Coal mill fault intelligent diagnosis and early warning system

Abstract

The invention belongs to the field of artificial intelligence, particularly relates to an intelligent fault diagnosis and early warning system for a coal mill, and aims to solve the problems that in the prior art, monitoring lags behind and depends on artificial experience, and early warning lacks foresight. The system comprises a multi-source sensing data fusion module, a dynamic working condition modeling module, a fault feature extraction and evolution analysis module, an intelligent diagnosis reasoning engine module and a grading early warning decision module, and through multi-source data fusion, working condition self-adaptive modeling, fault evolution trend tracking and mixed intelligent reasoning, accurate recognition and grading early warning of early faults of the coal mill are achieved. And the operation safety and economy of the power plant are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, and particularly relates to an intelligent fault diagnosis and early warning system for a coal mill. BACKGROUND

[0002] In thermal power generation and industrial coal-fired systems, the coal mill, as a key auxiliary equipment, its running stability is directly related to the efficiency and safety of the entire combustion system. The coal mill bears the core task of grinding raw coal into coal powder that meets the requirements of combustion. Its working environment is harsh, with frequent load fluctuations, and it is in high dust, high vibration and high temperature working conditions for a long time, which is prone to mechanical wear, coal blockage, coal breakage and bearing overheating and other faults. Therefore, real-time monitoring and fault warning of the running state of the coal mill is an important link to ensure the safe and economic operation of the power plant.

[0003] Among them, the intelligent fault diagnosis and early warning technology for the coal mill aims to realize the automatic identification and trend prediction of the abnormal state of the equipment through multi-source sensor data fusion and intelligent algorithms. This technology focuses on using multi-dimensional signals such as vibration, temperature, current, sound and operating parameters, combined with data analysis models, to discover potential fault signs in advance, so as to avoid unplanned shutdown and major equipment damage.

[0004] The existing technology relies on single sensor threshold alarm or simple statistical analysis in the aspect of coal mill state monitoring, which is difficult to effectively capture early weak fault characteristics; at the same time, the traditional diagnosis model has weak generalization ability, poor adaptability to different working conditions, and lacks deep fusion mechanism for multi-source heterogeneous data, resulting in high false alarm rate and early warning lag. In complex operating environment, the fault modes are diverse and strongly coupled, and the existing system often cannot accurately distinguish between normal fluctuations and real faults, and it is also difficult to realize precise positioning of fault types and quantitative evaluation of development trend. Therefore, there is an urgent need for an intelligent fault diagnosis and early warning system for the coal mill that can fuse multi-modal operating data and has strong robustness and self-adaptive ability. SUMMARY

[0005] The purpose of the present application is to provide an intelligent fault diagnosis and early warning system for a coal mill to solve the technical problems of current thermal power plant coal mill operation state monitoring means lag, fault identification relying on manual experience, and lack of forward-looking early warning mechanism. In the prior art, the operation monitoring of the coal mill relies on the threshold alarm of single sensor signals (such as vibration, temperature, current), which is difficult to accurately reflect the complex working condition changes inside the equipment; at the same time, the fault diagnosis model generally lacks deep fusion ability for multi-source heterogeneous data, which cannot effectively identify early weak fault characteristics, resulting in high fault omission rate and frequent false alarms, which seriously restricts the safety and economy of the unit operation.

[0006] The technical solution of the present invention is a coal mill fault intelligent diagnosis and early warning system, including a multi-source sensor data fusion module, a dynamic operating condition modeling module, a fault feature extraction and evolution analysis module, an intelligent diagnostic reasoning engine module, and a hierarchical early warning decision module.

[0007] The multi-source sensor data fusion module is used to synchronously collect multi-dimensional operating parameters of the coal mill body and its auxiliary systems. These operating parameters include, but are not limited to: coal mill cylinder vibration signal, main motor current and power factor, outlet air-coal mixing temperature, primary air pressure and volume, lubricating oil pressure and temperature, reducer bearing temperature, online monitoring value of coal powder fineness, real-time feedback value of coal feed rate, and ambient dust concentration. This module performs time alignment, noise suppression, and outlier removal on the original signals, and encapsulates the processed data stream into structured data frames with a unified timestamp for subsequent modules to call.

[0008] The dynamic operating condition modeling module is used to construct a dynamic operating state model of the coal mill under different load ranges based on the structured data frames output by the multi-source sensor data fusion module. This module adopts a sliding time window mechanism to calculate the deviation between the current operating condition and the historical steady-state operating condition in real time, and generates a comprehensive representation vector of the current operating state by combining the load command change rate, coal quality fluctuation index and equipment aging coefficient. The coal quality fluctuation index is dynamically estimated from the historical statistical variance of the calorific value, moisture and ash content of the coal fed into the furnace. The equipment aging coefficient is calculated by weighting the cumulative operating hours of the equipment, the number of start-ups and shutdowns and historical maintenance records.

[0009] The fault feature extraction and evolution analysis module receives the comprehensive representation vector output by the dynamic working condition modeling module and, in conjunction with a pre-set typical fault mode library, performs multi-scale feature extraction and trend evolution analysis. First, this module reconstructs the frequency band energy distribution of the vibration signal through wavelet packet decomposition to identify the characteristic frequency bands corresponding to bearing wear, gear tooth breakage, or cylinder liner detachment. Second, it uses a sliding window cross-correlation analysis method to detect abnormal phase lag between the main motor current and the coal feed rate to determine the risk of coal blockage or coal shortage. Simultaneously, this module continuously tracks the ratio of the lubricating oil temperature to the bearing temperature rise rate; when this ratio exceeds a pre-set threshold of 0.8 for three consecutive sampling periods, it is determined that the lubrication system efficiency has decreased. All extracted feature parameters are assigned time series labels and trend fitting is performed using an exponential smoothing algorithm to generate the evolution trajectory of each fault index.

[0010] The intelligent diagnostic reasoning engine module is used to perform multi-level fault reasoning based on the evolution trajectory output by the fault feature extraction and evolution analysis module. This module has a built-in three-layer reasoning structure: the first layer is a rule engine, which directly judges feature combinations that clearly violate the safe operation boundary based on industry standards and expert experience base; the second layer is a deep neural network classifier, whose input is a normalized multi-dimensional feature vector and whose output is the probability distribution of various faults. The deep neural network adopts a convolution-long short-term memory hybrid architecture, which can simultaneously capture spatial feature correlation and temporal dependency; the third layer is a causal graph reasoning unit, which is used to analyze the causal logic of fault propagation based on the equipment structure topology and energy transfer path when multiple fault probabilities are similar, eliminate non-root cause accompanying phenomena, and finally output the unique dominant fault type and its confidence level.

[0011] The tiered early warning decision module generates differentiated early warning commands based on the dominant fault type and confidence level output by the intelligent diagnostic inference engine module, combined with the current load status of the equipment and power grid dispatch instructions. This module sets up a three-tiered early warning mechanism: Level 1 warnings correspond to severe faults with a confidence level greater than or equal to 90% and a fault evolution rate exceeding 0.5 units / hour; the system immediately triggers an audible and visual alarm and pushes a shutdown and maintenance recommendation to the operator's terminal. Level 2 warnings correspond to potential faults with a confidence level between 70% and 90% or an evolution rate between 0.2 and 0.5 units / hour; the system highlights the risky equipment on the monitoring interface and suggests scheduling inspections in the next planned shutdown window. Level 3 warnings correspond to early anomalies with a confidence level below 70% but whose characteristic indicators continuously deviate from steady state for more than 24 hours; the system only records in the background log and activates a data augmentation acquisition mode, increasing the sampling frequency to twice the original frequency to accumulate more diagnostic evidence.

[0012] Furthermore, the load range in the dynamic operating condition modeling module is divided into three ranges: low load (0% to 40% of rated load), medium load (40% to 80% of rated load), and high load (80% to 100% of rated load). Each range maintains an independent steady-state reference model, and the system automatically switches to the matching model according to the real-time load.

[0013] Furthermore, the wavelet packet decomposition layer in the fault feature extraction and evolution analysis module is set to 5 layers, generating a total of 32 frequency band sub-nodes. The energy proportion of each sub-node is used as the input feature after dimensionality reduction by principal component analysis.

[0014] Furthermore, the deep neural network classifier in the intelligent diagnostic inference engine module automatically records the deviation between the actual running results and the predicted results after each model inference, and triggers an online fine-tuning mechanism during the low-end period in the early morning of each day, using an incremental learning algorithm to update the network weights, ensuring that the model continuously adapts to equipment aging and changes in coal quality.

[0015] Furthermore, the hierarchical early warning decision module is also integrated with the power plant equipment management system. When a level one or level two early warning is triggered, it automatically generates an electronic work order containing the fault type, location, historical trend chart, and maintenance suggestions, and pushes it to the mobile terminal of the maintenance team.

[0016] Furthermore, the system operates in a collaborative architecture between edge computing nodes and the cloud. Edge nodes are responsible for real-time data acquisition, feature extraction, and initial early warning, while the cloud is responsible for model training, knowledge base updates, and cross-unit fault mode mining. The two achieve low-latency data synchronization through encrypted message queues.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention achieves early identification and accurate warning of coal mill faults by constructing an intelligent system covering the entire chain of perception, modeling, diagnosis, and decision-making. The multi-source sensor data fusion module breaks through the limitations of traditional single-point monitoring, providing a high-dimensional, synchronous, and reliable data foundation for fault diagnosis. The dynamic operating condition modeling module effectively eliminates the interference of load fluctuations on fault characteristics, significantly improving the robustness of diagnosis. The fault feature extraction and evolution analysis module not only identifies static anomalies but also focuses on the dynamic evolution trend of feature parameters, enabling the system to predict progressive faults. The intelligent diagnostic reasoning engine adopts a hybrid architecture combining rule-based, data-driven, and causal reasoning, greatly improving accuracy and interpretability while ensuring diagnostic speed. The hierarchical early warning decision-making module deeply integrates technical judgment and operation management, achieving a leap from "alarm" to "decision support," avoiding excessive intervention or delayed response. Overall, this system transforms coal mill fault diagnosis from passive response to proactive prevention, significantly reducing unplanned downtime, extending the service life of key components, and improving the safety, reliability, and economy of power plant operation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical architecture of the intelligent fault diagnosis and early warning system for coal mills proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the intelligent diagnostic reasoning engine module in this invention. Detailed Implementation

[0019] Example 1 To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be described in detail below with reference to specific embodiments. This embodiment provides an intelligent fault diagnosis and early warning system for coal mills. Its core function is to address the technical challenges of lagging monitoring of coal mill operation in thermal power plants, excessive reliance on manual experience for fault identification, and a lack of forward-looking early warning mechanisms. This system deeply integrates multi-source heterogeneous data to construct a dynamic operating condition model, enabling accurate capture and evolution trend analysis of early, subtle fault characteristics. Based on this, it performs intelligent diagnostic reasoning, ultimately providing graded and differentiated early warnings and decision support.

[0020] Please refer to Figure 1 and Figure 2 This system architecture is based on a distributed computing environment that collaborates between an edge computing node and the cloud. The edge computing node is deployed at the power plant site, responsible for high-speed, real-time acquisition of coal mill operating data and performing preliminary data fusion, feature extraction, and basic early warning tasks, ensuring low-latency response. The cloud server cluster undertakes more complex tasks, including but not limited to model training, knowledge base updates, in-depth cross-unit fault mode mining, and long-term data storage and analysis. The edge computing node and cloud server synchronize data with low latency through encrypted message queues, ensuring data security and consistency. This collaborative architecture fully leverages the real-time capabilities of edge computing and the powerful capabilities of cloud computing, providing a solid technical foundation for intelligent diagnosis and early warning of coal mill faults.

[0021] The overall technical process of the system can be summarized as follows: First, the multi-source sensor data fusion module synchronously collects and preprocesses multi-dimensional operating parameters of the coal mill body and auxiliary systems; second, the dynamic operating condition modeling module constructs and evaluates the current operating status of the coal mill in real time based on the fused data; next, the fault feature extraction and evolution analysis module identifies potential fault features from the current state and tracks their evolution trajectory; subsequently, the intelligent diagnostic reasoning engine module performs multi-level fault reasoning based on these evolution trajectories to determine the fault type and confidence level; finally, the hierarchical early warning decision module generates and issues differentiated early warning instructions based on the diagnostic results and the current operating conditions.

[0022] Multi-source sensor data fusion module The multi-source sensor data fusion module, as the front end of the system, undertakes the comprehensive sensing and basic processing of coal mill operating data. This module synchronously collects multi-dimensional operating parameters of the coal mill itself and its auxiliary systems through various industrial-grade sensors and fieldbus interfaces. These parameters include, but are not limited to: coal mill cylinder vibration signals, main motor current and power factor, outlet air-coal mixing temperature, primary air pressure and volume, lubricating oil pressure and temperature, reducer bearing temperature, online monitoring values ​​of coal powder fineness, real-time feedback values ​​of coal feed rate, and ambient dust concentration.

[0023] Regarding the specific data acquisition mechanisms, the vibration signals of the coal mill cylinder are typically acquired using triaxial accelerometers installed at multiple locations on the cylinder surface, with a sampling frequency as high as 25.6 kHz to capture high-frequency vibration characteristics. The main motor current and power factor are acquired through high-precision current and voltage transformers and then via smart meters, with a sampling period of 100 milliseconds. The outlet air-coal mixture temperature, primary air pressure, and air volume data are measured in real-time at the coal mill outlet air duct and primary air duct using dedicated industrial-grade sensors (e.g., thermocouples, differential pressure transmitters, pitot tubes), with a sampling interval of 1 second. Lubricating oil pressure and temperature, and reducer bearing temperature are acquired by pressure sensors and platinum resistance temperature sensors installed at corresponding locations, with a sampling frequency of 1 Hz. Online monitoring values ​​of coal powder fineness are obtained through a laser particle size analyzer, with a data refresh rate of 1 minute. Real-time feedback values ​​of the coal feed rate are calculated by the weighing and speed sensors integrated into the coal feeder and acquired through a programmable logic controller interface, with a sampling frequency of 10 Hz. The concentration of ambient dust is monitored by an optical dust sensor with a sampling frequency of 1 Hz.

[0024] All acquired raw signals are first aggregated to the data preprocessing unit of the edge computing node through a unified hardware interface or communication protocol (such as Modbus TCP, OPC UA). The data preprocessing unit then performs three key processing steps on these raw signals: time alignment, noise suppression, and outlier removal.

[0025] Time alignment is fundamental to ensuring synchronous analysis of multi-source data. Since different sensors may have independent sampling clocks or transmission delays, the data preprocessing unit uses a network time protocol to synchronize the timestamps of each data source, refining all data points to millisecond-level timestamps. For data with different sampling frequencies, the unit uses linear interpolation or nearest-neighbor interpolation algorithms to unify them to a preset reference sampling frequency, ensuring that all parameters have corresponding values ​​at the same time point.

[0026] Noise suppression aims to eliminate random fluctuations and interference in the data. For high-frequency vibration signals, the unit employs Kalman filtering or wavelet thresholding denoising algorithms to remove background noise based on the signal's frequency characteristics and noise distribution model. For relatively stable signals such as temperature and pressure, the unit uses moving average filtering or exponential smoothing filtering to smooth the data curve and reduce instantaneous fluctuations. Filtering parameters (e.g., the Q and R matrices of the Kalman filter, the window size of the moving average, and the smoothing coefficient of the exponential smoothing) are optimized through historical data analysis and expert experience and can be dynamically adjusted according to operating conditions.

[0027] Outlier removal aims to identify and remove outliers in the data caused by obvious errors or sensor malfunctions. The unit employs a combination of statistical methods (e.g., the 3-sigma criterion) and machine learning methods (e.g., Isolation Forest or Local Outlier Factor algorithms). For transient spikes or drops outside the statistical range, the system marks them as outliers and, depending on their severity, adopts strategies such as replacement with adjacent normal values, interpolation, or direct deletion. Persistent outliers trigger sensor malfunction warnings, prompting maintenance personnel to investigate.

[0028] The data stream, after the above processing, is encapsulated into structured data frames with a unified timestamp. Each data frame uses an efficient binary serialization format (e.g., Protocol Buffers) and contains the original timestamp, unified timestamp, sensor identifier, data value, and data quality flag for each data point. The data quality flag indicates whether the data point has undergone interpolation, filtering, or outlier removal. These structured data frames are then stored in a high-speed time-series database on the edge computing nodes for use by subsequent modules. A circular buffer is used for storage to ensure data real-time performance while managing storage space.

[0029] Dynamic working condition modeling module The dynamic operating condition modeling module receives structured data frames output by the multi-source sensor data fusion module. Its core function is to construct a dynamic operating state model of the coal mill under different load ranges and calculate the deviation between the current operating condition and the historical steady-state operating condition in real time. This module can effectively eliminate the interference of factors such as load fluctuations, coal quality changes, and equipment aging on fault characteristics, thereby improving the robustness of subsequent diagnosis.

[0030] This module uses a sliding time window mechanism to process the input data. The length of the sliding time window is configurable, typically ranging from 10 to 30 minutes, with a step size of 1 minute. Within each time window, the system performs statistical analysis on the parameters in all structured data frames, calculating temporal features such as mean, standard deviation, peak value, and kurtosis. The set of these features constitutes the instantaneous operating condition feature vector within the current time window.

[0031] The operating load of the coal mill is a key factor affecting its operating condition. For accurate modeling, this module divides the load range into three categories: low load (0% to 40% of rated load), medium load (40% to 80% of rated load), and high load (80% to 100% of rated load). Each load range maintains an independent steady-state benchmark model. The steady-state benchmark model learns the normal correlation patterns and fluctuation ranges between parameters by statistically modeling and machine learning training (e.g., based on Gaussian mixture models or support vector machine regression) on historical data of the coal mill operating in a healthy state within a specific load range over a long period. The system automatically identifies the current load range of the coal mill based on real-time main motor power or coal feed rate and switches to the corresponding steady-state benchmark model.

[0032] The system calculates the deviation between the current operating condition and the matched steady-state reference model in real time. The deviation calculation method adopts Hotelling's method from multivariate statistical process control. The method combines statistical measures or principal component analysis with control charts. (Hotelling) Statistics can comprehensively reflect the degree of deviation between multidimensional data points and the mean, taking into account the correlation between variables. The higher the calculated deviation value, the greater the deviation between the current operating conditions and the historical steady state.

[0033] In addition to deviation, this module also combines load command change rate, coal quality fluctuation index and equipment aging coefficient to generate a comprehensive representation vector of the current operating status.

[0034] The load command change rate is calculated by taking into account the slope of the load command issued by the host computer within a sliding time window. This parameter reflects the severity of load changes.

[0035] The coal quality fluctuation index is dynamically estimated from the historical statistical variance of the calorific value, moisture content, and ash content of the coal fed into the furnace. The specific calculation method is as follows: First, the system obtains the calorific value, moisture content, and ash content data of the coal fed into the furnace over the past 24 hours from the fuel management system. Next, the standard deviations of these three parameters over the past 24 hours are calculated. The coal quality fluctuation index is then defined as the weighted average of these three standard deviations, with the weights set based on experience regarding their impact on the operation of the coal mill. A higher index value indicates greater coal quality fluctuation.

[0036] The equipment aging coefficient is calculated using a weighted average of cumulative operating hours, number of start-ups and shutdowns, and historical maintenance records. Cumulative operating hours and number of start-ups and shutdowns are obtained from the coal mill's operating log. Historical maintenance records are extracted from the equipment management system, including the replacement time, maintenance type, and severity of major components (such as grinding rollers, liners, and reducers). The aging coefficient is a value between 0 and 1, where 0 represents brand-new equipment and 1 represents the end of its lifespan. This coefficient is calculated using a multinomial regression model or a neural network model, taking operating hours, number of start-ups and shutdowns, and maintenance records as inputs and outputting the aging coefficient. For example, the coefficient increases by 0.01 for every 10,000 cumulative operating hours or 500 start-ups and shutdowns; after each major component overhaul, the coefficient resets to near its initial value, but there is a baseline increase reflecting the irreversible aging of the overall structure.

[0037] Ultimately, all these parameters—the deviation between the current operating condition and the steady state, the load command change rate, the coal quality fluctuation index, and the equipment aging coefficient—together constitute a comprehensive representation vector of the current operating state. This vector is encapsulated as a fixed-length numerical array with a timestamp, serving as input for subsequent modules.

[0038] Fault Feature Extraction and Evolution Analysis Module The fault feature extraction and evolution analysis module receives the comprehensive representation vector output by the dynamic operating condition modeling module and, in conjunction with a pre-set library of typical fault modes, performs multi-scale feature extraction and trend evolution analysis. This module not only focuses on instantaneous anomalies but, more importantly, tracks the dynamic evolution trend of fault indicators, enabling the system to predict progressive faults.

[0039] This module first reconstructs the frequency band energy distribution of the vibration signal through wavelet packet decomposition. Vibration signals are one of the most direct indicators of the health status of a coal mill. The system is preset to a 5-layer wavelet packet decomposition layer, which generates 32 different frequency band sub-nodes within the signal's frequency range. The wavelet basis function chosen is the db4 wavelet from the Daubechies wavelet family, due to its good trade-off between energy concentration and smoothness. Each sub-node represents a specific frequency range; for example, the second sub-node in the third layer might correspond to a specific harmonic band of the gear meshing frequency in a reducer, while the first sub-node in the fifth layer might correspond to the impact frequency of a bearing failure. After decomposition, the energy proportion of each frequency band sub-node is calculated. For example, the sum of the squares of each sub-node's coefficients is divided by the sum of the squares of all sub-node coefficients to obtain its proportion in the total vibration energy.

[0040] To reduce feature dimensionality and highlight key features, the energy proportion of each sub-node is then subjected to dimensionality reduction processing using principal component analysis (PCA). PCA retains principal components with a variance contribution rate of over 95%, and these dimensionality-reduced features serve as vibration-related input features. The system compares these features with a pre-defined library of typical fault modes to identify the characteristic frequency bands corresponding to bearing wear, gear tooth breakage, or cylinder liner detachment. For example, a sustained increase in the energy proportion of a specific high-frequency band may indicate early damage to the bearing or gear; an increase in the energy proportion of a specific low-frequency impact band may indicate liner loosening or detachment.

[0041] Secondly, this module utilizes a sliding window cross-correlation analysis method to detect abnormal phase lag between the main motor current and the coal feed rate, thereby assessing the risk of coal blockage or coal shortage. Under normal operating conditions, there should be a stable time-series relationship between the main motor current and the coal feed rate; that is, changes in the coal feed rate will lead to corresponding changes in the motor load current, but with a normal physical lag. The sliding window cross-correlation analysis identifies the lag corresponding to the maximum cross-correlation coefficient by calculating the cross-correlation coefficient of the two time series under different lags. If the observed lag consistently deviates from the average lag under normal operating conditions (e.g., deviating by more than three times its standard deviation), especially when the lag increases significantly, the system determines that there may be a risk of coal blockage, as coal blockage increases the motor load but the coal feed rate cannot respond in time. If the cross-correlation coefficient drops sharply, it may indicate a coal shortage. The sliding window size is set to 5 minutes, with a step size of 30 seconds.

[0042] Simultaneously, this module continuously tracks the ratio of the lubricating oil temperature to the bearing temperature rise rate. The temperature rise rate is calculated as follows: within each sampling period, the linear regression slope of the current temperature value and the temperature values ​​of the previous N sampling periods is calculated to obtain the instantaneous temperature rise rate. N is typically set to 60, corresponding to a 1-minute data window. The temperature rise rate ratio is the bearing temperature rise rate divided by the lubricating oil temperature rise rate. When this ratio exceeds a preset threshold of 0.8 for three consecutive sampling periods, it is determined that the lubrication system efficiency has decreased. For example, if the bearing temperature rises rapidly while the lubricating oil temperature does not change significantly, the ratio will increase rapidly, indicating poor bearing heat dissipation or lubrication failure. The judgment mechanism that the ratio exceeds the threshold for three consecutive sampling periods is to avoid misjudgments caused by single instantaneous fluctuations and enhance the stability of the judgment.

[0043] All extracted feature parameters, including energy features after wavelet packet decomposition and dimensionality reduction, the cross-correlation analysis results of main motor current and coal feed rate, and the temperature rise rate ratio, are assigned precise time series labels. Subsequently, these feature sequences are fitted with trends using an exponential smoothing algorithm. The core of the exponential smoothing algorithm lies in smoothing time series data through weighted averaging, where recent data has a higher weight. The smoothing coefficient alpha ranges from 0 to 1, typically set between 0.1 and 0.3, with the specific value optimized by minimizing the fitting error using historical data. Exponential smoothing effectively filters out short-term fluctuations and highlights long-term trends, thereby generating the evolution trajectories of various fault indicators. These evolution trajectories are stored in time series format for further analysis by the intelligent diagnostic inference engine module.

[0044] Intelligent diagnostic reasoning engine module The intelligent diagnostic reasoning engine module is the core intelligent unit of this system. Based on the evolutionary trajectory output by the fault feature extraction and evolution analysis module, it performs multi-level fault reasoning to determine the fault type and its confidence level of the coal mill. This module has a built-in three-layer reasoning structure, combining the advantages of rule-based, data-driven, and causal analysis to achieve rapid, accurate, and interpretable diagnosis.

[0045] The first layer is the rule engine. This engine incorporates a knowledge base provided by coal mill manufacturers, industry standards, and power plant operation experts. The knowledge base contains a series of "if-then" rules, such as "If the bearing temperature rise rate ratio exceeds 0.8 for three consecutive sampling periods and the high-frequency energy of bearing vibration increases significantly, then the diagnosis is poor bearing lubrication." The rule engine performs real-time pattern matching on the input evolution trajectory data. For feature combinations that clearly violate safe operating boundaries, the rule engine can make direct judgments and immediately output preliminary diagnostic results. Rule priorities and conflict resolution mechanisms are pre-configured using expert system methods to ensure the determinism and consistency of the diagnosis. For example, when multiple rules simultaneously meet the conditions, the system will prioritize the rule with the highest priority or combine the judgments of multiple rules.

[0046] The second layer is a deep neural network classifier. This classifier employs an advanced convolutional-long short-term memory hybrid architecture. Its input is a normalized multidimensional feature vector, which consists of the instantaneous values ​​and short-term trends of various evolutionary trajectories output by the fault feature extraction and evolution analysis module. First, convolutional layers (e.g., 2 layers, each containing 64 convolutional kernels with a kernel size of 3) are used to capture spatial feature correlations in the feature vector, such as simultaneously processing multiple frequency band features in the vibration spectrum. Next, the output of the convolutional layers is input to long short-term memory network layers (e.g., 2 layers, each containing 128 memory units) to capture temporal dependencies in the fault feature evolution trajectory, such as the continuous increase or periodic fluctuation of fault features. The top of the network is a fully connected layer and a softmax activation function, outputting the probability distribution of various faults (e.g., bearing wear, gear tooth breakage, coal blockage, liner detachment, and more than 10 other typical faults). The deep neural network classifier is trained offline using massive amounts of historical fault data, including normal operation data and various fault condition data, and the network structure and hyperparameters are optimized through cross-validation.

[0047] To ensure the model can continuously adapt to equipment aging and changes in coal quality, the deep neural network classifier automatically records the deviation between the actual operating results and the predicted results after each model inference. The predicted results are compared with the fault types actually confirmed by operators to form a deviation dataset. During the daily low-load period in the early morning, i.e., when the coal mill load is low or during maintenance shutdowns, the system automatically triggers an online fine-tuning mechanism. Online fine-tuning employs an incremental learning algorithm, using the newly collected deviation dataset to make small updates to the network weights. This mechanism can prevent performance drift caused by environmental changes during long-term operation, ensuring the model continuously adapts to the actual operating conditions of the equipment. The learning rate for incremental learning is typically set to 1% to 5% of the initial training learning rate to avoid catastrophic forgetting of historical knowledge.

[0048] The third layer is the causal graph reasoning unit. This unit is activated when the probabilities of multiple faults output by the second-layer deep neural network are similar (e.g., the probability difference is less than 5%). The causal graph reasoning unit incorporates a knowledge graph of the physical topology and energy transfer paths of the coal mill. The knowledge graph uses the various components of the coal mill (e.g., motors, reducers, bearings, grinding rollers, liners) as nodes, and the physical connections and energy / signal transfer paths between them as edges. The causal graph reasoning unit analyzes the propagation logic of faults in the equipment structure through Bayesian network reasoning or structural causal models (e.g., based on Pearl's do-calculus theory). For example, a bearing fault may cause abnormal vibration in the reducer, while a reducer fault may also affect the bearing. The causal graph reasoning unit can analyze the causal dependencies between components, combine the occurrence order and intensity of different fault characteristics, eliminate non-root cause accompanying phenomena, and finally output the unique dominant fault type and its confidence level. The confidence level is quantified by the posterior probability of the Bayesian network or the strength of the causal effect in the structural causal model.

[0049] Tiered early warning decision module The tiered early warning decision-making module is the final decision-making output of this system. It generates differentiated early warning commands based on the dominant fault type and confidence level output by the intelligent diagnostic inference engine module, combined with the current load status of the equipment and power grid dispatch instructions. This module establishes a three-level early warning mechanism, aiming to achieve deep integration of technical judgment and operation management, avoiding excessive intervention or delayed response.

[0050] A Level 1 warning corresponds to a severe fault with a confidence level greater than or equal to 90% and a fault evolution rate exceeding 0.5 units / hour. The evolution rate is obtained by differentiating the exponentially smoothed trajectory of the fault characteristics. When the system detects such an emergency, it will immediately trigger a multimodal audible and visual alarm and push a real-time pop-up alarm to the monitoring interface of the coal mill operator station via the industrial control interface of the edge computing node. It will also notify relevant on-duty personnel via SMS or email. Simultaneously, the system will automatically generate a shutdown and maintenance suggestion, including the dominant fault type, the possible location of the fault, a historical evolution trend graph, the time of the fault occurrence, and preliminary maintenance guidance. This suggestion will be pushed to the operator's terminal, and an electronic work order with a unique identifier will be automatically generated. This work order contains detailed fault information and is automatically submitted to the power plant equipment management system, which will then automatically dispatch it to the corresponding maintenance team's mobile terminal for rapid response.

[0051] Level 2 warnings correspond to potential faults with a confidence level between 70% and 90% or a fault evolution rate between 0.2 and 0.5 units / hour. These faults do not pose an immediate downtime risk, but require attention and should be scheduled for inspection in the near future. When a Level 2 warning is detected, the system will highlight the coal mill equipment at risk and its relevant parameters on the monitoring interface, and display the potential fault type and confidence level in a list format. Simultaneously, the system will suggest scheduling the inspection during the next planned shutdown window (e.g., weekend maintenance, monthly shutdown). Similar to Level 1 warnings, the system will automatically generate an electronic work order, including the potential fault type, recommended components for inspection, and historical trend data, and submit it to the power plant equipment management system for distribution to the maintenance team. However, its priority is lower than that of Level 1 warning work orders so that they can be processed during planned shutdown periods.

[0052] Level 3 early warning corresponds to early anomalies with a confidence level below 70% but whose characteristic indicators have consistently deviated from steady state for more than 24 hours. These anomalies are typically in the nascent stage of a fault and lack urgency, but warrant continued monitoring. In this situation, the system will not trigger audible or visual alarms or immediate push notifications, but will only record all relevant anomalies in detail in the background log. Furthermore, the system will activate a data augmentation acquisition mode, increasing the sampling frequency of the affected sensors to twice the original frequency; for example, if the original sampling frequency is 1 Hz, it will be increased to 2 Hz. This aims to accumulate more and denser diagnostic evidence, providing richer data support for subsequent fault evolution analysis and higher-level early warnings. Simultaneously, for Level 3 early warnings, the system can also generate a low-priority electronic work order in the background, reminding maintenance personnel to pay attention to the equipment during routine inspections.

[0053] The tiered early warning decision-making module is also deeply integrated with the power plant equipment management system. When a Level 1 or Level 2 early warning is triggered, in addition to the aforementioned on-site alarms and information pushes, the system will interface with the equipment management system through a standardized application programming interface (API) to automatically generate a structured electronic work order containing the fault type, fault impact range, equipment location, relevant historical trend graphs, and preliminary maintenance suggestions. These electronic work orders are automatically dispatched to the mobile terminals of the corresponding maintenance team leaders according to preset workflow rules, realizing a fully automated process from fault discovery to work order generation and task dispatch, significantly improving maintenance response efficiency and management transparency.

[0054] This embodiment constructs a closed-loop intelligent fault diagnosis and early warning system for coal mills through the collaborative work of the aforementioned units. This system can transform coal mill fault diagnosis from a passive response mode to a proactive prevention mode, significantly reducing the number of unplanned shutdowns, effectively extending the service life of key components, and thus significantly improving the safety, reliability, and economic benefits of power plant operation.

[0055] Compared with the prior art, the main advantages of the present invention are reflected in the following aspects: Firstly, at the data perception level, this system achieves comprehensive and high-precision acquisition and preprocessing of coal mill operating parameters through a multi-source sensor data fusion module. Unlike traditional monitoring using single or limited sensors, this system integrates multi-dimensional data such as vibration, electrical, temperature, pressure, flow rate, coal powder fineness, and coal feed rate. Through time alignment, advanced noise suppression, and intelligent outlier removal, it ensures high-dimensionality, synchronization, and reliability of the data, providing a solid data foundation for subsequent fault diagnosis and overcoming the problems of inaccurate diagnosis caused by insufficient data dimensions and poor data quality in traditional monitoring methods.

[0056] Secondly, at the operational condition modeling level, the dynamic operational condition modeling module innovatively introduces load range division, coal quality fluctuation index, and equipment aging coefficient. By constructing independent steady-state benchmark models under different load ranges and calculating the deviation between the current operational condition and the benchmark model in real time, the system effectively eliminates the interference of complex and variable operating conditions of the coal mill, such as load fluctuations and coal quality changes, on fault characteristics. This enables the system to more accurately distinguish between normal operating condition fluctuations and anomalies caused by faults, significantly improving the robustness and accuracy of diagnosis and solving the problem of high false alarm rate in traditional diagnostic models when operating conditions change.

[0057] Third, at the fault analysis level, the fault feature extraction and evolution analysis module not only identifies static anomalies but also focuses on the dynamic evolution trend of fault feature parameters. Through wavelet packet decomposition for multi-scale analysis of vibration signals, combined with sliding window cross-correlation analysis for time-series correlation analysis of electrical and coal feed signals, and the evaluation of lubrication status using the temperature rise rate ratio, the system can capture the unique fingerprints of early, subtle faults. In particular, by using an exponential smoothing algorithm to fit the trends of these features, the system possesses the ability to predict progressive faults (such as bearing fatigue and liner wear), elevating fault warning from "post-occurrence alarm" to "pre-occurrence warning."

[0058] Fourth, at the diagnostic reasoning level, the intelligent diagnostic reasoning engine module adopts a three-layer hybrid reasoning architecture combining a rule engine, a deep neural network classifier, and a causal graph reasoning unit. The rule engine utilizes expert experience for rapid deterministic diagnosis; the deep neural network classifier (a convolutional-long short-term memory hybrid architecture) learns complex fault patterns from massive amounts of historical data through a data-driven approach, simultaneously capturing spatial feature correlations and temporal dependencies, providing probability distributions for various faults, and possessing online fine-tuning capabilities to adapt to equipment changes; the causal graph reasoning unit, when multiple fault probabilities are similar, performs deep causal analysis based on the equipment's physical topology and energy transfer paths, eliminating accompanying phenomena and ultimately identifying the unique root cause fault, greatly improving the accuracy, speed, and interpretability of the diagnosis. This hybrid architecture overcomes the limitations of single diagnostic methods (such as pure rule-based or pure data-driven approaches), significantly improving accuracy and reliability while ensuring diagnostic speed.

[0059] Finally, at the early warning decision-making level, the tiered early warning decision-making module deeply integrates technical judgment with power plant operation management, constructing a differentiated three-level early warning mechanism. Based on the confidence level, evolution rate, and duration of the fault, the system can intelligently determine the urgency of the early warning and adopt different response strategies, including immediate audible and visual alarms and shutdown recommendations, planned maintenance arrangements and electronic work order generation, as well as background log recording and data augmentation acquisition. This tiered mechanism effectively avoids the excessive intervention or delayed response caused by traditional "one-size-fits-all" alarms, achieving a leap from simple "alarms" to efficient "decision support." This enables power plant operators to rationally arrange maintenance plans according to the early warning level, minimizing unplanned downtime and maintenance costs, and ensuring the economic efficiency and safety of unit operation.

[0060] This system operates within a collaborative architecture of edge computing nodes and the cloud. Edge nodes are responsible for real-time data acquisition, feature extraction, and initial early warning, ensuring low latency in diagnosis and alerting. The cloud handles model training, knowledge base updates, and cross-unit fault mode mining. Both utilize encrypted message queues for efficient and secure data synchronization, providing high availability and scalability for the entire system. This distributed deployment model balances the dual requirements of real-time performance and computing power, representing an advanced solution for intelligent diagnosis and early warning in complex industries.

[0061] Example 2 This embodiment further elaborates on the online fine-tuning mechanism of the deep neural network classifier in the intelligent diagnostic inference engine module. This mechanism is a key technology to ensure that the system can continuously adapt to changes in the coal mill's operating environment, coal quality fluctuations, and the aging process of the equipment itself during long-term operation.

[0062] After each pulverizer fault inference judgment, the deep neural network classifier compares the inference result (i.e., the probability distribution of various faults) with the actual operating results of the pulverizer (the actual fault types manually confirmed by operators at the control terminal, or the fault types discovered through subsequent maintenance) to calculate the deviation between the predicted and actual results. This deviation data is structured and stored as a deviation dataset, containing the timestamp of the inference, the input feature vector, the predicted probability distribution, and the actual fault label.

[0063] The system automatically triggers an online fine-tuning mechanism during the low-load period in the early morning each day, when the coal mill load is low or it is shut down. Choosing this time period for fine-tuning aims to avoid any potential interference with normal production operations and ensure sufficient computing resources for model updates. The fine-tuning mechanism employs an incremental learning-based algorithm. Its core idea is to use the latest biased dataset to make small, gradual updates to the weights of the neural network, rather than training the entire network from scratch.

[0064] The specific steps of the incremental learning algorithm are as follows: Data selection: Select data from the most recent period (e.g., the last 24 hours or the last 1000 records) from the accumulated bias dataset as the training samples for this fine-tuning.

[0065] Loss function calculation: For the selected training samples, forward propagation is performed using the weights of the current neural network to calculate the prediction result. Then, the error between the prediction result and the actual fault label is calculated using a preset loss function (e.g., cross-entropy loss function).

[0066] Gradient descent optimization: The gradient of the loss function with respect to the network weights is calculated using the backpropagation algorithm. Then, gradient descent optimization is performed using a small learning rate (e.g., 0.1 times the initial training learning rate) to update the network weights. This learning rate is carefully chosen to ensure that the model can absorb new knowledge while maximizing the retention of its generalization ability learned from historical data, avoiding "catastrophic forgetting," where new knowledge overwrites old knowledge, leading to performance degradation.

[0067] Weight Update: The updated network weights will be immediately applied to subsequent fault reasoning tasks.

[0068] Through the aforementioned online fine-tuning mechanism, the deep neural network classifier can continuously learn from new operational data and real-world feedback, constantly optimizing its internal parameters and decision boundaries. This process enables the model to: Adapting to Equipment Aging: As coal mill components wear and degrade, their fault characteristics may subtly change. Online fine-tuning can capture these long-term changes, allowing the model to still accurately identify faults on aging equipment.

[0069] Responding to changes in coal quality: The continuous changes in the type, moisture, ash content, and other properties of the coal fed into the furnace will affect the operating characteristics of the coal mill. A fine-tuning mechanism helps the model learn normal and abnormal patterns under different coal quality conditions, reducing false alarms and missed alarms.

[0070] Addressing new failure modes: Although the system possesses a large library of failure modes, previously unseen failure types or evolutionary patterns may still emerge. Incremental learning can gradually incorporate this new knowledge, enhancing the model's ability to identify unknown or rare failures.

[0071] Improve diagnostic accuracy: Through continuous error feedback and weight updates, the diagnostic accuracy of the model in actual operation will be continuously improved, ultimately achieving a higher prediction confidence.

[0072] This online fine-tuning mechanism is a key component in realizing the intelligence and robustness of this system. It ensures that the entire intelligent diagnosis and early warning system can self-optimize and continuously evolve with the operation cycle of the coal mill and changes in the external environment, and always maintain the best diagnostic performance.

[0073] Example 3 This embodiment further elaborates on the deployment details and working mechanism of the system running in a collaborative architecture between edge computing nodes and the cloud. This hybrid architecture fully utilizes the real-time response capability of edge computing and the powerful processing capability of cloud computing, providing an efficient and reliable operating environment for the intelligent fault diagnosis and early warning system for coal mills.

[0074] Edge computing nodes Edge computing nodes are dedicated industrial computers or embedded systems deployed on-site in power plants. Each coal mill or group of coal mills is typically equipped with one or more edge computing nodes to ensure proximity of data acquisition and low latency in processing. The main responsibilities of edge computing nodes include: Real-time data acquisition: Various sensors and control systems on the coal mill are directly connected via industrial bus (e.g., Modbus, Profibus) or Ethernet interface (e.g., OPC UA, MQTT). Edge nodes have built-in multi-source sensor data fusion modules responsible for high-frequency, synchronous acquisition of all operating parameters, including coal mill cylinder vibration signals, main motor current and power factor, outlet air-coal mixing temperature, primary air pressure and volume, lubricating oil pressure and temperature, reducer bearing temperature, online monitoring values ​​of coal powder fineness, real-time feedback values ​​of coal feed rate, and ambient dust concentration. The data acquisition frequency is dynamically adjusted based on sensor type and data importance to ensure timely acquisition of critical data.

[0075] Data preprocessing: Edge nodes are equipped with preprocessing subunits of multi-source sensor data fusion modules. Immediately after data acquisition, operations such as time alignment, advanced noise suppression (e.g., Kalman filtering, wavelet denoising), and intelligent outlier removal (e.g., the Isolation Forest algorithm) are performed. These preprocessing steps are completed at the point where the data is closest to the sensor, significantly reducing the load on subsequent data transmission and cloud processing, and improving data quality.

[0076] Fault Feature Extraction: The edge node incorporates a fault feature extraction and evolution analysis module. Based on the preprocessed data stream, the edge node calculates various fault features in real time, such as the frequency band energy after wavelet packet decomposition, the cross-correlation analysis results of the main motor current and coal feed rate, and the temperature rise rate ratio. These feature calculations are completed within millisecond response times, ensuring timely detection of early fault features.

[0077] Primary Early Warning: The edge node incorporates some functions of the first-layer rule engine and hierarchical early warning decision-making module of the intelligent diagnostic inference engine module. For clear and urgent fault situations directly determined by the rule engine, or severe faults where characteristic indicators rapidly evolve to reach the first-level early warning threshold, the edge node can directly trigger local audible and visual alarms and send immediate early warning information to the on-site operator terminal. This ensures timely response to critical security events even during network latency or cloud failures.

[0078] Data Caching and Transmission: Edge nodes have a built-in high-speed time-series database for local storage of raw data and processed feature data within a specific time period. This data is then securely transmitted asynchronously to the cloud server via encrypted message queues (e.g., based on MQTT or Kafka protocols). The message queues employ a priority mechanism to ensure that urgent alarms and the latest feature data are transmitted first. In the event of a network connection interruption, the edge node can cache the data and automatically resume transmission once the network is restored, ensuring data integrity.

[0079] cloud server Cloud server clusters are deployed in remote data centers and possess powerful computing, storage, and network resources. The primary responsibility of the cloud is to perform global, computationally intensive tasks. Model training and optimization: Historical operational and fault data from all edge nodes are centrally stored in the cloud. The deep neural network classifier of the intelligent diagnostic inference engine module undergoes large-scale offline training in the cloud. This training process typically takes hours or even days, and is accelerated using graphics processing units (GPUs) to learn deeper and more complex fault modes. Furthermore, the initial model training and optimization of hyperparameters such as learning rate and network structure via the online fine-tuning mechanism are also primarily completed in the cloud.

[0080] Knowledge base updates and management: The expert knowledge base required by the rule engine, and the physical structure topology and energy transfer path knowledge graph required by the causal graph reasoning unit, are all centrally managed and updated in the cloud. When new industry standards, expert experience, or equipment maintenance data are available, the cloud can update these knowledge bases and synchronize them to edge nodes through encrypted channels to ensure that all nodes use the latest diagnostic logic.

[0081] Cross-unit fault mode mining: The cloud aggregates operational data from all coal mills, enabling cross-unit comparative analysis. By applying advanced unsupervised learning or transfer learning algorithms, the cloud can identify common fault modes or batch defects that might be difficult to detect on a single coal mill. This global analytical capability helps identify common problems and provides decision support for equipment management throughout the power plant.

[0082] Long-term data storage and analysis: All raw data and processed feature data are stored in the cloud for long-term, highly reliable operation. This data can be used for trend analysis, fault diagnosis, performance evaluation, and to provide a data foundation for future research and development. The cloud provides powerful data query and visualization tools, facilitating in-depth analysis for users.

[0083] Advanced Diagnostic Inference: The entire three-layer inference structure of the intelligent diagnostic inference engine module, especially the computationally intensive deep neural network classifier and causal graph inference unit, runs in the cloud. Edge nodes send extracted feature data to the cloud, where complex diagnostic inference is executed, and the final fault type and confidence level are generated.

[0084] Global Early Warning Decision-Making: The hierarchical early warning decision-making module integrates diagnostic results from all coal mills in the cloud, combines grid dispatch instructions with the overall power plant's operating load, and performs global early warning decisions. The cloud can generate detailed electronic work orders and push them to the power plant's equipment management system via an integrated interface, enabling plant-wide equipment maintenance and management.

[0085] Encrypted message queue Edge computing nodes and cloud servers achieve low-latency data synchronization and command delivery via encrypted message queues. The message queues employ end-to-end encryption to ensure data confidentiality and integrity during transmission. Using a publish / subscribe model, edge nodes can publish real-time collected preprocessed data and extracted feature data to specific topics, which the cloud subscribes to and processes. Conversely, cloud-trained model updates and knowledge base update commands are published to edge nodes via message queues for subscription and loading. This asynchronous communication mechanism decouples the edge from the cloud, improving system scalability and fault tolerance.

[0086] This edge-cloud collaborative architecture overcomes the limitations of a single architecture through clear division of responsibilities and efficient and secure data communication mechanisms. Edge nodes ensure real-time response, meeting the requirements of low latency and high reliability in industrial settings; the cloud provides powerful computing and storage resources, supporting complex intelligent analysis and global optimization, together constructing an efficient, intelligent, and robust intelligent fault diagnosis and early warning system for coal mills.

Claims

1. An intelligent diagnosis and early warning system for coal mill failure, characterized in that, The method comprises the following steps: A multi-source sensor data fusion module is used to synchronously collect multi-dimensional operating parameters of the coal mill body and its auxiliary system, time-align the original signals, suppress noise, and remove outliers, and encapsulate the processed data stream into a structured data frame according to a unified timestamp; A dynamic working condition modeling module is used to construct a dynamic operating state model of the coal mill under different load intervals based on the structured data frame, to calculate the deviation degree of the current working condition and the historical steady-state working condition in real time by using a sliding time window mechanism, and to generate a comprehensive representation vector of the current operating state in combination with the load instruction change rate, the coal quality fluctuation index, and the equipment aging coefficient; A fault feature extraction and evolution analysis module is used to receive the comprehensive representation vector, to perform multi-scale feature extraction and trend evolution analysis in combination with a preset typical fault mode library, to detect the phase lag abnormality of the main motor current and the coal supply by using a wavelet packet decomposition and reconstruction method, and to track the temperature rise rate ratio of the lubricating oil temperature and the bearing temperature, all of which are assigned a time sequence label and trend-fitted by an exponential smoothing algorithm to generate an evolution trajectory of each fault index; An intelligent diagnosis reasoning engine module is used to perform multi-level fault reasoning based on the evolution trajectory, which has a three-layer reasoning structure: the first layer is a rule engine that directly judges the feature combination that clearly violates the safety operation boundary according to industry standards and an expert experience library; the second layer is a deep neural network classifier whose input is a normalized multi-dimensional feature vector and whose output is a probability distribution of each fault type, and the deep neural network classifier adopts a convolution-long short-term memory hybrid architecture; the third layer is a causal diagram reasoning unit that is used to exclude non-root accompanying phenomena and finally outputs a unique dominant fault type and its confidence level when multiple fault probabilities are similar, according to the device structure topology and energy transmission path analysis of the causal logic of fault propagation; A hierarchical early warning decision module is used to generate a differentiated early warning instruction according to the dominant fault type and confidence level in combination with the current load state of the device and the power grid scheduling instruction.

2. The intelligent diagnosis and early warning system for coal mill faults according to claim 1, characterized in that, The multi-dimensional operating parameters include the coal mill cylinder vibration signal, the main motor current and power factor, the outlet air-powder mixed temperature, the primary air pressure and flow, the lubricating oil pressure and temperature, the reducer bearing temperature, the online monitoring value of coal fineness, the real-time feedback value of coal supply, and the environmental dust concentration.

3. The intelligent diagnosis and early warning system for coal mill faults according to claim 1, characterized in that, The dynamic working condition modeling module divides the load interval into three intervals of low load, medium load, and high load, and independently maintains a set of steady-state benchmark models for each interval, and the system automatically switches and matches the models according to the real-time load.

4. The intelligent diagnosis and early warning system for coal mill failure according to claim 1, characterized in that, The coal quality fluctuation index is obtained by weighted averaging the standard deviations of the calorific value, moisture content, and ash content of the coal fed into the furnace in the past 24 hours; and the equipment aging coefficient is obtained by weighted calculation through a polynomial regression model or a neural network model according to the cumulative operating hours, start-stop times, and historical maintenance records of the equipment.

5. The intelligent diagnosis and early warning system for coal mill failure according to claim 1, characterized in that, The wavelet packet decomposition layer in the fault feature extraction and evolution analysis module is set to 5 layers, and 32 frequency band sub-nodes are generated, and the energy proportion of each sub-node is used as an input feature after dimension reduction by principal component analysis.

6. The intelligent diagnosis and early warning system for coal mill failure according to claim 1, characterized in that, The temperature rise rate ratio is the bearing temperature rise rate divided by the lubricating oil temperature rise rate, and when the temperature rise rate ratio exceeds the preset threshold value 0.8 for three consecutive sampling periods, it is determined that the lubricating system efficiency is reduced.

7. The intelligent diagnosis and early warning system for coal mill failure according to claim 1, characterized in that, The deep neural network classifier automatically records the deviation between the actual operation result and the predicted result after each model inference, and triggers an online fine-tuning mechanism at the low valley period in the early morning every day to update the network weight by using an incremental learning algorithm.

8. The intelligent diagnosis and early warning system for coal mill failure according to claim 1, characterized in that, The hierarchical early warning decision module sets a three-level early warning mechanism: the first level warning corresponds to a serious fault with a confidence greater than or equal to 90% and a fault evolution rate exceeding 0.5 units / hour; the second level warning corresponds to a potential fault with a confidence between 70% and 90% or an evolution rate between 0.2 and 0.5 units / hour; and the third level warning corresponds to an early abnormality with a confidence less than 70% but the feature index deviating from the steady state for more than 24 hours.

9. The intelligent diagnosis and early warning system for coal mill fault according to claim 8, characterized in that, When the first or second level warning is triggered, the system automatically generates an electronic work order containing the fault type, location, historical trend chart and maintenance suggestion, and pushes it to the mobile terminal of the maintenance team; When the third level warning is triggered, the system starts a data enhancement collection mode, and increases the sampling frequency of the affected sensors to twice the original frequency.

Images