _i1j1 from the input layer to the hidden layer, i1=1,2,...,p,j1=1,2,...,n.

The connection weight V _ti1 from the hidden layer to the output layer, i1=1,2,...,p,t=1,2,...,q.

The output threshold θ _i1 of each unit in the hidden layer, ${\mathrm{\θ}}_{i1}=\left[\begin{array}{c}{\mathrm{\θ}}_1\\ {\mathrm{\θ}}_2\\ \&Center\; Dot;\\ \&Center\; Dot;\\ \&Center\; Dot;\\ {\mathrm{\θ}}_{i1}\end{array}\right],$ i1=1,2,...,p.

The output threshold y _t of each unit in the output layer, ${\mathrm{the\; y}}_t=\left[\begin{array}{c}{\mathrm{the\; y}}_1\\ {\mathrm{the\; y}}_2\\ \&Center\; Dot;\\ \&Center\; Dot;\\ \&Center\; Dot;\\ {\mathrm{the\; y}}_t\end{array}\right],$ t=1,2,...,q.

α is the momentum factor of BP neural network, 0<α<1.

β is the learning rate of BP neural network, 0<β<1.

(2) The specific learning process of Fuzzy ART-BP hybrid neural network

1) Select a set of input and actual output sample pairs P=(a ₁ ,a ₂ ,...a _n ) ^T , T=(s ₁ ,s ₂ ,...s _q ) ^T , and compare the sample pairs The normalization process is then provided to the input layer and output layer of the Fuzzy ART-BP hybrid neural network respectively.

2) Complement the input vector P=(a ₁ ,a ₂ ,...a _n ) ^T of the input layer of the Fuzzy ART-BP hybrid neural network as the input vector I of the Fuzzy ART neural network, which can effectively To suppress the proliferation of the Fuzzy ART neural network category, the specific operation is shown in formula (1-1).

I=(a ₁ ,a ₂ ,a ₃ ,...a _n ,1-a ₁ ,1-a ₂ ,1-a ₃ ,...,1-a _n ) (1-1)

2) The input vector I is then classified by the "winner takes king" principle of the Fuzzy ART neural network.

3) Establish a BP neural network corresponding to the category number J of the input vector I, and initialize the BP neural network of category J, that is, give each connection weight W _i1j1 and V _ti1 , threshold θ _i1 and y _t ( -1, 1), and finally set the convergence error ε of the BP neural network algorithm;

4) The original input vector P corresponding to the input vector I obtained after complement operation and its expected output sample pair P=(a ₁ ,a ₂ ,...a _n ) ^T , T=(s ₁ , s ₂ ,...s _q ) ^T , input to the BP neural network of category J after normalization.

5) Use input layer sample data P=(a ₁ ,a ₂ ,...a _n ) ^T , connection weight W _i1j1 and threshold θ _i1 to calculate the input s _i1 of each unit in the hidden layer, and then use s _i1 to transfer The function calculates the output b _i1 of each unit in the hidden layer, and the transfer function uses the sigmoid function, and its form is: , the calculation formula of s _i1 is shown in formula (1-2), and the calculation formula of b _i1 is shown in formula (1-3).

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

6) Use the output b _i1 of the hidden layer and the connection weight V _ti1 threshold y _t to calculate the output l _t of each unit in the output layer, as shown in formula (1-4), and then use the transfer function to calculate the output l t of each unit in the output layer In response to c _t , the transfer function selects the sigmoid function, and its form is: , as shown in formula (1-5).

${\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; ti</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; =</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&Center\; Dot;</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

7) Use the actual output C=(c ₁ ,c ₂ ,...c _q ) ^T of the Fuzzy ART-BP hybrid neural network and the expected output T=(s ₁ ,s ₂ ,...s _q ) ^T of the network , calculate the error E, as shown in formula (1-6), if E is less than the set convergence error ε, the Fuzzy ART-BP hybrid neural network converges, the iteration ends and the weights W _i1j1 , V _ti1 and threshold θ _i1 are saved , y _t ; otherwise, continue to step 8), after modifying the weight and threshold matrix, continue to judge whether the error E is less than the set convergence error.

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

8) Using the actual output C=(c ₁ ,c ₂ ,...c _q ) ^T of the Fuzzy ART-BP hybrid neural network, the expected output of the network T=(s ₁ ,s ₂ ,...s _q ) ^T , calculate the generalization error d _t of each unit in the output layer, as shown in formula (1-7).

d _t =(s _t -c _t )·c _t ·(1-c _t ) (1-7)

9) Calculate the generalization error e _i1 of each unit in the hidden layer by using the generalization error d _t of each unit in the output layer of the BP network of the Fuzzy ART-BP hybrid neural network, the connection weight V _ti1 and the output b _i1 of the hidden layer, As shown in formula (1-8).

${\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; [</span>\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; \&Center\; Dot;</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; ti</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

10) Use the generalization error d _t of each unit in the output layer and the output b _i1 of each unit in the hidden layer to modify the connection weight V _ti1 , as shown in formula (1-9), and modify the threshold y _t as shown in formula (1-10 ) shown.

${\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; ti</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&alpha;V</span>}}_{\mathrm{<span\; class="notranslate">\; ti</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&beta;d</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&beta;d</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

11) Use the generalization error e _i1 of each unit in the hidden layer and the input P=(a ₁ ,a ₂ ,...a _n ) of each unit in the input layer to modify the connection weight W _i1j1 , as shown in formula (1-11 ), the threshold θ _i1 is modified as shown in formula (1-12).

${\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&alpha;w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&beta;e</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 11</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; \&beta;e</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; )</span>$

12) Use input layer sample data P=(a ₁ ,a ₂ ,...a _n ) ^T , connection weights and threshold Calculate the input of each unit of the hidden layer , such as formula (1-13), then use Calculate the output of each unit in the hidden layer through the transfer function , such as formula (1-14), the transfer function selects the sigmoid function, and its form is:

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&theta;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 13</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&Center\; Dot;</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 14</span>}<span\; class="notranslate">\; )</span>$

13) Use the output of the hidden layer , right to connect threshold . Calculate the output of each unit in the output layer , such as formula (1-15), and then use the transfer function to calculate the response of each unit in the output layer , the transfer function selects the sigmoid function, and its form is: Such as formula (1-16).

${\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; ti</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 15</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\left[\begin{array}{c}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ <span\; class="notranslate">\; \&Center\; Dot;</span>\\ <span\; class="notranslate">\; \&CenterDot;</span>\\ \mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; l</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>\end{array}\right]<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 16</span>}<span\; class="notranslate">\; )</span>$

14) Utilize the actual output of the network , the expected output of the network T=(s ₁ ,s ₂ ,...s _q ) ^T , calculate the objective function (error) E′, such as formula (1-17), if E′ is smaller than the set convergence error ε, Then the network converges, end the iteration and save the weight and threshold; otherwise return to step 8), modify the weight and threshold matrix, and continue the iterative calculation.

${\mathrm{<span\; class="notranslate">\; E.</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; q</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 17</span>}<span\; class="notranslate">\; )</span>$

15) Select the next training sample pair and provide it to the Fuzzy ART-BP hybrid neural network, and continue with steps 1)-15) until all training sample pairs are trained.

The above process can be represented by the learning and training flow chart of the Fuzzy ART-BP hybrid neural network in Figure (2).

Description of drawings

Figure 1 Schematic diagram of ink preset technology.

Figure 2 The learning and training flow chart of the Fuzzy ART-BP hybrid neural network.

Fig. 3 The cyan prediction results in ink key opening prediction.

Figure 4 The cyan prediction result map in the ink key opening verification.

Fig. 5 is a visual representation diagram of network prediction difference corresponding to each ink region after training 100 sets of data. Fig. 6 is a visual representation diagram of the network prediction difference corresponding to each ink region after training the network with 150 sets of data.

Detailed ways

Firstly, digitize the original training samples to obtain complete layout data information, and then generate dot matrix information after rasterization processing by RIP (raster image processor), and convert the generated dot matrix information through software to generate layout information-dot area ratio . The relative humidity of the printing site is 30%, the speed of the printing machine is 4000 sheets/hour, and the temperature is 25°C. Since the units of the collected data are inconsistent, in order to speed up the convergence of the training network, the data must be normalized to [0, 1]. Normalization processing, the normalization processing of the above three conditions are: 0.3, 0.4, 0.25. The full grid of the ink fountain used is 100. When performing normalization processing, divide the actual ink key opening by 100. The value of the dot area ratio is between [0-100%], and directly take the smaller value without normalization processing. value. The temperature of the test site, the humidity of the test site, and the speed of the printing machine are the main factors affecting the opening of the ink keys, and there will be mutual influence between adjacent ink keys, so the dot area ratio of a certain point cannot be used as the input of the BP neural network. Therefore, the number of neurons in the input layer is determined to be 23, including on-site temperature, on-site humidity, printing machine speed and the dot area ratio of 20 ink zones; the determination of the number of output layer nodes: the output layer nodes are 20 ink zone outlets in sequence The ink key opening corresponding to the area ratio; the number of hidden layer nodes is set at 23. Learning rate β=0.4, momentum factor α=0.9, expected error ε=10e-4; learning rate η=0.15 of Fuzzy ART neural network, warning threshold ρ=0.95 of Fuzzy ART neural network.

Fuzzy ART-BP hybrid neural network algorithm and BP neural network algorithm have the obvious advantage of being able to incrementally learn new sample data. The following is a test of the prediction accuracy for this function. Since the Fuzzy ART-BP hybrid neural network algorithm has a strong analysis ability for more data, its prediction results and BP neural network prediction results have been significantly improved. Therefore, on the basis of the existing data, Monte Carlo simulation is used to simulate more data and provide it to the Fuzzy ART-BP hybrid neural network for training and learning. Among them, the existing part of the data in cyan is normalized, as shown in the following table:

Network input data:

Network output data:

Under the same printing environment, that is, under the same printing machine speed, temperature and air humidity, the four colors corresponding to the same dot area ratio of the same printed sheet (CMYK is the initial letter of the names of the four printing inks: cyan Cyan, magenta Magenta , Yellow, Yellow, and K are the last letter of black.) The ink key openings are also different, so it is necessary to establish neural networks for the existing data and simulation data according to the four colors, and carry out training and learning of their respective neural networks. predict.

Use the training method of transitioning from small capacity to large capacity training samples to learn the ink preset technology based on the Fuzzy ART-BP hybrid neural network with a network instructor, that is, follow the detailed steps of the Fuzzy ART-BP hybrid neural network algorithm in the content of the invention Carry out tutor training and study, and verify the superiority of its ink preset technology with BP neural network. First, select 100 groups of samples to conduct network learning on the data based on BP neural network ink preset technology and Fuzzy ART-BP hybrid neural network ink preset technology, and wait for the convergence error of the network to reach the set convergence error ε=10e- After 4, save the weights and thresholds of the respective networks; then randomly select a set of data from the sample pair data and input them into the two ink preset systems that have been trained to predict the opening of the ink key. The cyan prediction results are shown in Figure 3 shown. Finally, another 50 groups of samples were selected to learn the ink preset technology of the two neural networks with the data, and the same group of data was used to verify the prediction ability of the network again. The cyan prediction effect is shown in Figure 4.

In order to see the prediction accuracy of the Fuzzy ART-BP neural network algorithm and the BP neural network algorithm more intuitively, this paper subtracts the real value of the ink key opening when the actual print is qualified and the predicted value of the two neural networks. It is called the network prediction difference, and Fig. 5 shows the visual representation of the network prediction difference corresponding to each ink area after training 100 sets of data. Figure 6 is a visual representation of the network prediction difference corresponding to each ink region after training 150 sets of data.

From Figure 3 to Figure 6, it can be seen that the prediction effect of ink preset technology based on Fuzzy ART-BP hybrid neural network is very impressive, and the prediction difference is basically controlled at about ±2%. As the amount of training sample data increases, the variation range of the prediction difference of the BP neural network increases, and some have reached ±3%, indicating that the newly learned samples have destroyed the model that the network has memorized, and do not have the ability of incremental learning. This leads to an increase in the error of the forecast results. Compared with the BP neural network, the Fuzzy ART-BP hybrid neural network can perform adaptive classification learning on the input data, and then perform BP neural network learning on the classified model, making the network more targeted to the training mode, reducing as much as possible After learning the new model, the extent of the network's memorized model is destroyed, and the prediction result is basically controlled at about ±2%, which improves the network prediction accuracy and enables the network to have the ability of incremental learning.

Claims (1)

1. there is a black key aperture Forecasting Methodology for incremental learning ability, comprise the following steps:

1) with field density namely evenly and without blank the surface color density that prints meet the printed sheet of GB printing standard for training sample, for four look CMYK, i.e. cyan Cyan, magenta Magenta, yellow Yellow and black Black, sets up corresponding Fuzzy ART-BP hybrid neural networks respectively;

2) by printing condition, i.e. as Fuzzy ART-BP hybrid neural networks input layer input data after the dot area percentage normalized of scene temperature, on-the-spot humidity, printing machine rotating speed and 20-30 ink-covered area; As Fuzzy ART-BP hybrid neural networks output layer input data after ink key aperture normalized; Input layer input data are through Fuzzy ART self-adaption cluster, and the weights preserving class are to database, then set up BP neural network for certain the class data after cluster, regulate the node in hidden layer of suitable BP neural network, be set as 21-35; Input layer, hidden layer and output layer 3 layers of BP neural network algorithm procedural training module are used to have tutor's training study to training sample, when the convergence error of BP neural network is less than 10e-4, BP neural network finally restrains, and preserves the weights and threshold of BP neural network algorithm Nonlinear Mapping to database;

3) can by the dot area percentage of printed sheet and printing condition for non-training sample, i.e. scene temperature, on-the-spot humidity and printing machine rotating speed, be defeated by the prediction module of Fuzzy ART-BP hybrid neural networks algorithm, Fuzzy ART-BP hybrid neural networks algorithm routine utilizes the weights and threshold stored to carry out prediction and calculation to the scene temperature of input, on-the-spot humidity, printing machine rotating speed and dot area percentage, thus the black key aperture that the dot area percentage doping non-training sample is corresponding; Ink key aperture predicted value is sent to printing machine control desk by network or storage medium, and control desk receives data and automatically controls corresponding black key and is inked to printing machine and completes printing.