Coal mill fault diagnosis and prediction method and system based on big data analysis

Abstract

The application provides a coal mill fault diagnosis and prediction method and system based on big data analysis, relates to the technical field of big data analysis, and comprises the following steps: collecting a coal mill bearing vibration signal through a vibration sensor, performing multi-scale analysis and extracting features by using wavelet decomposition, constructing a deep convolutional neural network model, introducing a maximum mean difference adaptive layer to align the features of source domain and target domain data, realizing the identification of fault type and degree, establishing a health state evaluation index containing a fault severity coefficient and a development trend coefficient, and calculating the remaining service life and giving an early warning. The application improves the accuracy and generalization ability of fault diagnosis, realizes accurate prediction and timely early warning of the coal mill bearing fault, and prolongs the service life of the equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of big data analysis, and in particular to a coal mill fault diagnosis and prediction method and system based on big data analysis. Background Art

[0002] Coal mills are key equipment in the pulverized coal preparation system of thermal power plants, and their operating status directly impacts the safe and stable operation of the power plant. Bearings, as core components of coal mills, have a high failure rate. Failures that are not detected and addressed promptly can lead to equipment downtime and even safety accidents. Therefore, real-time monitoring and fault diagnosis and prediction of coal mill bearings are crucial.

[0003] Existing coal mill fault diagnosis methods have the following main problems: 1. Traditional fault diagnosis methods rely primarily on manual judgment, making it difficult to accurately identify faults and provide timely warnings; 2. Conventional machine learning models have poor generalization capabilities when processing fault data under different operating conditions and are difficult to adapt to complex and changing working environments; 3. Existing life prediction methods often ignore the dynamic characteristics of fault development and cannot accurately assess the health status and remaining service life of the equipment.

[0004] The technical problem to be solved by the present invention is to provide a coal mill fault diagnosis and prediction method based on big data analysis. Through deep learning and transfer learning technology, intelligent diagnosis and accurate prediction of coal mill bearing faults can be achieved, the accuracy and reliability of fault diagnosis can be improved, and a scientific basis can be provided for equipment maintenance and management. Summary of the Invention

[0005] The embodiments of the present invention provide a coal mill fault diagnosis and prediction method and system based on big data analysis, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] Provides coal mill fault diagnosis and prediction methods based on big data analysis, including:

[0008] Vibration signals of coal mill bearings are collected in real time using vibration sensors, and the vibration signals are subjected to multi-scale analysis through wavelet decomposition to extract time domain features, frequency domain features, and time-frequency features of the vibration signals. The time domain features, frequency domain features, and time-frequency features are normalized to obtain a eigenvector matrix; a training dataset for a deep convolutional neural network model is constructed based on the eigenvector matrix, and the training dataset is divided into a source domain dataset and a target domain dataset;

[0009] Constructing a deep convolutional neural network model comprising a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; introducing a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively aligning the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer; pre-training the deep convolutional neural network model using the source domain dataset, and fine-tuning training using the target domain dataset;

[0010] The real-time collected vibration signal of the coal mill bearing is input into the trained deep convolutional neural network model to obtain the fault type and fault severity of the coal mill bearing; based on the fault type and the fault severity, a health status assessment index of the coal mill bearing is established, and the health status assessment index includes a fault severity coefficient and a fault development trend coefficient; the remaining service life of the coal mill bearing is calculated according to the health status assessment index, and a warning signal is issued when the remaining service life is lower than a preset health threshold.

[0011] Introducing a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively aligning the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer includes:

[0012] A maximum mean difference adaptive layer is constructed using a Gaussian kernel function; a multi-core mapping function is constructed based on the Gaussian kernel function, wherein the multi-core mapping function is obtained by linearly combining multiple Gaussian kernel functions with different bandwidths, each Gaussian kernel function corresponds to a weight coefficient, and the sum of the weight coefficients is 1;

[0013] Extracting deep feature representations of a source domain dataset and a target domain dataset, calculating channel attention weights of the source domain dataset and the target domain dataset, performing weighted processing on the deep feature representations according to the channel attention weights to obtain weighted source domain features and target domain features; inputting the weighted source domain features and the weighted target domain features into the multi-kernel mapping function, and calculating the maximum mean difference distance between the source domain features and the target domain features;

[0014] The product of the maximum mean difference distance and the dynamic balance parameter is used as the maximum mean difference loss, and the dynamic balance parameter gradually increases from an initial value to a preset value as the number of training rounds increases; the maximum mean difference loss is added to the total loss function of the deep neural network, and the network parameters of the deep neural network and the weight coefficient of the multi-core mapping function are updated through the back propagation algorithm until the maximum mean difference loss is minimized, thereby achieving adaptive alignment of source domain features and target domain features.

[0015] Inputting the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault severity of the coal mill bearing includes:

[0016] The Softmax activation function is used to calculate the probability distribution of the output layer of the deep convolutional neural network model, and a temperature scaling parameter is introduced to calibrate the Softmax output to obtain a calibrated fault type probability distribution. The deep convolutional neural network model is sampled multiple times using the Monte Carlo dropout method to obtain the distribution of fault severity prediction values. The mean and standard deviation of the fault severity prediction value distribution are calculated based on the distribution, and a fault severity confidence interval is constructed.

[0017] Establishing a fault type prior probability distribution model based on the Dirichlet distribution, introducing a time decay factor to dynamically adjust the impact of historical observation data, inputting the calibrated fault type probability distribution into the fault type prior probability distribution model, calculating the posterior probability using the Bayesian formula, and updating the fault type probability; combining the fault degree confidence interval with the updated fault type probability to obtain a comprehensive fault probability;

[0018] The comprehensive fault probability is time-series smoothed using an exponentially weighted moving average method to obtain a smoothed fault probability sequence; an operating condition adjustment factor is constructed based on the equipment load, speed, and temperature, and the operating condition adjustment factor changes dynamically with the equipment operating state; the local standard deviation of the smoothed fault probability sequence is weightedly combined with the operating condition adjustment factor to generate a dynamic fault threshold;

[0019] Perform continuity judgment and trend analysis on the smoothed fault probability sequence. When the fault probability at multiple consecutive time points exceeds the dynamic fault threshold and shows an upward trend, trigger an initial alarm signal. Collect diagnostic results of multiple sensors, calculate the consistency of the diagnostic results of the multiple sensors, and when the consistency exceeds a preset consistency threshold, convert the initial alarm signal into a final alarm signal. Perform fault diagnosis on the coal mill bearing based on the final alarm signal.

[0020] A health status evaluation index of a coal mill bearing is established based on the fault type and the fault severity. The health status evaluation index includes a fault severity coefficient including:

[0021] Based on historical maintenance data of the coal mill, the production decline rate, energy consumption increase rate, and vibration amplitude change rate at the time of failure are collected to construct a performance loss index, which is then converted into a performance impact score through principal component analysis. The stress concentration area is determined based on bearing stress and strain analysis, and the crack growth rate is calculated in combination with material fatigue properties. A Markov chain model is established to obtain the fault growth probability value. The spare parts replacement cost, labor cost, equipment downtime loss, associated fault repair expenditure, and production quality loss are fuzzy comprehensively evaluated to obtain a relative cost coefficient. The hierarchical analysis method is used to combine the performance impact score, the fault growth probability value, and the relative cost coefficient to form a fault type weight matrix.

[0022] Constructing a piecewise nonlinear normalization function to extract time domain features, frequency domain features, and time-frequency domain features from the vibration signal, combining the time domain features, the frequency domain features, the time-frequency domain features, temperature parameters, and acoustic emission parameters to form a feature vector, performing dimensionality reduction on the feature vector using principal component analysis, and establishing a mapping relationship between the feature vector and the fault degree through support vector regression;

[0023] The fault type weight matrix, the fault diagnosis probability and the fault degree processed by the piecewise nonlinear normalization function are weighted and combined to obtain an initial fault severity coefficient; the initial fault severity coefficient is subjected to exponential smoothing processing, and the smoothing coefficient is determined based on the minimization of prediction error criterion to obtain a final fault severity coefficient, and the final fault severity coefficient is used as the basis for fault assessment of the coal mill bearing.

[0024] A health status evaluation index of a coal mill bearing is established based on the fault type and the fault degree. The health status evaluation index includes a fault development trend coefficient including:

[0025] Determine the length of the sliding time window through autocorrelation analysis, and calculate the first-order difference sequence and the second-order difference sequence of the fault severity coefficient within the sliding time window; determine the polynomial fitting order based on the Akaike Information Criterion, perform local weighted polynomial fitting on the first-order difference sequence and the second-order difference sequence, and obtain a fault development trend curve;

[0026] Collecting coal mill speed parameters, load rate parameters, and feed rate parameters, performing dimensionality reduction processing on the speed parameters, load rate parameters, and feed rate parameters through principal component analysis to obtain operating condition characteristics, establishing a multivariate regression model of the operating condition characteristics and the fault development rate, and calculating the ratio of the actual operating condition development rate to the standard operating condition development rate to obtain the operating condition compensation factor;

[0027] Constructing a fuzzy neural network, inputting temperature parameters and humidity parameters into an input layer of the fuzzy neural network, performing fuzzy reasoning using a Gaussian fuzzy membership function in a hidden layer of the fuzzy neural network, obtaining an environmental impact coefficient in an output layer of the fuzzy neural network, and training the fuzzy neural network using a loss function with a regularization term;

[0028] Performing exponential weighted averaging on the first-order difference sequence to obtain a short-term variation characteristic, calculating the derivative of the fault development trend curve to obtain a long-term trend characteristic, and multiplying the working condition compensation factor by the environmental impact coefficient to obtain a comprehensive compensation term;

[0029] The short-term change characteristics, the long-term trend characteristics, and the comprehensive compensation items are weightedly combined to obtain an initial trend coefficient, and the Kalman filter algorithm is used to eliminate noise from the initial trend coefficient to obtain a final fault development trend coefficient.

[0030] Calculating the remaining service life of the coal mill bearing according to the health status assessment index includes:

[0031] Performing nonlinear weighted fusion of the fault severity coefficient and the fault development trend coefficient, introducing the cross term of the fault severity coefficient and the fault development trend coefficient, and establishing a comprehensive health index model, wherein the comprehensive health index model includes a linear term weight coefficient of the fault severity coefficient, a linear term weight coefficient of the fault development trend coefficient, and a cross term weight coefficient;

[0032] A Weibull process model is established based on the comprehensive health index, and a bearing force analysis model is established to calculate the stress distribution under different working conditions. The life loss caused by changes in working conditions is calculated based on Miner's linear cumulative damage theory. The life loss is input into the Weibull process model as a correction term, and a remaining life prediction equation considering the influence of working conditions is established to calculate the remaining service life of the coal mill bearing.

[0033] The method further comprises:

[0034] Using the predicted value of the remaining life prediction equation as a state variable and the health status assessment index as an observation variable, a state equation containing process noise and an observation equation containing measurement noise are established respectively, and an adaptive particle filter correction model is constructed;

[0035] Constructing an adaptive noise covariance matrix, wherein the adaptive noise covariance matrix adopts a recursive form with a forgetting factor, controls the decay rate of historical information by the forgetting factor, constructs an innovation covariance term based on the prediction error at the current moment, and dynamically estimates the process noise in the state equation and the measurement noise in the observation equation;

[0036] Inputting the adaptive noise covariance matrix into the adaptive particle filter correction model to correct the state prediction, calculating the likelihood value of each particle based on the observation data of the observation equation, and updating the weight of the particle;

[0037] The updated particle weights are used to perform weighted averaging on the state estimation values ​​to obtain a revised prediction value of the remaining life of the coal mill bearing. The revised prediction value of the remaining life of the coal mill bearing includes the influence of operating condition changes and random fluctuation characteristics during equipment operation.

[0038] According to a second aspect of the embodiments of the present invention,

[0039] Provide a coal mill fault diagnosis and prediction system based on big data analysis, including:

[0040] The first unit is configured to collect vibration signals of a coal mill bearing in real time using a vibration sensor, perform multi-scale analysis on the vibration signals through wavelet decomposition, extract time domain features, frequency domain features, and time-frequency features of the vibration signals, normalize the time domain features, frequency domain features, and time-frequency features to obtain a feature vector matrix; construct a training data set for a deep convolutional neural network model based on the feature vector matrix, and divide the training data set into a source domain data set and a target domain data set;

[0041] The second unit is used to construct a deep convolutional neural network model including a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; introduce a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively align the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer; pre-train the deep convolutional neural network model using the source domain dataset, and perform fine-tuning training using the target domain dataset;

[0042] The third unit is used to input the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault degree of the coal mill bearing; establish the health status assessment index of the coal mill bearing based on the fault type and the fault degree, and the health status assessment index includes a fault severity coefficient and a fault development trend coefficient; calculate the remaining service life of the coal mill bearing according to the health status assessment index, and issue a warning signal when the remaining service life is lower than a preset health threshold.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] According to a fourth aspect of the embodiments of the present invention,

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0050] The beneficial effects of this application are as follows:

[0051] 1. Improved diagnostic and prediction accuracy: By combining multi-scale feature extraction using wavelet decomposition with deep convolutional neural networks, we can more comprehensively and deeply mine the fault information contained in the vibration signals of coal mill bearings. In particular, by utilizing a maximum mean difference adaptive layer to adaptively align the feature distributions of source and target domain data, we effectively address the model generalization issues caused by data distribution differences under different operating conditions, thereby improving the accuracy of fault diagnosis and prediction.

[0052] 2. Implement health status assessment and life prediction for coal mill bearings: Based on the fault type and fault degree, health status assessment indicators (fault severity coefficient and fault development trend coefficient) are constructed, and the remaining service life is calculated based on this. This achieves a comprehensive assessment of the health status of coal mill bearings and can predict their remaining service life, providing a scientific basis for formulating reasonable maintenance plans.

[0053] 3. Improve equipment reliability and safety, and reduce maintenance costs: Real-time monitoring and fault warning functions can detect potential coal mill bearing failures in advance, avoiding unexpected downtime and significant economic losses caused by failures, and improving equipment reliability and safety. At the same time, by predicting the remaining service life, maintenance strategies can be optimized, avoiding excessive maintenance, and thus reducing maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the process of a coal mill fault diagnosis and prediction method based on big data analysis according to an embodiment of the present invention;

[0055] Figure 2 This is a structural diagram of a coal mill fault diagnosis and prediction system based on big data analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0057] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0058] Figure 1 FIG. 1 is a flow chart of a method for diagnosing and predicting coal mill faults based on big data analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] S101. Using a vibration sensor to collect vibration signals from a coal mill bearing in real time, performing multi-scale analysis on the vibration signals through wavelet decomposition to extract time domain features, frequency domain features, and time-frequency features of the vibration signals, and normalizing the time domain features, frequency domain features, and time-frequency features to obtain a feature vector matrix; constructing a training dataset for a deep convolutional neural network model based on the feature vector matrix, and dividing the training dataset into a source domain dataset and a target domain dataset;

[0060] S102. Construct a deep convolutional neural network model comprising a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; introduce a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively align the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer; pre-train the deep convolutional neural network model using the source domain dataset, and perform fine-tuning training using the target domain dataset;

[0061] S103. Input the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault degree of the coal mill bearing; establish a health status assessment index of the coal mill bearing based on the fault type and the fault degree, and the health status assessment index includes a fault severity coefficient and a fault development trend coefficient; calculate the remaining service life of the coal mill bearing according to the health status assessment index, and issue a warning signal when the remaining service life is lower than a preset health threshold.

[0062] The specific implementation of the coal mill fault diagnosis and prediction method is as follows:

[0063] First, vibration sensors, such as accelerometers, are placed at key locations on the mill's bearings to collect the mill's vibration signals in real time. The frequency of the vibration signals is set at 20 kHz, with a sampling duration of 10 seconds.

[0064] Next, the collected vibration signal is preprocessed, including removing noise and outliers. Wavelet decomposition is used to perform multi-scale analysis of the vibration signal. A suitable mother wavelet, such as the db4 wavelet, is selected to perform a five-layer decomposition of the vibration signal. Time domain features, frequency domain features, and time-frequency features are extracted from the decomposed wavelet coefficients of different frequency bands. Time domain features include mean, standard deviation, peak, kurtosis, etc. Frequency domain features include energy and frequency center of different frequency bands. Time-frequency features include wavelet energy entropy and wavelet packet energy entropy. For example, features such as a mean of 0.5g, a standard deviation of 0.1g, a kurtosis of 3.2, an energy of 0.8J in the 1kHz frequency band, and a wavelet energy entropy of 0.6 are extracted from the original vibration signal. All extracted features are normalized and scaled to between 0 and 1 to form a feature vector matrix.

[0065] The normalized eigenvector matrix is ​​then used to construct a training dataset for the deep convolutional neural network model. The dataset is divided into a source domain dataset and a target domain dataset. The source domain dataset contains a large amount of vibration data from the coal mill under normal operating conditions, as well as vibration data from some common fault conditions, such as bearing inner race faults, outer race faults, and rolling element faults. The target domain dataset contains a small amount of vibration data from the coal mill to be diagnosed, which may include unknown fault types. Assume that the source domain dataset contains 10,000 samples and the target domain dataset contains 100 samples.

[0066] A deep convolutional neural network model is constructed. The model contains multiple convolutional layers, pooling layers, and fully connected layers. Convolutional layers are used to extract features, pooling layers are used to reduce feature dimensions, and fully connected layers are used for classification. A maximum mean difference adaptive layer is introduced into the model to align the feature distributions of the source and target domain datasets and reduce the impact of domain shift. For example, the model contains three convolutional layers, each of which contains 32 convolution kernels of a 3x3 kernel size. The model also contains two pooling layers using maximum pooling with a pooling window size of 2x2. Finally, the model contains two fully connected layers. The first fully connected layer contains 128 neurons, and the second fully connected layer contains the number of neurons in the output layer. The number of neurons in the output layer is the same as the number of fault types, for example, 4, representing normal, inner race fault, outer race fault, and rolling element fault, respectively.

[0067] The deep convolutional neural network model was pre-trained using the source domain dataset. The Adam optimizer was used, with a learning rate of 0.001, a batch size of 32, and 100 training epochs. After pre-training, the model was fine-tuned using the target domain dataset. The learning rate for fine-tuning was set to 0.0001, and the number of training epochs was set to 50.

[0068] The real-time vibration signals collected from the coal mill bearings are fed into a trained deep convolutional neural network model to determine the bearing fault type and severity. For example, if the model output is [0.1, 0.8, 0.05, 0.05], there's an 80% probability that the bearing has an inner race fault.

[0069] Based on the fault type and severity, health assessment indicators for coal mill bearings were established, including the fault severity coefficient and the fault development trend coefficient. The fault severity coefficient indicates the severity of the current fault, ranging from 0 to 1, with larger values ​​indicating more severe faults. The fault development trend coefficient indicates the rate of fault development, ranging from -1 to 1, with positive values ​​indicating worsening faults and negative values ​​indicating improvement. For example, based on the model output and historical data, the fault severity coefficient was calculated to be 0.7 and the fault development trend coefficient was 0.2.

[0070] The remaining service life of a coal mill bearing is calculated based on health status assessment indicators. For example, based on the fault severity coefficient and fault development trend coefficient, as well as the coal mill's operating history data, the remaining service life of the bearing is predicted to be 1,000 hours.

[0071] When the remaining service life is lower than the preset health threshold, such as 500 hours, an early warning signal is issued to remind maintenance personnel to perform maintenance in time.

[0072] Beneficial effects:

[0073] 1. Improve the accuracy of fault diagnosis: Through the deep convolutional neural network model and the maximum mean difference adaptive layer, fault features can be effectively extracted, the impact of domain shift can be reduced, and the accuracy of fault diagnosis can be improved.

[0074] 2. Realize fault prediction and health management: Through health status assessment indicators and remaining service life prediction, fault prediction and health management of coal mill bearings can be realized to avoid unexpected downtime.

[0075] 3. Reduce maintenance costs: Through early warning, maintenance personnel can be guided to carry out targeted maintenance, avoid excessive maintenance, and reduce maintenance costs.

[0076] In an optional embodiment, introducing a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively aligning the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer includes:

[0077] A maximum mean difference adaptive layer is constructed using a Gaussian kernel function; a multi-core mapping function is constructed based on the Gaussian kernel function, wherein the multi-core mapping function is obtained by linearly combining multiple Gaussian kernel functions with different bandwidths, each Gaussian kernel function corresponds to a weight coefficient, and the sum of the weight coefficients is 1;

[0078] Extracting deep feature representations of a source domain dataset and a target domain dataset, calculating channel attention weights of the source domain dataset and the target domain dataset, performing weighted processing on the deep feature representations according to the channel attention weights to obtain weighted source domain features and target domain features; inputting the weighted source domain features and the weighted target domain features into the multi-kernel mapping function, and calculating the maximum mean difference distance between the source domain features and the target domain features;

[0079] The product of the maximum mean difference distance and the dynamic balance parameter is used as the maximum mean difference loss, and the dynamic balance parameter gradually increases from an initial value to a preset value as the number of training rounds increases; the maximum mean difference loss is added to the total loss function of the deep neural network, and the network parameters of the deep neural network and the weight coefficient of the multi-core mapping function are updated through the back propagation algorithm until the maximum mean difference loss is minimized, thereby achieving adaptive alignment of source domain features and target domain features.

[0080] To better achieve adaptive alignment of feature distributions between source and target datasets, a deep convolutional neural network model based on a maximum mean discrepancy (MMD) adaptive layer is proposed. By introducing a multi-kernel mapping function and a channel attention mechanism, this model effectively reduces the distribution discrepancy between the source and target domains, improving the model's generalization performance on the target domain.

[0081] First, a multi-kernel maximum mean difference adaptive layer is constructed. This adaptive layer uses a linear combination of multiple Gaussian kernel functions with different bandwidths to generate a multi-kernel mapping function. Each Gaussian kernel function is assigned a weight coefficient, and the sum of all weight coefficients is 1. For example, three Gaussian kernel functions with bandwidths of 1, 5, and 10 can be selected, with initial weight coefficients set to 0.2, 0.3, and 0.5, respectively. These parameters can be adjusted based on actual conditions.

[0082] Next, deep feature representations are extracted from both the source and target datasets. Assume the source dataset has 1,000 samples and the target dataset has 500 samples, and the extracted feature dimension is 256. Next, channel attention weights are calculated for both the source and target datasets. Channel attention weights can be calculated using methods such as global average pooling. For example, average pooling is performed on the features of each channel to obtain 256 values. These values ​​are then passed through a fully connected layer and a sigmoid activation function to obtain an attention weight for each channel, ranging from 0 to 1.

[0083] Based on the calculated channel attention weights, the deep feature representations of the source and target domains are weighted. The features of each channel are multiplied by the corresponding channel attention weight to obtain the weighted source domain features and target domain features.

[0084] The weighted source and target domain features are input into the constructed multi-kernel mapping function to calculate the maximum mean difference distance between the source and target domain features. This distance represents the degree of difference between the feature distributions of the two domains.

[0085] To balance the impact of the maximum mean difference loss and other losses (such as classification loss), a dynamic balance parameter is introduced. This parameter gradually increases from an initial value to a preset value as training rounds increase. For example, the initial value is set to 0.1 and the preset value is set to 1. As training progresses, the dynamic balance parameter gradually increases linearly from 0.1 to 1. The product of the maximum mean difference distance and the dynamic balance parameter is used as the maximum mean difference loss.

[0086] The maximum mean difference loss is added to the total loss function of the deep neural network. The network parameters of the deep neural network and the weight coefficients of the multi-core mapping function are updated through the backpropagation algorithm until the maximum mean difference loss is minimized. Assuming that the total loss function consists of the classification loss and the maximum mean difference loss, minimizing the total loss function can simultaneously optimize the model's classification performance and domain adaptability.

[0087] Through the above steps, adaptive alignment of source domain features and target domain features can be achieved, thereby improving the generalization performance of the model in the target domain.

[0088] The beneficial effects can be summarized into the following three aspects:

[0089] 1. Improved cross-domain generalization capability of the model: By minimizing the difference between the feature distributions of the source and target domains, it effectively solves the performance degradation problem caused by domain shift and improves the generalization performance of the model in the target domain.

[0090] 2. Enhanced model robustness: The multi-core maximum mean difference adaptation layer can better capture feature differences at different scales, making the model more robust to different types of domain shifts.

[0091] 3. Optimized model training efficiency: The channel attention mechanism can highlight important feature channels and reduce the impact of noise, thereby accelerating the training convergence of the model.

[0092] In an optional embodiment, inputting the real-time collected coal mill bearing vibration signal into a trained deep convolutional neural network model to obtain the fault type and fault severity of the coal mill bearing includes:

[0093] The Softmax activation function is used to calculate the probability distribution of the output layer of the deep convolutional neural network model, and a temperature scaling parameter is introduced to calibrate the Softmax output to obtain a calibrated fault type probability distribution. The deep convolutional neural network model is sampled multiple times using the Monte Carlo dropout method to obtain the distribution of fault severity prediction values. The mean and standard deviation of the fault severity prediction value distribution are calculated based on the distribution, and a fault severity confidence interval is constructed.

[0094] Establishing a fault type prior probability distribution model based on the Dirichlet distribution, introducing a time decay factor to dynamically adjust the impact of historical observation data, inputting the calibrated fault type probability distribution into the fault type prior probability distribution model, calculating the posterior probability using the Bayesian formula, and updating the fault type probability; combining the fault degree confidence interval with the updated fault type probability to obtain a comprehensive fault probability;

[0095] The comprehensive fault probability is time-series smoothed using an exponentially weighted moving average method to obtain a smoothed fault probability sequence; an operating condition adjustment factor is constructed based on the equipment load, speed, and temperature, and the operating condition adjustment factor changes dynamically with the equipment operating state; the local standard deviation of the smoothed fault probability sequence is weightedly combined with the operating condition adjustment factor to generate a dynamic fault threshold;

[0096] Perform continuity judgment and trend analysis on the smoothed fault probability sequence. When the fault probability at multiple consecutive time points exceeds the dynamic fault threshold and shows an upward trend, trigger an initial alarm signal. Collect diagnostic results of multiple sensors, calculate the consistency of the diagnostic results of the multiple sensors, and when the consistency exceeds a preset consistency threshold, convert the initial alarm signal into a final alarm signal. Perform fault diagnosis on the coal mill bearing based on the final alarm signal.

[0097] The deep convolutional neural network model training process is as follows: First, vibration signals of coal mill bearings under different operating conditions and fault states are collected and preprocessed, such as denoising and normalization. Then, a deep convolutional neural network model is constructed, consisting of multiple convolutional layers, pooling layers, and fully connected layers. The model is trained using the preprocessed vibration signal data, and the model parameters are optimized using a suitable loss function and optimization algorithm, enabling the model to accurately identify the fault type and severity of the coal mill bearings.

[0098] Next, the real-time coal mill bearing vibration signals were fed into the trained deep convolutional neural network model. The model's output layer used a softmax activation function to calculate the probability distribution of the fault type. To improve the reliability of the model output, a temperature scaling parameter was introduced to calibrate the softmax output. The temperature scaling parameter controls the smoothness of the probability distribution. A smaller temperature scaling parameter results in a sharper probability distribution, while a larger temperature scaling parameter results in a smoother one. For example, if the original probability distribution of the model output is [0.2, 0.3, 0.5], after calibration with a temperature scaling parameter of 0.5, the probability distribution becomes [0.15, 0.25, 0.6].

[0099] To assess the uncertainty of the fault severity, the deep convolutional neural network model is repeatedly sampled using the Monte Carlo dropout method. Specifically, during prediction, a subset of neurons is randomly disabled. Multiple predictions are then made to obtain multiple fault severity predictions. For example, 100 samples are taken to obtain 100 fault severity predictions. Based on these predictions, their mean and standard deviation are calculated, and a confidence interval for the fault severity is constructed. For example, if the mean of the 100 predictions is 0.8 and the standard deviation is 0.05, a 95% confidence interval can be constructed as [0.7, 0.9].

[0100] To incorporate historical observational data, a prior probability distribution model for fault types based on the Dirichlet distribution is established. This model estimates the probabilities of different fault types based on historical data. To dynamically adjust for the influence of historical data, a time decay factor is introduced. The time decay factor gradually decreases over time, making recent data have a greater impact on the prior probability. For example, assuming the current time is t, the time decay factor can be set to exp(-λ(t-t0)), where λ is the decay rate and t0 is the initial time. The calibrated fault type probability distribution is input into the prior probability distribution model, and the posterior probability is calculated using the Bayesian formula to update the fault type probability. For example, assuming the prior probability is [0.1, 0.2, 0.7] and the calibrated probability distribution is [0.15, 0.25, 0.6], the posterior probability can be calculated as [0.12, 0.22, 0.66].

[0101] The confidence interval for the fault severity is combined with the updated probability of the fault type to obtain the composite fault probability. For example, assuming the fault types are A, B, and C, with updated probabilities of 0.12, 0.22, and 0.66, respectively, and the confidence interval for the fault severity is [0.7, 0.9], the composite fault probability can be expressed as: A: [0.084, 0.108], B: [0.154, 0.198], C: [0.462, 0.594].

[0102] The exponentially weighted moving average method is used to smooth the time series of the comprehensive failure probability to obtain the smoothed failure probability sequence. For example, if the comprehensive failure probability at the current time point is 0.8, the smoothed failure probability at the previous time point is 0.7, and the attenuation factor is 0.9, then the smoothed failure probability at the current time point is 0.79.

[0103] An operating condition adjustment factor is constructed based on operating parameters such as equipment load, speed, and temperature. The operating condition adjustment factor changes dynamically with the equipment's operating status. For example, the higher the equipment load, the larger the operating condition adjustment factor. A dynamic fault threshold is generated by weighting the local standard deviation of the smoothed fault probability series with the operating condition adjustment factor. For example, if the local standard deviation is 0.05, the operating condition adjustment factor is 1.2, and the weights are 0.8 and 0.2, respectively, the dynamic fault threshold is 0.046.

[0104] Perform continuity assessment and trend analysis on the smoothed fault probability sequence. When the fault probability at multiple consecutive time points exceeds the dynamic fault threshold and shows an upward trend, an initial alarm signal is triggered. For example, if the fault probability at five consecutive time points exceeds the dynamic fault threshold and shows an upward trend, an initial alarm signal is triggered.

[0105] The diagnostic results of multiple sensors are collected and the consistency of these results is calculated. When the consistency exceeds a preset consistency threshold, the initial alarm signal is converted to a final alarm signal. For example, if the diagnostic results of all three sensors indicate a fault, the consistency is high, and the initial alarm signal is converted to a final alarm signal.

[0106] The coal mill bearing fault diagnosis is performed based on the final alarm signal.

[0107] The beneficial effects of this method can be summarized into the following three parts:

[0108] 1. Improve the accuracy of fault diagnosis: Through technologies such as deep learning models, temperature scaling parameters, Monte Carlo dropout methods, and Bayesian formulas, fault types can be identified and fault severity assessed more accurately.

[0109] 2. Enhance the reliability of fault diagnosis: Through technologies such as time attenuation factor, time series smoothing processing, working condition adjustment factor and multi-sensor diagnosis, the false alarm rate and missed alarm rate can be effectively reduced, and the reliability of fault diagnosis can be improved.

[0110] 3. Improve the efficiency of fault diagnosis: Through technologies such as dynamic fault thresholds and continuity judgment, faults can be discovered and alarms can be issued in a timely manner, thus avoiding equipment downtime and economic losses caused by faults.

[0111] In an optional embodiment, a health status assessment index of a coal mill bearing is established based on the fault type and the fault severity, wherein the health status assessment index includes a fault severity coefficient including:

[0112] Based on historical maintenance data of the coal mill, the production decline rate, energy consumption increase rate, and vibration amplitude change rate at the time of failure are collected to construct a performance loss index, which is then converted into a performance impact score through principal component analysis. The stress concentration area is determined based on bearing stress and strain analysis, and the crack growth rate is calculated in combination with material fatigue properties. A Markov chain model is established to obtain the fault growth probability value. The spare parts replacement cost, labor cost, equipment downtime loss, associated fault repair expenditure, and production quality loss are fuzzy comprehensively evaluated to obtain a relative cost coefficient. The hierarchical analysis method is used to combine the performance impact score, the fault growth probability value, and the relative cost coefficient to form a fault type weight matrix.

[0113] Constructing a piecewise nonlinear normalization function to extract time domain features, frequency domain features, and time-frequency domain features from the vibration signal, combining the time domain features, the frequency domain features, the time-frequency domain features, temperature parameters, and acoustic emission parameters to form a feature vector, performing dimensionality reduction on the feature vector using principal component analysis, and establishing a mapping relationship between the feature vector and the fault degree through support vector regression;

[0114] The fault type weight matrix, the fault diagnosis probability and the fault degree processed by the piecewise nonlinear normalization function are weighted and combined to obtain an initial fault severity coefficient; the initial fault severity coefficient is subjected to exponential smoothing processing, and the smoothing coefficient is determined based on the minimization of prediction error criterion to obtain a final fault severity coefficient, and the final fault severity coefficient is used as the basis for fault assessment of the coal mill bearing.

[0115] The specific implementation steps of the coal mill bearing health status assessment method are as follows:

[0116] First, collect historical maintenance records for the coal mill, including the type of each failure, the time of occurrence, the percentage decrease in production, the percentage increase in energy consumption, and the percentage change in vibration amplitude. For example, a bearing wear failure on January 10, 2023, resulted in a 5% decrease in production, an 8% increase in energy consumption, and a 15% increase in vibration amplitude. Collect similar data for multiple samples.

[0117] Then, based on the collected historical data, statistical methods were used to calculate the average production decline rate, energy consumption increase rate, and vibration amplitude change rate for each fault type. These average values ​​reflect the impact of different fault types on coal mill performance. For example, the average production decline rate for bearing wear faults was 6%, the average energy consumption increase rate was 9%, and the average vibration amplitude increase rate was 16%.

[0118] Next, the average values ​​of the production decline rate, energy consumption increase rate, and vibration amplitude change rate for each fault type are combined into a performance loss index. For example, the performance loss index for a bearing wear fault can be expressed as (6%, 9%, 16%).

[0119] Then, using principal component analysis, we converted the multidimensional performance loss indicators into a single performance impact score. For example, the performance impact score for a bearing wear fault is 0.8, while the performance impact score for a bearing fracture fault is 0.95. A higher score indicates a greater performance impact.

[0120] Simultaneously, stress-strain analysis is performed on the bearing to identify areas of stress concentration. For example, finite element analysis reveals that the inner ring raceway is a stress concentration area. The crack growth rate is then calculated based on the fatigue properties of the bearing material, such as the S-N curve. For example, the crack growth rate is calculated to be 1 mm per year at a specific stress level.

[0121] A Markov chain model is built to predict the probability of a fault developing from its current state to a more serious state. For example, the probability of a bearing with mild wear developing into severe wear is 0.1, and the probability of severe wear developing into fracture is 0.2.

[0122] Next, the relative cost factor for each failure type is evaluated. This requires considering spare part replacement costs, labor costs, equipment downtime losses, repair costs for other potential failures, and lost production quality. For example, the relative cost factor for bearing wear is 0.5, while the relative cost factor for bearing fracture is 1.

[0123] Using the analytic hierarchy process, the performance impact score, fault extension probability value, and relative cost coefficient are combined into a fault type weight matrix. For example, the weight of bearing wear is 0.3, and the weight of bearing fracture is 0.7.

[0124] Extract time-domain, frequency-domain, and time-frequency-domain features from the vibration signals of the coal mill bearings. Time-domain features include root mean square (RMS) values ​​and peak values; frequency-domain features include characteristic frequency amplitudes; and time-frequency-domain features include wavelet energy. Simultaneously, the bearing temperature and acoustic emission parameters are collected. All these features are combined with the temperature and acoustic emission parameters to form a feature vector.

[0125] Principal component analysis is used to reduce the dimensionality of the feature vector, for example, from a 10-dimensional feature vector to a 3-dimensional one. Support vector regression is then used to establish a mapping between the feature vector and the fault severity. For example, the feature vector (0.1, 0.2, 0.3) corresponds to a fault severity of 0.5. A piecewise nonlinear normalization function is used to process the fault severity, keeping it between 0 and 1.

[0126] The initial fault severity coefficient is obtained by weighting the fault type weight matrix, the fault diagnosis probability, and the fault severity processed by the piecewise nonlinear normalization function. This initial fault severity coefficient is then exponentially smoothed, for example, with a smoothing coefficient of 0.8, to obtain the final fault severity coefficient. This coefficient serves as the basis for fault assessment of the coal mill bearing.

[0127] The beneficial effects of this method are reflected in the following three aspects:

[0128] 1. Improve assessment accuracy: Comprehensively considering performance loss, failure development probability and cost factors, it can more comprehensively and accurately assess the health status of bearings.

[0129] 2. Achieve early warning: Through the crack growth rate and Markov model, the fault development trend can be predicted, early warning can be achieved, and major failures can be avoided.

[0130] 3. Optimize maintenance strategy: Based on the fault severity coefficient, a more reasonable maintenance strategy can be formulated to avoid over-maintenance or under-maintenance and reduce maintenance costs.

[0131] In an optional embodiment, a health status assessment index of a coal mill bearing is established based on the fault type and the fault degree, and the health status assessment index includes a fault development trend coefficient including:

[0132] Determine the length of the sliding time window through autocorrelation analysis, and calculate the first-order difference sequence and the second-order difference sequence of the fault severity coefficient within the sliding time window; determine the polynomial fitting order based on the Akaike Information Criterion, perform local weighted polynomial fitting on the first-order difference sequence and the second-order difference sequence, and obtain a fault development trend curve;

[0133] Collecting coal mill speed parameters, load rate parameters, and feed rate parameters, performing dimensionality reduction processing on the speed parameters, load rate parameters, and feed rate parameters through principal component analysis to obtain operating condition characteristics, establishing a multivariate regression model of the operating condition characteristics and the fault development rate, and calculating the ratio of the actual operating condition development rate to the standard operating condition development rate to obtain the operating condition compensation factor;

[0134] Constructing a fuzzy neural network, inputting temperature parameters and humidity parameters into an input layer of the fuzzy neural network, performing fuzzy reasoning using a Gaussian fuzzy membership function in a hidden layer of the fuzzy neural network, obtaining an environmental impact coefficient in an output layer of the fuzzy neural network, and training the fuzzy neural network using a loss function with a regularization term;

[0135] Performing exponential weighted averaging on the first-order difference sequence to obtain a short-term variation characteristic, calculating the derivative of the fault development trend curve to obtain a long-term trend characteristic, and multiplying the working condition compensation factor by the environmental impact coefficient to obtain a comprehensive compensation term;

[0136] The short-term change characteristics, the long-term trend characteristics, and the comprehensive compensation items are weightedly combined to obtain an initial trend coefficient, and the Kalman filter algorithm is used to eliminate noise from the initial trend coefficient to obtain a final fault development trend coefficient.

[0137] The core of the coal mill bearing health assessment method lies in constructing an indicator that reflects the development trend of faults. This indicator comprehensively considers multiple factors, including fault type, fault severity, and operating environment, aiming to more accurately predict the remaining life of the bearing and avoid unplanned downtime.

[0138] First, the vibration signals of the coal mill bearings are collected and analyzed based on known fault types. For example, envelope analysis can be used to extract the characteristic frequencies associated with a specific fault type. Then, based on the characteristic frequencies of the fault type, the fault severity coefficient is calculated. For example, the RMS value or peak value of the vibration signal can be used to quantify the fault severity.

[0139] Next, determine an appropriate sliding time window. The length of this window should be determined based on the autocorrelation characteristics of the vibration signal to ensure that the signal within the window fully reflects the development trend of the fault. For example, the autocorrelation function of the vibration signal can be calculated and the appropriate window length can be selected based on the decay characteristics of the autocorrelation function. In one example, if the autocorrelation function decays to near zero after 100 data points, the sliding time window length can be set to 100.

[0140] Within a sliding time window, the first-order and second-order difference sequences of the fault severity coefficients are calculated. The first-order difference reflects the rate of change of the fault severity, while the second-order difference reflects the rate of change, i.e., the acceleration of the fault development. For example, if the fault severity coefficients within a window are 1, 2, 4, 7, and 11, respectively, the first-order difference sequence is 1, 2, 3, 4, and the second-order difference sequence is 1, 1, 1.

[0141] A locally weighted polynomial fit is performed on the first-order and second-order difference sequences to obtain the fault development trend curve. The fitting order can be determined based on the Akaike Information Criterion (AIC), with the order that minimizes the AIC value being selected. The idea behind local weighting is that during the fitting process, samples closer to the current point are given greater weight, while samples farther away are given less weight. Assuming the fitting order is 3 based on the AIC criterion, a locally weighted cubic polynomial fit can be used to obtain the fault development trend curve.

[0142] At the same time, operating parameters such as the coal mill speed, load factor, and feed rate are collected. Principal component analysis is performed on these parameters to reduce their dimensionality to one or more key operating characteristics. For example, suppose the principal component analysis extracts two principal components, representing the coal mill load and feed conditions, respectively.

[0143] A multivariate regression model is established to correlate operating condition characteristics with fault development rate. The fault development rate can be derived by taking the derivative of the fault development trend curve. This model can be used to predict the speed of fault development under different operating conditions. For example, a simple linear regression model can be expressed as: Fault Development Rate = a * Load Condition + b * Feed Condition + c, where a, b, and c are regression coefficients.

[0144] The ratio of the actual operating rate to the standard operating rate is calculated to obtain the operating compensation factor. The standard operating condition can be the ideal operating condition for the coal mill design. For example, if the actual operating rate is 1.2 times the standard operating rate, the operating compensation factor is 1.2.

[0145] Ambient temperature and humidity parameters are collected and input into a fuzzy neural network. The network's hidden layer uses Gaussian fuzzy membership functions for fuzzy inference, while the output layer outputs the environmental impact coefficient. When training the fuzzy neural network, a loss function with a regularization term is used to prevent overfitting. For example, assuming a temperature of 25°C and a humidity of 60%, the fuzzy neural network calculates an environmental impact coefficient of 0.9.

[0146] Perform an exponentially weighted average on the first-order difference sequence to obtain the short-term variation characteristics. Exponentially weighted averaging prioritizes recent data and can better reflect the short-term variation trend of faults. For example, assuming an attenuation factor of 0.8 and the first-order difference sequence of 1, 2, 3, and 4, the short-term variation characteristic is 3.328.

[0147] Calculate the derivative of the fault development trend curve to obtain the long-term trend characteristics. This reflects the long-term development trend of the fault. Multiply the operating condition compensation factor by the environmental impact coefficient to obtain the comprehensive compensation term. For example, if the operating condition compensation factor is 1.2 and the environmental impact coefficient is 0.9, the comprehensive compensation term is 1.08.

[0148] The initial trend coefficient is calculated by weighting the short-term change characteristics, long-term trend characteristics, and comprehensive compensation items. The weights can be determined based on experience or expert knowledge. For example, assuming the weight of the short-term change characteristics is 0.4, the weight of the long-term trend characteristics is 0.5, and the weight of the comprehensive compensation item is 0.1, then the initial trend coefficient = 0.4 * short-term change characteristics + 0.5 * long-term trend characteristics + 0.1 * comprehensive compensation item.

[0149] Finally, the Kalman filter algorithm is used to eliminate noise from the initial trend coefficient to obtain the final fault development trend coefficient. Kalman filtering can effectively remove noise from the signal and improve the accuracy of the evaluation results.

[0150] The beneficial effects of this method can be summarized in the following three aspects:

[0151] 1. Improved assessment accuracy: This method comprehensively considers multiple factors such as fault type, fault severity, and operating environment, and can more accurately assess the health status of coal mill bearings.

[0152] 2. Enhanced reliability of assessment: The use of advanced technologies such as local weighted polynomial fitting, fuzzy neural network, and Kalman filtering effectively reduces the impact of noise and interference and improves the reliability of assessment results.

[0153] 3. Improved practicality of the assessment: This method can provide a quantitative failure development trend coefficient, which can be used to predict the remaining life of the bearing, guide maintenance work, avoid unexpected downtime, and thus improve production efficiency.

[0154] In an optional embodiment, calculating the remaining service life of the coal mill bearing according to the health status assessment index includes:

[0155] Performing nonlinear weighted fusion of the fault severity coefficient and the fault development trend coefficient, introducing the cross term of the fault severity coefficient and the fault development trend coefficient, and establishing a comprehensive health index model, wherein the comprehensive health index model includes a linear term weight coefficient of the fault severity coefficient, a linear term weight coefficient of the fault development trend coefficient, and a cross term weight coefficient;

[0156] A Weibull process model is established based on the comprehensive health index, and a bearing force analysis model is established to calculate the stress distribution under different working conditions. The life loss caused by changes in working conditions is calculated based on Miner's linear cumulative damage theory. The life loss is input into the Weibull process model as a correction term, and a remaining life prediction equation considering the influence of working conditions is established to calculate the remaining service life of the coal mill bearing.

[0157] The remaining service life assessment method of coal mill bearings is specifically implemented as follows:

[0158] First, collect the vibration signal of the coal mill bearing. For example, use an acceleration sensor to collect the vibration signal of the bearing during operation, set the sampling frequency to 20kHz, and set the sampling time to 10 seconds.

[0159] The collected vibration signals are then preprocessed. This includes trend removal, filtering, and noise reduction. For example, a high-pass filter is used to remove low-frequency noise, with a cutoff frequency set to 500 Hz. A wavelet threshold denoising method is used to remove high-frequency noise, using the db4 wavelet basis function and a decomposition level of 5.

[0160] Next, we extract bearing fault features. Based on the preprocessed vibration signal, we extract characteristic indicators reflecting the severity and development trend of the bearing fault. For example, extracted fault severity indicators include the root mean square value, peak value, and kurtosis; extracted fault development trend indicators include energy entropy and singular values. Assume that the extracted root mean square value is 0.8g, the peak value is 2.5g, the kurtosis is 4.2, the energy entropy is 0.7, and the singular value is 0.3.

[0161] Subsequently, the fault severity coefficient and fault development trend coefficient are calculated. Based on the extracted fault characteristic indicators, the fault severity coefficient and fault development trend coefficient are calculated separately. For example, a fuzzy membership function is used to map the root mean square value, peak value, and kurtosis to the [0, 1] interval, and then the fault severity coefficient is obtained by weighted averaging. A normalization method is used to map the energy entropy and singular values ​​to the [0, 1] interval, and then the fault development trend coefficient is obtained by weighted averaging. Assume that the calculated fault severity coefficient is 0.6 and the fault development trend coefficient is 0.4.

[0162] Next, a comprehensive health index model is established. The fault severity coefficient and the fault development trend coefficient are nonlinearly weighted and fused, and their cross-term is introduced to establish the comprehensive health index model. This model includes the linear weight coefficients for the fault severity coefficient, the linear weight coefficients for the fault development trend coefficient, and the cross-term weight coefficients. For example, assuming the linear weight coefficient for the fault severity coefficient is 0.7, the linear weight coefficient for the fault development trend coefficient is 0.3, and the cross-term weight coefficient is 0.2, the comprehensive health index is then 0.7*0.6+0.3*0.4+0.2*0.6*0.4=0.588.

[0163] Then, a Weibull process model is established. Based on the calculated comprehensive health index, a Weibull process model is established. The shape and scale parameters of the Weibull process model can be determined using the maximum likelihood estimation method. Assume that the shape parameter is 2.5 and the scale parameter is 10,000 hours.

[0164] Next, a bearing stress analysis model was established. Based on the actual operating conditions of the coal mill, this model was used to calculate the stress distribution under different operating conditions. For example, using finite element analysis, the maximum stress under rated operating conditions was calculated to be 500 MPa, 300 MPa under light load conditions, and 700 MPa under heavy load conditions.

[0165] Next, calculate the life loss due to changes in operating conditions. Based on Miner's linear cumulative damage theory, calculate the life loss due to changes in operating conditions. For example, assume that the coal mill operates at rated conditions for 80% of the time, under light load conditions for 10% of the time, and under heavy load conditions for 10% of the time. Based on the stress distribution under different operating conditions and the corresponding fatigue life curves, the life loss under each operating condition can be calculated and then accumulated to obtain the total life loss. Assume that the calculated life loss is 1000 hours.

[0166] Next, establish a remaining life prediction equation that accounts for the effects of operating conditions. The calculated life loss is input into the Weibull process model as a correction term to establish a remaining life prediction equation that accounts for operating conditions. For example, subtracting the life loss from the scale parameter of the Weibull process model yields a corrected scale parameter of 9000 hours.

[0167] Finally, the remaining useful life of the coal mill bearings is calculated using the modified Weibull process model. For example, assuming the current operating time is 5000 hours, the remaining useful life can be calculated using the Weibull distribution function.

[0168] The beneficial effects of this method can be summarized in the following three aspects:

[0169] 1. Improved remaining life prediction accuracy: By introducing the cross term of the fault severity coefficient and the fault development trend coefficient, and considering the impact of operating condition changes, the degradation process of the bearing can be reflected more accurately, thereby improving the remaining life prediction accuracy.

[0170] 2. Realize online assessment of remaining life: This method can perform real-time calculations based on online monitoring data, thereby realizing online assessment of remaining life and providing a decision-making basis for equipment maintenance.

[0171] 3. Strong practicability: This method does not require complex mathematical formulas, is easy to understand and apply, and can be widely used in various types of coal mill bearings.

[0172] In an optional embodiment, the method further includes:

[0173] Using the predicted value of the remaining life prediction equation as a state variable and the health status assessment index as an observation variable, a state equation containing process noise and an observation equation containing measurement noise are established respectively, and an adaptive particle filter correction model is constructed;

[0174] Constructing an adaptive noise covariance matrix, wherein the adaptive noise covariance matrix adopts a recursive form with a forgetting factor, controls the decay rate of historical information by the forgetting factor, constructs an innovation covariance term based on the prediction error at the current moment, and dynamically estimates the process noise in the state equation and the measurement noise in the observation equation;

[0175] Inputting the adaptive noise covariance matrix into the adaptive particle filter correction model to correct the state prediction, calculating the likelihood value of each particle based on the observation data of the observation equation, and updating the weight of the particle;

[0176] The updated particle weights are used to perform weighted averaging on the state estimation values ​​to obtain a revised prediction value of the remaining life of the coal mill bearing. The revised prediction value of the remaining life of the coal mill bearing includes the influence of operating condition changes and random fluctuation characteristics during equipment operation.

[0177] First, we collected operating data from the coal mill bearings, including vibration signals, temperature, and current, as well as health indicators reflecting the degree of bearing wear, such as energy and kurtosis. For example, we collected vibration signals every 10 minutes, collecting 1024 data points each time. Furthermore, we recorded bearing temperature and current every hour, and calculated energy and kurtosis based on the vibration signals as health indicators.

[0178] Next, using existing coal mill bearing life data, such as historical failure data or accelerated life test data, a prediction equation for the remaining life of the coal mill bearing is established. This equation can be a model established based on regression analysis, neural networks, and other methods. For example, a support vector machine method can be used to establish a prediction equation for the remaining life of the coal mill bearing based on historical failure data and corresponding health status assessment indicators. Assume that the remaining life prediction equation established based on historical data is: remaining life = 1000 - 5 * energy index - 2 * kurtosis index;

[0179] Then, an adaptive particle filter correction model is constructed, using the predicted value of the remaining life prediction equation as the state variable and the health status assessment indicator as the observation variable. The state equation describes the evolution of the remaining life over time and includes process noise, which accounts for the impact of unaccounted factors on the remaining life. The observation equation describes the relationship between the observation variable and the state variable and includes measurement noise, which accounts for the influence of measurement error. For example, the state equation can be set as: Remaining life at current moment = Remaining life at previous moment - Wear rate + Process noise; the observation equation can be set as: Energy index at current moment = Function related to remaining life at current moment + Measurement noise.

[0180] Construct an adaptive noise covariance matrix. This matrix uses a recursive form with a forgetting factor, which controls the decay rate of historical information. The forgetting factor is typically between 0.9 and 0.99. Based on the current forecast error, an innovation covariance term is constructed to dynamically estimate the process noise in the state equation and the measurement noise in the observation equation. For example, assuming the current forecast error is 20 hours, the forgetting factor is 0.95, and the process noise covariance at the previous moment is 10, the current process noise covariance can be updated to 0.95*10+(1-0.95)*20*20=29.5.

[0181] The adaptive noise covariance matrix is ​​input into the adaptive particle filter correction model to refine the state prediction. First, the state prediction is performed based on the state equation to obtain the predicted value for each particle. Then, based on the observed data from the observation equation, such as the current energy index and kurtosis index, the likelihood value of each particle is calculated. The likelihood value reflects the degree of consistency between the particle and the actual observed data. The particle weight is then updated based on the likelihood value, with particles with larger weights indicating a closer approximation to the true state.

[0182] The updated particle weights are used to perform a weighted average of the state estimates to obtain a revised prediction of the remaining life of the coal mill bearings. For example, assuming there are 100 particles, each representing a possible remaining life value, the weighted average of these 100 remaining life values ​​based on the updated weights yields a revised prediction of the remaining life of the coal mill bearings. This revised prediction accounts for the effects of varying operating conditions and random fluctuations during equipment operation.

[0183] The beneficial effects of this method can be summarized into the following three parts:

[0184] 1. Improve prediction accuracy: By modifying the model with adaptive particle filtering, the influence of process noise and measurement noise can be effectively reduced, and the accuracy of the prediction of the remaining life of the coal mill bearing can be improved.

[0185] 2. Adapt to changes in operating conditions: The adaptive noise covariance matrix can dynamically adjust the noise estimation according to the actual operating conditions, so that the prediction model can better adapt to changes in operating conditions and improve the reliability of the prediction.

[0186] 3. Extend bearing life: Accurate remaining life prediction can help formulate a more reasonable maintenance plan and avoid replacing bearings too early or too late, thereby extending the service life of bearings and reducing maintenance costs.

[0187] Figure 2 FIG. 1 is a schematic diagram of the structure of a coal mill fault diagnosis and prediction system based on big data analysis according to an embodiment of the present invention. Figure 2 As shown, the system includes:

[0188] The first unit is configured to collect vibration signals of a coal mill bearing in real time using a vibration sensor, perform multi-scale analysis on the vibration signals through wavelet decomposition, extract time domain features, frequency domain features, and time-frequency features of the vibration signals, normalize the time domain features, frequency domain features, and time-frequency features to obtain a feature vector matrix; construct a training data set for a deep convolutional neural network model based on the feature vector matrix, and divide the training data set into a source domain data set and a target domain data set;

[0189] The second unit is used to construct a deep convolutional neural network model including a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; introduce a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively align the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer; pre-train the deep convolutional neural network model using the source domain dataset, and perform fine-tuning training using the target domain dataset;

[0190] The third unit is used to input the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault degree of the coal mill bearing; establish the health status assessment index of the coal mill bearing based on the fault type and the fault degree, and the health status assessment index includes a fault severity coefficient and a fault development trend coefficient; calculate the remaining service life of the coal mill bearing according to the health status assessment index, and issue a warning signal when the remaining service life is lower than a preset health threshold.

[0191] According to a third aspect of the embodiments of the present invention,

[0192] An electronic device is provided, comprising:

[0193] processor;

[0194] a memory for storing processor-executable instructions;

[0195] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0196] According to a fourth aspect of the embodiments of the present invention,

[0197] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0198] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A coal mill fault diagnosis and prediction method based on big data analysis is characterized by: include: A vibration sensor is used to collect vibration signals of a coal mill bearing in real time, and the vibration signals are subjected to multi-scale analysis through wavelet decomposition to extract time domain features, frequency domain features, and time-frequency features of the vibration signals. The time domain features, frequency domain features, and time-frequency features are normalized to obtain a feature vector matrix. Constructing a training data set for a deep convolutional neural network model based on the feature vector matrix, and dividing the training data set into a source domain data set and a target domain data set; Constructing a deep convolutional neural network model comprising a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; introducing a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively aligning the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer; Pre-training the deep convolutional neural network model using the source domain dataset and performing fine-tuning training using the target domain dataset; Inputting the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault severity of the coal mill bearing; establishing a health status assessment index for the coal mill bearing based on the fault type and the fault severity, the health status assessment index including a fault severity coefficient and a fault development trend coefficient, the fault severity coefficient indicating the severity of the fault, and the fault development trend coefficient indicating the speed of fault development; Calculating the remaining service life of the coal mill bearing according to the health status assessment index, performing nonlinear weighted fusion on the fault severity coefficient and the fault development trend coefficient, introducing a cross term between the fault severity coefficient and the fault development trend coefficient, and establishing a comprehensive health index model, wherein the comprehensive health index model includes a linear term weight coefficient of the fault severity coefficient, a linear term weight coefficient of the fault development trend coefficient, and a cross term weight coefficient; A Weibull process model is established based on the comprehensive health index, and a bearing force analysis model is established to calculate the stress distribution under different working conditions. The life loss caused by changes in working conditions is calculated based on Miner's linear cumulative damage theory. The life loss is input into the Weibull process model as a correction term, and a remaining life prediction equation considering the influence of working conditions is established to calculate the remaining service life of the coal mill bearing, and a warning signal is issued when the remaining service life is lower than a preset health threshold.

2. The method according to claim 1, characterized in that Introducing a maximum mean difference adaptive layer into the deep convolutional neural network model, and adaptively aligning the feature distributions of the source domain dataset and the target domain dataset through the maximum mean difference adaptive layer includes: A maximum mean difference adaptive layer is constructed using a Gaussian kernel function; the maximum mean difference adaptive layer constructs a multi-core mapping function based on the Gaussian kernel function, the multi-core mapping function being obtained by linearly combining multiple Gaussian kernel functions with different bandwidths, each Gaussian kernel function corresponding to a weight coefficient, and the sum of the weight coefficients being 1; Extracting deep feature representations of a source domain dataset and a target domain dataset, calculating channel attention weights of the source domain dataset and the target domain dataset, performing weighted processing on the deep feature representations according to the channel attention weights to obtain weighted source domain features and target domain features; inputting the weighted source domain features and the weighted target domain features into the multi-kernel mapping function, and calculating the maximum mean difference distance between the source domain features and the target domain features; The product of the maximum mean difference distance and the dynamic balance parameter is used as the maximum mean difference loss, and the dynamic balance parameter gradually increases from an initial value to a preset value as the number of training rounds increases; the maximum mean difference loss is added to the total loss function of the deep neural network, and the network parameters of the deep neural network and the weight coefficient of the multi-core mapping function are updated through the back propagation algorithm until the maximum mean difference loss is minimized, thereby achieving adaptive alignment of source domain features and target domain features.

3. The method according to claim 1, characterized in that Inputting the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault severity of the coal mill bearing includes: The real-time collected coal mill bearing vibration signal is input into the trained deep convolutional neural network model. The Softmax activation function is used to calculate the probability distribution of the output layer of the deep convolutional neural network model. The temperature scaling parameter is introduced to calibrate the Softmax output to obtain the calibrated fault type probability distribution. The deep convolutional neural network model is sampled multiple times using the Monte Carlo dropout method to obtain the distribution of fault severity prediction values. The mean and standard deviation of the fault severity prediction value distribution are calculated based on the distribution of the fault severity prediction values, and a fault severity confidence interval is constructed. Establishing a fault type prior probability distribution model based on the Dirichlet distribution, introducing a time decay factor to dynamically adjust the impact of historical observation data, inputting the calibrated fault type probability distribution into the fault type prior probability distribution model, calculating the posterior probability using the Bayesian formula, and updating the fault type probability; combining the fault degree confidence interval with the updated fault type probability to obtain a comprehensive fault probability; The comprehensive fault probability is time-series smoothed using an exponentially weighted moving average method to obtain a smoothed fault probability sequence; an operating condition adjustment factor is constructed based on the equipment load, speed, and temperature, and the operating condition adjustment factor changes dynamically with the equipment operating state; the local standard deviation of the smoothed fault probability sequence is weightedly combined with the operating condition adjustment factor to generate a dynamic fault threshold; Perform continuity judgment and trend analysis on the smoothed fault probability sequence. When the fault probability at multiple consecutive time points exceeds the dynamic fault threshold and shows an upward trend, trigger an initial alarm signal. Collect diagnostic results of multiple sensors, calculate the consistency of the diagnostic results of the multiple sensors, and when the consistency exceeds a preset consistency threshold, convert the initial alarm signal into a final alarm signal. Perform fault diagnosis on the coal mill bearing based on the final alarm signal.

4. The method according to claim 1, wherein A health status evaluation index of a coal mill bearing is established based on the fault type and the fault severity. The health status evaluation index includes a fault severity coefficient including: Based on historical maintenance data of the coal mill, the production decline rate, energy consumption increase rate, and vibration amplitude change rate at the time of failure are collected to construct a performance loss index, which is then converted into a performance impact score through principal component analysis. The stress concentration area is determined based on bearing stress and strain analysis, and the crack growth rate is calculated in combination with material fatigue properties. A Markov chain model is established to obtain the fault growth probability value. The spare parts replacement cost, labor cost, equipment downtime loss, associated fault repair expenditure, and production quality loss are fuzzy comprehensively evaluated to obtain a relative cost coefficient. The hierarchical analysis method is used to combine the performance impact score, the fault growth probability value, and the relative cost coefficient to form a fault type weight matrix. Constructing a piecewise nonlinear normalization function to extract time domain features, frequency domain features, and time-frequency domain features from the vibration signal, combining the time domain features, the frequency domain features, the time-frequency domain features, temperature parameters, and acoustic emission parameters to form a feature vector, performing dimensionality reduction on the feature vector using principal component analysis, and establishing a mapping relationship between the feature vector and the fault degree through support vector regression; The fault type weight matrix, the fault diagnosis probability and the fault degree processed by the piecewise nonlinear normalization function are weighted and combined to obtain an initial fault severity coefficient; the initial fault severity coefficient is subjected to exponential smoothing processing, and the smoothing coefficient is determined based on the minimization of prediction error criterion to obtain a final fault severity coefficient, and the final fault severity coefficient is used as the basis for fault assessment of the coal mill bearing.

5. The method according to claim 1, wherein A health status evaluation index of a coal mill bearing is established based on the fault type and the fault degree. The health status evaluation index includes a fault development trend coefficient including: Determine the length of the sliding time window through autocorrelation analysis, and calculate the first-order difference sequence and the second-order difference sequence of the fault severity coefficient within the sliding time window; determine the polynomial fitting order based on the Akaike Information Criterion, perform local weighted polynomial fitting on the first-order difference sequence and the second-order difference sequence, and obtain a fault development trend curve; Collecting coal mill speed parameters, load rate parameters, and feed rate parameters, performing dimensionality reduction processing on the speed parameters, load rate parameters, and feed rate parameters through principal component analysis to obtain operating condition characteristics, establishing a multivariate regression model of the operating condition characteristics and the fault development rate, and calculating the ratio of the actual operating condition development rate to the standard operating condition development rate to obtain the operating condition compensation factor; Constructing a fuzzy neural network, inputting temperature parameters and humidity parameters into an input layer of the fuzzy neural network, performing fuzzy reasoning using a Gaussian fuzzy membership function in a hidden layer of the fuzzy neural network, obtaining an environmental impact coefficient in an output layer of the fuzzy neural network, and training the fuzzy neural network using a loss function with a regularization term; performing exponential weighted averaging on the first-order difference sequence to obtain short-term variation characteristics, calculating the derivative of the fault development trend curve to obtain long-term trend characteristics, and multiplying the operating condition compensation factor by the environmental impact coefficient to obtain a comprehensive compensation term; The short-term change characteristics, the long-term trend characteristics, and the comprehensive compensation items are weightedly combined to obtain an initial trend coefficient, and the Kalman filter algorithm is used to eliminate noise from the initial trend coefficient to obtain a final fault development trend coefficient.

6. The method according to claim 1, characterized in that The method further comprises: Using the predicted value of the remaining life prediction equation as a state variable and the health status assessment index as an observation variable, a state equation containing process noise and an observation equation containing measurement noise are established respectively, and an adaptive particle filter correction model is constructed; Constructing an adaptive noise covariance matrix, wherein the adaptive noise covariance matrix adopts a recursive form with a forgetting factor, controls the decay rate of historical information by the forgetting factor, constructs an innovation covariance term based on the prediction error at the current moment, and dynamically estimates the process noise in the state equation and the measurement noise in the observation equation; Inputting the adaptive noise covariance matrix into the adaptive particle filter correction model to correct the state prediction, calculating the likelihood value of each particle based on the observation data of the observation equation, and updating the weight of the particle; The updated particle weights are used to perform weighted averaging on the state estimation values ​​to obtain a revised prediction value of the remaining life of the coal mill bearing. The revised prediction value of the remaining life of the coal mill bearing includes the influence of operating condition changes and random fluctuation characteristics during equipment operation.

7. A coal mill fault diagnosis and prediction system based on big data analysis, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to collect vibration signals of a coal mill bearing in real time using a vibration sensor, perform multi-scale analysis on the vibration signals through wavelet decomposition, extract time domain features, frequency domain features, and time-frequency features of the vibration signals, and normalize the time domain features, frequency domain features, and time-frequency features to obtain a feature vector matrix; Constructing a training data set for a deep convolutional neural network model based on the feature vector matrix, and dividing the training data set into a source domain data set and a target domain data set; The second unit is configured to construct a deep convolutional neural network model comprising a convolutional layer, a pooling layer, and a fully connected layer, wherein the input layer of the deep convolutional neural network model receives the feature vector matrix, and the output layer outputs the fault type and fault severity; a maximum mean difference adaptive layer is introduced into the deep convolutional neural network model, and the feature distributions of the source domain dataset and the target domain dataset are adaptively aligned through the maximum mean difference adaptive layer; Pre-training the deep convolutional neural network model using the source domain dataset and performing fine-tuning training using the target domain dataset; The third unit is used to input the real-time collected vibration signal of the coal mill bearing into the trained deep convolutional neural network model to obtain the fault type and fault degree of the coal mill bearing; establish the health status assessment index of the coal mill bearing based on the fault type and the fault degree, and the health status assessment index includes a fault severity coefficient and a fault development trend coefficient; calculate the remaining service life of the coal mill bearing according to the health status assessment index, and issue a warning signal when the remaining service life is lower than a preset health threshold.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images