CN109146007B - Solid waste intelligent treatment method based on dynamic deep belief network - Google Patents

Abstract

The invention proposes a solid waste intelligent processing method based on a dynamic deep belief network, which belongs to the field of deep learning and solid waste intelligent processing. This method firstly proposes a DDBN using dynamic branch increase and decrease algorithm, which enables DDBN to increase hidden layer neurons and hidden layers according to the current training situation during the training process, and remove redundant neurons to effectively optimize the network structure of DDBN. Then, using DDBN to effectively extract the main features of the original data, use DDBN to effectively describe the state of the random, discrete, and nonlinear feature vectors of solid waste, so that the state features of the time series are easier to identify and ensure that the original data is not lost. The main information of the data. At the same time, according to the state description of the extracted solid waste, DDBN is used to predict the optimal combustion behavior suitable for its state, which reduces the waste of resources caused by blind combustion behavior and realizes the intelligent treatment of solid waste.

Description

An intelligent treatment method of solid waste based on dynamic deep belief network

technical field

The invention belongs to the fields of deep learning and intelligent processing of solid waste, and proposes a dynamic deep belief network (Dynamic Deep Belief Network, DDBN) model using a dynamic branch increase and decrement algorithm, which can effectively optimize the network structure of the deep belief network. To solve the problem of intelligent processing of bulk solid waste in the light industry.

Background technique

The development of light industry is currently facing great pressure on environmental protection and arduous task requirements for pollution reduction and emission reduction. With the development of the national economy, the market demand for fermentation and papermaking products has grown substantially. Although the pollution emission intensity per unit product of the industry has been significantly reduced in recent years, the total amount of solid waste emissions in the industry is still increasing due to the rapid expansion of production capacity. In order to achieve the relevant goals of energy conservation and emission reduction in the industry, it is necessary to study new methods of waste disposal, improve the pollution control and governance level of production enterprises through the application of new methods, and support the pollution reduction actions of the industry.

In recent years, deep learning has developed rapidly. Hinton et al. proposed the Deep Belief Network (DBN) and the unsupervised greedy layer-by-layer training algorithm in 2006, which solved the problem that deep neural networks are prone to fall into local optimality, causing deep Learn about the new wave of learning in academia. DBN obtains an abstract representation of the original data through multi-level feature transformation, thereby improving the accuracy of tasks such as classification and prediction. Because DBN has the advantages of automatic learning of features and data dimensionality reduction, it has become the most widely used deep learning network structure.

Taking advantage of the advantage that DBN can effectively extract the main features of the original data, DBN can be used to effectively describe the state of the random, discrete, and nonlinear feature vectors of solid waste, so that the state features of the time series can be more easily identified and ensure that the original data is not lost. The main information of the data. At the same time, according to the state description of the extracted solid waste, DBN is used to predict the optimal combustion behavior suitable for its state, which greatly reduces the waste of resources caused by blind combustion behavior and realizes the intelligent treatment of solid waste.

However, if the DBN is to solve a high-complexity problem, such as the above problem, then the DBN needs to increase the hidden layer neurons and hidden layers appropriately. However, the number of hidden layer neurons and hidden layers still needs to be selected by manual experiments, and the network structure is fixed during the training process. In this way, not only the error is larger, but also the calculation cost is higher and the efficiency is lower. Therefore, it is necessary to propose a new DBN structure design method, so that the DBN can dynamically increase and decrease branches according to the current training situation during the training process, optimize the network structure, and make it better solve the problem of intelligent disposal of solid waste.

SUMMARY OF THE INVENTION

In order to solve the problems existing in the prior art, the present invention proposes an intelligent processing method for solid waste based on a dynamic deep belief network (Dynamic Deep Belief Network, DDBN).

Technical scheme of the present invention:

An intelligent processing method for solid waste based on dynamic deep belief network, the steps are as follows:

Step 1. Measure the solid waste to obtain a solid waste data set, preprocess the solid waste data set, and divide it into a training data set and a test data set.

The preprocessing is: normalize the solid waste data set to [0, 1], and the normalization formula is:

为固废数据集的特征值，x_max和x_min分别为固废数据集所有特征的最大值和最小值，x是经过归一化后的固废数据集。in,

is the feature value of the solid waste data set, x _max and x _min are the maximum and minimum values of all the features of the solid waste data set, respectively, and x is the normalized solid waste data set.

Step 2. Input the training data set obtained after preprocessing in Step 1 into the DDBN model, and use the Contrastive Divergence (CD) algorithm to train each layer of restricted Boltzmann machine ( Restricted Boltzmann Machine, RBM), and in the training process, the network structure of the current RBM is optimized through the dynamic branch increase and decrease algorithm, and the network structure and parameter values of each RBM are obtained through iterative training, and finally the high value of the training data set is obtained. Hierarchical features; the parameter values are weights and biases. The specific operations are as follows:

Step 2.1, build a DDBN network model, and set the parameter values of DDBN: visual layer neurons, initial hidden layer neurons and hidden layer layers, learning rate, number of iterations and number of fine-tuning. Among them, the number of neurons in the visual layer is the feature dimension of the training data set.

In step 2.2, the training data set obtained after preprocessing is input into the first layer of RBM, and the CD algorithm is used to pre-train the RBM, and during the training process, the network structure of the current RBM is optimized through the dynamic branch increase and decrease algorithm.

(1) The energy function E(v, h; θ) of the RBM and the joint probability distribution P(v, h; θ) of the neurons in the visible layer and the hidden layer are:

Among them, v _i (1≤i≤I) and h _j (1≤j≤J) represent the neurons in the visible layer and the neurons in the hidden layer, respectively, w is the weight matrix between the visible layer and the hidden layer, b and c are the biases of the visual layer neurons and the hidden layer neurons, respectively, θ represents the parameters in the model, that is, θ={w,b,c}; Z is the pair of all possible visual layer and hidden layer neurons beg for peace.

Using the principle of Bayesian formula, the marginal probability distribution of the visible layer neuron v and the hidden layer neuron h is obtained according to formula (3):

Using the Bayesian formula, the conditional probability distribution of the visible layer neuron v and the hidden layer neuron h is derived:

Using formula (6) and formula (7), the approximate reconstruction P(v; θ) of the training sample is obtained by one-step Gibbs sampling using the contrastive divergence algorithm, and then the network parameter θ={w,b is updated according to the reconstruction error ,c}.

(2) During the training process, according to the current training situation, the network structure of the RBM is optimized through the dynamic branch increase and decrease algorithm.

Use the WD (Weight Distance, weight distance) method to monitor the change of the weight w:

WD _j [m]=Met(w _j [m],w _j [m-1]) (8)

Among them, w _j [m] is the weight vector of hidden layer neuron j after m iterations, and Met represents a metric function, such as Euclidean distance. The value of WD reflects the change in the weight vector of hidden layer neuron j in two iterations.

The branching conditions are considered from both local and global aspects.

The local condition is defined as:

是在第m次迭代中，第j个隐藏神经元对第n个输入样本的WD值，j＝1,2,3,..,J，J是隐藏神经元的个数，max(·)是最大值函数。in

is the WD value of the jth hidden neuron for the nth input sample in the mth iteration, j=1,2,3,..,J, J is the number of hidden neurons, max( ) is the maximum function.

Global conditions are defined as:

where N is the number of samples in the training dataset, and N' is the number of samples that increase the WD value of the jth neuron compared to the previous iteration, namely

The local and global conditions are considered by the hidden layer neuron j for a single input sample and all input samples, respectively. Multiplying these two conditions gives the branching condition:

max_WD _j [m]*iratio _j [m]>y(m) (11)

where y(m) is a curve used as a variable threshold, which is defined as:

where m is the current number of iterations, numepoches is the maximum number of iterations, u is the curvature of the curve, and y _max and y _min are the maximum and minimum values of the curve, respectively. When the jth neuron satisfies the formula (11), the neuron will be divided into two neurons, and each parameter of the new neuron is 0.

When the RBM training is completed, start pruning: use the standard deviation of the activation probability of the hidden layer neurons for all samples as the pruning condition. The standard deviation formula is:

where n=1,2,3,...,N,N is the number of samples in the input training data set, j represents the jth hidden layer neuron, P(n,j) represents the jth neuron pair nth The activation probability of the input sample, μ _j represents the average activation probability of the jth neuron for all input samples.

The pruning conditions are:

σ(j) < θ _A (14)

where θ _A is a threshold. When the jth neuron satisfies the formula (14), the neuron and all its parameters are removed. At the same time, a trade-off curve between the pruning rate and the prediction accuracy is made, and the value of θ _A is selected according to this curve, so that more redundant neurons are removed while retaining the original accuracy.

After pruning, the RBM is retrained so that the remaining neurons can compensate for the removed neurons, and then retrained for one iteration after pruning. Each iteration we update the threshold θ _A :

θ _A ←θ _A +δ[iter] (15)

The threshold in each iteration of pruning is updated by δ[iter] to remove more neurons. Each iteration is a greedy search, and according to the trade-off curve in each pruning, the optimal pruning rate can be found without loss of accuracy, so δ[iter] is set so that θ _A satisfies this iteration The pruning rate required for pruning.

Step 2.3, after the current RBM determines the network structure, use the energy function as the condition for adding a new RBM:

where E ^l is the total energy of the lth layer of RBM, obtained by formula (2), l=1, 2...L, L is the current number of layers of DDBN, mean( ) is the average function, and θ _L is a threshold. When the energy function satisfies the formula (16), a new layer of RBM is added, and the initialization of each parameter of the new RBM is the same as that of the first layer. Then use the output of the current RBM as the input of the newly added RBM.

Step 2.4, follow steps 2.2 and 2.3 to train the network cyclically to obtain the network structure of DDBN.

Step 3. Use the DDBN network structure and parameter values obtained in step 2 as the initial values of the fine-tuning stage, and use the top-down backpropagation algorithm to fine-tune the entire DDBN network to obtain the final DDBN network model. The specific operations are as follows:

Step 3.1, take the DDBN network structure and parameter value θ trained in step 2 as the initial value of the fine-tuning stage, and add an output layer after the last layer of RBM to predict the combustion behavior suitable for the training data set samples , including temperature, pressure, and gas flow, input the training dataset to start fine-tuning the entire DDBN network.

Step 3.2, use the forward propagation algorithm to calculate the activation probability of each hidden layer neuron.

Step 3.3, calculate the prediction result obtained by the forward propagation of the training sample, and compare it with the actual result to obtain the loss function:

Among them, t is the current number of fine-tuning, N is the number of samples in the training data set, and y _n and y' _n are the actual and predicted results of the nth training sample, respectively. The error of the actual result and the predicted result is back-propagated, and the weight w and the bias c are updated according to the formulas (18) and (19) using the gradient descent method:

where α is the learning rate.

Iteratively uses the gradient descent method to fine-tune the entire DDBN network top-down to reduce the value of J(t) until the maximum number of fine-tuning times is reached to obtain the final DDBN network model.

Step 4. Input the test data set into the final DDBN network model obtained in step 3, and finally output the prediction result. The specific operations are as follows:

Step 4.1, input the preprocessed test data set into the fine-tuned DDBN network model in step 3, and extract the main features of solid waste through RBM.

Step 4.2, input the main features of the test sample into the last output layer, and predict the combustion behavior suitable for the training sample, including temperature, pressure and gas flow.

Beneficial effects of the present invention: In order to make the model more capable of extracting features and predicting, a DDBN model using a dynamic branch increase and decrease algorithm is proposed, so that the DDBN can change the network structure according to the current training situation during the training process, including adding hidden layer neurons. Elements and hidden layers and removing redundant neurons can effectively optimize the network structure of DDBN, replacing manual experiments and overcoming the difficulty of network structure design. Then, using DDBN to effectively extract the main features of the original data, use DDBN to effectively describe the state of the random, discrete, and nonlinear feature vectors of solid waste, so that the state features of the time series are easier to identify and ensure that they are not lost. The main information of the raw data. At the same time, according to the state description of the extracted solid waste, DDBN is used to predict the optimal combustion behavior suitable for its state, including temperature, pressure and gas flow, which greatly reduces the waste of resources caused by blind combustion behavior and realizes the intelligentization of solid waste. deal with.

Description of drawings

1 is a schematic diagram of the operation of adding hidden layer neurons in the present invention;

2 is a schematic diagram of the operation of removing redundant neurons in the present invention;

3 is a schematic diagram of an operation of adding a hidden layer in the present invention;

Fig. 4 is the training process flow chart of DDBN model in the present invention;

Fig. 5 is the effect diagram after the solid waste is processed intelligently in the present invention;

Detailed ways

The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

As shown in Figure 4, a solid waste intelligent processing method based on dynamic deep belief network, the specific steps are as follows:

Step 1. Measure the GDP of solid waste, hazardous substances, solid waste volume, smelting waste residue, coal ash, slag, tailings and other data to obtain a solid waste data set, and perform data preprocessing to obtain training and test data set.

Since the data of solid waste are often not in the same order of magnitude, it is necessary to normalize the solid waste data set to [0, 1], which is beneficial to improve the training speed of the network. The normalization formula is:

为固废数据集的特征值，x_max和x_min分别为固废数据集所有特征的最大值和最小值，x是经过归一化后的固废数据集。in,

is the feature value of the solid waste data set, x _max and x _min are the maximum and minimum values of all the features of the solid waste data set, respectively, and x is the normalized solid waste data set.

Step 2. Input the training data set obtained after preprocessing into the DDBN model for pre-training, and use the Contrastive Divergence (CD) algorithm to train each layer of restricted Boltzmann machines independently from the bottom to the top. (Restricted Boltzmann Machine, RBM). And in the training process, the network structure of the current RBM is optimized through the dynamic branch increase and decrease algorithm, including adding new hidden layer neurons and removing redundant neurons. After the current RBM training is completed, if the DDBN meets the layer generation conditions, a new RBM is added, and the output of the current RBM is used as the input of the newly added RBM, and the initialization of the parameters of the new RBM is the same as the first layer. Through multiple iterations, the network structure and corresponding weights and biases of each RBM are obtained, and finally the high-level features of the data are obtained. The specific operations are as follows:

Step 2.1, build a DDBN network model, and set the parameter values of DDBN: the number of neurons in the visual layer is the feature dimension of the training data set, the initial number of neurons in the hidden layer and the number of hidden layers are set to 10 and 1, respectively, and the learning rate Set to 0.1, the number of pre-training iterations is 100, and the number of fine-tuning is 100.

In step 2.2, the training data set obtained after preprocessing is used as the input of the first-layer RBM, and the CD algorithm is used to pre-train the RBM, and during the training process, the network structure of the current RBM is optimized through the dynamic branch increase and decrease algorithm, including Add new hidden layer neurons and remove redundant neurons.

(1) RBM is a random network based on an energy model. Each group of neurons in the visible layer v and neurons in the hidden layer has a corresponding energy value. According to the definition of energy function and the principle of thermodynamic statistics, the joint probability distribution of v and h can be obtained. The energy function E(v, h; θ) of the RBM and the joint probability distribution P(v, h; θ) of v and h are:

Among them, v _i (1≤i≤I) and h _j (1≤j≤J) represent the neurons in the visible layer and the neurons in the hidden layer, respectively, w is the weight matrix between the visible layer and the hidden layer, b and c are the biases of the neurons in the visible layer and the neurons in the hidden layer, respectively, θ represents the parameters in the model, ie θ={w,b,c}, and Z is the pair of neurons in the visible layer and the hidden layer for all possible pairs beg for peace.

Using the principle of Bayesian formula, the marginal probability distribution of the visible layer neuron v and the hidden layer neuron h can be obtained according to formula (3):

The goal of RBM network training is to solve θ={w,b,c}, so that RBM can fit the input sample greatly under this parameter, so that P(v; θ) is the largest, that is, to solve the maximum likelihood estimation of the input sample . However, in order to obtain a maximum likelihood estimate, all possible cases need to be calculated, and the amount of computation grows exponentially, so RBM uses a contrastive divergence algorithm for estimation.

Using the Bayesian formula, the conditional probability distribution of the visible layer neuron v and the hidden layer neuron h is derived:

According to formula (6) and formula (7), the approximate reconstruction P(v; θ) of the training sample is obtained through one-step Gibbs sampling using the contrastive divergence algorithm, and then the network parameter θ={w,b is updated according to the reconstruction error ,c}.

(2) During the training process, according to the current training situation, the network structure of the RBM is optimized through the dynamic branch increase and decrease algorithm.

In DBN, the weight w plays a decisive role in network training. Therefore, the present invention proposes a method called WD (Weight Distance, weight distance) to monitor the change of the weight w:

WD _j [m]=Met(w _j [m],w _j [m-1]) (8)

where w _j [m] is the weight vector of hidden layer neuron j after m iterations, and Met is a metric function such as Euclidean distance. The value of WD reflects the change in the weight vector of hidden layer neuron j in two iterations. In general, the weight vector of neuron j will converge after a period of training, that is, the value of WD will become smaller and smaller. If some weight vectors fluctuate greatly after a long iteration, it should be considered that this is caused by the lack of hidden layer neurons to map input samples. In this case, the number of neurons needs to be increased to improve the performance of the network. The present invention considers branching conditions from both local and global aspects. The local condition is defined as:

是在第m次迭代中，第j个隐藏神经元对第n个输入样本的WD值，j＝1,2,3,..,J，J是隐藏神经元的个数，max(·)是最大值函数。全局条件定义为：in

is the WD value of the jth hidden neuron for the nth input sample in the mth iteration, j=1,2,3,..,J, J is the number of hidden neurons, max( ) is the maximum function. Global conditions are defined as:

where N is the number of training samples, and N' is the number of samples that increase the WD value of the jth neuron compared with the previous iteration, that is,

The local and global conditions are considered by the hidden layer neuron j for a single input sample and all input samples, respectively. Find the maximum WD value max_WD _j [m] of neuron j for a certain sample and the sample ratio iratio _j (m) that increases the WD value. Then multiply these two conditions to get the branching condition:

max_WD _j [m]*iratio _j [m]>y(m) (11)

where y(m) is a curve used as a variable threshold, which is defined as:

where m is the current number of iterations, numepoches is the maximum number of iterations, u is the curvature of the curve, and y _max and y _min are the maximum and minimum values of the curve, respectively. During the training process, if the network develops in a good direction, the values of max_WD and iratio will become smaller and smaller, so a curve y(m) is used as the variable threshold of the branching condition, and when h>0, y (m) is a monotonically decreasing curve. If the jth neuron satisfies formula (11), the neuron will be divided into two neurons, and each parameter of the new neuron will be 0.

When the RBM training is complete, start pruning. The purpose of RBM is to extract the main feature of the input sample, that is, the activation probability of the hidden layer neurons. These features are discriminative and facilitate further applied research on the data. If the activation probability of a neuron is close to the average value for all samples, it means that the features extracted by this neuron are not discriminative, that is, redundant neurons. In order to obtain a more concise and compact network structure, these redundant neurons need to be removed. The present invention uses the standard deviation to measure the dispersion degree of the activation probability of the same hidden layer neuron to all samples, and the standard deviation formula is:

Where n=1,2,3,...,N,N is the number of input samples, j represents the jth hidden layer neuron, P(n,j) represents the jth neuron's response to the nth input sample The activation probability, μ _j represents the average activation probability of the jth neuron for all input samples. A small standard deviation means that the values are close to the mean, i.e. the features extracted by this neuron are not discriminative, so this redundant neuron needs to be removed. The pruning conditions are:

σ(j) < θ _A (14)

where θ _A is a threshold. If the jth neuron satisfies Equation (14), then remove that neuron and all its parameters. At the same time, the present invention makes a trade-off curve between the pruning rate and the prediction accuracy, and selects the value of θ _A according to this curve, so as to remove more redundant neurons while retaining the original accuracy. Furthermore, after pruning, retraining the current RBM enables the remaining neurons to compensate for the removed neurons. This step is crucial, pruning and then retraining as one iteration. Iterative pruning removes fewer neurons each time, and retrains multiple times to compensate. After many such iterations, a higher pruning rate can be found without losing accuracy. The threshold θ _A is updated at each iteration:

θ _A ←θ _A +δ[iter] (15)

The threshold in each iteration of pruning is updated by δ[iter] to remove more neurons. Each iteration is a greedy search, and according to the trade-off curve in each pruning, the optimal pruning rate can be found without loss of accuracy, so δ[iter] is set so that θ _A satisfies this iteration The pruning rate required for pruning.

Step 2.3, after the current RBM determines the network structure, it starts to consider the growth of the hidden layer. According to formula (4), it can be found that P(v; θ) is inversely proportional to E(v, h; θ). So if you want to maximize P(v; θ), then the energy function E(v, h; θ) should be as small as possible. If the total energy of the DBN is greater than a threshold, it indicates that the DBN lacks the ability to represent data, and a new layer of RBM needs to be added at this time. Therefore, the present invention uses the energy function as a condition for adding new RBMs:

where E ^l is the total energy of the lth layer of RBM, obtained by formula (2), l=1, 2...L, L is the current number of layers of DDBN, mean( ) is the average function, and θ _L is a threshold. If the energy function satisfies formula (16), a new layer of RBM is added, and the initialization of each parameter of the new RBM is the same as that of the first layer, and then the output of the current RBM is used as the input of the newly added RBM.

In step 2.4, according to steps 2.2 and 2.3, the network is cyclically trained, and a deep DDBN network structure can be learned.

Step 3. Use fine-tuning to further optimize the DDBN. The network structure and parameter values obtained in the pre-training stage are used as the initial values of the fine-tuning stage to fine-tune the entire DDBN network. The invention uses the back-propagation algorithm to fine-tune the entire network, that is, the training error is back-propagated from top to bottom, and the network is optimized to obtain the final DDBN network model. The specific operations are as follows:

Step 3.1, take the network structure and parameter value θ trained in the pre-training stage as the initial value of the fine-tuning stage, and add an output layer after the last layer of RBM. The output layer has 3 neurons, the outputs represent temperature, pressure and gas flow, respectively, and are used to predict the combustion behavior suitable for the training sample. The input training data set is optimized into the network in the fine-tuning stage.

Step 3.2, use the forward propagation algorithm to calculate the activation probability of each hidden layer neuron.

Step 3.3, calculate the prediction result obtained by the forward propagation of the training sample, and compare it with the actual result to obtain the loss function:

Among them, t is the current number of fine-tuning, N is the number of training samples, y _n and y' _n are the actual and predicted results of the nth training sample, respectively. The error of the actual result and the predicted result is back-propagated, and the weight w and the bias c are updated according to the formulas (18) and (19) using the gradient descent method:

where α is the learning rate. Iteratively uses the gradient descent method to fine-tune the entire DDBN network top-down to reduce the value of J(t) until the maximum number of fine-tuning times is reached to obtain the final DDBN network model.

Step 4. Input the preprocessed test data set into the DDBN network model obtained in the fine-tuning stage, extract the main features of the test samples through RBM, and then input these main features into the last output layer, and the output values represent The values of temperature, pressure and gas flow rate predict the combustion behavior suitable for the training sample.

The collected solid waste data set is detected below by the method provided by the present invention. The dataset includes 1000 samples, including 800 training samples and 200 testing samples. Each sample has 7 features, so the number of neurons in the visual layer is set to 7; each sample has 3 outputs, namely temperature, pressure and gas flow, corresponding to its combustion behavior.

The test results show that the intelligent treatment method of solid waste based on dynamic deep confidence network saves 30% of the treatment time compared with the traditional manual control method, and the treatment effect also reaches the solid waste treatment index stipulated by the state. Therefore, the method proposed in the present invention can effectively treat solid waste, save time and cost, and realize efficient and intelligent treatment.

Claims (5)

1. A solid waste intelligent treatment method based on a dynamic deep belief network is characterized by comprising the following steps:

step 1, measuring solid waste to obtain a solid waste data set, preprocessing the solid waste data set, and dividing to obtain a training data set and a testing data set;

step 2, inputting the training data set obtained after the preprocessing in the step 1 into a DDBN model, using a contrast divergence algorithm to train each layer of restricted Boltzmann machine RBM from bottom to top independently without supervision, optimizing the network structure of the current RBM through a dynamic increase and decrease branch algorithm in the training process, obtaining the network structure and parameter values of each RBM through iterative training, and finally obtaining the high-level characteristics of the training data set; the parameter values are weight and bias;

step 3, taking the DDBN network structure and the parameter values obtained in the step 2 as initial values of a fine tuning stage, and fine tuning the whole DDBN network by using a top-down back propagation algorithm to obtain a final DDBN network model; the specific process is as follows:

step 3.1, taking the DDBN network structure and the parameter value theta trained in the step 2 as initial values of a fine tuning stage, adding an output layer after the last layer of RBM for predicting combustion behaviors including temperature, pressure and gas flow suitable for the training data set sample, and inputting the training data set to start fine tuning the whole DDBN network;

step 3.2, calculating the activation probability of each hidden layer neuron by using a forward propagation algorithm;

step 3.3, calculating a prediction result obtained by forward propagation of the training sample, and comparing the prediction result with an actual result to obtain a loss function:

where t is the current number of fine-tuning, N is the number of samples in the training dataset, y_nAnd y'_nRespectively obtaining an actual result and a predicted result of the nth training sample; and (3) reversely propagating errors of the actual result and the predicted result, and updating the weight w and the offset c according to the formulas (2) and (3) by using a gradient descent method:

wherein α is the learning rate;

iteratively using a gradient descent method to finely adjust the whole DDBN network from top to bottom to reduce the value of J (t) until the maximum fine adjustment times are reached to obtain a final DDBN network model;

and 4, inputting the test data set into the final DDBN network model obtained in the step 3, and finally outputting a prediction result.

2. The method for intelligently processing the solid waste based on the dynamic deep belief network as claimed in claim 1, wherein the preprocessing in the step 1 is: normalizing the solid waste data set to be between [0 and 1], wherein the normalization formula is as follows:

wherein,

characteristic value, x, of solid waste data set_maxAnd x_minThe maximum value and the minimum value of all characteristics of the solid waste data set are respectively, and x is the solid waste data set after normalization.

3. The method for intelligently processing the solid waste based on the dynamic deep belief network as claimed in claim 1 or 2, wherein the specific process of the step 2 is as follows:

step 2.1, constructing a DDBN network model, and setting the parameter values of DDBN: visual layer neurons, initial hidden layer neurons, hidden layer numbers, learning rate, iteration times and fine tuning times; wherein, the number of neurons in the visual layer is the feature dimension of the training data set;

step 2.2, inputting the training data set obtained after preprocessing into a first layer RBM, pre-training the RBM by using a CD algorithm, and optimizing the network structure of the current RBM by using a dynamic branch-increasing and-decreasing algorithm in the training process;

(1) the energy function E (v, h; θ) of the RBM and the joint probability distribution P (v, h; θ) of the visible and hidden layer neurons are:

wherein v is_i(1. ltoreq. I. ltoreq.I) and h_j(J is more than or equal to 1 and less than or equal to J) respectively represents a visual layer neuron and a hidden layer neuron, w is a weight matrix between the visual layer and the hidden layer, b and c are the bias of the visual layer neuron and the hidden layer neuron respectively, and theta represents a parameter in the model, namely theta is { w, b, c }; z is for all possible visible and hidden layersSumming the channel pairs;

and (3) solving the edge probability distribution of the visual layer neuron v and the hidden layer neuron h according to a formula (6) by utilizing the principle of a Bayes formula:

and deducing the conditional probability distribution of the visual layer neuron v and the hidden layer neuron h by using a Bayesian formula:

obtaining approximate reconstruction P (v; theta) of a training sample by using a contrast divergence algorithm through one-step Gibbs sampling by using a formula (9) and a formula (10), and then updating a network parameter theta to be { w, b, c } according to a reconstruction error;

(2) in the training process, according to the current training condition, optimizing the network structure of the RBM through a dynamic branch increasing and decreasing algorithm;

the change in weight w is monitored using the weight distance WD method:

WD_j[m]＝Met(w_j[m]，w_j[m-1]) (11)

wherein, w_j[m]The weight vector of the hidden layer neuron j after m iterations, wherein Met represents a measurement function, such as Euclidean distance; the value of WD reflects the change in the weight vector of hidden layer neuron j in two iterations;

the branch increasing condition is considered from the aspects of local and global;

the local conditions are defined as:

wherein

In the mth iteration, the WD value of the jth hidden neuron to the nth input sample, J is 1,2,3, J is the number of hidden neurons, and max (·) is a maximum function;

the global condition is defined as:

where N is the number of samples in the training data set and N' is the number of samples that increased the WD value of the jth neuron compared to the last iteration, i.e., the number of samples

The local and global conditions are considered for a single input sample and all input samples by the hidden layer neuron j, respectively; multiplying the two conditions to obtain the propagation condition:

max_WD_j[m]*iratio_j(m)＞y(m) (14)

where y (m) is a curve, which is used as a variable threshold, defined as:

where m is the current iteration number, numepochs is the maximum iteration number, u represents the curvature of the curve, y_maxAnd y_minMaximum and minimum values of the curve, respectively; when the jth neuron satisfies equation (14), then the neuron will be divided into two neurons, and each parameter of the new neuron will beThe numbers are all 0;

when the RBM training is completed, branch reduction is started: and (3) using the standard deviation of the hidden layer neurons on the activation probability of all samples as a branch-reducing condition, wherein the standard deviation formula is as follows:

where N is 1,2,3, …, N is the number of samples in the input training dataset, j represents the jth hidden layer neuron, P (N, j) represents the activation probability of the jth neuron on the nth input sample, μ_jRepresenting the average activation probability of the jth neuron on all input samples;

the branch reducing conditions are as follows:

σ(j)＜θ_A (17)

wherein theta is_AIs a threshold value; when the jth neuron satisfies formula (17), removing the neuron and all parameters thereof; at the same time, a trade-off curve relating to the ratio of branch reduction and the prediction accuracy is made, and theta is selected according to the curve_AA value of (a) such that the original accuracy is preserved while removing more redundant neurons;

after branch reduction, retraining the RBM to enable the remaining neurons to compensate the removed neurons, and retraining after branch reduction into one iteration; at each iteration we update the threshold θ_A：

θ_A←θ_A+δ[iter] (18)

By delta [ iter ]]To update the threshold in each iteration of the pruning to remove more neurons; each iteration is a greedy search, and according to a balance curve in each branch reduction, the optimal branch reduction rate can be found without losing the accuracy rate, so that the delta [ iter ]]Is set so that_AThe branch reducing rate required by the iterative branch reducing is met;

step 2.3, after the current RBM determines the network structure, using an energy function as a condition for adding a new RBM:

wherein E^lIs the total energy of the L-th layer RBM, which is found by equation (5), L is 1,2 … L, L is the current layer number of the DDBN, mean () is the average function, θ_LIs a threshold value; when the energy function meets the formula (19), a new layer of RBM is added, and the initialization of each parameter of the new RBM is the same as that of the first layer; then taking the output of the current RBM as the input of the newly added RBM;

and 2.4, training the network circularly according to the steps 2.2 and 2.3 to obtain the network structure of the DDBN.

4. The method for intelligently processing the solid waste based on the dynamic deep belief network as claimed in claim 1 or 2, wherein the specific process of the step 4 is as follows:

step 4.1, inputting the preprocessed test data set into the DDBN network model finely adjusted in the step 3, and extracting the main characteristics of the solid waste through RBM;

and 4.2, inputting the main characteristics of the test sample into the last output layer, and predicting the combustion behavior suitable for the training sample, including temperature, pressure and gas flow.

5. The method for intelligently processing the solid waste based on the dynamic deep belief network as claimed in claim 3, wherein the specific process of the step 4 is as follows:

step 4.1, inputting the preprocessed test data set into the DDBN network model finely adjusted in the step 3, and extracting the main characteristics of the solid waste through RBM;

and 4.2, inputting the main characteristics of the test sample into the last output layer, and predicting the combustion behavior suitable for the training sample, including temperature, pressure and gas flow.

Images