Method and device for predicting efficiency and service life of coal mill

Abstract

The invention relates to the crossing field of mechanical engineering and intelligent prediction technologies, particularly discloses a coal mill efficiency and service life prediction method and device, and aims to solve the problem that a traditional model is difficult to deal with nonlinear coupling prediction of equipment performance degradation under variable load and coal quality fluctuation. The method comprises the following steps: receiving a multi-source sensing data stream and constructing a structured feature matrix with aligned time sequences; efficiency degradation implicit features are extracted through a nonlinear dynamic encoder, and the interaction influence of grinding roller abrasion, lining plate fatigue and bearing degradation is quantified in combination with a multi-failure-mode coupling analysis module; and cooperatively predicting a network output efficiency attenuation curve and residual life probability distribution through a bidirectional attention mechanism. According to the method, through fusion of multi-source time sequence characteristics and multi-failure coupling modeling, limitation of a static threshold value and linear extrapolation is broken through, prediction precision and timeliness are remarkably improved, intelligent maintenance decision support is provided for a coal-fired power plant, non-planned shutdown risks are reduced, and operation economy and system reliability are optimized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the cross field of mechanical engineering and intelligent prediction technology, and particularly relates to a coal mill efficiency and life prediction method and device. BACKGROUND

[0002] With the continuous improvement of the requirements of the thermal power industry on energy utilization efficiency and intelligent equipment operation and maintenance, the coal mill, as the core equipment of the coal pulverizing system, its running state is directly related to the boiler combustion efficiency and the safety and stability of the unit. At present, large coal-fired power plants generally rely on regular maintenance and experience threshold alarm mechanism to manage the state of the coal mill, lack of dynamic quantitative evaluation ability of the equipment performance degradation trend and residual life, leading to preventive maintenance lag, frequent unplanned shutdown, seriously affecting the economic efficiency and reliability of the unit operation.

[0003] Among them, the coal mill efficiency and life prediction is a key link of equipment health management, and its core goal is to realize the early prediction of the coal mill output capacity attenuation trend and the key component wear limit through the fusion analysis of multi-source operating parameters. However, the existing prediction model is mostly based on single working condition parameter or static regression equation, which is difficult to capture the nonlinear degradation characteristics of the coal mill under variable load, coal quality fluctuation and environmental disturbance. At the same time, the historical operation data and real-time sensing information of the equipment lack a structured time sequence alignment mechanism, leading to single dimension of state feature extraction, which cannot support high-precision life attenuation trajectory modeling. In addition, the prediction logic excessively relies on linear extrapolation assumption, ignoring the coupling effect between the multi-failure modes of mill roller wear, lining fatigue and bearing deterioration, resulting in significant deviation between the prediction result and the real equipment state.

[0004] Therefore, there is an urgent need for a coal mill efficiency and life collaborative prediction method that can fuse multi-dimensional dynamic working conditions, identify nonlinear degradation paths, and quantify the interactive effects of multiple failure modes, in order to realize accurate prediction of equipment state and intelligent optimization of maintenance strategy. SUMMARY

[0005] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide a coal mill efficiency and life prediction method, which constructs a time sequence alignment structured representation model of multi-source operating parameters, extracts nonlinear degradation characteristics of the coal mill under variable load and coal quality fluctuation working conditions, and fuses the coupling influence factors of multiple failure modes of mill roller wear, lining fatigue and bearing deterioration, establishes a dynamic coupling prediction architecture, and realizes collaborative quantitative prediction of equipment efficiency attenuation trend and residual life trajectory. In this way, the limitations of traditional static threshold alarm and linear extrapolation model are broken through, the prediction accuracy and response timeliness are improved, the intelligent maintenance decision support driven by equipment state is provided for coal-fired power plants, the risk of unplanned shutdown is reduced, and the economic efficiency and system reliability of the unit operation are optimized.

[0006] According to an aspect of the present application, a coal mill efficiency and life prediction method is provided, which comprises: receiving a multi-source sensor data stream from a coal mill operation monitoring system, the multi-source sensor data stream comprising a coal mill current signal, an outlet temperature sequence, a coal fineness distribution, a coal supply time sequence, a vibration spectrum feature, and an ambient temperature and humidity parameter; performing timestamp alignment and working condition normalization processing on the multi-source sensor data stream to construct a structured time sequence feature matrix, the structured time sequence feature matrix containing synchronous evolution trajectories of device operating state variables and external disturbance variables; based on the structured time sequence feature matrix, using a nonlinear dynamic encoder to extract a coal mill efficiency degradation hidden feature vector, the nonlinear dynamic encoder capturing the nonlinear performance degradation pattern of the device under load fluctuation and coal quality change conditions by introducing a combination structure of a gating memory unit and a multi-scale convolution kernel; based on the structured time sequence feature matrix, using a multi-failure mode coupling analysis module to generate a life degradation driving factor vector, the multi-failure mode coupling analysis module quantifying the interactive enhancement effect between each failure mode by constructing a joint mapping relationship of a mill roller wear rate function, a liner stress accumulation function, and a bearing vibration energy entropy function; inputting the efficiency degradation hidden feature vector and the life degradation driving factor vector into a collaborative prediction network, the collaborative prediction network establishing dynamic association weights between efficiency features and life factors through a bidirectional attention mechanism, and outputting an efficiency prediction curve and a remaining life probability distribution of the coal mill within a specified time window in the future; based on the efficiency prediction curve and the remaining life probability distribution, generating a device maintenance priority instruction and an operation parameter optimization suggestion, the device maintenance priority instruction being used to trigger a preventive maintenance work order, and the operation parameter optimization suggestion being used to adjust the coal supply rate and the air-coal ratio to delay performance degradation.

[0007] The construction process of the nonlinear dynamic encoder comprises: dividing the structured time sequence feature matrix into multiple time segments, each time segment corresponding to a device operating cycle; applying local time sequence convolution operations to the feature sequences in each time segment to extract the transient change pattern of the device state in a short period; applying cross-period gating recursion operations to the convolution outputs of multiple time segments to establish an implicit state transmission path for long-period performance trends; weighting and fusing the cross-period implicit states through a feature channel attention mechanism to generate the final efficiency degradation hidden feature vector.

[0008] The construction process of the multi-failure mode coupling analysis module comprises: Separating the vibration high-frequency components and current fluctuation components related to the mill roller wear from the structured time-series feature matrix, calculating the energy dissipation gradient per unit time as the input variable of the mill roller wear rate function; Separating the temperature cycle number and stress amplitude sequence related to the liner fatigue from the structured time-series feature matrix, and constructing a stress accumulation function based on the Miner cumulative damage theory; Separating the vibration envelope spectrum entropy value and frequency band energy ratio related to the bearing degradation from the structured time-series feature matrix, and constructing a vibration energy entropy function; Nonlinearly weighting and superimposing the output values of the mill roller wear rate function, the stress accumulation function and the vibration energy entropy function to generate a life decay driving factor vector, and the nonlinear weighting coefficients are dynamically calibrated by a gradient boosting tree model based on a historical failure case library.

[0009] The construction process of the collaborative prediction network includes: Projecting the efficiency degradation implicit feature vector to a first feature space and projecting the life decay driving factor vector to a second feature space; Establishing a bidirectional cross-attention layer between the first feature space and the second feature space, the bidirectional cross-attention layer dynamically adjusts the influence strength of efficiency features on life factors and the feedback strength of life factors on efficiency features by calculating the mutual information weight between feature dimensions; Inputting the feature vectors modulated by the bidirectional cross-attention layer into a multilayer perceptron regression head, the multilayer perceptron regression head includes three fully connected hidden layers and an output layer, and the output layer respectively outputs the time series value of the efficiency prediction curve and the mean and variance parameters of the remaining life probability distribution; Applying a monotone decreasing constraint loss function to the efficiency prediction curve and a survival analysis likelihood loss function to the remaining life probability distribution, and jointly optimizing the parameters of the collaborative prediction network.

[0010] The generation logic of the equipment maintenance priority instruction includes: When the mean of the remaining life probability distribution is lower than a preset safety threshold and the variance is less than a first tolerance interval, a first-level maintenance instruction is generated, which triggers an immediate shutdown maintenance process; When the mean of the remaining life probability distribution is between the preset safety threshold and the warning threshold and the variance is between the first tolerance interval and the second tolerance interval, a second-level maintenance instruction is generated, which triggers a planned maintenance arrangement within seventy-two hours; When the mean of the remaining life probability distribution is higher than the warning threshold or the variance is greater than the second tolerance interval, a third-level maintenance instruction is generated, which only records state abnormalities and continues to monitor.

[0011] The generation logic of the operation parameter optimization suggestion comprises: When the efficiency prediction curve shows that the efficiency decay rate exceeds a set slope threshold within the next 24 hours, calculate the sensitivity gradient of the current coal feeding rate and the air-coal ratio to the efficiency decay; Based on the sensitivity gradient, generate a coal feeding rate reduction instruction and a primary air volume increase instruction, and the instruction amplitude is iteratively solved to an optimal value within the safe operation boundary of the device by a gradient descent algorithm; The coal feeding rate reduction instruction and the primary air volume increase instruction are packaged into a parameter adjustment message and sent to the coal mill distributed control system for execution.

[0012] A coal mill efficiency and life prediction device comprises: A data preprocessing module is configured to receive a multi-source sensor data stream from a coal mill operation monitoring system, wherein the multi-source sensor data stream comprises a coal mill current signal, an outlet temperature sequence, a coal fineness distribution, a coal feeding amount time sequence, a vibration spectrum feature, and an environmental temperature and humidity parameter, and to perform timestamp alignment and working condition normalization processing on the data stream to construct a structured time sequence feature matrix; An efficiency degradation feature extraction module is configured to extract a coal mill efficiency degradation hidden feature vector based on the structured time sequence feature matrix through a nonlinear dynamic encoder, wherein the nonlinear dynamic encoder adopts a combination structure of a gated memory unit and a multi-scale convolution kernel to capture the nonlinear mode of performance decay of the device under load fluctuation and coal quality change conditions; A life decay driving factor generation module is configured to generate a life decay driving factor vector based on the structured time sequence feature matrix through a multi-failure mode coupling analysis module, wherein the multi-failure mode coupling analysis module quantifies the interactive enhancement effect between each failure mode by constructing a joint mapping relationship of a mill roller wear rate function, a liner stress accumulation function, and a bearing vibration energy entropy function; A collaborative prediction network module is configured to input the efficiency degradation hidden feature vector and the life decay driving factor vector into a collaborative prediction network with a bidirectional attention mechanism to establish a dynamic correlation weight between the efficiency features and the life factors, and output an efficiency prediction curve and a remaining life probability distribution of the coal mill within a specified time window in the future; A maintenance decision and optimization suggestion generation module is configured to generate a device maintenance priority instruction and an operation parameter optimization suggestion based on the efficiency prediction curve and the remaining life probability distribution, wherein the device maintenance priority instruction is used to trigger a preventive maintenance work order, and the operation parameter optimization suggestion is used to adjust the coal feeding rate and the air-coal ratio to delay performance degradation.

[0013] Compared with the prior art, the coal mill efficiency and life prediction method provided by the application realizes the collaborative quantitative prediction of the equipment efficiency attenuation trend and the residual life trajectory by constructing a time sequence alignment structured representation model of multiple source operation parameters, extracting the nonlinear degradation characteristics of the coal mill under variable load and coal quality fluctuation conditions, and establishing a dynamic coupling prediction architecture by fusing the coupling influence factors of the mill roller wear, lining fatigue and bearing deterioration multi-failure modes. In this way, the limitations of traditional static threshold alarm and linear extrapolation model are broken through, the prediction accuracy and response timeliness are improved, the intelligent maintenance decision support of the equipment state driving is provided for the coal-fired power plant, the non-planned shutdown risk is reduced, and the economic efficiency and system reliability of the unit operation are optimized. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is the overall technical scheme architecture schematic diagram of the application; Figure 2 is the collaborative prediction architecture framework schematic diagram of the nonlinear dynamic encoder and the multi-failure mode coupling analysis module of the application. DETAILED DESCRIPTION

[0015] In the following description, specific terms and structures are used to facilitate understanding of the embodiments of the application. However, these descriptions are not intended to limit the application, and those skilled in the art can understand that various modifications and replacements can be made without departing from the spirit and scope of the application. The drawings are only used to illustrate the exemplary embodiments of the application, are not drawn to scale, and should not be interpreted as limiting the application.

[0016] The coal mill is the core equipment of the coal pulverizing system in a coal-fired power plant. Its operation efficiency and service life directly determine the fuel economy and system reliability of the unit. Current mainstream prediction methods rely on single sensor threshold alarms or life extrapolation models based on linear regression. They cannot effectively capture the nonlinear degradation behavior of the equipment under complex working conditions of variable load and coal quality fluctuations. It is also difficult to quantify the coupling enhancement effect between the multiple failure modes of mill roller wear, liner fatigue, and bearing degradation. Such methods have a lag response when faced with sudden working condition disturbances, resulting in large prediction errors. This leads to passive maintenance decisions, frequent unplanned shutdowns, and blind adjustments of operating parameters, which seriously restricts the intelligent operation and maintenance level of the power plant. To address the above technical defects, the present application proposes a coal mill efficiency and life prediction method. It builds a time series alignment structured representation model of multiple source operating parameters, extracts the nonlinear degradation characteristics of the coal mill under variable load and coal quality fluctuation conditions, and integrates the coupling influence factors of multiple failure modes of mill roller wear, liner fatigue, and bearing degradation. A dynamic coupling prediction architecture is established to realize the collaborative quantitative prediction of the equipment efficiency decay trend and the remaining life trajectory. This method breaks through the limitations of traditional static threshold alarms and linear extrapolation models, improves the prediction accuracy and response timeliness, and provides intelligent maintenance decision support for the equipment state of the coal-fired power plant. It reduces the risk of unplanned shutdowns, optimizes the operating economy and system reliability of the unit.

[0017] Please refer to Figure 1 and Figure 2In the above coal mill efficiency and service life prediction method, the step S110 receives a multi-source sensing data stream from a coal mill operation monitoring system, and the multi-source sensing data stream includes a coal mill current signal, an outlet temperature sequence, a coal fineness distribution, a coal supply time sequence, a vibration spectrum feature, and an environmental temperature and humidity parameter. It can be understood that receiving a multi-source sensing data stream aims to build a full-dimensional monitoring system covering the device ontology state, process parameters, and environmental disturbances, and to provide a complete data basis for subsequent nonlinear degradation feature extraction and multi-failure mode coupling analysis. Specifically, the coal mill current signal is collected by a high-precision current transformer installed in the main motor power supply circuit, with a sampling frequency of 1 kHz. The amplitude fluctuation directly reflects the change of the mill roller load and the dynamic thickness of the coal bed, and is the core index for judging the instantaneous energy consumption and mechanical resistance of the device. The outlet temperature sequence is obtained by a platinum resistance temperature sensor arranged on the wall surface of the mill outlet pipeline, with a sampling interval of 1 second. The value change represents the degree of coal drying and the matching state of the hot air system, and an abnormal temperature rise indicates the risk of coal accumulation or poor ventilation. The coal fineness distribution data is obtained by periodic measurement of an online laser particle size analyzer, with a particle size distribution histogram output every 30 seconds. The D90 value and distribution variance are used to quantify the grinding efficiency and coal quality adaptability, and an excessive fineness indicates abnormal mill roller gap or increased lining wear. The coal supply time sequence is calculated by a belt coal feeder weighing sensor and a speed encoder, with an update frequency of 5 Hz. The value stability directly affects the load balance and coal uniformity of the coal mill. The vibration spectrum feature is collected by a three-axis acceleration sensor installed on the key support points of the coal mill shell, and the amplitude spectrum and envelope spectrum in the frequency band of 0-2 kHz are output after fast Fourier transform. The high-frequency energy concentration and specific frequency band peak value are used to identify bearing damage, gear meshing abnormalities, and structural looseness. The environmental temperature and humidity parameters are monitored in real time by a temperature and humidity transmitter deployed in the coal mill plant, with a sampling period of 10 seconds. The values are used to correct the thermodynamic parameter calculation benchmark and eliminate the interference of environmental disturbances on the temperature and energy consumption model. All sensing data are transmitted to the central data acquisition server through the industrial Ethernet protocol. The original data packet contains a timestamp, a sensor identifier, a physical value, and a check code, ensuring data traceability and integrity.

[0018] In the above coal mill efficiency and service life prediction method, the step S120, the multi-source sensor data stream is timestamp aligned and working condition normalized, a structured time sequence feature matrix is constructed, and the structured time sequence feature matrix contains the synchronous evolution track of the device running state variable and the external disturbance variable. It can be understood that the timestamp alignment and working condition normalization process aims to eliminate the spatio-temporal dislocation problem caused by the differences in sampling frequency, transmission delay and non-uniform dimension of multi-source heterogeneous data, and construct a unified feature space with physical consistency and time sequence synchronization. Specifically, the timestamp alignment process first extracts the timestamp sequence of each data stream, takes the timestamp of the coal mill current signal as the reference time axis, resamples other data streams using the cubic spline interpolation method, unifies all data to a hundred hertz sampling rate, and the interpolation error is controlled within five percent of the original sampling interval. The working condition normalization process includes two sub-steps of physical dimension normalization and running condition normalization. The physical dimension normalization adopts the maximum and minimum value scaling method, linearly maps each type of physical quantity within the maximum and minimum value interval of its historical running data, and compresses all variable value ranges to zero to one interval. The running condition normalization is based on the ratio of the current coal supply to the rated coal supply, and proportionally corrects the variables related to current, temperature, vibration energy and load, eliminating the feature drift caused by load fluctuation. The construction of the structured time sequence feature matrix takes time as the row index and feature variable as the column index, each row corresponds to a complete state snapshot at a sampling time, and the column order is fixed as: normalized current value, normalized outlet temperature, coal fineness D90 value, coal fineness distribution variance, normalized coal supply, vibration frequency spectrum zero to five hundred hertz energy proportion, vibration frequency spectrum five hundred to one thousand hertz energy proportion, vibration frequency spectrum one thousand to two thousand hertz energy proportion, vibration envelope spectrum peak frequency, normalized ambient temperature, and normalized ambient humidity. The matrix is stored in the form of a floating-point two-dimensional array in the memory buffer for direct calling by the subsequent feature extraction module.

[0019] In the above coal mill efficiency and service life prediction method, in step S130, a nonlinear dynamic encoder is used to extract the mill efficiency degradation hidden feature vector based on the structured time sequence feature matrix. The nonlinear dynamic encoder captures the nonlinear performance degradation pattern of the device under load fluctuation and coal quality change conditions by introducing a combination structure of a gating memory unit and a multi-scale convolution kernel. It should be understood that the nonlinear dynamic encoder aims to automatically mine deep feature patterns strongly related to efficiency degradation from high-dimensional time sequence data, overcoming the lack of modeling ability of complex nonlinear relationships in artificial feature engineering. Specifically, the construction process of the nonlinear dynamic encoder includes: dividing the structured time sequence feature matrix into multiple time segments, each time segment corresponding to a device operating cycle; applying local time sequence convolution operation to the feature sequence in each time segment to extract the transient change pattern of the device state in the short cycle; applying cross-cycle gating recursion operation to the convolution output of multiple time segments to establish the hidden state transmission path of the long-cycle performance trend; and weighting and fusing the cross-cycle hidden state through a feature channel attention mechanism to generate the final efficiency degradation hidden feature vector.

[0020] Specifically, in step S131, the structured time sequence feature matrix is divided into multiple time segments, each time segment corresponding to a device operating cycle. It should be understood that dividing the time segments aims to decompose the continuous running process into independent analysis units with physical meaning, facilitating the capture of the cumulative impact of periodic load changes and coal quality switching on device performance. The device operating cycle is defined as the complete process from the stable execution of a coal feeding instruction to the next instruction change, and the start and end time is determined by the jump point where the amplitude change in the coal feeding amount time sequence exceeds five percent. The length of each time segment is not fixed, but the minimum duration is thirty seconds and the maximum is three hundred seconds, to ensure that at least three complete mechanical vibration cycles and thermodynamic response cycles are included. After the time segments are divided, the original matrix is cut into multiple sub-matrices, each of which is independently sent to the subsequent convolution module for processing.

[0021] Specifically, the step S132 applies a local temporal convolution operation to the feature sequence in each time slice to extract the transient change pattern of the device state in a short period. It should be understood that the local temporal convolution operation is intended to capture the dynamic response characteristics of the device in the time scale of seconds, such as current spikes, temperature surges, and vibration burst transient events. The convolution layer uses a one-dimensional convolution kernel that slides along the time axis, and the convolution kernel size is set to five, ten, and twenty-three scales, corresponding to high-frequency disturbances, medium-frequency fluctuations, and low-frequency trends, respectively. Each scale is configured with sixteen convolution kernels, the activation function is a rectified linear unit, the convolution step is one, and the padding method is zero padding to keep the output sequence length unchanged. After the convolution output is processed by a batch normalization layer, it is sent to a max pooling layer for down-sampling. The pooling window size is two, and the step is two. Finally, three scale feature maps are output, with dimensions of (time step divided by two, sixteen), (time step divided by two, sixteen), and (time step divided by two, sixteen), respectively. The three scale feature maps are concatenated in the channel dimension to form a multi-scale transient feature tensor.

[0022] Specifically, the step S133 applies a cross-period gated recurrent operation to the convolution output of multiple time slices to establish an implicit state transmission path for long-period performance trends. It should be understood that the cross-period gated recurrent operation is intended to model the gradual degradation process of device performance in the time scale of minutes to hours, capturing the long-term effects of mill roll passivation, backing plate slow accumulation of damage. The recurrent unit uses a gated recurrent unit structure, with a hidden state dimension of one hundred and twenty-eight. The multi-scale transient feature tensor corresponding to each time slice is flattened along the time axis and used as the input sequence of the gated recurrent unit in that period. The gated recurrent unit processes the input sequence in the order of time slices, and the final hidden state of the previous period is used as the initial hidden state of the next period, realizing cross-period state transmission. The output of the gated recurrent unit is a period-level implicit state vector corresponding to each time slice, with a dimension of one hundred and twenty-eight.

[0023] Specifically, the step S134 generates the final efficiency degradation implicit feature vector by weighting and fusing the cross-period implicit state through a feature channel attention mechanism. It should be understood that the feature channel attention mechanism is intended to adaptively strengthen the feature channels most relevant to efficiency degradation and suppress noise and redundant information. The period-level implicit state vectors of all time slices are stacked into a two-dimensional matrix and input into the channel attention module. The channel attention module first performs global average pooling on the matrix along the time axis to obtain a channel description vector, then generates a channel weight vector through two fully connected networks (the first layer has thirty-two neurons, and the second layer has one hundred and twenty-eight neurons), and finally multiplies the weight vector with the original implicit state matrix by channel to obtain the weighted feature matrix. The matrix is max-pooled along the time axis to output an efficiency degradation implicit feature vector with a dimension of one hundred and twenty-eight.

[0024] In the aforementioned method for predicting the efficiency and lifespan of a coal mill, step S140 involves generating a lifespan decay driving factor vector based on the structured time-series feature matrix using a multi-failure mode coupling analysis (MFCA) module. This MFCA module quantifies the interactive enhancement effect between various failure modes by constructing a joint mapping relationship between the roller wear rate function, the liner stress accumulation function, and the bearing vibration energy entropy function. It should be understood that the MFCA module aims to combine physical mechanism models with data-driven methods to accurately quantify the contribution of different component failures to the overall lifespan and their mutual coupling effects. Specifically, the construction process of the multi-failure mode coupling analysis module includes: separating the high-frequency vibration component and current fluctuation component related to roller wear from the structured time-series feature matrix, calculating the energy dissipation gradient per unit time, and using it as the input variable of the roller wear rate function; separating the temperature cycle number and stress amplitude sequence related to liner fatigue from the structured time-series feature matrix, and constructing a stress accumulation function based on Miner's cumulative damage theory; separating the vibration envelope spectrum entropy value and frequency band energy ratio related to bearing deterioration from the structured time-series feature matrix, and constructing a vibration energy entropy function; and performing nonlinear weighted superposition of the output values ​​of the roller wear rate function, stress accumulation function, and vibration energy entropy function to generate a life decay driving factor vector, wherein the nonlinear weighting coefficient is dynamically calibrated by a gradient boosting tree model from a historical failure case library.

[0025] Specifically, in step S141, the high-frequency vibration component and current fluctuation component related to roller wear are separated from the structured time-series feature matrix, and the energy dissipation gradient per unit time is calculated as the input variable of the roller wear rate function. It should be understood that roller wear is mainly manifested as increased frictional power consumption due to material loss at the contact surface and an increase in high-frequency vibration energy. The high-frequency vibration component refers to the energy proportion in the 1000-2000 Hz frequency band of the vibration spectrum; this frequency band reflects the high-frequency impact and frictional noise between the roller and the coal block. The current fluctuation component refers to the residual sequence after the normalized current value is processed by a fifth-order Butterworth high-pass filter with a filter cutoff frequency of 0.5 Hz, used to extract high-frequency current disturbances related to sudden changes in mechanical load. The formula for calculating the energy dissipation gradient is:

[0026] in, The energy dissipation gradient per unit time. To calculate the window length, a value of sixty seconds is used. For a moment The current fluctuation component, For a moment The high-frequency components of the vibration. This gradient value serves as the sole input to the roller wear rate function, and the function output is a dimensionless wear rate exponent.

[0027] Specifically, in step S142, the temperature cycle count and stress amplitude sequence related to liner fatigue are separated from the structured time-series feature matrix, and a stress accumulation function based on Miner's cumulative damage theory is constructed. It should be understood that liner fatigue damage originates from the periodic alternation of thermal and mechanical stress. The temperature cycle count is obtained by counting the complete peak-to-trough pairs crossing a preset threshold range (0.3 to 0.7) in the normalized outlet temperature sequence, with each cycle corresponding to one thermal expansion and contraction process. The stress amplitude sequence is obtained by mapping normalized current values, with the mapping relationship being a linear proportionality coefficient multiplied by the current value. The proportionality coefficient is calibrated based on the ratio of rated torque to rated current on the equipment nameplate. The stress accumulation function is constructed based on Miner's theory, and its expression is:

[0028] in, As a cumulative damage factor, This represents the total number of temperature cycles within the current time window. For the first The stress cycle number corresponding to this cycle is taken as one here. In the first The fatigue life of the material under the stress amplitude of each cycle is obtained from the SN curve of the liner material. This cumulative damage factor is output as a stress accumulation function.

[0029] Specifically, step S143 involves separating the vibration envelope spectrum entropy value and band energy ratio related to bearing degradation from the structured time-series feature matrix, and constructing a vibration energy entropy function. It should be understood that bearing degradation manifests as disordered vibration signal energy distribution and increased envelope spectrum complexity. The vibration envelope spectrum entropy value calculation process is as follows: first, the original vibration signal is subjected to a Hilbert transform to extract the envelope signal; then, the envelope signal is subjected to a fast Fourier transform to obtain the envelope spectrum; finally, the information entropy of the envelope spectrum amplitude sequence is calculated. The band energy ratio is defined as the ratio of the sum of the energy of the bearing characteristic frequency and its sideband frequencies (such as the inner race fault frequency ± rotation frequency) in the vibration spectrum to the sum of the energy across the entire frequency band. The vibration energy entropy function output is the product of the envelope spectrum entropy value and the band energy ratio; this product value characterizes the degree of bearing degradation.

[0030] Specifically, the step S144, the grinding roller wear rate function, stress accumulation function and the output value of vibration energy entropy function are nonlinearly weighted and superimposed to generate a life attenuation driving factor vector, and the nonlinear weighting coefficient is dynamically calibrated by the historical failure case library through the gradient boosting tree model. It can be understood that the nonlinear weighted superposition aims to fuse the multi-failure mode contribution and reflect the coupling enhancement effect. The life attenuation driving factor vector has three dimensions, and each component is a grinding roller wear rate index, a cumulative damage factor and a bearing degradation index. The nonlinear weighting coefficient is obtained by offline training: collecting the whole process data of the historical equipment from commissioning to failure, labeling the failure time and the dominant failure mode, taking the average value of the three indexes one hour before failure as the input feature, and taking whether failure occurs as the label to train the gradient boosting tree classification model; after the model training is completed, the importance scores of each feature when the root node is split are extracted and normalized as fixed weighting coefficients. The final life attenuation driving factor vector is the Hadamard product of the three index components and the corresponding weighting coefficients.

[0031] In the above coal mill efficiency and life prediction method, the step S150 inputs the efficiency degradation implicit feature vector and the life attenuation driving factor vector into a collaborative prediction network, the collaborative prediction network establishes dynamic association weights between efficiency features and life factors through a bidirectional attention mechanism, and outputs an efficiency prediction curve and a residual life probability distribution of the coal mill in a specified time window in the future. It can be understood that the collaborative prediction network aims to jointly model the internal correlation between efficiency degradation and life loss, and realize double-target collaborative prediction. Specifically, the construction process of the collaborative prediction network includes: projecting the efficiency degradation implicit feature vector into a first feature space and projecting the life attenuation driving factor vector into a second feature space; a bidirectional cross-attention layer is established between the first feature space and the second feature space, the bidirectional cross-attention layer dynamically adjusts the influence strength of efficiency features on life factors and the feedback strength of life factors on efficiency features by calculating the mutual information weight between feature dimensions; the feature vectors modulated by the bidirectional cross-attention layer are input into a multilayer perceptron regression head, the multilayer perceptron regression head includes three fully connected hidden layers and an output layer, the output layer outputs time series values of the efficiency prediction curve and mean and variance parameters of the residual life probability distribution respectively; a monotone decreasing constraint loss function is applied to the efficiency prediction curve, and a survival analysis likelihood loss function is applied to the residual life probability distribution, and the parameters of the collaborative prediction network are jointly optimized.

[0032] Specifically, the step S151 projects the efficiency degradation implicit feature vector to a first feature space and projects the life attenuation driving factor vector to a second feature space. It should be understood that the feature space projection aims to map heterogeneous features to a unified dimension, facilitating subsequent interactive calculation. The efficiency degradation implicit feature vector is projected to a first feature space with a dimension of sixty-four through a fully connected network, and the activation function is hyperbolic tangent. The life attenuation driving factor vector is projected to a second feature space with a dimension of sixty-four through another fully connected network, and the activation function is hyperbolic tangent. After projection, two sixty-four-dimensional feature vectors are obtained.

[0033] Specifically, the step S152 establishes a bidirectional cross-attention layer between the first feature space and the second feature space, which dynamically adjusts the influence strength of efficiency features on life factors and the feedback strength of life factors on efficiency features by calculating the mutual information weight between feature dimensions. It should be understood that the bidirectional cross-attention layer aims to explicitly model the nonlinear dependence between efficiency features and life factors. The calculation process is as follows: first, calculate the dot product similarity matrix of the first feature space vector and the second feature space vector, with a dimension of sixty-four by sixty-four; perform soft-max normalization along the row direction of the matrix to obtain the attention weight matrix of efficiency features on life factors; perform soft-max normalization along the column direction to obtain the attention weight matrix of life factors on efficiency features; multiply the efficiency feature vector and the life factor attention weight matrix to obtain the efficiency feature modulated by the life factor; multiply the life factor vector and the efficiency feature attention weight matrix to obtain the life factor modulated by the efficiency feature; add the original feature and the modulated feature to complete the bidirectional attention fusion.

[0034] Specifically, the step S153 inputs the feature vector modulated by the bidirectional cross-attention layer into a multi-layer perceptron regression head, which includes three fully connected hidden layers and an output layer that respectively outputs the time series value of the efficiency prediction curve and the mean and variance parameters of the remaining life probability distribution. It should be understood that the multi-layer perceptron regression head aims to map the fused features to the prediction target space. The modulated feature vector is input into the first hidden layer after splicing, with two hundred and fifty-six neurons and a rectified linear unit activation function; the second hidden layer has one hundred and twenty-eight neurons and a rectified linear unit activation function; the third hidden layer has sixty-four neurons and a rectified linear unit activation function; the output layer includes two branches: the efficiency prediction branch outputs the efficiency value at twelve future time points (each point interval is one hour), with twelve output nodes; the life prediction branch outputs the mean and variance of the remaining life distribution, with two output nodes. The efficiency value output has no activation function, the mean output has an exponential function activation function, and the variance output has a soft plus function activation function.

[0035] Specifically, the step S154 applies a monotonically decreasing constraint loss function to the efficiency prediction curve and a survival analysis likelihood loss function to the remaining life probability distribution, and jointly optimizes the parameters of the collaborative prediction network. It should be understood that the monotonically decreasing constraint ensures that the efficiency prediction conforms to the physical degradation law, and the survival analysis likelihood loss improves the robustness of life prediction using censored data. The monotonically decreasing constraint loss function is defined as the sum of the negative parts of the differences between adjacent points on the prediction curve. The survival analysis likelihood loss function is based on the Weibull distribution assumption and calculates the log-likelihood using the current running time of the device and the label indicating whether it has failed. The total loss function is the weighted sum of the two losses, and the weight coefficients are determined by grid search on the validation set.

[0036] In the above coal mill efficiency and life prediction method, the step S160 generates a device maintenance priority instruction and an operation parameter optimization suggestion based on the efficiency prediction curve and the remaining life probability distribution, the device maintenance priority instruction is used to trigger a preventive maintenance work order, and the operation parameter optimization suggestion is used to adjust the coal feeding rate and the air-coal ratio to delay performance degradation. It should be understood that this step aims to convert the prediction results into executable operation and maintenance decisions to realize the closed loop from prediction to action. Specifically, the generation logic of the device maintenance priority instruction includes: when the mean of the remaining life probability distribution is lower than a preset safety threshold and the variance is less than a first tolerance interval, a first-level maintenance instruction is generated, which triggers an immediate shutdown maintenance process; when the mean of the remaining life probability distribution is between the preset safety threshold and a warning threshold and the variance is between the first tolerance interval and a second tolerance interval, a second-level maintenance instruction is generated, which triggers a planned maintenance arrangement within seventy-two hours; when the mean of the remaining life probability distribution is higher than the warning threshold or the variance is greater than the second tolerance interval, a third-level maintenance instruction is generated, which only records state abnormalities and continues to monitor. The preset safety threshold is seven hundred and twenty hours, the warning threshold is one thousand four hundred and forty hours, the first tolerance interval is zero to fifty hours, and the second tolerance interval is fifty to one hundred hours.

[0037] The generation logic of the operation parameter optimization suggestion comprises: when the efficiency prediction curve shows that the efficiency attenuation rate exceeds a set slope threshold in the next 24 hours, calculating the sensitivity gradient of the current coal feeding rate and the air-coal ratio to the efficiency attenuation; based on the sensitivity gradient, generating a coal feeding rate down-regulation instruction and a primary air quantity up-regulation instruction, the instruction amplitude being iteratively solved to an optimal value within the safe operation boundary of the equipment by a gradient descent algorithm; and encapsulating the coal feeding rate down-regulation instruction and the primary air quantity up-regulation instruction into a parameter adjustment message and sending the parameter adjustment message to the coal mill distributed control system for execution. The set slope threshold is 0.5% per hour. The sensitivity gradient is calculated by a finite difference method: the coal feeding rate and the primary air quantity are respectively fine-tuned by 2%, and the efficiency attenuation rate change is obtained by running the collaborative prediction network, and the ratio of the change to the fine-tuning amount is the sensitivity gradient. The initial learning rate of the gradient descent algorithm is 0.1, the iteration number is 10, and the constraint conditions are that the coal feeding rate is not less than 80% of the rated value and the primary air quantity is not higher than 120% of the rated value. The parameter adjustment message is encapsulated by using an industrial control protocol, and contains an instruction type, a target device identifier, a parameter name, a target value, and an execution timestamp.

[0038] A coal mill efficiency and life prediction device comprises: A data preprocessing module is configured to receive a multi-source sensor data stream from a coal mill operation monitoring system, the multi-source sensor data stream comprising a coal mill current signal, an outlet temperature sequence, a coal fineness distribution, a coal feeding amount time sequence, a vibration spectrum feature, and an environmental temperature and humidity parameter, and perform timestamp alignment and working condition normalization processing on the data stream to construct a structured time sequence feature matrix. An efficiency degradation feature extraction module is configured to extract a coal mill efficiency degradation hidden feature vector based on the structured time sequence feature matrix by using a nonlinear dynamic encoder, the nonlinear dynamic encoder adopting a combination structure of a gated memory unit and a multi-scale convolution kernel to capture the nonlinear mode of performance attenuation of the device under load fluctuation and coal quality change conditions. A life attenuation driving factor generation module is configured to generate a life attenuation driving factor vector based on the structured time sequence feature matrix by using a multi-failure mode coupling analysis module, the multi-failure mode coupling analysis module quantifying the interactive enhancement effect between failure modes by constructing a joint mapping relationship of a mill roller wear rate function, a liner stress accumulation function, and a bearing vibration energy entropy function. A collaborative prediction network module is configured to input the efficiency degradation hidden feature vector and the life attenuation driving factor vector into a collaborative prediction network with a bidirectional attention mechanism, establish a dynamic correlation weight between the efficiency feature and the life factor, and output an efficiency prediction curve and a residual life probability distribution of the coal mill in a specified time window in the future. A maintenance decision and optimization suggestion generation module is configured to generate device maintenance priority instructions and operation parameter optimization suggestions based on the efficiency prediction curve and the residual life probability distribution, wherein the device maintenance priority instructions are used to trigger a preventive maintenance work order, and the operation parameter optimization suggestions are used to adjust the coal feeding rate and the air-coal ratio to delay performance degradation.

[0039] In summary, the embodiment describes in detail the complete implementation process of the coal mill efficiency and life prediction method, covering six core links of data collection, feature alignment, nonlinear coding, failure coupling analysis, collaborative prediction and decision generation. The method deeply integrates data-driven and mechanism model, builds a dynamic coupling prediction architecture, realizes high-precision, high-time-efficiency and multi-target collaborative prediction of equipment performance degradation, and provides solid technical support for intelligent operation and maintenance of coal-fired power plants.

Claims

1. A method of predicting the efficiency and life of a coal mill, characterized by, The method comprises the following steps: Receiving multi-source sensing data streams from the coal mill operation monitoring system, which includes mill current signals, outlet temperature sequences, coal fineness distributions, coal supply time series, vibration spectrum features, and environmental temperature and humidity parameters; Timestamp alignment and working condition normalization are performed on the multi-source sensing data streams to construct a structured time series feature matrix, which contains the synchronous evolution trajectories of device operating state variables and external disturbance variables; Based on the structured time series feature matrix, a nonlinear dynamic encoder is used to extract an efficiency degradation hidden feature vector, which captures the nonlinear performance degradation patterns of the device under load fluctuations and coal quality changes through the combination of a gating memory unit and a multi-scale convolution kernel; Based on the structured time series feature matrix, a multi-failure mode coupling analysis module is used to generate a life degradation driving factor vector, which quantifies the interactive enhancement effect between failure modes through the joint mapping relationship of mill roller wear rate function, liner stress accumulation function, and bearing vibration energy entropy function; The efficiency degradation hidden feature vector and the life degradation driving factor vector are input into a collaborative prediction network, which establishes dynamic association weights between efficiency features and life factors through a bidirectional attention mechanism, and outputs efficiency prediction curves and residual life probability distributions of the coal mill in a specified future time window; Based on the efficiency prediction curve and the residual life probability distribution, device maintenance priority instructions and operation parameter optimization suggestions are generated, the device maintenance priority instructions are used to trigger preventive maintenance work orders, and the operation parameter optimization suggestions are used to adjust the coal supply rate and the air-coal ratio to delay performance degradation.

2. The coal mill efficiency and life prediction method of claim 1, wherein, The construction process of the nonlinear dynamic encoder includes: Divide the structured time series feature matrix into multiple time segments, each corresponding to a device operating cycle; Apply local time series convolution operation to the feature sequence in each time segment to extract the transient change pattern of the device state in the short cycle; Apply cross-cycle gating recursion operation to the convolution output of multiple time segments to establish an implicit state transmission path for long-cycle performance trends; Weighted fusion of cross-cycle implicit states is performed through feature channel attention mechanism to generate the final efficiency degradation hidden feature vector.

3. The coal mill efficiency and life prediction method of claim 2, wherein, The local time series convolution operation uses one-dimensional convolution kernels sliding along the time axis, with five, ten, and twenty-three scales for convolution kernel size, corresponding to high-frequency disturbances, medium-frequency fluctuations, and low-frequency trends, respectively. Each scale is configured with sixteen convolution kernels, the activation function is a rectified linear unit, the convolution step is one, and the padding method is zero padding. The convolution output is processed by a batch normalization layer and then sent to a max pooling layer for downsampling. The pooling window size is two and the step is two. Finally, three scale feature maps are output and concatenated in the channel dimension to form a multi-scale transient feature tensor.

4. The coal mill efficiency and life prediction method of claim 3, wherein, The cross-period gating recurrent operation adopts a gating recurrent unit structure, the hidden state dimension is one hundred and twenty-eight, and the multi-scale transient feature tensor corresponding to each time segment is flattened along the time axis and taken as the input sequence of the gating recurrent unit in the period, the final hidden state of the previous period is taken as the initial hidden state of the next period, the state transmission across periods is realized, and the output is a period-level hidden state vector corresponding to each time segment.

5. The coal mill efficiency and life prediction method as claimed in claim 1, wherein, The construction process of the multi-failure mode coupling analysis module includes: Separating the vibration high-frequency component and the current fluctuation component related to the mill roll wear from the structured time sequence feature matrix, and calculating the energy dissipation gradient per unit time as an input variable of the mill roll wear rate function; Separating the temperature cycle number and stress amplitude sequence related to the liner fatigue from the structured time sequence feature matrix, and constructing a stress accumulation function based on the Miner cumulative damage theory; Separating the vibration envelope spectrum entropy value and frequency band energy ratio related to bearing degradation from the structured time sequence feature matrix, and constructing a vibration energy entropy function; Nonlinearly weighting and superimposing the output values of the mill roll wear rate function, the stress accumulation function and the vibration energy entropy function to generate a life attenuation driving factor vector, and the nonlinear weighting coefficients are dynamically calibrated by a gradient boosting tree model through a historical failure case library.

6. The coal mill efficiency and life prediction method as claimed in claim 5, wherein, The energy dissipation gradient calculation formula is the integral mean of the product of the square of the current fluctuation component and the vibration high-frequency component in a unit time window, the calculation window length is sixty seconds, the vibration high-frequency component refers to the energy proportion of the one thousand to two thousand hertz frequency band in the vibration spectrum, and the current fluctuation component refers to the residual sequence after the normalized current value is processed by a fifth-order Butterworth high-pass filter, and the filter cutoff frequency is zero point five hertz.

7. The coal mill efficiency and life prediction method as claimed in claim 1, wherein, The construction process of the collaborative prediction network includes: Projecting the efficiency degradation hidden feature vector into a first feature space and projecting the life attenuation driving factor vector into a second feature space; Establishing a bidirectional cross-attention layer between the first feature space and the second feature space, the bidirectional cross-attention layer dynamically adjusts the influence strength of the efficiency feature on the life factor and the feedback strength of the life factor on the efficiency feature by calculating the mutual information weight between the feature dimensions; Inputting the feature vectors modulated by the bidirectional cross-attention layer into a multilayer perceptron regression head, the multilayer perceptron regression head includes three fully connected hidden layers and an output layer, and the output layer outputs the time series value of the efficiency prediction curve and the mean and variance parameters of the residual life probability distribution, respectively; Applying a monotone decreasing constraint loss function to the efficiency prediction curve and a survival analysis likelihood loss function to the residual life probability distribution, and jointly optimizing the parameters of the collaborative prediction network.

8. The coal mill efficiency and life prediction method as claimed in claim 7, wherein, The bidirectional cross attention layer calculates a dot product similarity matrix of the first feature space vector and the second feature space vector, performs soft maximum value normalization on the matrix in a row direction to obtain an efficiency feature to life factor attention weight matrix, performs soft maximum value normalization on the matrix in a column direction to obtain a life factor to efficiency feature attention weight matrix, multiplies the efficiency feature vector and the life factor attention weight matrix to obtain a life factor modulated efficiency feature, multiplies the life factor vector and the efficiency feature attention weight matrix to obtain an efficiency feature modulated life factor, and adds the original features and the modulated features to complete bidirectional attention fusion.

9. The coal mill efficiency and life prediction method as claimed in claim 1, wherein, The device maintenance priority instruction generation logic comprises: when the mean of the remaining life probability distribution is lower than seven hundred and twenty hours and the variance is less than fifty hours, generating a first-level maintenance instruction, the first-level maintenance instruction triggering an immediate shutdown maintenance process; when the mean of the remaining life probability distribution is between seven hundred and twenty hours and one thousand four hundred and forty hours and the variance is between fifty hours and one hundred hours, generating a second-level maintenance instruction, the second-level maintenance instruction triggering a planned maintenance arrangement within seventy-two hours; when the mean of the remaining life probability distribution is higher than one thousand four hundred and forty hours or the variance is greater than one hundred hours, generating a third-level maintenance instruction, the third-level maintenance instruction only recording a state anomaly and continuously monitoring.

10. An apparatus for applying the method of predicting the efficiency and life of a coal pulverizer according to any one of claims 1 to 9, characterized by, Comprise: A data preprocessing module for receiving a multi-source sensor data stream from a coal mill operation monitoring system, the multi-source sensor data stream comprising a coal mill current signal, an outlet temperature sequence, a coal fineness distribution, a coal supply time sequence, a vibration spectrum feature, and an environmental temperature and humidity parameter, and performing timestamp alignment and working condition normalization processing on the data stream to construct a structured time sequence feature matrix; An efficiency degradation feature extraction module for extracting a coal mill efficiency degradation hidden feature vector based on the structured time sequence feature matrix through a nonlinear dynamic encoder, the nonlinear dynamic encoder adopting a combination structure of a gated memory unit and a multi-scale convolution kernel to capture the nonlinear pattern of performance degradation of the device under load fluctuation and coal quality change conditions; A life degradation driving factor generation module for generating a life degradation driving factor vector based on the structured time sequence feature matrix through a multi-failure mode coupling analysis module, the multi-failure mode coupling analysis module quantifying the interactive enhancement effect between each failure mode by constructing a joint mapping relationship of a mill roller wear rate function, a liner stress accumulation function, and a bearing vibration energy entropy function; A collaborative prediction network module for inputting the efficiency degradation hidden feature vector and the life degradation driving factor vector into a collaborative prediction network with a bidirectional attention mechanism, establishing a dynamic correlation weight between the efficiency feature and the life factor, and outputting an efficiency prediction curve and a remaining life probability distribution of the coal mill in a specified time window in the future. A maintenance decision and optimization suggestion generation module is configured to generate a device maintenance priority instruction and an operation parameter optimization suggestion based on the efficiency prediction curve and the remaining life probability distribution, wherein the device maintenance priority instruction is used to trigger a preventive maintenance work order, and the operation parameter optimization suggestion is used to adjust the coal feeding rate and the air-coal ratio to delay performance degradation.

Images