Coal mill material level voiceprint monitoring system based on CLDNN

Through CLDNN deep learning technology and Φ-OTDR fiber optic sensing system, real-time monitoring and prediction of steel ball coal mill material levels is achieved, the problem of inaccurate monitoring of material levels in the existing technology is solved, and the powder making system of thermal power plants is improved Automatic control capability and operating efficiency.

CN118913402AInactive Publication Date: 2024-11-08NAT ENERGY SICHUAN HUAYINGSHAN POWER GENERATION CO LTD

Patent Information

Application Number
CN202410640517.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-05-22
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing technology cannot accurately monitor and control the material level of the steel ball coal mill, resulting in unreliable automatic control of the powder making system of thermal power plants and requires manual operation, which affects operating efficiency and safety.

Method used

The soundprint monitoring system of coal mill material level based on CLDNN deep learning is adopted. The soundprint signal is collected through the acoustic sensing module, combined with the Φ-OTDR fiber sensing system, and the improved CLDNN network model is used for feature extraction and classification identification to realize material Real-time monitoring and prediction of bit state.

Benefits of technology

It improves the operating efficiency of coal mills, reduces energy consumption, enhances safety, and realizes intelligent and efficient control of the power plant powder making system, which significantly improves accuracy and reliability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118913402A_ABST
    Figure CN118913402A_ABST
Patent Text Reader

Abstract

The invention provides a coal mill material level voiceprint monitoring system based on a CLDNN. The coal mill material level voiceprint monitoring system comprises an acoustic sensing module, a preprocessing module and a feature extraction and classification recognition module. The acoustic sensing module collects voiceprint signals by using a phi-OTDR optical fiber sensing system, and constructs a space-time matrix; the preprocessing module preprocesses the space-time matrix, including normalization, framing, windowing and the like; the feature extraction and classification recognition module adopts an improved CLDNN network to carry out automatic feature extraction on the preprocessed voiceprint signals. And during training, inputting the preprocessed space-time matrix into the CLDNN for training, optimizing model parameters, and performing material level identification, classification and prediction on newly collected voiceprint signals by using the trained model. According to the method, rich material level state features contained in the voiceprint signals are automatically extracted by using the CLDNN deep learning algorithm, accurate recognition and prediction of the material level of the coal mill are realized, and the utilization efficiency of energy is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of signal detection and signal processing measurement, and in particular aims at intelligent recognition based on collected coal mill material level soundprint information. Background Art

[0002] With the deepening reform of my country's power market, "separation of power plants from the grid and bidding for access to the grid" has made power plants strive to improve power generation efficiency and reduce power generation costs to strengthen their market competitiveness. However, an important factor affecting the economic operation of thermal power plants is that many important technical and economic parameters are difficult to measure online in real time, such as flue gas oxygen content, ball mill material level, fly ash carbon content, etc. These important parameters that can directly reflect power generation efficiency and operational safety cannot be directly measured using conventional sensors due to technical and economic reasons. Therefore, power plants are now facing the following dilemma: on the one hand, power plants can only use offline analysis methods such as manual sampling and testing for some difficult-to-measure parameters. Manual sampling and testing has a large lag, a long sampling cycle, and poor representativeness, and cannot be directly used for guidance and control; on the other hand, many production links of thermal power plants are forced to adopt conservative operating conditions to ensure production safety and stability, sacrificing economy. For example, the pulverizing system of thermal power plants.

[0003] The pulverizing system is a very important auxiliary system in the boiler system of a thermal power plant. The safety and economy of its operation affect the economy and safety of the thermal power plant. The ball mill is an important equipment in the intermediate storage pulverizing system. It has many advantages such as simple maintenance, reliable operation, low maintenance cost, strong adaptability, etc. and is adopted by the vast majority of small and medium-sized thermal power plants. However, the ball mill is also a strongly coupled, nonlinear, multivariable, and large-delay object. In addition, the changes in raw coal moisture and coal quality, as well as the wear of lining plates and steel balls, make the working conditions of the ball mill very complicated, making it very difficult to monitor and control the ball mill online.

[0004] The operation of the ball mill under ideal working conditions can effectively save energy and cost. The ideal working condition is to operate economically under the condition of maximum coal grinding output and minimum unit consumption of pulverizing. Both theoretical research and operation practice show that the optimal operation of the ball mill pulverizing system is closely related to the amount of coal stored in the drum of the ball mill. That is, when the output of the ball mill is insufficient and the amount of coal stored in its corresponding drum is too small, the consumption of steel balls is large and the power consumption will also increase; when there are too many, the grinding capacity will be reduced, and the fineness requirements of the coal powder cannot be guaranteed. At the same time, it is easy to cause powder spraying (coal powder spraying outside) and even full grinding, blocking and other accidents. Therefore, in order to realize the automatic control and optimized operation of the ball mill, reduce the unit consumption of pulverizing, and increase the economy of the ball mill pulverizing system operation, the key is to solve the problem of measuring the amount of coal stored in the drum of the ball mill, that is, the material level.

[0005] Nowadays, power plants often use a simple differential pressure signal or current signal to characterize the amount of coal stored in the drum of the ball mill, and then indirectly reflect the level of the mill. However, due to the complex characteristics of the ball mill and the serious coupling phenomenon between various parameters, a simple signal often cannot accurately characterize the change of the ball mill level.

[0006] Although ball mills are widely used, they cannot be reasonably and economically controlled according to the material level or pulverizing output of the ball mills, which results in the inability of automatic control of the pulverizing system of thermal power plants to be put into operation reliably for a long time, and even the automatic control is abandoned and replaced by manual control. Moreover, in order to ensure safe operation and prevent accidents such as powder spraying, mill blockage and fullness, the operators often operate the ball mills under undesirable material level conditions with insufficient coal storage and excessive air volume.

[0007] Coal pulverizer is an important auxiliary equipment of coal power plant, and its performance directly affects the combustion efficiency and economic operation of the power plant. Traditional material level monitoring systems mostly rely on manual observation or simple instrument measurement, which has certain delays and inaccuracies.

[0008] The following is an analysis of the shortcomings of the prior art:

[0009] The use of a single signal such as differential pressure or current cannot accurately characterize the complex changes in the material level state inside the coal mill, which can easily lead to inaccurate control. Most of the existing coal mill material level monitoring relies on differential pressure signals or motor current signals. However, due to the complex movement of materials inside the coal mill and the coupling influence of multiple factors such as feed amount, air volume, temperature, etc., a single differential pressure or current signal is difficult to accurately reflect the actual material level changes. For example, when the material level is low, the differential pressure is smaller due to the large air volume; when the material level is high, the motor current may decrease due to the increased material resistance. This complex coupling effect makes it difficult for a single signal to accurately describe the material level state, thus affecting the precise control of the coal mill.

[0010] Traditional manual observation and simple instrument measurement methods have delays and inaccuracies, making it difficult to achieve real-time and accurate monitoring of the coal mill level. At present, the monitoring of coal mill levels in power plants mostly relies on manual observation by operators or simple instrument measurements. This method has obvious limitations: on the one hand, manual observation has the defects of strong subjectivity and long reaction time, and it is difficult to accurately reflect the instantaneous changes of the material level in real time; on the other hand, single differential pressure, current and other instrument measurements are also difficult to fully characterize the complex changes in the material level state. The delay and inaccuracy of these traditional monitoring methods directly restrict the real-time and accurate control of coal mills by power plants.

[0011] Since it is impossible to accurately monitor and control the material level of the coal mill, the power plant can only operate the coal mill under conservative and undesirable conditions, and cannot achieve economical and efficient automatic control. Due to the limitations of existing material level monitoring technology, power plants can only adopt conservative control strategies and operate the coal mill under undesirable conditions with insufficient coal storage and large air volume for a long time to prevent accidents such as dust injection, mill blockage, and full magic. Although this approach can ensure safe operation, it greatly reduces the operating efficiency of the coal mill and increases energy consumption. If the real-time and accurate monitoring and prediction of the coal mill material level can be achieved, the operating parameters can be dynamically optimized according to the material level conditions, the comprehensive energy efficiency of the coal mill can be improved, and the intelligent and efficient control of the power plant pulverizing system can be achieved. Summary of the invention

[0012] In view of the above-mentioned defects of the prior art, the present invention proposes a coal mill material level soundprint monitoring method and system based on CLDNN deep learning, which aims to realize real-time and accurate monitoring and prediction of the coal mill material level status through intelligent soundprint signal analysis technology, provide support for the intelligent control of the power plant pulverizing system, thereby improving the operating efficiency of the coal mill, reducing energy consumption and enhancing safety.

[0013] The specific technical solutions are as follows:

[0014] A coal mill material level soundprint monitoring system based on CLDNN includes an acoustic sensing module, a preprocessing module, and a feature extraction and classification recognition module. The acoustic sensing module extracts the original soundprint signal during the operation of the coal mill and constructs a space-time matrix; the feature extraction and classification recognition module automatically extracts features from the preprocessed soundprint signal, identifies and classifies according to the features, and outputs the current material level status of the coal mill.

[0015] Furthermore, the acoustic sensing module uses the Φ-OTDR optical fiber sensing system to collect voiceprint signals. The Φ-OTDR optical fiber sensing system specifically adjusts the laser light source in the OTDR system to a narrow linewidth laser and adds a pulse adjuster thereafter.

[0016] Furthermore, the Φ-OTDR optical fiber sensing system uses the coherence characteristics of light to detect the phase change of the corresponding position at the far end of the OTDR system sensing optical fiber. The original voiceprint signal collected by the Φ-OTDR optical fiber sensing system is represented as an M×N space-time matrix X={x ij}, where M is the total number of sampling time steps, N is the number of fiber optic sensor units, and x ij is the sensor signal at the jth spatial position on the optical fiber path at the i-th moment.

[0017] Furthermore, the feature extraction and classification recognition module adopts the CLDNN network model, which includes CNN, LSTM and DNN modules connected in sequence. The CNN module is responsible for extracting the local time-frequency features of the voiceprint signal and capturing key acoustic patterns; the LSTM module is used to model the evolution and correlation of these features in the time dimension; the DNN module classifies different material level states based on the outputs of the first two modules.

[0018] Furthermore, the feature extraction and classification recognition module adopts an improved CLDNN network model. The improvement is that:

[0019] In the CNN module, a progressive receptive field expansion strategy is adopted, that is, small convolution kernels are first used to extract fine-grained local features, and then larger convolution kernels are used layer by layer to fuse these local features;

[0020] In the LSTM module, an attention mechanism is introduced to adaptively focus on key acoustic events at different moments. Specifically, the hidden state at the current moment is weighted summed with the hidden state at all historical moments to obtain a dynamic context vector, which is then concatenated with the input at the current moment as the input of the LSTM unit at the current moment.

[0021] In the DNN module, residual connection is used, that is, the input is directly added to the output of an intermediate layer through identity mapping, and then sent to the subsequent layer.

[0022] Furthermore, the improved CLDNN network model includes two LSTMs. An encoder LSTM network encodes the voiceprint signal of L historical time steps to obtain a semantically rich context vector. A decoder LSTM network then uses the vector as the initial state to iteratively generate a material level state sequence for the next K time steps. The two LSTMs constitute a Seq2Seq network based on sequence to sequence.

[0023] Furthermore, the loss function of the Seq2Seq network is the mean square error between the material level output of the decoder at each time step and the true future material level sequence, which is optimized through end-to-end backpropagation.

[0024] Furthermore, the preprocessing module includes median filtering, wavelet threshold denoising, normalization, and framing and windowing.

[0025] Furthermore, a monitoring and diagnosis module is set up. The monitoring and diagnosis module receives the output results of the feature extraction and classification recognition module, and monitors the material level status of the coal mill in real time in combination with historical monitoring data. When an abnormal state is detected, an alarm is triggered, and the possible causes of the fault are analyzed and diagnosed.

[0026] Beneficial Effects

[0027] (1) In the existing technology, it is difficult to accurately describe the internal material level state of the coal mill with a single differential pressure or current signal, which is affected by the coupling of multiple factors. Acoustic sensing technology is used to fully perceive the operating state of the coal mill. Compared with a single differential pressure or current signal, the soundprint signal can better reflect the complex dynamic changes of the material level state, laying a good foundation for subsequent accurate identification.

[0028] (2) Compared with the traditional OTDR system, the Φ-OTDR optical fiber sensing system introduces a narrow linewidth laser, which makes the interference of backscattered light stronger, thereby improving the detection sensitivity by 1-2 orders of magnitude. At the same time, the Φ-OTDR directly detects phase information and has stronger anti-fading ability.

[0029] (3) The CLDNN network of the present invention has been improved and optimized on the basis of the classical network:

[0030] In the CNN module, a progressive receptive field expansion strategy is adopted, which helps to obtain robust high-level semantic expression without losing detailed information;

[0031] In the LSTM module, an attention mechanism is introduced to enable the network to adaptively focus on key acoustic events at different times. This attention mechanism enhances the long-term modeling capability of LSTM.

[0032] In the DNN module, residual connection is used. This structure can alleviate the degradation problem of deep networks and make training more stable and efficient.

[0033] Based on the above improvements, the CLDNN network of the present invention achieves better performance than the general network architecture in voiceprint recognition tasks.

[0034] (4) In the improved CLDNN deep learning network proposed in the present invention, the CNN module is responsible for extracting the local time-frequency features of the voiceprint signal and capturing the key acoustic patterns; the LSTM module can model the evolution and correlation of these features in the time dimension; and the DNN module classifies different material level states based on the outputs of the first two modules. The three modules work closely together to fully tap the discriminative information contained in the voiceprint signal.

[0035] The present invention uses an improved CLDNN deep learning network to automatically extract features from preprocessed voiceprint signals. Compared with traditional Gaussian mixture models and hidden Markov models, this model has achieved higher accuracy (more than 5% improvement), stronger robustness (performance degradation in noisy environments does not exceed 3%), and faster training and reasoning speeds (20% and 50% improvement, respectively) in voiceprint recognition tasks.

[0036] Compared with manual observation and simple instrument measurement, the present invention uses CLDNN deep learning algorithm to automatically extract time-frequency domain features and dynamic features from voiceprint signals, which can effectively depict the complex change pattern of material level status and greatly improve the accuracy and reliability of material level identification. This provides reliable technical support for power plants to achieve accurate real-time monitoring of coal mill material levels.

[0037] (5) The Sequence-to-Sequence (Seq2Seq) model proposed in the present invention can not only identify the material level status at the current moment, but also predict the material level change in the future based on the historical voiceprint signal change trend. This is of great significance for early warning and optimized scheduling.

[0038] (6) The present invention combines the material level monitoring results with fault diagnosis to achieve comprehensive intelligent monitoring and analysis of the coal mill operating status, thereby improving the reliability and economy of equipment operation.

[0039] In the prior art, due to the lack of accurate material level monitoring means, power plants can only adopt conservative and undesirable control strategies, which limits the operating efficiency of coal mills. The present invention combines material level monitoring with fault diagnosis to achieve comprehensive monitoring and analysis and diagnosis of coal mill operating conditions, which can not only improve the reliability of the equipment, but also dynamically optimize operating parameters according to material level, significantly improve the comprehensive energy efficiency of coal mills, and realize intelligent and efficient control of power plant pulverizing systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 , Structure diagram of coal mill material level soundprint monitoring system based on CLDNN;

[0041] Figure 2 , OTDR system structure diagram;

[0042] Figure 3 , schematic diagram of OTDR scattering measurement curve changes;

[0043] Figure 4 , Φ-OTDR system structure diagram;

[0044] Figure 5 ,Data flow diagram in improved deep learning network;

[0045] Figure 6 , basic structure of improved deep learning network;

[0046] Figure 7 , the basic process of convolutional layer; DETAILED DESCRIPTION

[0047] The present invention provides a coal mill material level soundprint monitoring system based on CLDNN, such as Figure 1As shown. The system mainly includes the following modules:

[0048] (1) Acoustic sensing module: This module is based on a distributed fiber optic sensing system based on Φ-OTDR technology, which collects the soundprint signals generated during the operation of the coal mill. The distributed fiber optic sensing system based on Φ-OTDR technology includes multiple high-sensitivity fiber optic microphone sensors. In order to obtain comprehensive sound field information, the sensors are arranged at equal intervals along the circumference of the coal mill, and the appropriate installation position and number are determined according to the size of the coal mill. The reason for using a fiber optic microphone is that it has the advantages of wide-band response, low noise, and strong anti-interference ability, and can obtain high-quality soundprint signals.

[0049] In the traditional OTDR system, the pulse laser sends the pulse laser into the optical fiber path for elastic scattering, thereby generating Rayleigh scattered light with the same frequency as the incident light. The incident light and the Rayleigh scattered light interfere with each other and are received by the photodetector through the optical circulator, and finally the relationship between the intensity change of the Rayleigh scattered light and the monitoring distance is obtained. The traditional OTDR system structure is as follows: Figure 2 shown.

[0050] like Figure 3 The following is a schematic diagram of the change of the OTDR scattering measurement curve. It can be seen from the figure that the intensity of the backward Rayleigh scattered light in the optical fiber decreases with the increase of distance. If the optical signal encounters abnormal conditions such as bending loss and connection points at a certain point in the optical fiber path when propagating in the optical fiber, the corresponding backward Rayleigh scattering curve at the abnormal location will undergo a mutation, and the mutation will produce corresponding changes according to the specific situation of the abnormality. Figure 3 The intensity curve of the scattered light sometimes rises suddenly and sometimes drops sharply.

[0051] The intensity detection OTDR system can use the light intensity scattered back from the far end of a long-distance optical fiber for detection, but studies have shown that this method is difficult to measure directly because the collected back Rayleigh scattered light intensity signal is weak. The present invention further proposes a phase-sensitive optical time domain reflectometer system, namely a Φ-OTDR (Phase-sensitive Optical Time Domain Reflectometer) system, which can use the coherence characteristics of light to detect the phase change of the corresponding position at the far end of the sensing optical fiber of the OTDR system.

[0052] The structure of the phase-sensitive optical time domain reflectometer (Φ-OTDR) system is similar to that of the OTDR system. The system mainly consists of a narrow linewidth light source, a pulse adjuster, a circulator, a photodetector, a signal acquisition and processing system, and a long-distance sensing fiber. The laser light source of the system is adjusted to a narrow linewidth laser, which enhances the pulse interference of the backward Rayleigh scattered light in the sensing fiber, thereby improving the sensitivity of the system. The main structure of the system is as follows: Figure 4 shown.

[0053] When the optical fiber signal is emitted in a periodic form, the signal collected by the system acoustic sensor module is the backward Rayleigh scattered light propagating along the optical fiber direction, where the scattering trajectory decays in an exponential form. The signal matrix collected by the kth pulse can be written as:

[0054] X k =[t k1 , t k2 ,…,t kN ]

[0055] As the acquisition pulse time increases, M laser scattering measurement trajectories can be accumulated, from which a space-time matrix XY with M as time and N as space can be constructed. The expression is as follows:

[0056] XY={t ki}, k=1, 2,...,M; i=1, 2,...,N

[0057] In the above expression, k is the trigger pulse cycle sequence, and i is the spatial acquisition point sequence. Therefore, the signal collected at the spatial position i of the sensing optical fiber is expressed by a matrix as follows:

[0058] Y i ={t ki}, k = 1, 2, ..., M

[0059] In summary, the signals collected by the Φ-OTDR optical fiber sensing system can form a time-space response matrix, where X k Y is a transverse spatial sequence, that is, it corresponds to the sensor signal at each spatial position on the optical fiber path at a certain moment. The signal amplitude of each point in the spatial sequence corresponds to the spatial position information in the matrix. i It is a longitudinal time series, which represents the sequence signal corresponding to the entire monitoring time at a certain spatial position detected by the sensing optical fiber.

[0060] (2) Preprocessing module

[0061] Before inputting the collected original voiceprint spatiotemporal matrix into the CLDNN model, the present invention specially designs a set of preprocessing processes to improve the data quality. Specifically, it includes:

[0062] Median filtering: Remove impulse noise at each time and space point through a sliding window to smooth the signal.

[0063] Wavelet threshold denoising: Utilizing the multi-scale characteristics of wavelet transform, the high-frequency components of the signal are adaptively processed by soft threshold to remove high-frequency noise.

[0064] Normalization: Perform maximum and minimum normalization on the signal amplitude at each time step to eliminate the scale differences between different signals.

[0065] Framing and windowing: The signal is framed according to a fixed length, and Hamming window weighting is performed within each frame to enhance the stability of the signal.

[0066] The above preprocessing makes the voiceprint data input into CLDNN more regular and pure, which is conducive to subsequent feature learning and classification. Experiments show that using preprocessed voiceprint data to train CLDNN can increase the accuracy of material level recognition by an average of 2-3 percentage points.

[0067] (3) Feature extraction and classification recognition module

[0068] The present invention designs an improved CLDNN deep learning network to automatically extract features from preprocessed voiceprint signals. According to the experimental data collection results, the time domain features of the voiceprint signals monitored by the system have contextual correlation characteristics. Compared with the early developed speech signal recognition technology implemented by Gaussian mixture model / hidden Markov model, the present invention combines CNN, LSTM and DNN into one network, and uses the network to directly learn signal features from the time series data of the signal. The data flow diagram in the improved deep learning network is as follows: Figure 5 As shown in the figure. This neural network structure connects CNN to reduce the changes in the signal frequency domain, and then inputs the output of CNN into several layers of LSTM for time series modeling. Finally, the output of LSTM is input into the fully connected DNN layer, and the feature space will be mapped to the output layer that is easier to classify, generating a feature representation that is easier to separate.

[0069] like Figure 6 The figure shows the basic structure of the improved deep learning network. According to the orderliness of the time domain characteristics of the intrusion event signal, the signal recognition process of the system mainly includes two parts: signal standardization and neural network recognition and classification. The present invention processes the data using the Max Abs Scaler method to control the eigenvalue of the data within the range of [-1,1]. The mathematical expression is as follows:

[0070]

[0071] In the above formula, x and x' are the signal matrices before and after standardization, respectively, and max is the maximum value of a row or column of the signal matrix, depending on whether the standardization is performed in row compression or column compression. This method has the characteristics of not destroying the data structure, not causing data movement and aggregation, and not destroying the sparsity of the data. This characteristic determines that this method can be used in signal acquisition of optical fiber sensing systems.

[0072] The standardized signal will be input into the improved deep learning network for deep learning. First, the signal is input into the CNN network. Through convolution, the CNN network effectively avoids complex data reconstruction in signal feature extraction and classification. By weighting the data parameters, the complexity of the data can be reduced. The CNN network mainly includes convolution layers and pooling layers, in which the convolution layer distinguishes the input data through filters, and the role of the pooling layer can reduce the connection between the convolution layers, thereby optimizing the amount of calculation of the entire neural network, so the CNN network can effectively reduce the frequency change. The LSTM network can model the signal time series characteristics, connect the CNN network to reduce the signal frequency domain changes, and then insert the output results of the CNN network into several LSTM network layers for time series modeling. Finally, the output results of the LSTM network are inserted into the fully connected layer for analysis, and then the feature space is mapped to the output layer. This layer can be more easily classified to create a more easily separable feature representation. Therefore, the recognition model established in this way has better recognition and classification characteristics than a single neural network.

[0073] Depend on Figure 6 As shown, in this embodiment, the improved deep learning network structure is mainly constructed by combining the features of the CNN network and the LSTM network. The network consists of three CNN layers, a linear layer (dim red), an LSTM layer and a fully connected layer. First, the intrusion signal collected by the system is marked and standardized. Max Abs Scaler is the standardization technology used in this study. The marked and standardized signal is first input into the CNN layer. Figure 7As shown in the figure, this method has three CNN layers. Each CNN layer contains a frequency-time filter (1x5), a maximum pooling layer of 2, and a step size of 2. The output size of the CNN layer is large, and even if a linear layer is added to the CNN layer, the image accuracy will not be reduced. Through experiments, it can be known that the most suitable output size of the linear layer is 256. 256 outputs can construct the input of the LSTM layer. By passing the frequency signal to the LSTM layer, the time modeling analysis of the signal can be realized. The LSTM layer used includes 64 units and 256 hidden units. Finally, the result of the fully connected layer is used as the output of the signal. Through the above steps, the system can simply perform deep learning modeling on the collected intrusion signals to accelerate the creation of the neural network model. Then, the system can identify and classify the collected voiceprint signals through the model.

[0074] The improved CLDNN network model can also include two LSTMs. An encoder LSTM network encodes the voiceprint signal of the historical L time steps to obtain a semantically rich context vector; then a decoder LSTM network uses the vector as the initial state to iteratively generate the material level state sequence of the next K time steps; the two LSTMs constitute a Seq2Seq network based on sequence to sequence.

[0075] The spatiotemporal response matrix obtained by the Φ-OTDR optical fiber sensing system can be automatically extracted through the CLDNN deep learning network mentioned above, and voiceprint signal samples representing different material level states are collected. After the samples are manually labeled by experts, a training data set is formed for subsequent network training and parameter optimization. The trained CLDNN classifier can identify the material level state of the newly collected voiceprint signal in real time, and predict the material level loading according to the update speed of the material level state, such as giving the possible time from 20% material level to 50% material level.

[0076] (4) Monitoring and diagnosis module

[0077] This module receives the output results of the classification and recognition module and monitors the material level status of the coal mill in real time in combination with historical monitoring data. When an abnormal state is detected, the module will trigger an alarm and analyze and diagnose the possible causes of the fault in combination with the fault diagnosis algorithm to provide decision support for power plant production management.

[0078] In a 48-hour blind test, the system's material level recognition accuracy for five typical working conditions was above 97%, fully demonstrating its practical performance. The following table lists the detailed test results:

[0079] Working conditions Recognition accuracy No load 99.5% 25% material level 97.8% 50% material level 98.2% 75% material level 97.6% Fully loaded 99.2%

[0080] In addition, the system can also estimate the absolute value of the material level by analyzing the change trend of the voiceprint signal over time. Experiments show that the relative error of the estimation is within ±5%.

[0081] The present invention utilizes the CLDNN deep learning network to automatically extract the rich material level status features contained in the voiceprint signal, including time-frequency domain features and dynamic features, laying the foundation for subsequent material level identification. Based on the trained CLDNN classifier, real-time and accurate material level status identification and prediction of the newly collected voiceprint signal is realized, providing a basis for the intelligent control of the power plant's pulverizing system. The material level monitoring results are combined with the fault diagnosis algorithm to realize comprehensive intelligent monitoring and analysis and diagnosis of the coal mill's operating conditions, thereby improving the reliability and economy of equipment operation. The overall technical solution of the present invention utilizes acoustic sensing and deep learning algorithms to realize accurate real-time monitoring of the coal mill's material level status, providing effective intelligent support for the power plant's pulverizing system, and has important engineering application value. There are indeed other alternative technical solutions for the coal mill material level voiceprint monitoring system based on CLDNN proposed in this application, mainly including:

[0082] (1) Alternative sensor types: In addition to using fiber optic microphones, other types of acoustic sensors can also be used, such as condenser microphones, dynamic microphones, etc. Different types of sensors have their own advantages and disadvantages, and can be selected according to the needs of specific application scenarios.

[0083] (2) Alternatives to signal processing algorithms: In addition to the CLDNN deep learning network, you can also try to use other machine learning algorithms, such as support vector machine (SVM), random forest, neural network, etc., to extract voiceprint features and identify material level status. Different algorithms have their own advantages, and comparative tests can be performed to select the best solution.

[0084] (3) Alternatives to the monitoring mode: In addition to real-time monitoring of material level status, offline analysis can also be used. That is, first collect voiceprint data samples, train and tune the deep learning model in a laboratory environment, and then deploy the trained model to the on-site system for real-time reasoning. The advantage of offline training is that it can make full use of large-scale data samples without being limited by the real-time computing power of the site.

[0085] (4) Alternatives to system integration architecture: In addition to the multi-module integration architecture proposed in this application, other system integration methods can also be tried, such as edge computing architecture, cloud-edge collaborative architecture, etc. Different architectures will have different characteristics and focuses in terms of computing resource allocation, real-time performance, scalability, etc.

[0086] In summary, for the coal mill material level soundprint monitoring system proposed in the present invention, the above-mentioned alternatives are all worthy of consideration and in-depth study. By comparing and evaluating, the technical solution that best suits the actual application scenario can be selected to further enrich and improve the protection scope of the present invention.

Claims

1. A coal mill material level soundprint monitoring system based on CLDNN, Features: It includes an acoustic sensing module, a preprocessing module, and a feature extraction and classification recognition module. The acoustic sensing module extracts the original soundprint signal during the operation of the coal mill and constructs a space-time matrix; the feature extraction and classification recognition module automatically extracts features from the preprocessed soundprint signal, identifies and classifies it according to the features, and outputs the current material level status of the coal mill.

2. According to the coal mill material level soundprint monitoring system based on CLDNN described in right 1, Features: The acoustic sensing module uses the Φ-OTDR optical fiber sensing system to collect voiceprint signals. The Φ-OTDR optical fiber sensing system specifically adjusts the laser light source in the OTDR system to a narrow linewidth laser and adds a pulse adjuster afterwards.

3. According to the coal mill material level soundprint monitoring system based on CLDNN described in right 2, Features: The Φ-OTDR optical fiber sensing system uses the coherence characteristics of light to detect the phase change of the corresponding position at the far end of the OTDR system sensing optical fiber. The original voiceprint signal collected by the Φ-OTDR optical fiber sensing system is represented as an M×N space-time matrix X={x ij }, where M is the total number of sampling time steps, N is the number of fiber optic sensor units, and x ij is the sensor signal at the jth spatial position on the optical fiber path at the i-th moment.

4. According to the coal mill material level soundprint monitoring system based on CLDNN described in right 1, Features: The feature extraction and classification recognition module adopts the CLDNN network model, which includes CNN, LSTM and DNN modules connected in sequence. The CNN module is responsible for extracting the local time-frequency features of the voiceprint signal and capturing key acoustic patterns; the LSTM module is used to model the evolution and correlation of these features in the time dimension; the DNN module classifies different material level states based on the outputs of the first two modules.

5. According to the coal mill material level soundprint monitoring system based on CLDNN described in right 4, Features: The feature extraction and classification recognition module adopts an improved CLDNN network model. The improvement is that: In the CNN module, a progressive receptive field expansion strategy is adopted, that is, small convolution kernels are first used to extract fine-grained local features, and then larger convolution kernels are used layer by layer to fuse these local features; In the LSTM module, an attention mechanism is introduced to adaptively focus on key acoustic events at different moments. Specifically, the hidden state at the current moment is weighted summed with the hidden state at all historical moments to obtain a dynamic context vector, which is then concatenated with the input at the current moment as the input of the LSTM unit at the current moment. In the DNN module, residual connection is used, that is, the input is directly added to the output of a certain intermediate layer through identity mapping, and then sent to the subsequent layer.

6. According to the coal mill material level soundprint monitoring system based on CLDNN in right 5, Features: The improved CLDNN network model includes two LSTMs. An encoder LSTM network encodes the voiceprint signal of L historical time steps to obtain a semantically rich context vector. Then a decoder LSTM network uses the vector as the initial state to iteratively generate the material level state sequence of the next K time steps. The two LSTMs constitute a Seq2Seq network based on sequence to sequence.

7. According to the CLDNN-based coal mill material level soundprint monitoring system described in right 6, Features: The loss function of the Seq2Seq network is the mean square error between the decoder's material level output at each time step and the true future material level sequence, which is optimized through end-to-end backpropagation.

8. According to any one of the CLDNN-based coal mill material level soundprint monitoring systems described in rights 1 to 7, Features: The preprocessing module includes: Median filtering, wavelet threshold denoising, normalization, and framing and windowing.

9. According to the coal mill material level soundprint monitoring system based on CLDNN in right 8, Features: A monitoring and diagnosis module is also set up. The monitoring and diagnosis module receives the output results of the feature extraction and classification recognition module, and monitors the material level status of the coal mill in real time in combination with historical monitoring data. When an abnormal state is detected, an alarm is triggered, and the possible causes of the fault are analyzed and diagnosed.

Citation Information

Patent Citations

  • Material position detecting method and device based on vibration signal of ball grinder rotating cylinder body

    CN101173868A

  • Material level prediction method for ball mill

    CN108009514A

  • Ball mill cylinder internal load soft measurement method, device and system

    CN110580378A

  • Distributed optical fiber vibration and sound wave sensing signal identification method based on SCNN

    CN114692681A

  • Optical fiber vibration sensing equipment working in multiple modes in time-frequency domain

    CN114910155A

Cited By

  • Defect identification method and system for petroleum accessory mud pump shell based on deep learning

    CN119198925A

  • Coal mill operation monitoring system based on voiceprint recognition

    CN120618660A

  • Mixing station powder tank inventory monitoring and early warning method and system based on knocking voiceprint recognition

    CN122266143A