Coal preparation plant vibrating screen state analysis method based on deep learning

Abstract

The invention discloses a coal preparation plant vibrating screen state analysis method based on deep learning, belongs to the field of coal preparation plant vibrating screen state analysis, and solves the problem of how to accurately monitor the fault state of a coal preparation plant vibrating screen. Through the LSTM and ST-GNN combined network model, the combination of features in time and space is enhanced so as to improve the recognition and prediction precision. Compared with a traditional time sequence model, the time-space diagram neural network which can be used for extracting a complex time-space dependency relationship is fused, so that the model has more accurate recognition and prediction effects; reliable data are provided for daily regular inspection and health management of the vibrating screen, abnormity of the vibrating screen can be effectively judged, and economic losses caused by abnormal shutdown are completely eradicated.

Description

Technical Field

[0001] The present invention belongs to the technical field of vibrating screen state analysis in a coal preparation plant, and relates to a vibrating screen state analysis method in a coal preparation plant based on deep learning. Background Art

[0002] Coal is an important material basis for social and economic development and a pillar industry of the national economy. It has always occupied a major position in my country's economic development and has strongly supported the stable and rapid development of the national economy and society.

[0003] As a key equipment in coal preparation plants, the stable and efficient operation of vibrating screens directly affects the production efficiency and product quality of coal preparation plants. However, due to the harsh working environment and high working intensity of vibrating screens, the possibility of failure is high, which brings great risks to the normal production of coal preparation plants.

[0004] Existing fault diagnosis methods mainly rely on manual experience, judging through the current test data of the vibrating screen equipment, or troubleshooting after the vibrating screen has a serious fault and stops running. Existing methods have monitoring delays and cannot monitor the operating status of the vibrating screen in real time. At the same time, they are very cumbersome during daily maintenance, affecting work efficiency.

[0005] With the increasing development and improvement of artificial intelligence technology, fault detection methods based on deep learning have been widely used in various fields. Using deep learning fault diagnosis technology to realize the detection of vibrating screen status, accurate detection and alarm of faults is an important means to solve many problems existing in traditional methods. Summary of the invention

[0006] The technical solution of the present invention is used to solve the problem of how to accurately monitor the fault state of the vibrating screen in a coal preparation plant.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] A method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning comprises the following steps:

[0009] S1, collecting audio signals of the vibrating screen operation status, preprocessing the collected audio signals, obtaining an audio fault detection data set after screening, and dividing the data set into a training set and a test set;

[0010] S2, using ST-GNN to extract spatial features from the spatiotemporal graph of the audio signal, using LSTM to extract temporal features from the time series data of the audio, and fusing the spatial features and temporal features of the extracted audio signal;

[0011] S3. Train the ST-GNN-LSTM network model and obtain the final weight after the training is completed, so as to test the field data.

[0012] Furthermore, the method for collecting the audio signal of the operating status of the vibrating screen described in step S1 is as follows: multiple aluminum ribbon pickups are distributed and arranged in various parts of the vibrating screen, and the induced current generated by the sound wave vibration is converted into an analog electrical signal to realize multi-angle and multi-directional collection of the audio signal during the operation of the vibrating screen.

[0013] Furthermore, the method for preprocessing the collected audio signal described in step S1 is as follows:

[0014] (1) Missing value processing: By analyzing the frequency spectrum of the audio signal, the missing audio segments are reconstructed in the frequency domain to obtain a complete vibrating screen audio signal dataset;

[0015] (2) Decomposition and reconstruction: Use the EMD algorithm to decompose and reconstruct the audio signal, decomposing the input audio signal into multiple intrinsic mode functions and a residual, namely:

[0016]

[0017] Where I(n) represents the input signal, IMF m (n) represents the mth eigenmode function, Res M (n) represents the residual;

[0018] The decomposition process of the EMD algorithm is as follows:

[0019] ①First, obtain the local extreme point of the time series data X(t) of the audio signal;

[0020] ② Connect the maximum points of X(t) to form the upper envelope, connect the minimum points to form the lower envelope, the mean line h1(t), and calculate the difference m1(t) between X(t) and h1(t);

[0021] ③ Determine whether the time series m1(t) satisfies the condition that the number of extreme points is equal to the number of zero points or the difference is 1, and the average of the upper and lower envelopes is 0. If so, the signal is an IMF component. If not, repeat steps ① and ②.

[0022] ④ Let r1(t) be h1(t), replace X(t) in steps ① and ② with S1(t), and repeat steps ① to ③;

[0023] ⑤When r1(t) is a monotonic sequence or a constant sequence, the process ends and multiple IMF components and residuals are obtained;

[0024] (3) Normalization processing: adjust the amplitude of the audio signal so that its RMS value reaches the preset standard.

[0025] Furthermore, the method of extracting spatial features from the spatiotemporal graph of the audio signal using ST-GNN in step S2 is as follows: the relationship and dependency between nodes are captured by graph convolution operation, and the graph convolution neural network is used to process the graph structure data, and the feature representation of each node is updated by convolution operation, so as to extract the spatial feature F GCN .

[0026] Furthermore, the method of extracting time domain features from the time series data of the audio using LSTM in step S2 is as follows: LSTM processes the time series data through an input gate, a forget gate, and an output gate to extract the time domain features F LSTM .

[0027] Furthermore, the method for fusing the spatial features and time domain features of the extracted audio signal described in step S2 is as follows:

[0028] The spatial features F extracted by ST-GNN GCN And the time domain features F extracted by LSTM LSTM After weighted fusion, the fused feature F is as follows:

[0029] F=ω GCN ·F GCN +ω LSTM ·F LSTM

[0030] Among them, ω GCN and ω LSTM They are the spatial features F GCN and the time domain feature F LSTM The weight of ω GCN +ω LSTM =1.

[0031] Furthermore, the method for training the ST-GNN-LSTM network model described in step S3 is as follows:

[0032] The fused feature F is passed through the fully connected layer F P To process:

[0033] F p =σ(W·F+b)

[0034] Where W and b are the weights and biases of the fully connected layer, and σ is the activation function;

[0035] Classification is performed through the softmax layer;

[0036] Train the ST-GNN-LSTM network model using the MIMII dataset and save the obtained pre-trained weights;

[0037] The audio fault detection dataset is then used to train the ST-GNN-LSTM network model and adjust its parameters.

[0038] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above-mentioned coal preparation plant vibrating screen state analysis method based on deep learning, and the processor is configured to execute the program stored in the memory.

[0039] A storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, executes the steps of the above-mentioned method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning.

[0040] The advantages of the present invention are:

[0041] The present invention collects audio signals from the working of the vibrating screen, pre-processes the audio signals using the EMD algorithm, and strengthens the combination of features in time and space through a network model combining LSTM and ST-GNN to improve recognition and prediction accuracy. Compared with traditional time series models, the present invention integrates a spatiotemporal graph neural network that can be used to extract complex spatiotemporal dependencies, so that the model has more accurate recognition and prediction effects. The present invention provides reliable data for daily regular inspections and health management of the vibrating screen, can effectively judge abnormalities in the vibrating screen, and prevent economic losses caused by abnormal shutdowns. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of the method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning of the present invention;

[0043] Figure 2 Installation diagram of the aluminum ribbon pickup for real-time collection of field data of the present invention;

[0044] Figure 3 The analysis results of the present invention are monitored in real time in the host computer. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments:

[0047] Embodiment 1

[0048] like Figure 1 As shown, a method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning comprises the following steps:

[0049] 1. Collect the audio signal of the vibrating screen operation status, pre-process the collected audio signal, and obtain the audio fault detection data set after screening. The data set is divided into a training set and a test set according to a ratio of 7:3.

[0050] The method for collecting the audio signal of the vibrating screen operation status is as follows:

[0051] like Figure 2 As shown in the figure, multiple aluminum ribbon pickups are distributed in key parts of the vibrating screen, and the induced current generated by the sound wave vibration is converted into an analog electrical signal to realize the multi-angle and multi-directional collection of audio signals during the operation of the vibrating screen. The comprehensive processing of these data helps to comprehensively evaluate the working status of the vibrating screen and predict potential failures, thereby improving the safety and efficiency of the vibrating screen.

[0052] The method for preprocessing the collected audio signal is as follows:

[0053] 1.1. Missing value processing: By analyzing the frequency spectrum of the audio signal, the missing audio segments are reconstructed in the frequency domain to obtain a complete vibrating screen audio signal dataset.

[0054] 1.2 Decomposition and Reconstruction: Use the EMD algorithm to decompose and reconstruct the audio. The EMD algorithm is a time-frequency domain signal processing method used to process non-stationary and nonlinear signal sequences.

[0055] The input audio signal is decomposed into multiple eigenmode functions and a residual, namely:

[0056]

[0057] Where I(n) represents the input signal, IMF m (n) represents the mth eigenmode function, Res M (n) represents the residual.

[0058] The decomposition process of the EMD algorithm is to extract multiple IMF components. The steps are as follows:

[0059] (1) First, obtain the local extreme point of the time series data X(t) of the audio signal;

[0060] (2) Connect the maximum points of X(t) to form the upper envelope, connect the minimum points to form the lower envelope, and the mean line h1(t). Calculate the difference m1(t) between X(t) and h1(t);

[0061] (3) Determine whether the time series m1(t) satisfies the condition that the number of extreme points is equal to the number of zero points or the difference is 1, and the average of the upper and lower envelopes is 0. If so, the signal is an IMF component. If not, repeat steps (1) and (2).

[0062] (4) Let r1(t) be h1(t), replace X(t) in steps (1) and (2) with S1(t), and repeat steps (1) to (3);

[0063] (5) When r1(t) is a monotonic sequence or a constant sequence, the process ends and multiple IMF components and residuals are obtained.

[0064] 1.3. Normalization. Adjust the amplitude of the audio signal so that its root mean square (RMS) value reaches a preset standard, such as -20dB. RMS is a way to measure the energy of an audio signal and focuses more on the overall loudness of the signal. The purpose of normalization is mainly to eliminate differences between recording devices or recording conditions to ensure consistency and stability of data during processing and analysis.

[0065] 2. Feature extraction and fusion of audio signals

[0066] 2.1 Spatial feature extraction

[0067] When processing spatiotemporal graph data, ST-GNN (Spatiotemporal Graph Neural Network) aims to capture the spatial and temporal dependencies of data. Graph Convolutional Neural Network (GCN) is used to capture spatial dependencies. For a graph G = (V, E), where V is a set of nodes and E is a set of edges. Each node v∈V has a feature vector xv. The graph convolution operation can be expressed as:

[0068]

[0069] Where H(l) is the node feature of the lth layer, for the 0th layer, i.e. the input layer, H(0) = X, W(l) is the weight matrix of the lth layer, and σ is the activation function is the adjacency matrix A plus the identity matrix, yes The diagonal matrix of the degree matrix of .

[0070] 2.2 Time Domain Feature Extraction

[0071] The collected audio data is stored in time series. LSTM (Long Short-Term Memory) is a recurrent neural network with a gating mechanism that can handle data with long-term dependencies. LSTM uses gating units to implement information memory and forgetting. A typical LSTM consists of three gates: input gate, forget gate, and output gate, as well as a cell state.

[0072] The forget gate determines which information will be discarded and retains information that is relevant to the calculation results.

[0073] f t =σ(W f ·[h t-1 ,x t ]+b f ) (3)

[0074] Where ft is the output of the forget gate at time step t, σ is the sigmoid function, Wf and bf are the weight and bias of the forget gate respectively, ht-1 is the output of the previous time step, and xt is the input of the current time step.

[0075] The input gate determines what new information is stored in the cell state and is calculated in two parts.

[0076] i t =σ(W i ·[h t-1 ,x t ]+b i ) (4)

[0077]

[0078] Where it is the output of the input gate, (C_t)~ is the candidate cell state, Wi, Bi, Wc and Bc are the corresponding weights and biases.

[0079] The old cell state Ct-1 is updated through the output of the forget gate and the input gate to generate a new cell state Ct.

[0080]

[0081] Where Ct is the cell state at time step t and * denotes element-wise multiplication.

[0082] The output gate determines the hidden state of the next time step, which is the output. The hidden state is obtained by processing the current cell state.

[0083] o t =σ(W O ·[ht-1 ,x t ]+b O ) (7)

[0084] h t =o t *tanh(C t ) (8)

[0085] Where ot is the output of the output gate, ht is the output at time step t, and Wo and bo are the weight and bias of the output gate.

[0086] 2.3 Fusion

[0087] ST-GNN is used to extract spatial features from the spatiotemporal graph of audio signals. It captures the relationships and dependencies between nodes (parts of the audio signal) through graph convolution operations. It uses graph convolutional neural networks (GCNs) to process graph structure data and updates the feature representation of each node through convolution operations to extract spatial features F. GCN LSTM is used to extract time domain features from audio time series data. It captures the dynamic changes of audio signals over time, including long-term dependencies, through its unique gating mechanism. LSTM processes time series data through input gates, forget gates, and output gates to extract time domain features F. LSTM .

[0088] The spatial features F extracted by ST-GNN GCN And the time domain features F extracted by LSTM LSTM Perform weighted fusion. Weight ω GCN and ω LSTM Determine according to the importance of the feature, and obtain the fusion feature through weighted summation:

[0089] F=ω GCN ·F GCN +ω LSTM ·F LSTM (9)

[0090] where ω GCN and ω LSTM are the weights of GCN and LSTM features, and ω GCN +ω LSTM = 1. During the training process, the gradient of each feature to the loss function is calculated through back propagation, and the weight is adjusted according to the size of the gradient. Features with larger gradients may have a greater impact on the loss, so they can be given higher weights. Finally, we determined the weight ω GCN =0.4 and ω LSTM =0.6.

[0091] 3. Establishment of sensitive information database

[0092] In order to realize the functions of real-time monitoring of the on-site sound of the vibrating screen, reproduction and comparison of abnormal sounds of the vibrating screen, the on-site operation sound of the vibrating screen is stored, and the classified distributed storage of audio files is realized, which is convenient for the host computer comprehensive management system to call the audio files in real time and realize the long-term operation status judgment of the vibrating screen.

[0093] 4. Network model training and testing

[0094] The above fused feature F is processed through a fully connected layer:

[0095] F p =σ(W·F+b) (10)

[0096] Where W and b are the weights and biases of the fully connected layer, and σ is the activation function.

[0097] Classification is performed through a softmax layer.

[0098] The ST-GNN-LSTM network model is trained using the MIMII dataset, and the pre-trained weights are saved. The ST-GNN-LSTM network model is then trained using the audio signal dataset and its parameters are adjusted. After the training is completed, the final weights are obtained and tested on the field data.

[0099] At the same time, if Figure 3 As shown, the host computer receives and displays the detected vibrating screen status information in real time, thereby improving the monitoring accuracy of the vibrating screen and ensuring the healthy operation of the vibrating screen.

[0100] Embodiment 2

[0101] An electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the coal preparation plant vibrating screen state analysis method based on deep learning in Example 1, and the processor is configured to execute the program stored in the memory.

[0102] Embodiment 3

[0103] A storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning in Example 1 are executed.

[0104] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning, characterized in that: The following steps are involved: S1, collecting audio signals of the vibrating screen operation status, preprocessing the collected audio signals, obtaining an audio fault detection data set after screening, and dividing the data set into a training set and a test set; S2, using ST-GNN to extract spatial features from the spatiotemporal graph of the audio signal, using LSTM to extract temporal features from the time series data of the audio, and fusing the spatial features and temporal features of the extracted audio signal; S3. Train the ST-GNN-LSTM network model and obtain the final weight after the training is completed, so as to test the field data.

2. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method for collecting the audio signal of the operating status of the vibrating screen described in step S1 is as follows: multiple aluminum ribbon pickups are distributed and arranged in various parts of the vibrating screen, and the induced current generated by the sound wave vibration is converted into an analog electrical signal to realize multi-angle and multi-directional collection of the audio signal during the operation of the vibrating screen.

3. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method for preprocessing the collected audio signal described in step S1 is as follows: (1) Missing value processing: By analyzing the frequency spectrum of the audio signal, the missing audio segments are reconstructed in the frequency domain to obtain a complete vibrating screen audio signal dataset; (2) Decomposition and reconstruction: Use the EMD algorithm to decompose and reconstruct the audio signal, decomposing the input audio signal into multiple intrinsic mode functions and a residual, namely: Where I(n) represents the input signal, IMF m (n) represents the mth eigenmode function, Res M (n) represents the residual; The decomposition process of the EMD algorithm is as follows: ①First, obtain the local extreme point of the time series data X(t) of the audio signal; ② Connect the maximum points of X(t) to form the upper envelope, connect the minimum points to form the lower envelope, the mean line h1(t), and calculate the difference m1(t) between X(t) and h1(t); ③ Determine whether the time series m1(t) satisfies the condition that the number of extreme points is equal to the number of zero points or the difference is 1, and the average of the upper and lower envelopes is 0. If so, the signal is an IMF component. If not, repeat steps ① and ②. ④ Let r1(t) be h1(t), replace X(t) in steps ① and ② with S1(t), and repeat steps ① to ③; ⑤When r1(t) is a monotonic sequence or a constant sequence, the process ends and multiple IMF components and residuals are obtained; (3) Normalization processing: adjust the amplitude of the audio signal so that its RMS value reaches the preset standard.

4. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method of extracting spatial features from the spatiotemporal graph of the audio signal using ST-GNN described in step S2 is as follows: the mutual relationship and dependency between nodes are captured by graph convolution operation, and the graph convolution neural network is used to process the graph structure data, and the feature representation of each node is updated by convolution operation, so as to extract the spatial feature F GCN .

5. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method of extracting time domain features from the time series data of audio using LSTM in step S2 is as follows: LSTM processes the time series data through an input gate, a forget gate, and an output gate to extract the time domain features F LSTM .

6. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method for fusing the spatial features and time domain features of the extracted audio signal described in step S2 is as follows: The spatial features F extracted by ST-GNN GCN And the time domain features F extracted by LSTM LSTM After weighted fusion, the fused feature F is as follows: F=ω GCN ·F GCN +oh LSTM ·F LSTM Among them, ω GCN and ω LSTM They are the spatial features F GCN and the time domain feature F LSTM The weight of ω GCN +ω LSTM =1.

7. The method for analyzing the state of a vibrating screen in a coal preparation plant based on deep learning according to claim 1, characterized in that: The method for training the ST-GNN-LSTM network model described in step S3 is as follows: The fused feature F is passed through the fully connected layer F P To process: F p =σ(W·F+b) Where W and b are the weights and biases of the fully connected layer, and σ is the activation function; Classification is performed through the softmax layer; Train the ST-GNN-LSTM network model using the MIMII dataset and save the obtained pre-trained weights; The audio fault detection dataset is then used to train the ST-GNN-LSTM network model and adjust its parameters.

8. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the deep learning-based coal preparation plant vibrating screen state analysis method as described in any one of claims 1 to 7, and the processor is configured to execute the program stored in the memory.

9. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the deep learning-based coal preparation plant vibrating screen state analysis method according to any one of claims 1 to 7 are executed.

Images