Coal machine equipment safety early warning method based on industrial real-time data

Abstract

This invention relates to a safety early warning method for coal mining equipment based on real-time industrial data, belonging to the field of fault diagnosis technology. The method includes: acquiring real-time frequency domain feature sequences at multiple monitoring times and obtaining reference frequency domain feature vectors; obtaining real-time potential feature vectors for each monitoring time using a SafetyCAE feature extraction network; calculating the Mahalanobis distance between the real-time potential feature vectors and the reference frequency domain feature vectors; calculating the prediction uncertainty factor for each monitoring time; calculating a comprehensive safety early warning factor based on the Mahalanobis distance and the prediction uncertainty factor; performing component-level and whole-machine-level risk diagnosis on the coal mining equipment based on the comprehensive safety early warning factor; inputting the comprehensive safety early warning factor into the fault diagnosis isolated forest iForest model, and outputting the anomaly score and anomaly label for each monitoring time. This invention has stronger forward-looking capabilities, realizing a shift from traditional post-event diagnosis to pre-event safety early warning.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method for early warning of safety of coal mining equipment based on real-time industrial data. Background Technology

[0002] In the current field of intelligent operation and maintenance of coal mining machinery, fault diagnosis and life prediction have become relatively mature technologies. Fault diagnosis mainly solves the problem of classifying known faults, usually based on deep learning models to perform pattern recognition on signal spectra; while life prediction attempts to give a specific countdown to equipment failure through regression analysis. However, in the extreme and harsh working conditions of underground coal mines, the above two methods have a gap in engineering practicality: diagnosis is often too late, and prediction is often inaccurate.

[0003] The most pressing need in coal mining operations is not merely knowing whether something is broken or how many hours it can be used, but rather a stable, continuous early warning factor that reflects changes in safety margins. Traditional fault diagnosis techniques are essentially reactive mechanisms; the algorithm can only identify and classify equipment when it has already suffered significant physical damage and exhibits a marked frequency of fault characteristics. In heavy-duty industrial settings like underground coal mines, once a clear fault is diagnosed, the equipment is often already in an irreversible damage stage, leaving extremely short window periods for on-site personnel to shut down and repair it, making it difficult to avoid huge economic losses from production interruptions. Furthermore, remaining life prediction methods heavily rely on the completeness of the full lifecycle degradation model and the stability of the load conditions. However, in the underground coal mine environment, frequent load fluctuations and strong random noise can interfere with the fitting of degradation trends. This causes life prediction models to often output fluctuating remaining hours, easily leading to inaccurate predictions in practical engineering applications. Whether for diagnosis or prediction, existing methods mostly focus on changes in signal energy amplitude, neglecting the cognitive uncertainty in the model's decision-making process. In the early stages of a fault, the physical characteristics are extremely vague and drowned out by background noise. Traditional methods struggle to distinguish between normal operating condition fluctuations and early performance degradation, leaving the system often in a blind spot when the equipment enters the irreversible degradation process, unable to provide forward-looking risk warnings.

[0004] Patent CN116413031B discloses an adaptive prediction method for the remaining life of bearings in coal mining machinery. This method quantitatively estimates the actual damage level of the bearing by collecting and extracting real-time vibration signal characteristics. Then, based on the deviation between the actual damage and the damage predicted by the bearing's physical model, an adaptive filtering algorithm is used to adaptively update the physical model parameters. Finally, a remaining life prediction is given based on the updated model. This method effectively solves the technical obstacles of inaccurate traditional physical models caused by the unpredictable operating environment of coal mining equipment under harsh conditions, and the short prediction time due to relying solely on real-time signals. It integrates the advantages of both real-time signals and physical models, effectively limiting the uncertainty of prediction and improving the accuracy of life prediction, providing a reliable basis for preventative maintenance of equipment. However, the main drawback of this solution is that its core is still "remaining life prediction," relying on the updating of the bearing's physical degradation model and damage parameters. When coal mining equipment is in an environment with high noise, frequent load changes, and multiple operating conditions, the physical model parameters are prone to drift, leading to unstable remaining life prediction results. In addition, this scheme focuses on estimating the extent of damage that has already occurred to the bearings, and is mainly aimed at bearing components. It is not adaptable enough to different types of coal mining equipment, different key components, and overall machine-level safety early warning, and it does not fully consider the cognitive uncertainty of the model in the early degradation stage.

[0005] Patent CN116430272B discloses a method for diagnosing anomalies in coal mine equipment based on electrical disturbances. This method introduces a sliding window mechanism to acquire real-time and historical data of various electrical parameters and calculates the rate of change of their weighted average. When the rate of change exceeds a threshold, it triggers the storage of relevant disturbance waveforms. Finally, features are extracted from the disturbance waveforms and input into a neural network prediction model to predict and determine the fault type. This method effectively addresses the limitations of traditional coal mine inspections or single-state monitoring methods, which can only reflect existing faults and cannot detect future equipment anomalies in advance. By combining implicit electrical disturbances with neural network algorithms, it achieves efficient and accurate prediction of coal mine equipment fault types, thereby reducing the fault rate, improving mining safety, and reducing production costs. However, the drawback of this scheme is that it mainly relies on the rate of change of electrical parameter disturbances and preset thresholds to trigger anomaly diagnosis. This represents fault type judgment after the disturbance has clearly appeared, making it difficult to continuously and smoothly characterize the safety margin of the early performance degradation process of equipment. Meanwhile, this solution focuses on identifying electrical disturbance waveforms and fault types, but it does not adequately consider the deep degradation characteristics in vibration signals, model cognitive uncertainties, and the integration of component-level and whole-machine-level risks. It is easily affected by transient load disturbances and threshold settings. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a safety early warning method for coal mining equipment based on real-time industrial data. The technical solution of this invention is as follows: A method for early warning of safety of coal mining equipment based on real-time industrial data, comprising: S1, acquire real-time industrial data during the operation of coal mining equipment; S2, extract the real-time frequency domain feature sequence of industrial real-time data at multiple monitoring times; S3, the SafetyCAE feature extraction network is trained unsupervised using absolute health period data during the healthy operation phase of coal mining equipment, and the trained SafetyCAE feature extraction network is obtained. The reference frequency domain feature vector of the absolute health period data used in training the SafetyCAE feature extraction network is also obtained. S4. Input the real-time frequency domain feature sequence of each monitoring moment into the trained SafetyCAE feature extraction network, and the SafetyCAE feature extraction network obtains the real-time potential feature vector of each monitoring moment. S5, calculate the Mahalanobis distance between the real-time potential feature vector and the reference frequency domain feature vector at each monitoring time; S6, Calculate the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time potential feature vector at each monitoring time based on the Monte Carlo dropout method; S7, calculate the comprehensive safety warning factor for each monitoring moment based on Mahalanobis distance and the prediction uncertainty factor for each monitoring moment; S8 performs component-level and whole-machine-level risk diagnosis of coal mining equipment based on comprehensive safety early warning factors at each monitoring time. S9. Obtain the reference comprehensive safety early warning factor sequence for the healthy operation phase of the coal mining equipment, and train the fault diagnosis isolated forest iForest model based on the reference comprehensive safety early warning factor sequence. Input the comprehensive safety early warning factor at each monitoring time into the trained fault diagnosis isolated forest iForest model, and output the abnormal score and abnormal label at each monitoring time from the trained fault diagnosis isolated forest iForest model.

[0007] Preferably, S2 includes: S21, synchronize the initial industrial real-time data from different acquisition channels, different key components and different control units of coal mining equipment according to a unified time base to obtain time-aligned industrial real-time data; S22, according to the preset sliding window, the time-aligned industrial real-time data is divided into real-time data segments of multiple monitoring times; S23, perform abnormal data cleaning on the real-time data segments at each monitoring time to obtain clean real-time data segments at multiple monitoring times; S24. Use Fast Fourier Transform to convert the clean real-time data segments of each monitoring moment to the frequency domain to obtain the frequency domain feature sequence of each monitoring moment. S25, normalize the frequency domain feature sequence at each monitoring time to obtain the real-time frequency domain feature sequence at each monitoring time.

[0008] Preferably, S3 includes: S31, acquire multiple training frequency domain feature sequences of absolute health period data of coal mining equipment during the healthy operation phase; S32, the SafetyCAE feature extraction network is trained unsupervised by multiple training frequency domain feature sequences to obtain the trained SafetyCAE feature extraction network; S33, input multiple training frequency domain feature sequences into the trained SafetyCAE feature extraction network respectively, and obtain multiple health potential feature vectors from the trained SafetyCAE feature extraction network; S34, calculate the mean vector of multiple health potential feature vectors, and use it as a reference frequency domain feature vector.

[0009] Preferably, S5 includes: S51, obtain the covariance matrix of multiple potential health feature vectors; S52, calculate the real-time potential feature vector at the i-th monitoring time based on the covariance matrix. and reference frequency domain eigenvectors Mahalanobis distance between The calculation formula is: (1); In formula (1), Let be the covariance matrix of multiple potential health eigenvectors.

[0010] Preferably, the encoding layer of the SafetyCAE feature extraction network includes a Dropout layer, and S6, when calculating the prediction uncertainty factor of the trained SafetyCAE feature extraction network for any real-time potential feature vector at any monitoring time based on the Monte Carlo dropout method, includes: S61, Keep the Dropout layer active, and perform T Monte Carlo samplings on the real-time potential feature vector at any monitoring time; S62, Calculate the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time latent feature vector at any given monitoring time based on the results of T Monte Carlo samplings. The calculation formula is: (2); In formula (2), The output of the SafetyCAE feature extraction network trained for the t-th Monte Carlo sampling is... The mean of the output of the SafetyCAE feature extraction network trained by T Monte Carlo samplings.

[0011] Preferably, S7 calculates the comprehensive safety warning factor at the i-th monitoring time. When, the calculation formula is: (3); In formula (3), For risk sensitivity coefficient, The real-time potential feature vector at the i-th monitoring time. and reference frequency domain eigenvectors Mahalanobis distance between them The uncertainty factor for the prediction of the real-time potential feature vector at the i-th monitoring time by the trained SafetyCAE feature extraction network.

[0012] Preferably, S8 includes: S81, based on the comprehensive safety warning factor at each monitoring moment, the growth rate of the comprehensive safety warning factor at each monitoring moment, the duration of the anomaly, the trend slope and the intensity of the mutation, divides the operating status of a single key component into four levels: normal, attention, warning and danger. S82 weights and aggregates component-level comprehensive safety early warning factors according to the component importance, historical failure sensitivity, and current operating load of each key component to generate the overall system comprehensive safety index.

[0013] Preferably, S82 generates the overall safety index of the machine. When, the calculation formula is: (4); In formula (4), The number of key components in coal mining equipment The normalized weight of the k-th key component is... Let $\frac{k}{k}$ be the component risk value of the kth critical component. in, and (5); (6); In formula (5), This indicates the importance of the k-th critical component. This represents the historical failure sensitivity of the k-th critical component. The load factor represents the current operating load of the k-th critical component; In formula (6), Let k be the normalized component comprehensive safety early warning factor. Let be the normalized outlier score of the k-th component. The duration of the normalized anomaly for the k-th component.

[0014] Preferably, S9 includes: S91, obtain the reference comprehensive safety early warning factor sequence of the coal mining equipment in the healthy operation stage and input it into the fault diagnosis isolated forest iForest model to establish the abnormal detection benchmark of the coal mining equipment in the healthy state; S92, the comprehensive safety early warning factors of each monitoring moment are input into the trained fault diagnosis isolated forest iForest model in chronological order, and the trained fault diagnosis isolated forest iForest model outputs the abnormal score and abnormal label of each monitoring moment in real time. S93: When the number of times the fault diagnosis isolated forest iForest model determines an anomaly in a preset number of consecutive monitoring times exceeds a preset consensus threshold, an early warning is triggered. The comprehensive safety early warning factor corresponding to the first monitoring time determined to be abnormal is anchored as the early warning threshold under the current operating cycle. The device number, key component number, early warning level, first trigger time, corresponding comprehensive safety early warning factor, anomaly duration, anomaly score, and associated original waveform segment are output. The results are then pushed to the host computer, scheduling platform, mobile terminal, or maintenance work order system.

[0015] Preferably, S9 further includes: When coal mining equipment undergoes maintenance, component replacement, sensor reinstallation, or operating condition switching, it automatically calls the latest confirmed healthy real-time industrial data to update the benchmark and early warning threshold of the trained fault diagnosis isolated forest iForest model, so that the same coal mining equipment can obtain personalized safety boundaries that match its own status at different stages of its life cycle.

[0016] All of the above-mentioned optional technical solutions can be combined arbitrarily, and the present invention will not provide a detailed description of the structure after each combination.

[0017] By means of the above solution, the beneficial effects of the present invention are as follows: By calculating the Mahalanobis distance between the real-time potential feature vector and the reference frequency domain feature vector at each monitoring moment, and calculating the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time potential feature vector at each monitoring moment based on the Monte Carlo dropout method, this invention calculates a comprehensive safety early warning factor based on the Mahalanobis distance and the prediction uncertainty factor. Since the Mahalanobis distance characterizes the geometric deviation of the current state from the center of the healthy cluster, and the prediction uncertainty factor characterizes the closeness of the current state to the healthy pattern, this invention performs component-level and whole-machine-level risk diagnosis on coal mining equipment based on the comprehensive safety early warning factor. This allows for the early detection of abnormal trends when the coal mining equipment has not yet shown significant fault characteristic frequencies and only exhibits weak performance degradation signs, demonstrating stronger foresight and realizing a shift from traditional post-event diagnosis to pre-event safety early warning. Furthermore, by performing component-level and whole-machine-level risk diagnosis on the coal mining equipment, a holistic, continuous, and personalized characterization of the equipment's operating status can be achieved, thus realizing comprehensive fault diagnosis and early warning for the coal mining equipment.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention.

[0020] Figure 2 This is a schematic diagram of the SafetyCAE feature extraction network in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] like Figure 1 As shown in the figure, the coal mining equipment safety early warning method based on industrial real-time data provided in this embodiment of the invention includes: S1 acquires real-time industrial data during the operation of coal mining equipment.

[0023] The coal mining equipment can be underground equipment such as coal mining machines, scraper conveyors, and tunneling machines. In practice, monitoring sensors are deployed at the operating site of the coal mining equipment to collect real-time industrial data characterizing the equipment's operating status. This real-time industrial data is preferably vibration signals.

[0024] S2 extracts the real-time frequency domain feature sequence of industrial real-time data at multiple monitoring times.

[0025] Step S2 aims to transform the high-noise raw time-domain signal acquired by the sensors of the coal mining equipment into a frequency-domain feature space that can characterize physical degradation.

[0026] In one specific embodiment, S2 includes: S21 synchronizes the initial real-time industrial data from different acquisition channels, key components, and control units of the coal mining equipment according to a unified time reference to obtain time-aligned real-time industrial data.

[0027] The key components vary depending on the type of coal mining equipment. For example, key components of a coal mining machine include gears, bearings, and cutting tools. Each initial piece of real-time industrial data carries the data source, indicating which key component it originates from, facilitating subsequent component-level risk diagnosis.

[0028] S22, according to the preset sliding window, divides the time-aligned industrial real-time data into real-time data segments of multiple monitoring times.

[0029] Specifically, during the real-time monitoring of coal mining equipment, the continuous time series collected by the sensors is truncated using a sliding window. The preset length of the sliding window is set to N=1024 sampling points. Adjacent windows can be advanced with a fixed step size or by using overlapping sliding windows to generate continuous samples, thereby improving the ability to capture short-term abnormal shocks.

[0030] S23, perform abnormal data cleaning on the real-time data segments at each monitoring time to obtain clean real-time data segments at multiple monitoring times.

[0031] Data cleaning includes outlier removal and noise reduction.

[0032] S24. The clean real-time data segments at each monitoring moment are converted to the frequency domain using the Fast Fourier Transform to obtain the frequency domain feature sequence at each monitoring moment.

[0033] Specifically, the Fast Fourier Transform (FFT) is used to convert a clean real-time data segment x(n) at a certain monitoring moment into a frequency domain feature sequence X(k). The calculation formula is as follows: ; Where n is the discrete time-domain index and j is the imaginary unit.

[0034] S25, normalize the frequency domain feature sequence at each monitoring time to obtain the real-time frequency domain feature sequence at each monitoring time.

[0035] Specifically, for the frequency domain feature sequence at any monitoring time, the normalization process includes absolute value conversion of the FFT spectrum and removal of the DC component, and the use of maximum-minimum normalization to map the feature values ​​to the [0, 1] interval to eliminate the influence of dimensions.

[0036] S3. The SafetyCAE feature extraction network is trained unsupervised using absolute health period data from the healthy operation phase of the coal mining equipment. The trained SafetyCAE feature extraction network is obtained, and the reference frequency domain feature vector of the absolute health period data used in training the SafetyCAE feature extraction network is acquired.

[0037] Among them, the SafetyCAE feature extraction network uses the absolute health period data of the coal mining equipment in the early stage of operation for unsupervised training, with the aim of learning the essential topological structure under the healthy state.

[0038] In one specific embodiment, S3 includes: S31, acquire multiple training frequency domain feature sequences of absolute health period data of coal mining equipment during the healthy operation phase.

[0039] This step requires first acquiring absolute health data for the coal mining equipment's healthy operation phase. Then, using the same method as acquiring real-time frequency domain feature sequences for each monitoring moment in S2, multiple training frequency domain feature sequences are obtained. Furthermore, each data record is bound to the equipment number, key component number, sensor number, acquisition time, operating condition label, and operating batch identifier to ensure that different equipment instances and different operating phases can be distinguished during subsequent modeling. Further, after acquiring multiple training frequency domain feature sequences, a health benchmark library is established according to three levels: "equipment type (e.g., coal mining machine) — equipment instance (specific model of the coal mining machine) — key components (bearings, gears, etc.)". For common behavioral patterns shared among equipment of the same model, a group health benchmark is established; for individual differences in a single piece of equipment due to variations in installation, break-in, and load history, an individual health benchmark is established; for significantly different dynamic characteristics between components, a component-specific health benchmark is established. Through the collaboration of these three levels of benchmarks, the model's transferability, personalization capabilities, and overall applicability to the entire machine can be simultaneously considered.

[0040] Absolute health period data includes real-world data from the initial commissioning phase of coal mining equipment (e.g., the first 10% of its lifecycle), the initial post-maintenance recovery phase, and the manually confirmed normal operation phase. This data, combined with work order records, inspection records, alarm records, load records, and production plan information, automatically identifies absolute health period samples, thus avoiding complete reliance on manual delineation of health intervals. Segments with minor noise disturbances but no actual damage can be included as extended health samples in the training process to enhance the model's adaptability to complex on-site conditions.

[0041] S32, the SafetyCAE feature extraction network is trained unsupervised using multiple training frequency domain feature sequences to obtain the trained SafetyCAE feature extraction network.

[0042] The SafetyCAE feature extraction network is a one-dimensional convolutional autoencoder that incorporates an attention mechanism. It consists of two symmetrical parts: an encoder and a decoder. The encoder is responsible for compressing the high-dimensional frequency domain input sequence layer by layer and mapping it to a low-dimensional latent feature space. The decoder is responsible for reconstructing the latent features into a reconstructed spectral sequence. By comparing the input with the reconstruction result, the model learns the inherent manifold structure of the healthy frequency domain.

[0043] The encoder consists of multiple one-dimensional convolutional layers, batch normalization (BN) layers, ReLU activation functions, dropout layers with a dropout rate of 0.2, and a one-dimensional convolutional block attention (CBAM) module. The dropout layers improve the model's generalization ability and support subsequent uncertainty measurement, while the CBAM module automatically enhances feature bands sensitive to weakly damaged signals and suppresses background noise interference by reweighting features in both channel and spatial dimensions. The decoder employs the inverse structure of the encoder, including at least one set of one-dimensional transposed convolutional layers, batch normalization layers, ReLU activation functions, and dropout layers. At the output, a one-dimensional transposed convolutional layer and a Tanh activation function are applied to progressively upsample the latent feature vectors to reconstruct the spectral sequence. By progressively upsampling the latent feature vectors through multiple nonlinear mappings to reconstruct the spectral sequence, the latent space simultaneously retains the main topological information and local structural information of the healthy state.

[0044] The goal of unsupervised training of the SafetyCAE feature extraction network is to minimize the reconstruction error between the original input signal and the reconstructed signal, typically using mean squared error (MSE) as the loss function. : ; Where W and b are the network weights and biases, and A is the number of iterations. The predicted value obtained in the c-th iteration is... This represents the mean of the predicted values ​​obtained through A iterations. This training process forces the model to learn and lock onto the essential topological structure and manifold characteristics of the frequency domain signal under healthy conditions.

[0045] During model training, the training loss, validation loss, and reconstruction error distribution are recorded. Once the loss function converges and the reconstruction of healthy samples is stable, the model parameters are frozen and the current model version is output. If necessary, early stopping strategies, learning rate decay, or batch size adjustments can be used to improve training stability.

[0046] S33, input multiple training frequency domain feature sequences into the trained SafetyCAE feature extraction network respectively, and obtain multiple health potential feature vectors from the trained SafetyCAE feature extraction network.

[0047] Specifically, after training the SafetyCAE feature extraction network, all the trained frequency domain feature sequences are input back into the trained SafetyCAE feature extraction network, and the encoder in the trained SafetyCAE feature extraction network obtains multiple health potential feature vectors.

[0048] S34, calculate the mean vector of multiple health potential feature vectors, and use it as a reference frequency domain feature vector.

[0049] S4. Input the real-time frequency domain feature sequence of each monitoring moment into the trained SafetyCAE feature extraction network, and the SafetyCAE feature extraction network obtains the real-time potential feature vector of each monitoring moment.

[0050] Specifically, the encoder of the SafetyCAE feature extraction network obtains the real-time potential feature vector for each monitoring moment.

[0051] S5 calculates the Mahalanobis distance between the real-time potential feature vector and the reference frequency domain feature vector at each monitoring time.

[0052] In one specific embodiment, S5 includes: S51, obtain the covariance matrix of multiple health potential feature vectors.

[0053] Among them, the mean vector and covariance matrix of multiple potential health feature vectors are used for subsequent migration and individual fine-tuning of similar equipment. When a new coal mining equipment is connected to the system, the group health benchmark of similar equipment can be called first for initialization, and then the real field health data of the coal mining equipment in the early stage of operation can be used for individual fine-tuning to shorten the cold start cycle after the new coal mining equipment is put into operation.

[0054] S52, calculate the real-time potential feature vector at the i-th monitoring time based on the covariance matrix. and reference frequency domain eigenvectors Mahalanobis distance between The calculation formula is: (1); In formula (1), Let be the covariance matrix of multiple potential health eigenvectors.

[0055] Mahalanobis distance weakens the correlation between dimensions of deep features by introducing the inverse covariance matrix, thus characterizing the geometric deviation of the current state from the center of the healthy cluster. When the coal mining equipment is still in a healthy state, the MD value fluctuates within a small range; when the coal mining equipment undergoes degradation processes such as wear, loosening, and crack propagation, the real-time latent feature vector gradually deviates from the healthy cluster, and the MD value continues to rise.

[0056] S6, based on the Monte Carlo dropout method, calculates the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time potential feature vector at each monitoring time.

[0057] In one specific embodiment, the encoding layer of the SafetyCAE feature extraction network includes a Dropout layer. S6, when calculating the prediction uncertainty factor of the trained SafetyCAE feature extraction network for any real-time potential feature vector at any monitoring time based on the Monte Carlo dropout method, includes: S61, Keep the Dropout layer active, and perform T Monte Carlo samplings on the real-time potential feature vector at any monitoring time; S62, Calculate the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time latent feature vector at any given monitoring time based on the results of T Monte Carlo samplings. The calculation formula is: (2); In formula (2), The output of the SafetyCAE feature extraction network trained for the t-th Monte Carlo sampling is... The mean of the output of the SafetyCAE feature extraction network trained by T Monte Carlo samplings.

[0058] When the real-time potential feature vector at a certain monitoring moment is close to the health pattern learned by the model, the results of multiple reconstructions tend to be consistent and the U value is small. When the sample is in an early degradation, mixed pattern or unclear boundary region, the dispersion of the results of multiple forward propagation increases and the U value increases significantly, which can be used as a quantitative indicator of the model's hesitation in recognizing the current state.

[0059] S7 calculates the comprehensive safety warning factor for each monitoring moment based on Mahalanobis distance and the prediction uncertainty factor for each monitoring moment.

[0060] In one specific embodiment, the prediction uncertainty is regarded as an amplifier of physical degradation, and S7 calculates the comprehensive safety warning factor at the i-th monitoring time. When, the calculation formula is: (3); In formula (3), For risk sensitivity coefficient, The real-time potential feature vector at the i-th monitoring time. and reference frequency domain eigenvectors Mahalanobis distance between them This is the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time potential feature vector at the i-th monitoring time. This formula means that in the early stage of fault initiation, even if the physical degradation MD is small, if the feature ambiguity is large (prediction uncertainty U increases abruptly), the model can still sensitively amplify abnormal signals to achieve early warning.

[0061] Furthermore, to avoid false increases in the comprehensive safety warning factor due to single-point fluctuations, time smoothing and trend constraints are introduced after the comprehensive safety warning calculation is completed. Moving median smoothing, exponential smoothing, or local trend estimation processing are performed on the HI sequence of continuous time slices, and the growth rate, duration of abnormality, local slope, and intensity of mutation of the comprehensive safety warning factor are recorded simultaneously to improve the monotonicity, stability, and interpretability of the comprehensive safety warning factor.

[0062] Among them, let the first The comprehensive safety early warning factor after smoothing at each monitoring time is: The sampling time interval is Then the first The growth rate of the comprehensive safety early warning factor at each monitoring time point It can be represented as: ; To determine the duration of the statistical anomaly, let... For the first Anomaly markers at each monitoring time, Indicates an anomaly. This indicates normal. Among them, The fault diagnosis isolated forest iForest model outputs the comprehensive safety warning factor at each monitoring time through the output of the fault diagnosis isolated forest iForest model described below. After inputting the model, the corresponding anomaly marker is output by the model. . A threshold of interest is set. Based on this, consecutive anomaly counts are performed. It can be recursively represented as: ; Then the first Abnormal duration of comprehensive safety early warning factors at each monitoring moment for: .

[0063] To describe a local upward trend, one can look at the most recent If a linear fit is performed on the comprehensive safety early warning factor sequence within each monitoring point, then the trend slope of the comprehensive safety early warning factor at the i-th monitoring time is... It can be represented as: ; in, .

[0064] To describe the degree of single-point mutation, the standard deviation of the reference comprehensive safety warning factor sequence in the healthy phase can be standardized, and then the mutation intensity of the comprehensive safety warning factor at the i-th monitoring time point can be calculated. It can be represented as: ; in, The standard deviation of the comprehensive safety warning factor sequence for the healthy stage; ; Where M represents the number of reference comprehensive security early warning factors in the reference comprehensive security early warning factor sequence.

[0065] S8 performs component-level and whole-machine-level risk diagnosis of coal mining equipment based on comprehensive safety early warning factors at each monitoring time.

[0066] In one specific embodiment, S8 includes: S81 classifies the operating status of a single key component into four levels: normal, attention, warning, and danger, based on the comprehensive safety warning factor at each monitoring time, the growth rate of the comprehensive safety warning factor at each monitoring time, the duration of the abnormality, the trend slope, and the intensity of the mutation.

[0067] Among them, "normal" indicates that no obvious signs of degradation have been found, "attention" indicates that there is slight drift, "warning" indicates that there is a persistent abnormality, and "danger" indicates that the equipment has entered a stage of rapid deterioration and needs to be scheduled for maintenance, reduced load operation, or shutdown for inspection.

[0068] Specifically, for each monitoring time point, the normal state requires the comprehensive safety warning factor to not exceed the 95th percentile of the reference comprehensive safety warning factor sequence for the healthy stage, with fewer than 3 consecutive abnormal points, a trend slope close to zero, and a mutation intensity of less than 2; the attention state requires the comprehensive safety warning factor to exceed the upper limit of normal but not reach the lower limit of warning, or have 3-5 consecutive abnormal points, a trend slope greater than 0 and less than 0.05, and a mutation intensity greater than 2 and less than 3; the warning state is defined as the comprehensive safety warning factor exceeding the warning threshold (e.g., 2 standard deviations of the mean of the reference comprehensive safety warning factor sequence for the healthy stage), or having more than 5 but less than 10 consecutive abnormal points, a trend slope greater than 0.05 and less than 0.15, and a mutation intensity greater than 3 and less than 5; the dangerous state requires the comprehensive safety warning factor to exceed the alarm threshold (e.g., 5 standard deviations of the mean of the reference comprehensive safety warning factor sequence for the healthy stage), or having 10 or more consecutive abnormal points, a trend slope greater than 0.15, and a mutation intensity greater than 5. Among these, consecutive abnormal points are determined by the number of consecutive abnormal points at each monitoring time point. The value of is determined.

[0069] S82 weights and aggregates component-level comprehensive safety early warning factors according to the component importance, historical failure sensitivity, and current operating load of each key component to generate the overall system comprehensive safety index.

[0070] Specifically, for component-level monitoring results deployed on multiple key components, the component-level comprehensive safety early warning factors are weighted and aggregated according to component importance, historical failure sensitivity, and current operating load to generate a comprehensive safety index for the entire machine. The comprehensive safety index can be output using a weighted average, a weighted maximum value, or a master-slave component linkage rule, thereby achieving a unified expression of risk from local anomalies to the overall coal mining equipment risk.

[0071] For any given monitoring time, assume the coal mining equipment includes The first key component, The component risk value of each key component is Its normalized weight is Among them, normalized weights Based on the importance of the component Sensitivity to historical failures and the load factor of the current operating load Perform normalization calculations: (5); in, This indicates the importance of the k-th critical component. This represents the historical failure sensitivity of the k-th critical component. This represents the load factor of the current operating load of the k-th critical component.

[0072] Component risk value It can be obtained by weighting the normalized component comprehensive safety warning factor, anomaly score, and anomaly duration: (6); in, Let k be the normalized component comprehensive safety early warning factor. The normalized anomaly score for the k-th component (output by the subsequent fault diagnosis isolated forest iForest model). The normalized anomaly duration for the k-th component. and The values ​​are the fusion weights for each item, and are taken as empirical values. and This is the normalized value of the comprehensive safety warning factor, anomaly score, and anomaly duration for the k-th component.

[0073] For any given monitoring moment, the overall safety index of the machine A weighted average method can be used to calculate: (4).

[0074] Furthermore, S8 may also include: when the overall risk of a key component or the whole machine exceeds a threshold (the component-level risk diagnosis result is a warning or danger, and the overall safety index of the whole machine exceeds a specified threshold), synchronously outputting the original waveform segment, spectrum segment, main abnormal frequency band, HI change trend (trend slope) and the degree of deviation from the historical health benchmark within the corresponding time window, providing on-site personnel with traceable abnormal evidence.

[0075] S9. Obtain the reference comprehensive safety early warning factor sequence for the healthy operation phase of the coal mining equipment, and train the fault diagnosis isolated forest iForest model based on the reference comprehensive safety early warning factor sequence. Input the comprehensive safety early warning factor at each monitoring time into the trained fault diagnosis isolated forest iForest model, and output the abnormal score and abnormal label at each monitoring time from the trained fault diagnosis isolated forest iForest model.

[0076] In one specific embodiment, S9 includes: S91. Obtain the reference comprehensive safety early warning factor sequence for the healthy operation phase of coal mining equipment and input it into the fault diagnosis isolated forest iForest model to establish an anomaly detection benchmark under the healthy state of coal mining equipment.

[0077] The method for obtaining the reference comprehensive safety early warning factor during the healthy operation phase of the coal mining equipment is the same as the principle for calculating the comprehensive safety early warning factor at each monitoring time in step S7, and will not be repeated here. Multiple reference comprehensive safety early warning factors form a reference comprehensive safety early warning factor sequence.

[0078] Because the healthy phase samples are concentrated and have a limited range of fluctuations, the iForest model for fault diagnosis isolation forest can learn the normal distribution boundary of healthy factors, providing personalized references for subsequent online detection.

[0079] S92, in chronological order, inputs the comprehensive safety warning factors of each monitoring moment into the trained fault diagnosis isolated forest iForest model, and the trained fault diagnosis isolated forest iForest model outputs the anomaly score and anomaly label of each monitoring moment in real time.

[0080] This process relies only on data collected at the current moment and before, without invoking future information, thus meeting the real-time requirements of online early warning scenarios.

[0081] S93: When the number of times the fault diagnosis isolated forest iForest model determines an anomaly in a preset number of consecutive monitoring moments exceeds a preset consensus threshold, an early warning is triggered. The comprehensive safety early warning factor corresponding to the first monitoring moment determined to be abnormal is anchored as the early warning threshold under the current operating cycle. The device number, key component number, early warning level, first trigger time, corresponding comprehensive safety early warning factor, anomaly duration, anomaly score, and associated original waveform segment are output. The results are pushed to the host computer, scheduling platform, mobile terminal, or maintenance work order system for on-site personnel to quickly locate the source of risk.

[0082] For example, by introducing a sliding window size W=5, when the number of times the fault diagnosis isolated forest iForest model identifies anomalies in the past 5 consecutive monitoring points exceeds a preset consensus threshold, a system warning is triggered, and the HI value corresponding to the first trigger time is anchored as the warning threshold for that operating cycle. Continuity constraints can significantly suppress false alarms caused by transient noise, local load disturbances, and occasional sampling anomalies.

[0083] Furthermore, S9 also includes: when the coal mining equipment undergoes maintenance, component replacement, sensor reinstallation, or operating condition switching, it automatically calls the latest confirmed healthy real-time industrial data to update the benchmark (reference anomaly score) and early warning threshold of the trained fault diagnosis isolated forest iForest model, so that the same coal mining equipment can obtain a personalized safety boundary that matches its own state at different stages of its life cycle.

[0084] In addition, after on-site personnel complete the inspection and confirmation, they write the inspection conclusions, fault types, component handling methods, and restoration results back to the system database. Based on this feedback, the system continuously corrects the healthy sample database, the abnormal sample database, and model parameters, thus forming a closed-loop iterative mechanism of "on-site data collection—online early warning—inspection confirmation—model update".

[0085] Through the above technical solutions, the embodiments of the present invention no longer stop at the static judgment of whether a single component has failed, but are driven by real field data collection to continuously track, dynamically model and grade the operational risks of key components and the whole machine of coal mining equipment. It can identify irreversible degradation trends earlier in complex noise environments and variable working conditions, and provide technical support for proactive maintenance and safe production of coal mining equipment.

[0086] In summary, the method provided by the embodiments of the present invention has the following beneficial effects: (1) A safety early warning technology route for coal mining equipment, which is different from the traditional equipment safety operation and maintenance method, is proposed.

[0087] Existing equipment safety operation and maintenance methods mostly focus on single equipment or single key components, typically emphasizing post-fault type diagnosis or remaining life prediction based on ideal degradation models. Overall, these methods still rely on post-fault identification or estimation based on fixed patterns, making them ill-suited to the actual needs of coal mines under conditions of high noise, fluctuating loads, and complex operating conditions. This invention, however, addresses the entire coal mining equipment system. Based on real-world data collection, it moves beyond isolated analysis of single components. Instead, it constructs a hierarchical health benchmark of "equipment type—equipment instance—key component," providing a holistic, continuous, and personalized characterization of the coal mining equipment's operating status. This achieves a shift from traditional post-fault diagnosis to proactive safety early warning.

[0088] (2) The present invention has the advantages of early warning, strong noise resistance and good adaptability.

[0089] This invention integrates deep feature space deviation (represented by Mahalanobis distance) with model cognitive uncertainty (represented by prediction uncertainty factor), enabling it to detect abnormal trends in coal mining equipment before significant fault frequency is observed and only slight performance degradation is shown. This provides stronger foresight compared to traditional methods. Furthermore, by combining real-world data modeling, adaptive health benchmarks, isolated forest anomaly detection, and a sliding window consensus mechanism, this invention effectively suppresses false alarms caused by noise accumulation, transient disturbances, and fluctuations in operating conditions under complex coal mine conditions. Compared to fixed threshold or single-feature early warning methods, it offers better stability, sensitivity, and engineering applicability. It also allows for dynamic adjustment of early warning boundaries after equipment maintenance, component replacement, and changes in operating conditions, making it more suitable for the safe operation and maintenance needs of coal mining equipment during long-term online operation.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for safety early warning of coal mining equipment based on real-time industrial data, characterized in that, include: S1, acquire real-time industrial data during the operation of coal mining equipment; S2, extract the real-time frequency domain feature sequence of industrial real-time data at multiple monitoring times; S3, the SafetyCAE feature extraction network is trained unsupervised using absolute health period data during the healthy operation phase of coal mining equipment, and the trained SafetyCAE feature extraction network is obtained. The reference frequency domain feature vector of the absolute health period data used in training the SafetyCAE feature extraction network is also obtained. S4. Input the real-time frequency domain feature sequence of each monitoring moment into the trained SafetyCAE feature extraction network, and the SafetyCAE feature extraction network obtains the real-time potential feature vector of each monitoring moment. S5, calculate the Mahalanobis distance between the real-time potential feature vector and the reference frequency domain feature vector at each monitoring time; S6, Calculate the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time potential feature vector at each monitoring time based on the Monte Carlo dropout method; S7, calculate the comprehensive safety warning factor for each monitoring moment based on Mahalanobis distance and the prediction uncertainty factor for each monitoring moment; S8 performs component-level and whole-machine-level risk diagnosis of coal mining equipment based on comprehensive safety early warning factors at each monitoring time. S9. Obtain the reference comprehensive safety early warning factor sequence for the healthy operation phase of the coal mining equipment, and train the fault diagnosis isolated forest iForest model based on the reference comprehensive safety early warning factor sequence. Input the comprehensive safety early warning factor at each monitoring time into the trained fault diagnosis isolated forest iForest model, and output the abnormal score and abnormal label at each monitoring time from the trained fault diagnosis isolated forest iForest model.

2. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, S2 includes: S21, synchronize the initial real-time industrial data from different acquisition channels, key components and control units of coal mining equipment according to a unified time base to obtain time-aligned real-time industrial data; S22, according to the preset sliding window, the time-aligned industrial real-time data is divided into real-time data segments of multiple monitoring times; S23, perform abnormal data cleaning on the real-time data segments at each monitoring time to obtain clean real-time data segments at multiple monitoring times; S24. Use Fast Fourier Transform to convert the clean real-time data segments of each monitoring moment to the frequency domain to obtain the frequency domain feature sequence of each monitoring moment. S25, normalize the frequency domain feature sequence at each monitoring time to obtain the real-time frequency domain feature sequence at each monitoring time.

3. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, S3 includes: S31, acquire multiple training frequency domain feature sequences of absolute health period data of coal mining equipment during the healthy operation phase; S32, the SafetyCAE feature extraction network is trained unsupervised by multiple training frequency domain feature sequences to obtain the trained SafetyCAE feature extraction network; S33, input multiple training frequency domain feature sequences into the trained SafetyCAE feature extraction network respectively, and obtain multiple health potential feature vectors from the trained SafetyCAE feature extraction network; S34, calculate the mean vector of multiple health potential feature vectors, and use it as a reference frequency domain feature vector.

4. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 3, characterized in that, S5 includes: S51, obtain the covariance matrix of multiple potential health feature vectors; S52, calculate the real-time potential feature vector at the i-th monitoring time based on the covariance matrix. and reference frequency domain eigenvectors Mahalanobis distance between The calculation formula is: （1）； In formula (1), Let be the covariance matrix of multiple potential health eigenvectors.

5. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, The encoding layer of the SafetyCAE feature extraction network includes a Dropout layer. S6, when calculating the prediction uncertainty factor of the trained SafetyCAE feature extraction network for any real-time latent feature vector based on the Monte Carlo dropout method, includes: S61, Keep the Dropout layer active, and perform T Monte Carlo samplings on the real-time potential feature vector at any monitoring time; S62, Calculate the prediction uncertainty factor of the trained SafetyCAE feature extraction network for the real-time latent feature vector at any given monitoring time based on the results of T Monte Carlo samplings. The calculation formula is: （2）； In formula (2), The output of the SafetyCAE feature extraction network trained for the t-th Monte Carlo sampling is... The mean of the output of the SafetyCAE feature extraction network trained by T Monte Carlo samplings.

6. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, The S7 calculates the comprehensive safety early warning factor at the i-th monitoring time. When, the calculation formula is: （3）； In formula (3), For risk sensitivity coefficient, The real-time potential feature vector at the i-th monitoring time. and reference frequency domain eigenvectors Mahalanobis distance between them The uncertainty factor for the prediction of the real-time potential feature vector at the i-th monitoring time by the trained SafetyCAE feature extraction network.

7. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, S8 includes: S81, based on the comprehensive safety warning factor at each monitoring moment, the growth rate of the comprehensive safety warning factor at each monitoring moment, the duration of the anomaly, the trend slope and the intensity of the mutation, divides the operating status of a single key component into four levels: normal, attention, warning and danger. S82 weights and aggregates component-level comprehensive safety early warning factors according to the component importance, historical failure sensitivity, and current operating load of each key component to generate the overall system comprehensive safety index.

8. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 7, characterized in that, The S82 generates the overall machine safety index. When, the calculation formula is: （4）； In formula (4), The number of key components in coal mining equipment The normalized weight of the k-th key component is... Let $\frac{k}{k}$ be the component risk value of the kth critical component. in, and (5); （6）； In formula (5), This indicates the importance of the k-th critical component. This represents the historical failure sensitivity of the k-th critical component. The load factor represents the current operating load of the k-th critical component; In formula (6), Let k be the normalized component comprehensive safety early warning factor. Let be the normalized outlier score of the k-th component. The duration of the normalized anomaly for the k-th component.

9. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 1, characterized in that, S9 includes: S91, obtain the reference comprehensive safety early warning factor sequence of the coal mining equipment in the healthy operation stage and input it into the fault diagnosis isolated forest iForest model to establish an anomaly detection benchmark in the healthy state of the coal mining equipment; S92, the comprehensive safety early warning factors of each monitoring moment are input into the trained fault diagnosis isolated forest iForest model in chronological order, and the trained fault diagnosis isolated forest iForest model outputs the abnormal score and abnormal label of each monitoring moment in real time. S93: When the number of times the fault diagnosis isolated forest iForest model determines an anomaly in a preset number of consecutive monitoring times exceeds a preset consensus threshold, an early warning is triggered. The comprehensive safety early warning factor corresponding to the first monitoring time determined to be abnormal is anchored as the early warning threshold under the current operating cycle. The device number, key component number, early warning level, first trigger time, corresponding comprehensive safety early warning factor, anomaly duration, anomaly score, and associated original waveform segment are output. The results are then pushed to the host computer, scheduling platform, mobile terminal, or maintenance work order system.

10. The method for safety early warning of coal mining equipment based on real-time industrial data according to claim 9, characterized in that, S9 further includes: When coal mining equipment undergoes maintenance, component replacement, sensor reinstallation, or operating condition switching, it automatically calls the latest confirmed healthy real-time industrial data to update the benchmark and early warning threshold of the trained fault diagnosis isolated forest iForest model, so that the same coal mining equipment can obtain personalized safety boundaries that match its own status at different stages of its life cycle.

Images