Vibrating screen screening efficiency optimization method based on H-RPCA-GRU and CNN-LSTM-Attention coupling model

Abstract

The invention relates to a method for optimizing screening efficiency of a vibrating screen. Acceleration vectors of the vibrating screen under different process parameters are acquired for multiple times; performing H-RPCA de-noising on the acceleration vector; training the GRU network by taking the process parameters as input data and the acceleration vectors as labels; acquiring different process parameters and corresponding screening efficiency of the vibrating screen in real time, and predicting the GRU network to obtain an acceleration vector; training a CNN-LSTM-Attention model by taking the predicted acceleration vector as input data and the screening efficiency as a label; preset process parameters are collected and input into the trained GRU network to predict an acceleration vector; and inputting the predicted acceleration vector into the trained CNN-LSTM-Attention coupling model to predict the screening efficiency, generating a genetic algorithm fitness function according to the predicted screening efficiency, energy consumption and production cost, and optimizing process parameters.

Description

Technical Field

[0001] This invention relates to the field of intelligent industrial screening and process optimization technology, and in particular to a method for optimizing the screening efficiency of a vibrating screen based on a coupled model of H-RPCA-GRU and CNN-LSTM-Attention. Background Technology

[0002] In modern industrial production, material screening is a crucial step in raw material pretreatment, product grading, and quality control. Its efficiency and accuracy directly impact the overall production system's capacity, energy consumption, product quality, and economic benefits. Vibrating screens, as core equipment for efficient material grading and screening, are widely used in mining, coal, metallurgy, chemical, building materials, grain, and environmental protection industries. With the expansion of industrial production scale and the increasing complexity of processes, achieving accurate prediction and intelligent optimization of vibrating screen screening efficiency has become a vital issue for enhancing the level of intelligent manufacturing and core competitiveness in industrial production.

[0003] For a long time, the prediction and optimization of screening efficiency of vibrating screening equipment has mainly relied on the following methods: (1) Empirical method and semi-empirical formula: By summarizing a large amount of production practice experience, empirical formulas or charts are established to guide the selection and adjustment of screening parameters. This type of method relies on the experience accumulation of operators, is difficult to adapt to new equipment, complex materials or variable working conditions, and lacks theoretical support, resulting in limited prediction accuracy. (2) Static modeling method: Based on physical mechanisms, mechanical analysis or statistical regression, a static relationship model between process parameters and screening efficiency is established, such as multiple linear regression, response surface methodology, etc. This type of method usually assumes stable working conditions and constant parameters, and is difficult to reflect the dynamic, nonlinear, and strongly coupled relationship between process parameters and screening efficiency in actual production. (3) Traditional machine learning method: In recent years, some studies have attempted to use traditional machine learning algorithms such as support vector machines, decision trees, and random forests to model and predict screening efficiency. Although these methods have improved the modeling flexibility to a certain extent, their feature extraction capabilities are limited, making it difficult to fully explore the deep-seated correlation between complex process parameters and screening efficiency, and their ability to process dynamic signal data is insufficient.

[0004] Traditional methods often simplify the relationship between process parameters and screening efficiency into static, linear, or weakly nonlinear models, making it difficult to capture the dynamic impact of parameter changes on screening efficiency under actual operating conditions. For example, a small change in excitation force can lead to abrupt changes in the screen surface motion pattern, thereby affecting material distribution and screening efficiency; fluctuations in feed rate can cause changes in material layer thickness, material flowability, and other aspects, thus affecting the stability and efficiency of the screening process. Static modeling methods cannot reflect these complex dynamic processes, resulting in insufficient model prediction accuracy and difficulty in providing effective support for intelligent adjustment and optimization.

[0005] With the development of sensor technology, vibration signals, as key dynamic parameters reflecting equipment operating status, structural health, and material flow characteristics, have been widely used in equipment condition monitoring, fault diagnosis, and performance evaluation. By analyzing vibration signals in the time domain, frequency domain, and time-frequency domain, abnormal equipment conditions (such as loosening, cracks, imbalance, and blockage) can be identified, enabling real-time monitoring of equipment health.

[0006] However, in the field of screening efficiency prediction and optimization, installing vibration signal sensors on every vibrating screen would inevitably increase costs. Therefore, there is an urgent need for a low-cost and high-efficiency method for optimizing the screening efficiency of vibrating screens. Summary of the Invention

[0007] Based on the above analysis, the present invention aims to provide a method for optimizing the screening efficiency of a vibrating screen based on a coupled model of H-RPCA-GRU and CNN-LSTM-Attention, in order to solve the problem that there is an urgent need in the field of vibrating screening for a low-cost and high-efficiency method for optimizing the screening efficiency of a vibrating screen.

[0008] This invention provides a method for optimizing the screening efficiency of a vibrating screen based on a coupled H-RPCA-GRU and CNN-LSTM-Attention model. The method involves repeatedly acquiring acceleration vectors from accelerometers installed at multiple locations on the vibrating screen under different process parameters. These process parameters include excitation force, excitation frequency, excitation direction angle, screen inclination angle α, and feed rate Q. The acceleration vectors are then denoised using H-RPCA to obtain a denoised acceleration vector. The process parameters are preprocessed and concatenated into a vector, and the denoised acceleration vector is used as a label to form a first training sample set. The GRU network is then trained based on this first training sample set.

[0009] The different process parameters and corresponding screening efficiencies of the vibrating screen are collected in real time. The process parameters are preprocessed and concatenated into vectors, which are then input into a trained GRU network. The GRU network predicts the acceleration vectors at multiple corresponding positions. A second training sample set is generated based on the predicted acceleration vectors at each position and the screening efficiency. The CNN-LSTM-Attention coupled model is trained based on the second training sample set.

[0010] After collecting and preprocessing the preset process parameters, they are concatenated into vectors and input into a trained GRU network to predict the acceleration vectors at each position. The predicted acceleration vectors at each position are then input into a trained CNN-LSTM-Attention coupled model to predict the screening efficiency. Based on the predicted screening efficiency, energy consumption, and production cost, a genetic algorithm fitness function is generated to optimize the process parameters.

[0011] Furthermore, the H-RPCA denoising includes: performing sliding window segmentation on the acceleration vectors collected from acceleration sensors installed at multiple locations on the vibrating screen to obtain the Hankel matrix; iterating the Hankel matrix using the nonholonomic augmented Lagrange multiplier method until the iteration process converges, generating a low-rank matrix L of the Hankel matrix under convergence conditions; and obtaining the denoised acceleration vector based on the elements in the low-rank matrix L of the Hankel matrix under convergence conditions.

[0012] Furthermore, the acceleration vectors collected from multiple locations on the vibrating screen are divided into a sliding window to obtain the Hankel matrix, which includes: the window size is m, the step size is 1, and the collected data of each window is used as the column vector of the Hankel matrix; wherein, m is 1.3-1.6 times the period length of the elements in the acceleration vector.

[0013] Furthermore, the calculation formula for iteratively applying the nonholonomic augmented Lagrange multiplier method to the Hankel matrix includes:

[0014]

[0015] Y k+1 =Y k +μ k (HL k +S k+1 ),

[0016] in, The singular value soft thresholding operator performs singular value decomposition on the input matrix and max(σ) on the singular values. i After -τ,0), the matrix containing the singular values ​​is replaced; σ i For the i-th singular value, H is the input Hankel matrix;

[0017] This is an element-level soft thresholding operator used to threshold each element 'a' in the matrix within parentheses. ij Perform the following calculations:

[0018]

[0019] Where sgn() is the symbolic function;

[0020] μ k To increment the penalty parameter, μ k+1 =ρμ k ρ = 1.5~1.8;

[0021] Y k S is the Lagrange operator matrix for the k-th iteration; kLet L be the sparse matrix in the k-th iteration. k Let be the low-rank matrix in the k-th iteration.

[0022] Furthermore, based on the elements of the low-rank matrix L of the Hankel matrix under the aforementioned convergence condition, the denoised acceleration vector is obtained, including calculating the t-th element in the denoised acceleration vector using the following formula:

[0023]

[0024] Where N is the number of elements in the denoised acceleration vector; L i,j Let n be the element in the i-th row and j-th column of the low-rank matrix L of the Hankel matrix under the convergence condition; the line connecting the positions of each element in each average calculation is parallel to the anti-diagonal of the m×m submatrix contained in the low-rank matrix L, where m is the row number of the low-rank matrix L of the Hankel matrix, and n is the number of rows. t Let t be the number of elements used to calculate the average in the t-th iteration.

[0025] Furthermore, a fitness function for the genetic algorithm is generated based on the predicted screening efficiency, energy consumption, and production cost. The formula for calculating the fitness function is as follows:

[0026] Fitness = w1 (t) ·S E (E)+w2 (t) ·S C (C)+w3 (t) ·S P (P),

[0027] Among them, S E (E) is the screening efficiency correlation function, S C (C) is the energy consumption normalized score function, S P (P) is the normalized score function for production costs, w1 (t) w2 (t) w3 (t) w1 is the weight that is dynamically adjusted during the t-th generation of evolution. (t) +w2 (t) +w3 (t) =1.

[0028] Furthermore, the formula for calculating the screening efficiency correlation function is as follows:

[0029] S E (E)=A E ,

[0030] Where E is the normalized screening efficiency, and A ranges from [2, 5.5].

[0031] The formula for calculating the energy consumption normalized score function is:

[0032] S C (C) = log a1 C,

[0033] Where C is the normalized energy consumption, and a1 ranges from [0.1, 1 / 3];

[0034] The formula for calculating the normalized score function of production costs is:

[0035] S P (P) = log a2 P,

[0036] Where P is the normalized generation cost, and a2 ranges from [0.5, 2 / 3].

[0037] Furthermore, the production cost includes material costs due to multiple screenings, machine wear costs under different excitation forces, excitation frequencies, excitation direction angles, screen inclination angles α, and feed rates Q, and labor costs.

[0038] Furthermore, the formula for internal information transmission within a GRU network is:

[0039] r t =σ(W r ·[h t-1 ,x t ]),

[0040] z t =σ(W z ·[h t-1 ,x t ]),

[0041]

[0042] Where, x t h is a vector composed of preprocessed excitation force, excitation frequency, excitation direction angle, screen inclination angle α, and feed rate Q. t-1 Here is the state vector of the hidden layer, [] represents vector concatenation, and r t and z t , respectively, are the outputs of the reset gate and the update gate; W is the weight matrix of the hidden states outside the reset gate and the update gate; W r and W z are the weight matrices for the reset gate and the update gate, respectively; tanh is the state activation function; and h t σ represents the candidate acceleration output and the predicted acceleration output at time t, respectively, and σ is the activation function of the reset gate and update gate.

[0043] Furthermore, the output of the last hidden layer state information in the LSTM model is used as the input of the Attention model, and the Attention model outputs the predicted screening efficiency.

[0044] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0045] 1. This invention uses a GRU network to predict the acceleration at different positions of the vibrating screen; the predicted acceleration signal is input into the CNN-LSTM-Attention coupled model to predict the screening efficiency, which takes into account the influence of dynamic data on screening efficiency and avoids the installation of multiple acceleration sensors (acceleration sensors are only installed during training to collect acceleration data as labels), and the screening efficiency prediction results are accurate.

[0046] 2. The vibrating screen screening efficiency optimization method of the present invention comprehensively considers screening efficiency, energy consumption, and production cost factors. According to the importance of screening efficiency, energy consumption, and production cost, their respective indexes and logarithmic weight coefficient calculation formulas are set, so that screening efficiency is the most important, and its influence increases exponentially, while production cost is the second most important but its influence is higher than that of energy consumption. The vibrating screen efficiency is reasonably optimized through this fitness function.

[0047] 3. The vibrating screen screening efficiency optimization method of the present invention denoises the acceleration during GRU training by employing an H-RPCA denoising method based on sliding window segmentation. The denoising m is 1.3-1.6 times the period length of the elements in the acceleration vector, covering the entire period and extracting complete contextual information without obtaining excessive redundant information. This method iterates the Hankel matrix using the nonholonomic augmented Lagrange multiplier method until the iteration process converges, generating a low-rank matrix L of the Hankel matrix under convergence conditions. The denoised acceleration vector is then obtained based on the elements of the low-rank matrix L of the Hankel matrix under convergence conditions. This method demonstrates significant denoising effectiveness, thereby making the prediction of screening efficiency more accurate.

[0048] 4. The method for optimizing the screening efficiency of the vibrating screen in this invention uses the output of the last hidden layer state information in the LSTM model as the input of the Attention model, and the Attention model outputs the predicted screening efficiency.

[0049] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0050] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0051] Figure 1 This is a flowchart of a method for optimizing the screening efficiency of a vibrating screen based on a coupled H-RPCA-GRU and CNN-LSTM-Attention model, according to the present invention.

[0052] Figure 2 This is a schematic diagram of the LSTM structure of the present invention;

[0053] Figure 3 This is a flowchart of the screening efficiency optimization method of the present invention. Detailed Implementation

[0054] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0055] A specific embodiment of the present invention discloses a method for optimizing the screening efficiency of a vibrating screen based on a coupled model of H-RPCA-GRU and CNN-LSTM-Attention, such as... Figure 1 As shown. Specifically, it includes steps S1-S4.

[0056] Step S1: Obtain the acceleration vectors of acceleration sensors installed at multiple locations on the vibrating screen under different process parameters. The process parameters include excitation force, excitation frequency, excitation direction angle, screen surface inclination angle α, and feed rate Q. The acceleration vectors are denoised using H-RPCA to obtain denoised acceleration vectors. The process parameters are preprocessed and then concatenated into vectors. The denoised acceleration vectors are used as labels to form the first training sample set.

[0057] Specifically, the data acquisition system employs a high-precision A / D conversion module, with a sampling frequency preferably between 1kHz and 10kHz to ensure the timeliness and accuracy of the parameter data. Preferably, the process parameter acquisition points can be flexibly configured according to actual needs to adapt to different types of vibrating screens and varying operating conditions. The acquired parameters can be represented as vectors:

[0058] P = [F, f, θ, α, Q]

[0059] Where F is the excitation force, f is the excitation frequency, θ is the excitation direction angle, α is the screen surface inclination angle, and Q is the feeding rate.

[0060] The vibration acceleration signal acquisition uses a high-precision triaxial accelerometer, which is installed in key parts such as the screen box, screen mesh, and support to acquire vibration acceleration signals in real time.

[0061] The raw signal collected by each sensor can be represented as:

[0062] a i (t), i = 1, 2, ..., N

[0063] Where N is the number of sensors, a i (t) represents the acceleration value of the i-th sensor at time t. The sensor is connected to the edge computing unit via an industrial data acquisition card, and the acquisition frequency is preferably 2kHz to 5kHz to ensure the integrity and representativeness of the signal. Preferably, the number and location of vibration signal acquisition points can be flexibly adjusted according to actual needs.

[0064] Data synchronization and preprocessing. To ensure the timing consistency between process parameters and vibration signals, the system employs a unified clock synchronization mechanism to align the timestamps of multi-source data. Preprocessing operations such as denoising, normalization, and outlier removal are performed on the acquired raw signals. Preferably, the Z-score normalization method is used to normalize each parameter.

[0065]

[0066] Where x is the original data, μ is the mean, and σ is the standard deviation.

[0067] The H-RPCA denoising process includes: obtaining a Hankel matrix by sliding window segmentation of acceleration vectors collected from acceleration sensors installed at multiple locations on the vibrating screen; iterating the Hankel matrix using the nonholonomic augmented Lagrange multiplier method until the iteration process converges, generating a low-rank matrix L of the Hankel matrix under convergence conditions; and obtaining the denoised acceleration vector based on the elements of the low-rank matrix L of the Hankel matrix under convergence conditions.

[0068] The Hankel matrix is ​​obtained by sliding window segmentation of acceleration vectors collected from acceleration sensors installed at multiple locations on the vibrating screen. The window size is m, the step size is 1, and the collected data of each window is used as the column vector of the Hankel matrix. Wherein, m is 1.3-1.6 times the period length of the elements in the acceleration vector.

[0069] Specifically, the acceleration data collected from the acceleration sensors installed at various locations in the sieve box, sieve mesh, and support all have a length of N. The sliding window segmentation yields a Hankel matrix consisting of m rows and N-m+1 columns.

[0070] The formula for iteratively calculating the Hankel matrix using the nonholonomic augmented Lagrange multiplier method includes:

[0071]

[0072] Y k+1 =Y k +μ k (HL k +S k+1 ),

[0073] in, The singular value soft thresholding operator performs singular value decomposition on the input matrix and max(σ) on the singular values. i After -τ,0), the matrix containing the singular values ​​is replaced; σ i For the i-th singular value, H is the input Hankel matrix;

[0074] This is an element-level soft thresholding operator used to threshold each element 'a' in the matrix within parentheses. ij Perform the following calculations:

[0075]

[0076] Where sgn() is the symbolic function;

[0077] μ k To increment the penalty parameter, μ k+1 =ρμ k ρ = 1.5~1.8;

[0078] Y k S is the Lagrange operator matrix for the k-th iteration; k Let L be the sparse matrix in the k-th iteration. k Let be the low-rank matrix in the k-th iteration.

[0079] The denoised acceleration vector is obtained from the elements of the low-rank matrix L of the Hankel matrix under the aforementioned convergence condition, including calculating the t-th element in the denoised acceleration vector using the following formula:

[0080]

[0081] Where N is the number of elements in the denoised acceleration vector; L i,j Let n be the element in the i-th row and j-th column of the low-rank matrix L of the Hankel matrix under the convergence condition; the line connecting the positions of each element in each average calculation is parallel to the anti-diagonal of the m×m submatrix contained in the low-rank matrix L, where m is the row number of the low-rank matrix L of the Hankel matrix, and n is the number of rows. tLet t be the number of elements used to calculate the average in the t-th iteration.

[0082] For each m×m submatrix contained in the low-rank matrix L of the Hankel matrix under the convergence condition, the average value of the anti-diagonal elements is obtained sequentially. In the first m×m submatrix from the left in the low-rank matrix L, the average value of the elements in the matrix covered by the lines whose anti-diagonal lines are successively shifted to the left by 1 to m-1 element distances is also required. In the rightmost m×m submatrix of the low-rank matrix L, the average value of the elements in the matrix covered by each shifted line whose anti-diagonal lines are successively shifted to the right by 1 to m-1 element distances is also required. The average values ​​of all the elements covered by the lines are obtained and arranged in order from left to right to form the denoised acceleration vector, where m is the row number of the low-rank matrix L of the Hankel matrix.

[0083] Step S2: Train the GRU network based on the first training sample set.

[0084] F is the excitation force, f is the excitation frequency, θ is the excitation direction angle, α is the screen surface inclination angle, and Q is the normalized feed rate vector. This vector is then input into a gated recurrent unit (GRU) network to extract time-dependent features.

[0085] The formula for internal information transmission in a GRU network is:

[0086] r t =σ(W r ·[h t-1 ,x t ]),

[0087] z t =σ(W z ·[h t-1 ,x t ]),

[0088]

[0089] Where, x t h is a vector composed of preprocessed excitation force, excitation frequency, excitation direction angle, screen inclination angle α, and feed rate Q. t-1 Here is the state vector of the hidden layer, [] represents vector concatenation, and r t and z t , respectively, are the outputs of the reset gate and the update gate; W is the weight matrix of the hidden states outside the reset gate and the update gate; W r and W z are the weight matrices for the reset gate and the update gate, respectively; tanh is the state activation function; and h t σ represents the candidate acceleration output and the predicted acceleration output at time t, respectively, and σ is the activation function of the reset gate and update gate.

[0090] The Particle Swarm Optimization (PSO) algorithm is used to train the internal hyperparameters of the GRU network. During training, the PSO algorithm is used to determine the hidden state dimension and learning rate of the GRU network, including steps S21-S25.

[0091] Step S21: Define the particle swarm.

[0092] Set the particle dimension to 2, which represents the hidden layer dimension and learning rate of GRU. Choose MAE as the objective function. The smaller the MAE, the closer the predicted value is to the true value.

[0093] Step S22: Initialize the particle swarm.

[0094] Initialize two particles in the particle swarm, and randomly select a hidden layer dimension (32 to 512) and a learning rate (0.0001 to 0.1).

[0095] Step S23: Evaluate the fitness of each particle.

[0096] The GRU model is trained using hyperparameters corresponding to two particle combinations. Cross-validation is used during training to evaluate the model's performance, and the predicted acceleration and the error of actual acceleration (MAE) are returned.

[0097] Step S24: Update the position and velocity of the particles.

[0098] The particle swarm optimization algorithm has two main update rules: position update and velocity update.

[0099] The speed update formula is:

[0100]

[0101] The position update formula is:

[0102]

[0103] in: It is the velocity of particle i in wheel t. It is the position of particle i in wheel t (i.e., the hyperparameter). This is the optimal position of particle i on wheel t. t represents the global optimal position of the entire particle swarm in round t, w is the inertial weight, which controls the degree to which particles approach their individual and global optimal positions, and r1 and r2 are random numbers between 0 and 1.

[0104] The velocity and position of each particle are updated according to the above formula. The particle's goal is to minimize the objective function MAE, so the particle will move towards the direction of the optimal hyperparameter.

[0105] Different fitness function (MAE) values ​​are generated based on the different positions of the particles. The MAE values ​​generated by different particle positions are compared, and the minimum MAE value is selected. This yields the trained GRU network.

[0106] Step S25, Termination Condition.

[0107] The returned error MAE decreased to the threshold ∈ ≤ 0.01.

[0108] The positions of the two particles after termination are the corresponding optimal hyperparameters, which are the optimal combination of hidden layer dimension and learning rate.

[0109] Step S3: Collect different process parameters and corresponding screening efficiencies of the vibrating screen in real time, preprocess the process parameters and concatenate them into vectors, input them into the trained GRU network, and the GRU network predicts the acceleration vectors at multiple corresponding positions; generate a second training sample set based on the predicted acceleration vectors at each position and the screening efficiency; train the CNN-LSTM-Attention coupled model based on the second training sample set.

[0110] During training, a genetic optimization algorithm is used to determine the number of convolutional kernels and layers in the CNN network, the number of layers and the hidden state dimension of each layer in the LSTM network, and the learning rate of the CNN-LSTM-Attention coupled model.

[0111] The output of the last hidden layer state information in the LSTM model is used as the input of the Attention model, and the Attention model outputs the screening efficiency of the prediction.

[0112] CNN spatial feature extraction includes: performing convolution operations on temporal features to extract spatial structure features. The output of the convolutional layer is:

[0113] y = f(W*x + b)

[0114] Where * represents the convolution operation, W is the convolution kernel, x is the input matrix, b is the bias, and f is the activation function.

[0115] Long Short-Term Memory (LSTM) networks are employed to further model long-term dependencies, and an attention mechanism is introduced to weight key features. A schematic diagram of the LSTM structure of this invention is shown below. Figure 2 As shown.

[0116] LSTM introduces three gating units: the input gate, the forget gate, and the output gate. These gates regulate the flow of information and determine which information should be retained or discarded. The operation of these gates helps to learn long-term dependencies in the input sequence.

[0117] The formula for calculating the input gate is as follows:

[0118] i t =σ(W t ·[h t-1 ,x t ]+b t ),

[0119] The formula for calculating the forgetting gate is as follows:

[0120] f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0121] The formula for calculating the updated cell state is as follows:

[0122] C t =f t ×C t-1 +i t ×tanh(W C ·[h t-1 ,x t ]+b C C t ,

[0123] The formula for calculating the output gate is as follows:

[0124] O t =σ(W O ·[h t-1 ,x t ]+b O ),

[0125] The formula for calculating the hidden layer is as follows:

[0126] h t =O t ×tanh(C t ),

[0127] W t W f W c and W o These are the weight matrices for the input gate, forget gate, cell state, and output gate, respectively. t b f b C and b O These represent the bias terms of the weight matrices for the input gate, forget gate, cell state, and output gate, respectively; σ represents the sigmoid function; and C... t The vector representing the cell state, h t x represents the output of the LSTM unit at the current time step. t This represents the input at time t.

[0128] The formula for calculating the Attention weights is:

[0129]

[0130] e t =v T tanh(W h h t '+b h ),

[0131] Among them, a t It is the final attention allocation weight; e t It is the relevance score of the t-th input, used to measure the relevance of the output to the current input h. t The degree of relevance; v T This is a weight vector used to map the tanh-activated vector to scalar scores; tanh() is the hyperbolic tangent activation function; b h It is a bias vector; exp() is the exponentiation operation; W h It is a weight matrix.

[0132] The specific steps are as follows: For each input h t Perform a linear transformation to obtain W. h h t '+b h Applying tanh activation to the linear transformation result yields tanh(W). h h t '+b h ); using the weight vector v T The inner product of the activation results yields the relevance score e. t For all e t Perform exponential operations and normalization to obtain the attention weight α. t .

[0133] Screening efficiency prediction output. High-precision screening efficiency prediction results are output through the fully connected layer, providing a basis for intelligent adjustment and optimization decisions in the production line. The prediction output is:

[0134]

[0135] Among them, h att For Attention-weighted features, f FC This is a mapping for fully connected layers.

[0136] If the predicted results do not meet expectations, the system can automatically provide feedback and optimize process parameters to achieve adaptive intelligent adjustment of the screening system.

[0137] like Figure 3As shown, the deep learning coupled modeling structure of this invention includes a GRU temporal modeling module, a CNN spatial feature extraction module, an LSTM temporal modeling module, an Attention key feature focusing module, and an output layer. Multi-source feature samples are first input into the GRU module to extract temporally dependent features; subsequently, the CNN module extracts the spatial structure of the features, identifying key patterns and local changes; the LSTM module further models long-term dependencies, improving the ability to model complex dynamic processes; the Attention module assigns weights to features at different time steps, focusing on key dynamic information affecting screening efficiency. Finally, the output layer generates a high-precision screening efficiency prediction result. This structure achieves deep fusion and collaborative optimization of multi-source information, significantly improving the accuracy of prediction and the intelligence level of the system.

[0138] Step S4: Collect the preset process parameters, preprocess them, and then concatenate them into vectors. Input the vectors into the trained GRU network to predict the acceleration vectors at each position. Input the predicted acceleration vectors at each position into the trained CNN-LSTM-Attention coupled model to predict the screening efficiency. Genetic algorithm fitness function is generated based on the predicted screening efficiency, energy consumption, and production cost to optimize the process parameters.

[0139] Based on the predicted screening efficiency, energy consumption, and production cost, a genetic algorithm fitness function is generated. The formula for calculating the fitness function is as follows:

[0140] Fitness = w1 (t) ·S E (E)+w2 (t) ·S C (C)+w3 (t) ·S P (P),

[0141] Among them, S E (E) is the screening efficiency correlation function, S C (C) is the energy consumption normalized score function, S P (P) is the normalized score function for production costs, w1 (t) w2 (t) w3 (t) w1 is the weight that is dynamically adjusted during the t-th generation of evolution. (t) +w2 (t) +w3 (t) =1.

[0142] The formula for calculating the screening efficiency correlation function is as follows:

[0143] S E (E)=A E ,

[0144] Where E is the normalized screening efficiency, and A ranges from [2, 5.5].

[0145] The formula for calculating the energy consumption normalized score function is:

[0146] S C (C) = log a1 C,

[0147] Where C is the normalized energy consumption, and a1 ranges from [0.1, 1 / 3];

[0148] The formula for calculating the normalized score function of production costs is:

[0149] S P (P) = log a2 P,

[0150] Where P is the normalized generation cost, and a2 ranges from [0.5, 2 / 3].

[0151] Specifically, screening efficiency is calculated using an exponential function; an increase in screening efficiency leads to a rapid increase in the calculated result of the screening efficiency-related function. C represents the normalized energy consumption; as energy consumption increases, the fitness function decreases. Similarly, as production costs increase, the fitness function decreases. However, considering that production costs are more important to manufacturers than energy consumption, the fitness function should decay rapidly as production costs increase; conversely, energy consumption has a smaller impact on manufacturers than production costs, so the fitness function should decay slowly as energy consumption increases.

[0152] The production costs include material costs due to multiple screenings, machine wear costs under different excitation forces, excitation frequencies, excitation direction angles, screen inclination angles α, and feed rates Q, and labor costs.

[0153] The input process parameters are optimized using a genetic algorithm, and the optimized parameters are used as the control parameters of the vibrating screen. Specifically, this includes:

[0154] A1: Based on the predicted screening efficiency, sample data with the same model as the current vibrating screen and a screening efficiency greater than the predicted screening efficiency are selected from the first training sample set to form a first similar data set. The process parameters to be tested and the input parameter data of each sample in the first similar data set are used as individuals in the initial population, wherein: the excitation force F, excitation frequency f, excitation direction angle γ, screen surface inclination angle α, and feed rate Q are sequentially concatenated in real number form to form a gene string of length 5 [F,f,γ,α,Q]. Each gene is set with upper and lower limits according to the equipment safety limits and production permit range; values ​​exceeding the limits are reduced to the legal range by boundary truncation in subsequent operations.

[0155] A2: Determine if the current initial population size has reached the preset number of individuals Np. If it has, proceed to step A4; otherwise, execute step A3.

[0156] A3: Using a random generation algorithm, new individuals are generated by uniformly sampling values ​​within the allowable range of each input process parameter until the population size meets the preset number of individuals Np.

[0157] A4: For each individual in the initial population, perform the following operations in sequence to obtain its fitness function value: (1) Predict the screening efficiency SE corresponding to the individual using the trained H-RPCA-GRU and CNN-LSTM-Attention coupled model. (2) Calculate the energy consumption SC based on the real-time energy consumption of the equipment, and calculate the production cost SP. (3) Substitute into the fitness evaluation function.

[0158] Fitness = w1 (t) ·S E (E)+w2 (t) ·S C (C)+w3 (t) ·S P (P),

[0159] Wherein: S E (E)=A E E is the normalized screening efficiency, A ranges from [2, 5.5], and S... C (C) = log a1 C, where C is the normalized energy consumption, a1 ranges from [0.1, 1 / 3], S P (P) = log a1 P, where P is the normalized energy consumption, a2 ranges from [0.5, 2 / 3], and w1 (t) w2 (t) w3 (t) w1 is the weight that is dynamically adjusted during the t-th generation of evolution. (t) +w2 (t) +w3 (t) =1.

[0160] A5: Use the tournament selection algorithm to select individuals with high fitness function values ​​from the current population to enter the next generation.

[0161] A6: Perform crossover and mutation operations on the population entering the next generation to produce offspring.

[0162] A7: Determine whether the current iteration meets the preset termination condition. If it does, end the iteration; otherwise, return to A5.

[0163] A8: After the iteration is completed, select the individual with the largest fitness function value from the final population, and use its corresponding excitation force F*, excitation frequency f*, excitation direction angle γ*, screen surface inclination angle α*, and feed rate Q* as optimized parameters, and set them as the control parameters of the vibrating screen to be tested for actual operation.

[0164] Compared with existing technologies, the screening efficiency optimization method provided in this embodiment uses a GRU network to predict the acceleration at different positions of the vibrating screen; the predicted acceleration signal is input into a CNN-LSTM-Attention coupled model to predict screening efficiency, which takes into account the influence of dynamic data on screening efficiency and avoids the need to install multiple acceleration sensors (acceleration sensors are only installed during training to collect acceleration data as labels), resulting in accurate screening efficiency prediction results. The screening efficiency optimization method for vibrating screens provided in this embodiment comprehensively considers screening efficiency, energy consumption, and production cost factors. It sets exponential and logarithmic weight coefficient calculation formulas according to the importance of screening efficiency, energy consumption, and production cost, making screening efficiency the most important factor with an exponentially increasing influence, followed by production cost, but with a higher influence than energy consumption. This fitness function is used to reasonably optimize the vibrating screen efficiency. The screening efficiency optimization method provided in this embodiment denoises the acceleration during GRU training using an H-RPCA denoising method based on sliding window segmentation. The denoising m is 1.3-1.6 times the period length of the elements in the acceleration vector, covering the entire period and extracting complete contextual information without obtaining excessive redundant information. This method iterates the Hankel matrix using the nonholonomic augmented Lagrange multiplier method until the iteration process converges, generating a low-rank matrix L of the Hankel matrix under convergence conditions. The denoised acceleration vector is then obtained based on the elements of the low-rank matrix L of the Hankel matrix under convergence conditions. This method achieves significant denoising, resulting in more accurate predictions of screening efficiency. The vibration screen screening efficiency optimization method provided in this embodiment uses the output of the last hidden layer state information in the LSTM model as the input to the Attention model, which outputs the predicted screening efficiency.

[0165] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0166] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for optimizing the screening efficiency of a vibrating screen based on a coupled model of H-RPCA-GRU and CNN-LSTM-Attention, characterized in that, Acceleration vectors from accelerometers installed at multiple locations on the vibrating screen are acquired multiple times under different process parameters. The process parameters include excitation force, excitation frequency, excitation direction angle, screen surface inclination angle α, and feed rate Q. The acceleration vectors are then denoised using H-RPCA to obtain denoised acceleration vectors. The process parameters are preprocessed and then concatenated into vectors. The denoised acceleration vectors are used as labels to form the first training sample set. The GRU network is trained based on the first training sample set; The different process parameters and corresponding screening efficiencies of the vibrating screen are collected in real time. The process parameters are preprocessed and concatenated into vectors, which are then input into a trained GRU network. The GRU network predicts the acceleration vectors at multiple corresponding positions. A second training sample set is generated based on the predicted acceleration vectors at each position and the screening efficiency. The CNN-LSTM-Attention coupled model was trained based on the second training sample set; After collecting and preprocessing the preset process parameters, they are concatenated into vectors and input into a trained GRU network to predict the acceleration vectors at each position. The predicted acceleration vectors at each position are then input into a trained CNN-LSTM-Attention coupled model to predict the screening efficiency. Based on the predicted screening efficiency, energy consumption, and production cost, a genetic algorithm fitness function is generated to optimize the process parameters.

2. The method for optimizing the screening efficiency of a vibrating screen according to claim 1, characterized in that, The H-RPCA denoising process includes: obtaining a Hankel matrix by sliding window segmentation of acceleration vectors collected from acceleration sensors installed at multiple locations on the vibrating screen; iterating the Hankel matrix using the nonholonomic augmented Lagrange multiplier method until the iteration process converges, generating a low-rank matrix L of the Hankel matrix under convergence conditions; and obtaining the denoised acceleration vector based on the elements of the low-rank matrix L of the Hankel matrix under convergence conditions.

3. The method for optimizing the screening efficiency of a vibrating screen according to claim 2, characterized in that, The Hankel matrix is ​​obtained by sliding window segmentation of acceleration vectors collected from acceleration sensors installed at multiple locations on the vibrating screen. The window size is m, the step size is 1, and the collected data of each window is used as the column vector of the Hankel matrix. Wherein, m is 1.3-1.6 times the period length of the elements in the acceleration vector.

4. The method for optimizing the screening efficiency of a vibrating screen according to claim 3, characterized in that, The formula for iteratively calculating the Hankel matrix using the nonholonomic augmented Lagrange multiplier method includes: in, The singular value soft thresholding operator performs singular value decomposition on the input matrix and max(σ) on the singular values. i After -τ,0), the matrix containing the singular values ​​is replaced; σ i For the i-th singular value, H is the input Hankel matrix; This is an element-level soft thresholding operator used to threshold each element 'a' in the matrix within parentheses. ij Perform the following calculations: Where sgn() is the symbolic function; μ k To increment the penalty parameter, μ k+1 =ρμ k ρ = 1.5~1.8; Y k S is the Lagrange operator matrix for the k-th iteration; k Let L be the sparse matrix in the k-th iteration. k Let be the low-rank matrix in the k-th iteration.

5. The method for optimizing the screening efficiency of a vibrating screen according to claim 4, characterized in that, The denoised acceleration vector is obtained from the elements of the low-rank matrix L of the Hankel matrix under the aforementioned convergence condition, including calculating the t-th element in the denoised acceleration vector using the following formula: Where N is the number of elements in the denoised acceleration vector; L i,j Let n be the element in the i-th row and j-th column of the low-rank matrix L of the Hankel matrix under the convergence condition; the line connecting the positions of each element in each average calculation is parallel to the anti-diagonal of the m×m submatrix contained in the low-rank matrix L, where m is the row number of the low-rank matrix L of the Hankel matrix, and n is the number of rows. t Let t be the number of elements used to calculate the average in the t-th iteration.

6. The method for optimizing the screening efficiency of a vibrating screen according to claim 1, characterized in that, Based on the predicted screening efficiency, energy consumption, and production cost, a genetic algorithm fitness function is generated. The formula for calculating the fitness function is as follows: Fitness=w1 (t) ·S E (E)+w2 (t) ·S C (C)+w3 (t) ·S P (P), Among them, S E (E) is the screening efficiency correlation function, S C (C) is the energy consumption normalized score function, S P (P) is the normalized score function for production costs, w1 (t) w2 (t) w3 (t) w1 is the weight that is dynamically adjusted during the t-th generation of evolution. (t) +w2 (t) +w3 (t) =1.

7. The method for optimizing the screening efficiency of a vibrating screen according to claim 6, characterized in that, The formula for calculating the screening efficiency correlation function is as follows: S E (E)=A E , Where E is the normalized screening efficiency, and A ranges from [2, 5.5]. The formula for calculating the energy consumption normalized score function is: S C (C)=log a1 C, Where C is the normalized energy consumption, and a1 ranges from [0.1, 1 / 3]; The formula for calculating the normalized score function of production costs is: S P (P)=log a2 P, Where P is the normalized generation cost, and a2 ranges from [0.5, 2 / 3].

8. The method for optimizing the screening efficiency of a vibrating screen according to claim 7, characterized in that, The production costs include material costs due to multiple screenings, machine wear costs under different excitation forces, excitation frequencies, excitation direction angles, screen inclination angles α, and feed rates Q, and labor costs.

9. The method for optimizing the screening efficiency of a vibrating screen according to claim 1, characterized in that, The formula for internal information transmission in a GRU network is: r t =σ(W r ·[h t-1 ,x t ]), z t =σ(W z ·[h t-1 ,x t ]), Where, x t h is a vector composed of preprocessed excitation force, excitation frequency, excitation direction angle, screen inclination angle α, and feed rate Q. t-1 Here is the state vector of the hidden layer, [] represents vector concatenation, and r t and z t , respectively, are the outputs of the reset gate and the update gate; W is the weight matrix of the hidden states outside the reset gate and the update gate; W r and W z are the weight matrices for the reset gate and the update gate, respectively; tanh is the state activation function; and h t σ represents the candidate acceleration output and the predicted acceleration output at time t, respectively, and σ is the activation function of the reset gate and update gate.

10. The method for optimizing the screening efficiency of a vibrating screen according to claim 1, characterized in that, The output of the last hidden layer state information in the LSTM model is used as the input of the Attention model, and the Attention model outputs the screening efficiency of the prediction.

Images