A coal mill internal multi-source heterogeneous data fusion early warning method and system
By using a multi-source heterogeneous data fusion early warning method for coal mills, digital antigen vectors and immune detector sets are used to identify abnormal states of coal mills. Combined with a danger signal index for collaborative judgment, this method solves the problems of insufficient data fusion and poor adaptability to operating conditions in coal mill early warning methods. It achieves accurate fault identification and control and reduces equipment maintenance costs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HEBEI GUOHUA CANGDONG POWER CO LTD
- Filing Date
- 2026-03-16
- Publication Date
- 2026-07-07
AI Technical Summary
Existing early warning methods for coal mills suffer from insufficient multi-source data fusion, poor adaptability to operating conditions, low early warning accuracy, and lack of fault tracing and precise control, leading to increased equipment maintenance costs.
A multi-source heterogeneous data fusion early warning method is adopted inside the coal mill. By acquiring structured process parameters and unstructured physical field data, a digital antigen vector is constructed. An immune detector set is used to identify abnormal states, and a hazard signal index is combined for collaborative judgment to execute a feedback control strategy.
It improves the sensitivity of early fault identification, reduces false alarm and false alarm rates, enhances adaptability to operating conditions, enables accurate fault tracing and control, and reduces equipment maintenance costs.
Smart Images
Figure CN121847312B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment fault diagnosis technology, and more specifically, to a method and system for early warning of multi-source heterogeneous data fusion inside a coal mill. Background Technology
[0002] Coal mills are key equipment in industries such as thermal power generation and metallurgy, and their operating status directly affects production efficiency and system safety. The internal working conditions of coal mills are complex, and they are prone to failures such as coal blockage, grinding roller wear, and metal impact. Early failure characteristics are hidden and affected by multiple coupled factors, so reliable condition monitoring and early warning technologies are required.
[0003] In existing technologies, one method for early warning of coal mills relies solely on threshold judgment based on structured process parameters (such as mill outlet temperature and coal feed rate). This method depends on a single data dimension and cannot capture early fault characteristics reflected by unstructured physical field signals such as mechanical vibration and impact within the coal mill, resulting in incomplete fault characterization and insufficient early warning sensitivity. Another technology uses an anomaly detection algorithm with a fixed threshold, identifying anomalies through a preset parameter range. However, the operating load of a coal mill often fluctuates with operating conditions, and the fixed threshold cannot adapt to changes in the normal fluctuation range of parameters caused by load variations. This leads to a significant increase in false alarm rate during high load fluctuations and a prominent problem of missed alarms during low load conditions. Furthermore, most existing technologies can only achieve fault warning, lacking the ability to accurately trace the fault location and failing to provide feedback control strategies that match the fault type, thus failing to promptly curb fault development and increasing equipment maintenance costs.
[0004] In view of this, a method and system for early warning of multi-source heterogeneous data fusion inside a coal mill is proposed. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for early warning of multi-source heterogeneous data fusion inside a coal mill, so as to solve the technical problems of insufficient multi-source data fusion, poor adaptability to operating conditions, low early warning accuracy, and lack of fault tracing and precise control in existing early warning methods for coal mills.
[0006] To solve the above-mentioned technical problems, the present invention provides a method for early warning of multi-source heterogeneous data fusion inside a coal mill, comprising the following steps:
[0007] S1. Acquire multi-source heterogeneous data during the operation of the coal mill, wherein the multi-source heterogeneous data includes structured process parameter data and unstructured physical field data;
[0008] S2. Preprocess and extract features from the multi-source heterogeneous data to construct a digital antigen vector representing the current operating status of the coal mill;
[0009] S3. Construct a self-space based on the historical healthy operation data of the coal mill, and generate an immune detector set covering the non-self-space based on the negative selection principle; the immune detector set is used to identify abnormal states that deviate from the self-space.
[0010] S4. Calculate the Euclidean distance between the digital antigen vector at the current moment and each detector in the immune detector set in real time, and calculate the danger signal index of the coal mill operation;
[0011] S5. Based on the immune recognition result of whether the digital antigen vector is recognized by the immune detector set and the danger signal index, a joint determination is made. When the digital antigen vector is recognized by the immune detector set and the danger signal index exceeds the preset threshold, an early warning signal is triggered and the corresponding feedback control strategy is executed.
[0012] As a further improvement to this technical solution, in step S1, the structured process parameter data includes at least the mill bowl differential pressure, the coal mill current, the hydraulic oil station oil pressure, the mill outlet temperature, and the coal feed rate; the unstructured physical field data includes at least the vibration acceleration signal on the surface of the coal mill cylinder and the acoustic emission signal inside the coal mill.
[0013] As a further improvement to this technical solution, the process of constructing the digital antigen vector in step S2 is as follows:
[0014] For the structured process parameter data, statistical features within a preset time window are calculated. The statistical features include at least one of mean, variance, skewness, and kurtosis, generating a first type of feature component that reflects the changes in process parameters.
[0015] For the vibration acceleration signal and acoustic emission signal, wavelet packet decomposition is used to extract the energy spectrum of different frequency bands, and the Mel frequency cepstral coefficients are calculated to generate a second type of characteristic component that reflects the change of physical state.
[0016] After normalizing the first type of feature components and the second type of feature components, feature fusion mapping is performed by concatenating and splicing them to form the digital antigen vector.
[0017] As a further improvement to this technical solution, step S3, which involves generating an immune detector set covering the non-self space based on the negative selection principle, includes:
[0018] Define the set of digital antigen vectors under the healthy state of a coal mill as its own set;
[0019] Candidate detectors are randomly generated in the normalized feature space, and each candidate detector is defined as a hypersphere with a center vector and a recognition radius.
[0020] The Euclidean distance between the candidate detector and all sample points in its own set is calculated using the real-valued negation selection algorithm.
[0021] If a candidate detector covers any autologous sample point, the candidate detector is discarded; if a candidate detector does not cover any autologous sample point, it is retained and added to the immune detector set.
[0022] As a further improvement to this technical solution, the recognition radius adopts a dynamic adaptive adjustment strategy, including:
[0023] Obtain the real-time load command of the coal mill and calculate the standard deviation of the load command within a preset sliding time window. Define the standard deviation as the load fluctuation index. Based on formula Calculate the recognition radius of the candidate detector at the current time. ,in, This indicates the preset reference radius of the coal mill under steady-state operating conditions. This represents the preset radius adjustment gain coefficient; the length of the preset sliding time window and the radius adjustment gain coefficient. Determined through calibration using historical operational data;
[0024] When the coal mill is in an unsteady operating condition, the load fluctuation index... When the radius is increased, the calculated recognition radius It increases accordingly;
[0025] When the coal mill is in steady-state operation, the load fluctuation index is... When reduced, the calculated recognition radius It decreases accordingly.
[0026] As a further improvement to this technical solution, in step S4, the method for calculating the danger signal index is as follows:
[0027] The specific energy consumption of the coal mill, the amount of slag discharged, and the deviation of the outlet powder pipe wind speed were selected as hazardous indicators to characterize the degree of damage to the equipment performance.
[0028] Calculate the deviation of the current collected value of each of the above-mentioned hazard indicators from the rated reference value;
[0029] Based on preset weighting coefficients, the deviation of each hazard indicator is weighted and summed to obtain the hazard signal index; wherein, the weighting coefficients are determined based on the correlation analysis between historical fault data and operational data.
[0030] As a further improvement to this technical solution, the logic for collaborative determination in step S5 is as follows:
[0031] Set danger threshold The danger threshold Determined based on statistical values of the danger signal index from historical fault data;
[0032] If the Euclidean distance between the current digital antigen vector and any immune detector is less than the recognition radius of that detector. If so, an immune recognition event is determined to have occurred;
[0033] Only when an immune recognition event occurs, and the calculated danger signal index is greater than [value missing] When this occurs, it is determined to be a valid intrusion, and a fault warning is generated.
[0034] As a further improvement to this technical solution, after triggering the warning signal in step S5, a fault tracing step S6 is also included:
[0035] Based on the mechanistic model of the coal mill or the feature analysis of historical fault data, a mapping table between the feature dimensions of the digital antigen vector and the physical components of the coal mill is pre-established.
[0036] From all the immune detectors that recognize the current digital antigen vector, the one with the smallest Euclidean distance is selected as the winning immune detector;
[0037] Calculate the component distances between the center vector of the superior immunodetector and the current digital antigen vector in each feature dimension, and identify the abnormal feature dimensions with the highest contribution of component distances.
[0038] Based on the mapping table, find the physical component or parameter source corresponding to the abnormal feature dimension and output the fault location result.
[0039] As a further improvement to this technical solution, the feedback control strategy in step S5 includes at least one of the following:
[0040] When the warning type is coal blockage risk, the feeder speed is automatically reduced and the primary air pressure is increased;
[0041] When the warning type is abnormal roller loading, the proportional relief valve setting of the hydraulic loading system is automatically adjusted;
[0042] When the warning type is "metal impact noise", an emergency trip signal is triggered.
[0043] A multi-source heterogeneous data fusion early warning system for the inside of a coal mill, the system being used to implement the aforementioned multi-source heterogeneous data fusion early warning method for the inside of a coal mill, comprising:
[0044] The data acquisition module is used to collect structured process parameter data and unstructured physical field data of the coal mill in real time;
[0045] The feature construction module, connected to the data acquisition module, is used to preprocess and extract features from the acquired data to construct a digital antigen vector characterizing the current operating status of the coal mill.
[0046] The model training module, connected to the data acquisition module, is used to generate an immune detector set and maintain the self-space based on historical health data and the principle of negative selection.
[0047] The collaborative monitoring module is connected to the model training module and the feature construction module respectively. It is used to calculate the Euclidean distance between the digital antigen vector and the immune detector set, as well as the danger signal index of the coal mill operation, and to execute collaborative judgment logic based on danger theory.
[0048] The early warning and control module, connected to the collaborative monitoring module, is used to output early warning information when the collaborative judgment result is abnormal, and to send adjustment commands to the coal mill control system according to the fault type.
[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0050] 1. The multi-source heterogeneous data fusion early warning method and system for coal mills integrates structured process parameters and unstructured physical field data to construct digital antigen vectors, which can capture multi-dimensional features of early coal mill faults, solve the problem of insufficient representation by a single data dimension, improve the sensitivity of early fault identification, and enhance the comprehensiveness of feature representation.
[0051] 2. In the multi-source heterogeneous data fusion early warning method and system inside the coal mill, the recognition radius of the immune detector is adjusted based on the load fluctuation index, so that the abnormal detection threshold can match the parameter fluctuation range under different working conditions, thereby reducing the false alarm rate and false alarm rate caused by load fluctuation and enhancing the adaptability to working conditions.
[0052] 3. In the multi-source heterogeneous data fusion early warning method and system inside the coal mill, invalid abnormal signals are filtered through dual judgment logic to ensure the reliability of the early warning results, avoid the limitations of a single judgment method, and improve the accuracy of the early warning.
[0053] 4. In the multi-source heterogeneous data fusion early warning method and system inside the coal mill, the fault source is located by mapping the feature dimension to the physical component, and a precise control strategy is executed in combination with the fault type to effectively curb the development of the fault and reduce the equipment maintenance cost. Attached Figure Description
[0054] Figure 1 This is a flowchart illustrating the overall method of the present invention;
[0055] Figure 2 This is an overall system block diagram of the present invention. Detailed Implementation
[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0057] Example 1: Currently, as the core equipment of the pulverizing system in coal-fired power plants, the operating status of the coal mill directly affects the stability of boiler combustion and the economy of the unit. Due to the harsh operating conditions of high temperature, high vibration and heavy load for a long time, the coal mill is prone to failures such as coal blockage, grinding roller wear and internal foreign object impact. In the existing technology, whether it is based on the threshold judgment of structured process parameters (such as mill outlet temperature and coal feed rate) or the anomaly detection algorithm with fixed threshold, there are technical problems such as insufficient multi-source data fusion, poor adaptability to operating conditions, low early warning accuracy and lack of fault tracing and precise control.
[0058] In view of this, please refer to Figure 1 As shown, one of the objectives of this invention is to provide a method for early warning of multi-source heterogeneous data fusion inside a coal mill, the method comprising the following steps:
[0059] S1. Acquire multi-source heterogeneous data during the operation of the coal mill, wherein the multi-source heterogeneous data includes structured process parameter data and unstructured physical field data;
[0060] S2. Preprocess and extract features from the multi-source heterogeneous data to construct a digital antigen vector representing the current operating status of the coal mill;
[0061] S3. Construct a self-space based on the historical healthy operation data of the coal mill, and generate an immune detector set covering the non-self-space based on the negative selection principle; the immune detector set is used to identify abnormal states that deviate from the self-space.
[0062] S4. Calculate the Euclidean distance between the digital antigen vector at the current moment and each detector in the immune detector set in real time, and calculate the danger signal index of the coal mill operation;
[0063] S5. Based on the immune recognition result of whether the digital antigen vector is recognized by the immune detector set and the danger signal index, a joint determination is made. When the digital antigen vector is recognized by the immune detector set and the danger signal index exceeds the preset threshold, an early warning signal is triggered and the corresponding feedback control strategy is executed.
[0064] This embodiment describes the steps of a multi-source heterogeneous data fusion early warning method for a medium-speed coal mill (such as the HP1003 type medium-speed coal mill), as follows:
[0065] Considering that the formation and development of coal mill faults can simultaneously affect process parameters and internal physical field states, acquiring only a single type of data cannot fully reflect the equipment's operating status. Therefore, it is necessary to simultaneously collect structured process parameter data and unstructured physical field data. Thus, step S1 involves acquiring multi-source heterogeneous data; the specific operation process is as follows:
[0066] Structured process parameter data are collected through the existing monitoring system of the coal mill, including mill bowl differential pressure, coal mill current, hydraulic oil station oil pressure, mill outlet temperature, and coal feed rate. Differential pressure is collected using a pressure sensor with a range of 0–10 kPa and a sampling frequency of 1 Hz; mill current is collected using a current transformer with a range of 0–500 A and a sampling frequency of 1 Hz; hydraulic oil station oil pressure is collected using a pressure sensor with a range of 0–30 MPa and a sampling frequency of 1 Hz; mill outlet temperature is collected using a temperature sensor with a range of 0–150 °C and a sampling frequency of 1 Hz; and coal feed rate is collected using a weighing feeder's metering module with a range of 0–60 t / h and a sampling frequency of 1 Hz.
[0067] Unstructured physical field data were acquired using additionally deployed sensors, including vibration acceleration signals and acoustic emission signals from the surface of the coal mill cylinder. Two piezoelectric vibration acceleration sensors, with a range of 0–50g and a sampling frequency of 50kHz, were symmetrically arranged on both sides of the coal mill cylinder; two acoustic emission sensors, with a range of 0–100dB and a sampling frequency of 1MHz, were arranged at the mill roller bearing housing. All sensor signals were converted into digital signals by a data acquisition card and then transmitted to the data processing unit.
[0068] It should be noted that the range and sampling frequency of the sensor are determined based on the rated operating parameters and fault characteristic frequencies of the HP1003 coal mill. For example, the vibration frequency of the coal mill during normal operation is mainly concentrated in the range of 10 to 1000 Hz, while an impact signal of 2000 to 5000 Hz may occur during a fault. Therefore, the sampling frequency of the vibration sensor is set to 50 kHz to ensure that the Nyquist sampling theorem is satisfied and to avoid signal distortion.
[0069] Through the above operations, multi-source data were collected synchronously, covering multiple dimensions of information such as the operating parameters, mechanical vibration, and internal impact of the coal mill, providing a data foundation for subsequent comprehensive characterization of the equipment status.
[0070] Considering the significant differences in dimensionality and type among multi-source heterogeneous data, direct fusion would lead to feature redundancy or loss of effective information. Therefore, it is necessary to construct a unified feature vector that accurately represents the device status through targeted feature extraction and fusion methods. Thus, step S2, digital antigen vector construction, is described below:
[0071] Feature extraction of structured process parameter data: A preset time window of 5 seconds is set (the 5-second time window length is based on historical fault data statistics; this length captures short-term parameter trends while avoiding excessive feature fluctuations due to a short window), and a sliding step size of 1 second. The mean, variance, skewness, and kurtosis of each parameter within each time window are calculated. Taking the mill bowl differential pressure as an example, if the collected data within a certain time window is [3.2 kPa, 3.3 kPa, ..., 3.1 kPa] (a total of 5 data points), the calculated mean is 3.2 kPa, variance is 0.005 kPa², skewness is 0.1, and kurtosis is 2.8, forming four feature components. Similarly, the mill current, hydraulic oil station pressure, mill outlet temperature, and coal feed rate each generate four feature components, for a total of 20 first-class feature components.
[0072] Feature extraction from unstructured physical field data: Vibration acceleration signals and acoustic emission signals are subjected to 3-level wavelet packet decomposition (the number of decomposition levels is determined based on the signal's frequency range to ensure coverage of the main frequency bands of fault features), resulting in 8 frequency bands (1–3.125kHz, 3.125–6.25kHz, ..., 37.5–40.625kHz). The energy value of each frequency band is calculated, yielding 8 energy features. Mel-frequency cepstral coefficients (MFCCs) are then calculated on the decomposed signals, extracting 12 MFCC coefficients and 1 frame energy (12 MFCC coefficients are a commonly used effective dimension in speech signal processing, capable of capturing nonlinear frequency features), for a total of 13 features. Each signal generates 8 + 13 = 21 feature components, resulting in a total of 42 second-type feature components for both signals.
[0073] Feature fusion and normalization: employing The normalization method maps the first and second type feature components to the [0,1] interval, respectively. The normalization formula is as follows: ,in, These are the original eigenvalues. and These are the minimum and maximum values of the feature in historical health data, respectively; the 20 normalized first-class feature components and 42 second-class feature components are concatenated in sequence to form a 62-dimensional digital antigen vector;
[0074] Through the above operations, the trend characteristics of structured data and the frequency domain characteristics of unstructured data are fully preserved by categorical feature extraction. Normalization process eliminates the differences in data dimensions. The digital antigen vector formed by cascading and fusion can comprehensively and uniformly represent the operating status of the coal mill, providing a high-quality feature foundation for subsequent anomaly identification.
[0075] Considering the negative selection principle of the biological immune system, specific identification of fault states can be achieved by constructing a self-space representing healthy states and an immune detector set covering abnormal states. However, it is necessary to address the poor adaptability of fixed detector parameters under fluctuating operating conditions. Therefore, step S3 involves the construction of the self-space and immune detector set; the specific operation process is as follows:
[0076] Self-space construction: Select the fault-free historical data of HP1003 coal mill running continuously for 30 days. During this period, the coal mill load is stable at 70% to 100% of the rated load (50t / h), and there is no excessive fluctuation of any operating parameters. There are zero fault records. 10,000 samples are extracted from this historical data at 1-second intervals. Each sample is processed in step S2 to form a 62-dimensional digital antigen vector, which forms a self-set, i.e., self-space.
[0077] Generation of immune detector sets:
[0078] 500 candidate detectors are randomly generated in the [0,1] normalized feature space. Each candidate detector is defined as a hypersphere containing a center vector (62-dimensional) and a recognition radius. ;
[0079] The real-valued negation selection algorithm is used to calculate the Euclidean distance between the center vector of each candidate detector and all 10,000 sample points in its own set. The distance calculation formula is as follows: ( =1 to 62, The center vector of the candidate detector is the dimensional components, For the self-sample vector of the th (dimensional components).
[0080] If the Euclidean distance between a candidate detector and any self-sample point is less than its recognition radius If the detector is found to cover the self-sample, it is discarded; if the Euclidean distance to all self-sample points is greater than or equal to... If the detector is selected, it will be retained and added to the immune detector set, resulting in approximately 300 effective immune detectors.
[0081] If the self-set is dominated by nonlinear properties, then the formula for calculating the kernel distance between the candidate detector and all sample points in the self-set using the real-valued negation selection algorithm is: ; where, in the formula, Candidate detectors With self-sample nuclear distance, Indicate candidate detector The inner product of itself in higher-dimensional space. Indicate candidate detector With self-sample In the inner product of higher-dimensional space, Represents self-sample The inner product of itself in a higher-dimensional space;
[0082] If a candidate detector covers any autologous sample point, the candidate detector is discarded; if a candidate detector does not cover any autologous sample point, it is retained and added to the immune detector set.
[0083] Dynamic adjustment of recognition radius:
[0084] Set the preset sliding time window length to 10 seconds, acquire the real-time load command of the coal mill, calculate the standard deviation of the load command within this window, and define it as the load fluctuation index. ;
[0085] Set the preset reference radius under steady-state conditions =0.08, radius adjustment gain coefficient =0.3, this parameter was determined through calibration using 200 sets of historical load fluctuation data: three sets of data were selected with load fluctuation ranges of ±5% (steady state), ±5% to ±15% (transient state), and above ±15% (fluctuation state), and different values were tested respectively. False alarm rate at value, when When the value is 0.3, the false alarm rate is low under all operating conditions (the experimentally measured false alarm rate is less than 3%), which meets the usage requirements;
[0086] The formula for calculating the identification radius is as follows: For example, when the load fluctuation index When = 0.05 (steady state), ;when When =0.2 (wave dynamics), ;
[0087] Through the above steps, the self-space is constructed based on a large amount of health data, ensuring accurate representation of normal operating conditions; the immune detector set is generated through a negative selection algorithm, which can specifically identify abnormal features that deviate from the healthy state; the dynamically adjusted recognition radius enables the detector to adapt to different load fluctuation conditions, maintain high recognition accuracy in steady state, expand the recognition range in fluctuating dynamic state, avoid false alarms and missed alarms, and improve the working condition adaptability of the algorithm.
[0088] Since immune recognition alone may misclassify deviations from normal parameters caused by fluctuations in operating conditions as abnormal, it is necessary to introduce a hazard signal index that reflects the degree of damage to equipment performance to effectively screen abnormal signals. Therefore, steps S4, Euclidean distance calculation and hazard signal index calculation, are performed as follows:
[0089] Euclidean distance calculation: Real-time acquisition of coal mill operating data, processing in step S2 to generate a digital antigen vector at the current moment, calculating the Euclidean distance between this vector and the center vector of each detector in the immune detector set, and recording the distance values of all detectors;
[0090] Danger signal index calculation: The coal mill unit consumption, slag discharge rate and outlet powder pipe wind speed deviation are selected as danger indicators. Among them, the coal mill unit consumption reflects the degree of energy loss, the slag discharge rate reflects the grinding efficiency, and the outlet powder pipe wind speed deviation reflects the degree of material conveying abnormality. It should be noted that the selection of danger indicators is based on the coal mill failure mechanism analysis. The unit consumption, slag discharge rate and outlet powder pipe wind speed deviation are the three parameters that change most significantly during the failure development process.
[0091] The rated baseline values for each hazardous indicator are determined as follows: the rated baseline value for unit consumption is 30 kWh / t, the rated baseline value for slag discharge is 5 t / h, and the rated baseline value for outlet powder pipe wind speed is 25 m / s. These baseline values are determined based on the design parameters and rated operating data of the HP1003 coal mill.
[0092] Calculate the deviation of each indicator: The formula for calculating the deviation is as follows: (The deviation calculation formula uses relative deviation to avoid the imbalance of indicator weights caused by absolute numerical differences.) This is the current collected value. This is the rated baseline value, for example, when the current unit consumption is 33 kWh / t. When the current slag discharge rate is 6t / h, When the current value of the outlet powder pipe air velocity is 22 m / s, .
[0093] Weighting coefficients were determined based on Pearson correlation analysis of historical fault data and operational data (the correlation analysis ensured that the danger signal index accurately reflected the severity of the fault). The correlation coefficient between unit consumption and fault severity was 0.85, slag discharge rate was 0.72, and outlet powder pipe velocity deviation was 0.78. Therefore, the weighting coefficients were set as follows: =0.4、 =0.3、 =0.3, the weighting coefficient satisfies .
[0094] The formula for calculating the danger signal index is as follows: Substituting the example data above, we get ;
[0095] Through the above operations, Euclidean distance calculation provides a quantitative basis for immune identification. The danger signal index integrates key indicators related to equipment performance and can effectively distinguish between normal deviations caused by operating condition fluctuations and abnormal deviations caused by faults, providing a dual quantitative standard for subsequent collaborative judgment.
[0096] By using the synergistic judgment of immune recognition results and danger signal indices to filter out invalid abnormal signals and ensure the accuracy of early warnings, and by implementing targeted feedback control for different fault types to promptly curb fault development, step S5, the synergistic judgment and feedback control strategy, is as follows:
[0097] Collaborative decision-making logic:
[0098] Set danger threshold =0.65. This threshold is determined based on the statistical analysis of the danger signal index of 50 historical fault data. The 95th percentile of the danger signal index at the time of the fault is taken to ensure that more than 95% of the faults can be effectively identified, while reducing false alarms.
[0099] If the Euclidean distance between the current digital antigen vector and any immune detector is less than the recognition radius of that detector. If an immune recognition event is detected, then an immune recognition event is determined to have occurred; if a danger signal index is also present... If an immune recognition event occurs, it is considered a valid intrusion, and a fault warning is generated; if only an immune recognition event occurs but... or However, if no immune recognition event occurs, no warning will be generated; only abnormal signals will be recorded.
[0100] Explanation of the collaborative judgment principle: The negative selection algorithm of the artificial immune system can effectively detect abnormal patterns that deviate from the healthy state, but it may respond to some operating condition fluctuations without serious consequences (i.e., "false positives"). Therefore, this embodiment introduces a danger signal index as a second criterion. This index is derived from macroscopic parameters that directly reflect the performance degradation of the equipment (such as unit consumption and slag discharge). Its core idea is: only when the immune system detects an abnormality and the equipment simultaneously shows a clear performance impairment signal can it be determined as a valid fault intrusion that requires warning. This dual verification mechanism significantly improves the accuracy and reliability of the warning.
[0101] Feedback control strategy:
[0102] Once a fault warning is triggered, and after the subsequent fault tracing in step S6 is completed and a precise fault location result (such as "coal blockage risk") is output, the system will execute or fine-tune the corresponding precise feedback control strategy, for example:
[0103] Coal blockage risk warning: When the warning type is coal blockage risk, the coal feeder speed is automatically reduced from the current speed to 80% of the rated speed (for example, when the rated speed is 1000 r / min, it is reduced to 800 r / min). At the same time, the primary air pressure is increased from the current air pressure to 110% of the rated air pressure (for example, when the rated air pressure is 6 kPa, it is increased to 6.6 kPa). By reducing the coal feed rate and enhancing ventilation, the coal blockage trend is alleviated.
[0104] Grinding roller loading abnormality warning: When the warning type is grinding roller loading abnormality, the proportional relief valve setting value of the hydraulic loading system is automatically adjusted according to the feedback signal of the hydraulic oil station oil pressure. If the oil pressure is higher than the rated value (for example, rated 15MPa), the setting value is reduced by 1 to 2MPa; if the oil pressure is lower than the rated value, it is increased by 1 to 2MPa to ensure that the grinding roller loading force is within the normal range.
[0105] Metal impact noise warning: When the warning type is metal impact noise, it is judged as a serious fault and an emergency trip signal is immediately triggered to cut off the power supply to the coal feeding system and the coal mill drive to avoid further damage to the equipment.
[0106] Through the above operations, the collaborative judgment logic reduces the false alarm rate and false negative rate of the single judgment method by filtering under dual conditions, and improves the reliability of the early warning results. The targeted feedback control strategy can take precise adjustment measures according to the fault type. The risk of coal blockage is mitigated by adjusting parameters, abnormal loading is repaired by calibrating the hydraulic system, and serious faults are prevented from losing money by tripping protection, thus realizing the hierarchical processing and effective control of faults.
[0107] Considering that existing technologies can only provide fault warnings but cannot pinpoint the exact location of the fault, making it difficult for maintenance personnel to quickly locate the problem and prolonging downtime, it is necessary to establish a mapping relationship between feature dimensions and physical components to achieve accurate fault tracing. Therefore, step S6, the fault tracing step, is as follows:
[0108] Mapping Relationship Table Construction: Based on the coal mill mechanism model and historical fault data feature analysis, a mapping relationship table between the feature dimensions of digital antigen vectors and physical components is constructed. Specific construction methods include:
[0109] The first step is data preparation: collect historical cases containing various typical failures (such as coal blockage, grinding roller wear, metal impact, etc.) and their corresponding digital antigen vector samples;
[0110] The second step is feature importance analysis: using feature importance analysis algorithms (such as feature gain calculation based on tree models or filtering feature selection methods) to quantify the contribution of each feature dimension to distinguishing different fault types;
[0111] The third step is knowledge mapping: combining domain knowledge (for example, the time-domain statistical characteristics of the differential pressure of the grinding bowl directly reflect the internal material level, and the vibration energy of a specific frequency band is related to the harmonics of the rotation frequency of the grinding roller), the high-contribution feature dimensions are classified and mapped to potential faulty components or root cause parameters.
[0112] The fourth step is to form a mapping relationship table: the above mapping relationships are stored in a structured manner to form a queryable mapping relationship table;
[0113] Taking the HP1003 coal mill as an example, the mapping relationship of some key features is explained:
[0114] If the feature dimension category is the time domain feature of process parameters, and the specific features are the variance and kurtosis of the mill bowl differential pressure, then the potential faulty component / root cause is the material level inside the coal mill, and the main associated fault modes are coal blockage and uneven coal feeding.
[0115] If the feature dimension category is the frequency domain feature of vibration signal, specifically the energy in the 1000-2000Hz frequency band, then the potential faulty components / root causes mapped are grinding rollers and grinding disc liners, and the main associated fault modes are wear and detachment.
[0116] If the feature dimension category is acoustic emission signal feature, and the specific feature is MFCC coefficient 1 to 3, then the mapped potential faulty component / root cause is internal metal component, and the main associated fault modes are impact and crack.
[0117] If the feature dimension category is hydraulic parameter features, specifically the mean oil pressure and volatility, then the potential faulty component / root cause is the hydraulic loading system, and the main associated fault modes are relief valve failure and leakage.
[0118] Subsequent fault tracing will be based on this mapping table; it should be further noted that the above mapping table is an example, and the actual mapping relationship needs to be determined based on the specific device model and historical data.
[0119] Selection of the superior immune detector: From all immune detectors that identify the current digital antigen vector, the detector with the smallest Euclidean distance to the current vector is selected as the superior immune detector. This detector has the highest specificity for identifying the current abnormal features.
[0120] Anomaly Feature Dimension Identification: Calculate the component distances between the center vector of the superior immunodetector and the current digital antigen vector across 62 feature dimensions. The formula for calculating the component distances is as follows: ( =1 to 62), calculate the proportion (contribution) of each component distance to the total Euclidean distance, and select the top 3 feature dimensions with the highest contribution as the anomalous feature dimensions; for example, if the total Euclidean distance is 0.09, the component distance contribution of the differential pressure variance of the grinding bowl is 45%, and the component distance contribution of the vibration acceleration energy in the 1000-2000Hz frequency band is 30%, then these two dimensions are determined to be the main anomalous features;
[0121] Fault location output: Based on the mapping relationship table, find the physical component corresponding to the abnormal feature dimension and output the fault location result; for example, if the abnormal feature dimension is the differential pressure variance of the grinding bowl, then output "abnormal material level inside the coal mill, suspected coal blockage"; if the abnormal feature dimension is the vibration acceleration energy in the 1000-2000Hz frequency band, then output "grinding roller wear fault".
[0122] Through the above operations, the mapping relationship table established by using mechanism analysis and historical data training achieves a precise correspondence between feature dimensions and physical components; the selection of superior immune detectors and component distance contribution analysis can focus on the main abnormal features and improve the accuracy of fault tracing; the fault location results provide maintenance personnel with a clear direction for repair, shorten equipment downtime for maintenance, and reduce maintenance costs.
[0123] Example 2: This example shares the same core technical solution as Example 1, except that some parameter thresholds are adjusted based on the structural characteristics and operating parameters of the MPS250 coal mill. The specific differences are as follows:
[0124] In step S1, the structured process parameters are increased to include “pressure difference between inlet and outlet of coal mill” (range 0-5 kPa, sampling frequency 1 Hz). In the unstructured physical field data, the vibration sensor is placed at the fan impeller bearing seat and the sampling frequency is adjusted to 60 kHz (because the fan impeller of MPS250 coal mill has a higher speed, the fault characteristic frequency can reach 20 kHz).
[0125] In step S2, the time window for the structured process parameters is adjusted to 8 seconds (the MPS250 coal mill has a relatively slow parameter response speed), the number of wavelet packet decomposition layers is adjusted to 4 layers (to cover higher fault feature frequencies), and the digital antigen vector dimension is 68 dimensions (the newly added inlet and outlet pressure difference contributes 4 statistical features).
[0126] In step S3, the number of self-set samples is 12,000 (based on longer fault-free operation data of the MPS250 coal mill), and the baseline radius is... =0.09, radius adjustment gain coefficient =0.35 (calibrated with 200 sets of load fluctuation data, false alarm rate is less than 3%)
[0127] In step S4, the rated baseline values for the hazardous indicators are adjusted as follows: unit consumption 35 kWh / t, slag discharge 8 t / h, outlet powder pipe wind velocity 28 m / s, and the weighting coefficient is adjusted as follows. =0.45、 =0.25、 =0.3 (Correlation analysis of failure mechanism based on MPS250 coal mill);
[0128] In step S5, the danger threshold =0.7 (based on historical data of 50 failures of MPS250 coal mill), the feeder speed adjustment range in the feedback control strategy is to reduce to 75% of the rated speed and increase the primary air pressure to 115% of the rated air pressure (to meet the ventilation requirements of MPS250 coal mill).
[0129] This embodiment maintains the universality of the core technical solution by adjusting the parameter thresholds adapted to the MPS250 coal mill, while ensuring the accuracy and adaptability of early warning on different types of coal mills, thus verifying the scalability of the technical solution of the present invention.
[0130] Example 3: To implement the methods of Examples 1 and 2, please refer to... Figure 2 As shown, the purpose of this embodiment is to provide a multi-source heterogeneous data fusion early warning system for coal mills, which includes:
[0131] The data acquisition module is used to collect structured process parameter data and unstructured physical field data of the coal mill in real time;
[0132] The data acquisition module includes a pressure sensor, a current transformer, a temperature sensor, a weighing coal feeder metering module, a piezoelectric vibration acceleration sensor, an acoustic emission sensor, and a data acquisition card. The sensor models and their arrangement are consistent with the corresponding embodiments. The data acquisition card uses an NI cDAQ-9178, which supports multi-channel synchronous acquisition and has a configurable sampling frequency.
[0133] The feature construction module, connected to the data acquisition module, is used to preprocess and extract features from the acquired data, constructing a digital antigen vector representing the current operating status of the coal mill. The feature construction module uses an industrial control computer (Intel Core i7-12700 CPU, 16GB memory) to run a feature extraction and fusion program written in Python, performing statistical feature calculation, wavelet packet decomposition, MFCC extraction, normalization, and vector concatenation. The program runs for 1 second (synchronized with the data acquisition frequency).
[0134] The model training module, connected to the data acquisition module, is used to generate an immune detector set and maintain the self-space based on historical health data and the negative selection principle. The model training module and the feature construction module share an industrial control computer and have a built-in self-space database and immune detector set generation algorithm.
[0135] The collaborative monitoring module is connected to the model training module and the feature construction module respectively. It is used to calculate the Euclidean distance between the digital antigen vector and the immune detector set, as well as the danger signal index of the coal mill operation, and execute collaborative judgment logic based on danger theory. The collaborative monitoring module uses an FPGA chip (Xilinx XC7K325T) to realize the rapid calculation of Euclidean distance and danger signal index to ensure real-time performance. At the same time, it has built-in collaborative judgment logic.
[0136] The early warning and control module, connected to the collaborative monitoring module, is used to output early warning information when the collaborative judgment result is abnormal, and to send adjustment commands to the coal mill control system according to the fault type. The early warning and control module includes an audible and visual alarm (alarm volume ≥85dB, alarm light flashing red), an industrial touch screen (displaying early warning information and fault location results), and a PLC controller (Siemens S7-1500). The PLC controller communicates with the coal mill control system and outputs feedback control commands.
[0137] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for early warning by fusing multi-source heterogeneous data inside a coal mill, characterized in that, Includes the following steps: S1. Acquire multi-source heterogeneous data during the operation of the coal mill, wherein the multi-source heterogeneous data includes structured process parameter data and unstructured physical field data; S2. Preprocess and extract features from the multi-source heterogeneous data to construct a digital antigen vector representing the current operating status of the coal mill; S3. Construct a self-space based on the historical health operation data of the coal mill, and generate an immune detector set covering the non-self-space based on the principle of negative selection; The immune detector set is used to identify abnormal states that deviate from the self-space; The steps for generating an immune detector set covering the non-self space based on the negation selection principle include: defining the set of digital antigen vectors in the healthy state of the coal mill as the self set; randomly generating candidate detectors in the normalized feature space, each candidate detector being defined as a hypersphere with a center vector and a recognition radius; calculating the Euclidean distance between the candidate detector and all sample points in the self set using the real-valued negation selection algorithm; discarding a candidate detector if it covers any self sample point; and retaining a candidate detector and adding it to the immune detector set if it does not cover any self sample point. S4. Calculate the Euclidean distance between the digital antigen vector at the current moment and each detector in the immune detector set in real time, and calculate the danger signal index of the coal mill operation; S5. Based on the immune recognition result of whether the digital antigen vector is recognized by the immune detector set and the danger signal index, a joint determination is made. When the digital antigen vector is recognized by the immune detector set and the danger signal index exceeds the preset threshold, an early warning signal is triggered and the corresponding feedback control strategy is executed.
2. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 1, characterized in that, In step S1, the structured process parameter data includes at least the mill bowl differential pressure, mill current, hydraulic oil station oil pressure, mill outlet temperature, and coal feed rate; the unstructured physical field data includes at least the vibration acceleration signal on the surface of the mill cylinder and the acoustic emission signal inside the mill.
3. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 2, characterized in that, The process of constructing the digital antigen vector in step S2 is as follows: For the structured process parameter data, statistical features within a preset time window are calculated. The statistical features include at least one of mean, variance, skewness, and kurtosis, generating a first type of feature component that reflects the changes in process parameters. For the vibration acceleration signal and acoustic emission signal, wavelet packet decomposition is used to extract the energy spectrum of different frequency bands, and the Mel frequency cepstral coefficients are calculated to generate a second type of characteristic component that reflects the change of physical state. After normalizing the first type of feature components and the second type of feature components, feature fusion mapping is performed by concatenating and splicing them to form the digital antigen vector.
4. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 1, characterized in that, The recognition radius employs a dynamic adaptive adjustment strategy, including: Obtain the real-time load command of the coal mill and calculate the standard deviation of the load command within a preset sliding time window. Define the standard deviation as the load fluctuation index. Based on formula Calculate the recognition radius of the candidate detector at the current time. ,in, This indicates the preset reference radius of the coal mill under steady-state operating conditions. This represents the preset radius adjustment gain coefficient; the length of the preset sliding time window and the radius adjustment gain coefficient. Determined through calibration using historical operational data; When the coal mill is in an unsteady operating condition, the load fluctuation index... When the radius is increased, the calculated recognition radius It increases accordingly; When the coal mill is in steady-state operation, the load fluctuation index is... When reduced, the calculated recognition radius It decreases accordingly.
5. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 1, characterized in that, In step S4, the method for calculating the danger signal index is as follows: The specific energy consumption of the coal mill, the amount of slag discharged, and the deviation of the outlet powder pipe wind speed were selected as hazardous indicators to characterize the degree of damage to the equipment performance. Calculate the deviation of the current collected value of each of the above-mentioned hazard indicators from the rated reference value; Based on preset weighting coefficients, the deviation of each hazard indicator is weighted and summed to obtain the hazard signal index; wherein, the weighting coefficients are determined based on the correlation analysis between historical fault data and operational data.
6. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 4, characterized in that, In step S5, the logic for collaborative determination is as follows: Set danger threshold The danger threshold Determined based on statistical values of the danger signal index from historical fault data; If the Euclidean distance between the current digital antigen vector and any immune detector is less than the recognition radius of that detector. If so, an immune recognition event is determined to have occurred; Only when an immune recognition event occurs, and the calculated danger signal index is greater than [value missing] When this occurs, it is determined to be a valid intrusion, and a fault warning is generated.
7. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 1, characterized in that, After the warning signal is triggered in step S5, the fault tracing step S6 is also included: Based on the mechanistic model of the coal mill or the feature analysis of historical fault data, a mapping table between the feature dimensions of the digital antigen vector and the physical components of the coal mill is pre-established. From all the immune detectors that recognize the current digital antigen vector, the one with the smallest Euclidean distance is selected as the winning immune detector; Calculate the component distances between the center vector of the superior immunodetector and the current digital antigen vector in each feature dimension, and identify the abnormal feature dimensions with the highest contribution of component distances. Based on the mapping table, find the physical component or parameter source corresponding to the abnormal feature dimension and output the fault location result.
8. The method for early warning of multi-source heterogeneous data fusion inside a coal mill according to claim 1, characterized in that, The feedback control strategy in step S5 includes at least one of the following: When the warning type is coal blockage risk, the feeder speed is automatically reduced and the primary air pressure is increased; When the warning type is abnormal roller loading, the proportional relief valve setting of the hydraulic loading system is automatically adjusted; When the warning type is "metal impact noise", an emergency trip signal is triggered.
9. A multi-source heterogeneous data fusion early warning system for a coal mill, wherein the multi-source heterogeneous data fusion early warning system for a coal mill is used to implement the multi-source heterogeneous data fusion early warning method for a coal mill as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to collect structured process parameter data and unstructured physical field data of the coal mill in real time; The feature construction module, connected to the data acquisition module, is used to preprocess and extract features from the acquired data to construct a digital antigen vector characterizing the current operating status of the coal mill. The model training module, connected to the data acquisition module, is used to generate an immune detector set and maintain the self-space based on historical health data and the principle of negative selection. The collaborative monitoring module is connected to the model training module and the feature construction module respectively. It is used to calculate the Euclidean distance between the digital antigen vector and the immune detector set, as well as the danger signal index of the coal mill operation, and to execute collaborative judgment logic based on danger theory. The early warning and control module, connected to the collaborative monitoring module, is used to output early warning information when the collaborative judgment result is abnormal, and to send adjustment commands to the coal mill control system according to the fault type.