A coal mine equipment fault prediction method based on deep learning

Abstract

This invention discloses a deep learning-based method for predicting coal mine equipment faults, belonging to the field of fault prediction technology. The method includes: collecting historical multi-source time-series operational data of coal mine equipment; using a deep autoencoder to map historical operational states to a state space; constructing a healthy data cloud and various fault data clouds; and establishing typical fault cloud routes based on the connection relationships of discrete state points; mapping real-time collected operational data to the state space to generate real-time state points; matching and judging these points with typical fault cloud routes to determine the fault evolution path of the equipment and generate intervention information. This invention solves the technical problem that existing coal mine equipment fault detection methods rely on periodic manual inspections or static analysis, making it difficult to achieve early fault identification and targeted intervention. It achieves the technical effect of continuously judging real-time states based on deep learning, enabling forward-looking prediction and proactive intervention of coal mine equipment fault evolution trends.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, and specifically to a method for predicting faults in coal mine equipment based on deep learning. Background Technology

[0002] With the continuous improvement of coal mine production scale and automation level, various mining, transportation, and auxiliary equipment operate under complex conditions such as high load, strong vibration, and high dust levels for extended periods. Equipment failures occur frequently, and their evolution is often insidious and gradual. Sudden failures in critical equipment can easily lead to production shutdowns or even safety accidents, causing significant economic losses and safety risks. However, existing methods for coal mine equipment fault diagnosis and prediction largely rely on manual experience or monitoring based on single sensor thresholds, limiting their ability to analyze complex fault modes. While some machine learning-based methods introduce data-driven thinking, they often focus on static state identification or single-moment classification, lacking monitoring of the continuous evolution of equipment from a healthy state to a faulty state, making it difficult to achieve early warning and targeted intervention. Summary of the Invention

[0003] This application provides a deep learning-based method for predicting coal mine equipment faults, which addresses the technical problem that existing coal mine equipment fault detection methods rely on regular manual inspections or static analysis, making it difficult to achieve early identification and targeted intervention of faults.

[0004] This application provides a deep learning-based method for predicting coal mine equipment faults. The method includes: collecting historical multi-source time-series operational data of the coal mine equipment, including normal state data and fault data of different known fault types; using a trained deep autoencoder to map the historical multi-source time-series operational data into points in a state space, constructing a healthy data cloud and multiple fault data cloud sets, and analyzing the connection paths of discrete state points in the healthy data cloud and the multiple fault data cloud sets to establish multiple typical fault cloud routes; collecting the current operational data of the coal mine equipment in real time and mapping it into points in a state space to generate real-time state points; determining the matching status of the real-time state points to the multiple typical fault cloud routes, and if a preset matching status is met, determining that the coal mine equipment has entered the corresponding fault cloud route, issuing an early warning signal, and generating active intervention suggestions adapted to the corresponding fault cloud route.

[0005] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application provides a deep learning-based method for predicting coal mine equipment faults, relating to the field of fault prediction technology. It maps multi-source time-series operational data of coal mine equipment to a unified state space using a deep autoencoder, constructing health and fault data clouds. Typical fault cloud routes are extracted to show the evolution of equipment from a healthy state to different fault states. Based on this, the method identifies and judges the equipment fault evolution path by matching real-time state points with typical routes. This solves the technical problem of existing coal mine equipment fault detection methods relying on periodic manual inspections or static analysis, which makes it difficult to achieve early fault identification and targeted intervention. It achieves the technical effect of continuously judging real-time states by constructing typical fault cloud routes, enabling forward-looking prediction and proactive intervention of coal mine equipment fault evolution trends. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 A schematic flowchart of a deep learning-based coal mine equipment fault prediction method provided in this application embodiment; Figure 2 This is a flowchart illustrating the process of establishing multiple typical fault cloud routes in a deep learning-based coal mine equipment fault prediction method provided in this application embodiment. Detailed Implementation

[0008] This application provides a deep learning-based method for predicting coal mine equipment faults, which addresses the technical problem that existing coal mine equipment fault detection methods rely on regular manual inspections or static analysis, making it difficult to achieve early identification and targeted intervention of faults.

[0009] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0010] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0011] Example 1, as Figure 1 As shown, this application provides a deep learning-based method for predicting coal mine equipment faults, the method comprising: P10: Collect historical multi-source time-series operating data of coal mine equipment, including normal state data and fault data of different known fault types.

[0012] Specifically, the first step is to comprehensively collect historical operational data from coal mine equipment. This data originates from multi-source time-series data of the equipment under different operating conditions, including status data during normal operation and fault data generated under known fault types. Multi-source time-series data refers to a sequence of data obtained from multiple different data sources and arranged in chronological order. These data sources may include various sensors, monitoring systems, and other acquisition devices installed on the equipment, capable of monitoring and recording the equipment's operating parameters in real time, such as temperature, pressure, vibration frequency, and current intensity. These parameters change over time, forming continuous time-series data.

[0013] In collecting data, it is crucial to distinguish between normal state data and fault data. Normal state data refers to data generated when the equipment is operating normally or without faults. This data reflects the equipment's standard operating state and constitutes the system's baseline for the equipment's healthy operation. This data is essential for training subsequent deep learning models, providing a "normal" paradigm to help the model learn to recognize the equipment's operating modes when there are no anomalies or faults. The collected normal state data should include the equipment's diverse performance under different environments and loads to ensure that the model can identify the equipment's normal operating range.

[0014] In contrast, fault data refers to the data generated when equipment experiences known faults. Each type of fault may produce different characteristics in sensor data. For example, mechanical faults may cause significant abnormal fluctuations in vibration sensor data, while electrical faults may manifest as increased temperature or abnormal current. Therefore, fault data needs to be obtained from equipment failures or experimental simulations, and each data point must be clearly labeled with its corresponding fault type. The purpose of this is to provide clear labels for subsequent model training, enabling deep learning models to learn the characteristics of different fault modes based on this known fault data and to identify similar faults in actual operation.

[0015] When collecting data, it is also necessary to consider the equipment's operating cycle and working environment. Equipment may undergo various state changes, such as long-term operation, different load conditions, and equipment aging. Data collection should cover these different stages of state changes to ensure the representativeness of the collected fault data. Furthermore, the data collection period should be long enough to capture the dynamic changes of the equipment under various fault types, covering the entire fault process and providing rich information for subsequent fault prediction model training.

[0016] Furthermore, to ensure data accuracy and integrity, equipment calibration and sensor maintenance during the acquisition process are crucial. Each sensor should be calibrated regularly to ensure the accuracy of the data it provides and to avoid data distortion due to equipment deviations. Simultaneously, the acquired data should be a sufficient time series to reflect trends in equipment status changes. For example, collecting data over a longer operating cycle can help identify subtle fault signals in the equipment, rather than relying solely on short-term observations.

[0017] This data collection process yields comprehensive equipment operation data, providing a foundation for subsequent model training and fault prediction. The collected data can not only be used to train deep learning models but also provide essential information for fault diagnosis and early warning systems.

[0018] P20: Using a trained deep autoencoder, the historical multi-source time-series running data is mapped to points in the state space, constructing a healthy data cloud and multiple sets of fault data clouds, and analyzing the connection paths of the discrete state points of the healthy data cloud and the multiple sets of fault data clouds to establish multiple typical fault cloud routes.

[0019] Optionally, a pre-trained deep autoencoder can be used to process historical multi-source time-series runtime data. A deep autoencoder is a neural network model consisting of an encoder and a decoder. The encoder compresses complex high-dimensional time-series data into a low-dimensional state space to help identify potential features of the device state, while the decoder restores it to the original data. Specifically, each piece of historical runtime data, after being encoded by the autoencoder, is mapped to a point in the state space. This point retains the key features of the original data in the low-dimensional space, effectively reflecting the device's operating state at a given moment.

[0020] Next, based on the points in these mapped state spaces, we construct a health data cloud and multiple sets of fault data clouds. First, we group the points mapped from all normal state data into a set, resulting in the health data cloud. These points typically cluster together in the state space, forming a relatively dense region. This region reflects the state distribution characteristics of the device during normal operation and serves as an important reference benchmark for subsequent fault detection and prediction. Simultaneously, for each known fault type, we aggregate the points mapped from the corresponding fault data, forming multiple sets of fault data clouds. Each set of fault data clouds corresponds to a specific fault mode, and they exhibit their own distribution characteristics in the state space, showing a certain degree of distinction from the health data cloud.

[0021] After constructing the health data cloud and fault data cloud, further analysis is needed on the connection paths between discrete state points within these clouds. These connection paths reveal the evolution of equipment status from normal to faulty. Therefore, through in-depth analysis of a large amount of historical data, common evolution paths, i.e., typical fault cloud routes, can be identified and extracted. These routes show how, under specific fault modes, equipment state points gradually move from the health data cloud to the corresponding fault data cloud. For example, under certain mechanical fault modes, the equipment state points may gradually deviate from the normal range along a specific path, eventually entering the fault zone. By establishing these typical fault cloud routes, important references can be provided for subsequent fault prediction. When the real-time monitored state points move along these routes, early warnings of potential equipment fault types can be given, thus enabling proactive prevention.

[0022] Furthermore, in the construction of the depth autoencoder, step P20 of this embodiment further includes: P21: Construct a deep network consisting of a symmetrical encoder and decoder; the encoder contains three convolutional layers for layer-by-layer feature extraction, and the decoder contains three deconvolutional layers for input reconstruction. P22: Use normal state data from historical multi-source time-series running data as input, and perform unsupervised pre-training of the deep network with the primary objective of minimizing reconstruction error. P23: On the pre-trained model, use fault data of various known fault types, and perform supervised fine-tuning with the secondary objective of minimizing intra-class distance and maximizing inter-class distance among labeled samples in the state space. The convergence objective is to obtain the deep autoencoder by achieving a separation clarity of different state data greater than a preset threshold in the low-dimensional state space.

[0023] It should be understood that in order to construct a deep autoencoder that can effectively map the operating state of coal mine equipment, a deep network architecture needs to be designed and trained. This network consists of a symmetric encoder and decoder.

[0024] The encoder section comprises three convolutional layers for extracting features from the input data layer by layer. First, the first convolutional layer receives multi-source time-series input data, typically multi-dimensional time series data such as sensor data on temperature, pressure, and vibration. This layer extracts local features from the data through a sliding convolutional kernel operation and introduces non-linearity using an activation function (such as ReLU) to output a feature map. Next, the second convolutional layer further convolves the output feature map of the first layer to extract higher-level features. Simultaneously, pooling operations, such as max pooling, reduce the spatial dimensionality of the feature map, decreasing computational cost and enhancing feature robustness. Finally, the third convolutional layer continues feature extraction from the output of the second layer, outputting low-dimensional feature representations that effectively capture key information from the input data.

[0025] The decoder consists of three deconvolutional layers, which reconstruct the low-dimensional features output by the encoder into high-dimensional data similar to the input data. The first deconvolutional layer receives the output features from the encoder, gradually recovers the spatial dimension of the feature map through deconvolution operations, and introduces non-linearity through an activation function. The second deconvolutional layer continues to operate on the output of the first layer, further recovering the dimension of the feature map until it approximates the dimension of the original input data. The third deconvolutional layer finally outputs the reconstructed data, whose dimension is consistent with the input data. Through this symmetrical encoder-decoder structure, the deep autoencoder can extract key features of the data during the encoding process and restore the original data as much as possible during the decoding process.

[0026] When training a deep autoencoder, unsupervised pre-training is first performed using normal-state data from historical multi-source time-series datasets as input. The goal of pre-training is to minimize the reconstruction error, i.e., by adjusting the network weights to make the decoder output data as close as possible to the original input data. This process can be implemented using the backpropagation algorithm. For example, the error between the reconstructed data and the original data, such as the mean squared error, can be calculated, and then the network weights can be updated using gradient descent to gradually reduce the error. Unsupervised pre-training enables the network to learn the feature representations of normal-state data, providing a good initial model for subsequent supervised fine-tuning.

[0027] After unsupervised pre-training, supervised fine-tuning is performed using fault data of various known fault types. The goal of fine-tuning is to minimize the intra-class distance of labeled samples in the state space while maximizing the inter-class distance. Specifically, minimizing the intra-class distance means that data points of the same fault type are clustered together as much as possible in the state space, while maximizing the inter-class distance means that data points of different fault types are kept as far apart as possible. This process can be achieved by adjusting the network weights and optimizing the distribution of the state space, so that data of different states can be clearly separated in the low-dimensional space. Furthermore, the fine-tuning process uses a convergence objective where the separation clarity of data of different states in the low-dimensional state space is greater than a preset threshold. In practice, a loss function can be defined that considers both intra-class and inter-class distances, and the network weights can be optimized using gradient descent until the convergence condition is met, i.e., the separation clarity of the data is greater than the preset threshold, indicating that the model can accurately distinguish different fault states.

[0028] This process not only ensures that the model can effectively extract features, but also clearly separates different fault states in a low-dimensional space, making it better suited for fault prediction tasks and providing a solid foundation for subsequent fault prediction.

[0029] Furthermore, such as Figure 2 As shown, the connection paths of discrete state points of the health data cloud and the multiple fault data cloud sets are analyzed, and multiple typical fault cloud routes are established. Step P20 in this embodiment further includes: P24: Connect the health data cloud clusters with the cloud cluster data in the multiple fault data cloud cluster sets to establish multiple health-fault transition cloud cluster sets; P25: For the multiple health-fault transition cloud clusters, a continuous sequence of state points starting from the window marked as normal and continuing until the window where the fault is confirmed constitutes multiple initial degradation trajectory sets; P26: Use a dynamic time warping algorithm to align the initial degradation trajectories in the multiple initial degradation trajectory sets, and define typical fault cloud cluster routes from the health cloud clusters to the fault cloud clusters of the corresponding fault types based on the aligned trajectories, generating the multiple typical fault cloud cluster routes, and each typical fault cloud cluster route carries a fault type label.

[0030] Optionally, in order to establish multiple typical fault cloud routes, the connection paths of discrete state points of health data cloud and multiple fault data cloud sets can be further refined.

[0031] First, the healthy data cloud is connected with the cloud data from multiple sets of fault data clouds to construct a set of health-fault transition clouds. This process begins by calculating the distance between healthy and fault data points in the state space. Euclidean distance or other suitable distance metrics can be used to calculate the distance between each pair of healthy and fault data points, and a distance matrix is ​​obtained through matrix operations, where each element represents the distance between healthy and fault data points. Next, based on the distance matrix, fault data points that are closer to the healthy data cloud and healthy data points that are closer to the fault data cloud are identified, forming the transition region between the healthy and fault states. These points can be filtered by setting a distance threshold; for example, points with a distance less than the threshold are selected as candidate points for the transition region. Then, the points in the selected transition region are combined with the healthy and fault data clouds respectively to form multiple sets of health-fault transition clouds. Each set of transition clouds contains intermediate state points from the healthy state to a specific fault state, and these points form a continuous transition path in the state space.

[0032] Next, for each health-failure transition cloud set, a continuous sequence of state points from the normal state to the fault state is extracted to form an initial degradation trajectory set. For example, firstly, based on the time-series data of the device operation, the data is divided into multiple time windows. Each time window contains a certain number of state points, and the window size can be adjusted according to the device's operating characteristics and data sampling frequency. Then, starting from the window marked as normal, the process proceeds window by window until the window where the fault is confirmed ends. In each window, a representative point in the state space is selected, such as the mean or median of all points within the window, forming a continuous sequence of state points. These sequences reflect the gradual evolution of the device state from a healthy state to a fault state. Finally, all extracted continuous state point sequences are categorized into an initial degradation trajectory set. Each set corresponds to a specific fault type and contains degradation trajectories for multiple instances, reflecting the state change process of different devices under similar fault modes.

[0033] Because different devices or even the same device under different operating conditions exhibit varying degradation rates, significant time differences may exist between degradation trajectories. Therefore, Dynamic Time Warping (DTW) is required to align multiple initial degradation trajectories. Specifically, the DTW algorithm adjusts the state points in the trajectories along the time axis through nonlinear alignment, enabling comparisons between different trajectories on the same time scale. In this process, the local similarity of each pair of degradation trajectories is first calculated, for example, using Euclidean distance as a metric. Then, the time axes of the two trajectories are aligned by calculating the global minimum distance path, which is the optimal time alignment path. The aligned trajectories eliminate degradation rate differences, allowing for better correspondence between time points between trajectories, thereby improving the accuracy of fault mode identification.

[0034] Next, cluster analysis is performed on the aligned trajectories to calculate the trajectory centerline for each fault type. For example, the centerline can be calculated by averaging or medianing the aligned trajectories. In the state space, the trajectory centerline serves as a typical path for fault development, connecting the health data cloud and the fault data cloud. By aggregating the centerlines of multiple trajectories, typical fault cloud routes can be extracted. These routes reflect the typical degradation patterns of the device under different fault types. Each typical fault cloud route carries a fault type label, indicating that the route applies to a specific type of fault and helping the fault prediction system determine which stage of fault development the device may currently be in.

[0035] These typical fault cloud routes can provide clear reference paths for subsequent fault prediction, enabling the system to provide early warnings of possible equipment fault types based on the real-time monitored status point locations and movement trends, thereby achieving accurate prediction and effective management of coal mine equipment faults.

[0036] Furthermore, the aligned trajectory is defined by the trajectory centerline of the same fault type as a typical fault cloud route pointing from the healthy cloud to the fault cloud corresponding to the fault type. Step P26 in this embodiment of the application also includes: P26-1: After aligning multiple initial degradation trajectory sets, fit the centerlines corresponding to the cubic B-spline curves to obtain multiple parameterized curves; P26-2: According to the correspondence, calculate the vertical distance from other degradation trajectory points in the multiple initial degradation trajectory sets to the corresponding parameterized curves, and statistically obtain the width distribution at different progress levels; P26-3: Generate the multiple typical fault cloud routes using the multiple parameterized curves and the corresponding width distributions.

[0037] In one possible embodiment of this application, in order to accurately define the typical fault cloud route from the healthy cloud to the fault cloud corresponding to the fault type, the definition process of the trajectory centerline can be further refined.

[0038] After aligning multiple initial degradation trajectory sets, a cubic B-spline curve is used to fit the corresponding centerline to obtain a smooth parametric curve. Specifically, a B-spline curve is a piecewise polynomial function used to generate smooth curves, suitable for handling situations with many data points and irregular variations. In practice, several key points are selected from the aligned trajectory set as control points for the cubic B-spline curve. These control points can include the starting point (center of the healthy cloud), the ending point (center of the fault cloud), and several intermediate key state points. Then, using the mathematical formula of the cubic B-spline curve, a smooth parametric curve is fitted through these control points. This curve effectively represents the typical evolution path of the equipment state from healthy to faulty and has good mathematical properties, facilitating subsequent analysis and calculation. Repeating the above fitting process for each initial degradation trajectory set yields multiple parametric curves, each corresponding to a specific fault type.

[0039] Next, according to the correspondence, the vertical distances from other degenerate trajectory points within multiple initial degenerate trajectory sets to their corresponding parametric curves are calculated, and the width distribution at different progress levels is statistically obtained. To do this, it is first necessary to calculate the vertical distance from each trajectory point within each initial degenerate trajectory set to its corresponding parametric curve. The vertical distance typically refers to the shortest distance from the trajectory point to the fitted curve, i.e., the orthogonal distance from the trajectory point to the curve, which can be calculated using a mathematical formula. The purpose of this step is to measure the degree of deviation of each data point's position within the trajectory, thereby reflecting the degree of fit between the trajectory point and the fitted curve.

[0040] Next, the parameterized curve is divided into several segments according to the progress of equipment state evolution. For example, the path from the center of the healthy cloud cluster to the center of the faulty cloud cluster can be divided into several equal parts, with each segment representing a stage of equipment state evolution. Within each progress segment, the vertical distance distribution from all trajectory points to the parameterized curve is statistically analyzed. For example, the mean, standard deviation, and other statistics of all vertical distances within each progress segment are calculated to obtain the width distribution at different progress stages. This width distribution describes the dispersion of the trajectory points from the fitted curve at each time progress stage, i.e., the extent of trajectory expansion.

[0041] Finally, based on the fitted parametric curves and the calculated width distribution, multiple typical fault cloud routes are generated. Specifically, according to the parametric curve and corresponding width distribution of each trajectory, the width change of the fault cloud route at different time stages is calculated. This means combining each parametric curve with its corresponding width distribution to form a typical fault cloud route with width information. This route not only describes the central path of equipment status from healthy to faulty but also reflects the distribution width of status points at different stages, providing a more comprehensive representation of the characteristics of equipment status evolution. Repeating the above process for each fault type generates multiple typical fault cloud routes. Each route carries a corresponding fault type marker to identify the specific fault mode corresponding to that route. These routes not only help accurately predict equipment fault types but also provide maintenance personnel with specific fault intervention paths, ensuring that appropriate preventative measures can be taken before a fault occurs. In practical applications, by comparing the distance and distribution between the real-time monitored status points and these typical fault cloud routes, early warnings of potential equipment fault types can be provided, along with corresponding decision support, thereby achieving accurate prediction and effective management of coal mine equipment faults.

[0042] P30: Real-time acquisition of the current operating data of the coal mining equipment and mapping it to points in the state space to generate real-time state points.

[0043] Specifically, to achieve real-time status monitoring of coal mine equipment, it is necessary to collect the current operating data of the equipment in real time and map it into the state space to generate real-time state points. This process is a key step in the fault prediction system, ensuring timely detection of changes in the operating status of the equipment.

[0044] First, a real-time data acquisition system needs to be established to continuously monitor the operating status of coal mine equipment. This system includes multiple sensors distributed at key locations on the equipment to acquire real-time operating parameters such as temperature, pressure, vibration frequency, and current intensity. These sensors convert physical signals into electrical signals, which are then digitized by a data acquisition card. The data acquisition card transmits these digitized signals to a data processing center for further analysis and processing.

[0045] During data acquisition, it is crucial to ensure the integrity and accuracy of the data. Therefore, the data acquisition system should possess an automatic calibration function to ensure accurate sensor readings. Simultaneously, the system should be able to handle noise and interference during data transmission to ensure data reliability. The acquired real-time operational data typically exists in time-series format, with each time point corresponding to a set of device operating parameters.

[0046] Next, the collected real-time operational data is input into the previously trained deep autoencoder, compressing the high-dimensional real-time operational data into a low-dimensional state space. Specifically, the real-time operational data is used as input, and features are extracted layer by layer through the encoder's convolutional layers. Each convolutional layer performs a convolution operation on the input data and introduces non-linearity through an activation function (such as ReLU). After processing by three convolutional layers, a low-dimensional feature representation, namely the real-time state point, is finally obtained. This real-time state point reflects the device's operational state at the current moment in the state space.

[0047] The generated real-time status points will be stored in the state space for subsequent comparative analysis with the previously constructed health data cloud and fault data cloud to determine whether the device's current state is close to a known fault mode. By continuously performing the above real-time acquisition, preprocessing, encoding, and mapping process, the real-time status points can be continuously updated, enabling the system to dynamically reflect the changing trajectory of the device's operating state in the state space.

[0048] P40: Determine the matching status of the real-time status point to the multiple typical fault cloud routes. If the preset matching status is met, determine that the coal mine equipment has entered the corresponding fault cloud route, issue an early warning signal, and generate an active intervention suggestion adapted to the corresponding fault cloud route.

[0049] Furthermore, step P40 in this embodiment of the application also includes: P41: Perform distance matching status judgment between the real-time status point and the multiple typical fault cloud cluster routes to generate a first status matching result; P42: Perform direction matching status judgment between the real-time status point and the multiple typical fault cloud cluster routes to generate a second status matching result; P43: Combine the first status matching result and the second status matching result to determine whether the preset matching state is satisfied, and generate the corresponding fault cloud cluster route. The preset matching state is that the shortest vertical distance is less than the typical width of the typical fault cloud cluster route at the advancement progress point, and the movement direction vector is consistent with the tangent direction.

[0050] It should be understood that after obtaining the real-time status point, it is necessary to match and judge the real-time status point with multiple established typical fault cloud routes to determine whether the current operating status of the coal mine equipment has entered a certain fault evolution path.

[0051] First, distance matching is performed between the real-time state point and multiple typical fault cloud routes to determine the state, generating a first state matching result. Specifically, each typical fault cloud route has been represented as a parameterized curve in step P26 above, and has a corresponding typical width distribution at different progress positions of the curve. Therefore, for the real-time state point generated at the current moment, the following calculations can be performed sequentially for each typical fault cloud route.

[0052] First, for each typical fault cloud route, the shortest vertical distance from the real-time status point to the curve is calculated using the mathematical expression of its parameterized curve. For example, by traversing the parameterized curve or using projection calculation, the point with the smallest distance to the real-time status point is searched on the parameterized curve of the typical fault cloud route. Then, the shortest vertical distance from the real-time status point to this point is obtained by calculating Euclidean distance, which is the shortest normal distance from the real-time status point to the parameterized curve at that location. Simultaneously, based on the parameter value of this point on the parameterized curve, the current advancement progress is determined, and the typical width value corresponding to this advancement progress is read. If the shortest vertical distance is less than the typical width at this advancement progress, the real-time status point is considered to match the typical fault cloud route in the distance dimension, and the first state matching result for this route is considered satisfied; if it is greater than or equal to the typical width, it is considered not satisfied. The above distance judgment is performed on all typical fault cloud routes to obtain multiple first state matching results.

[0053] Next, the real-time state points are matched with multiple typical fault cloud routes to determine their state and generate a second state matching result. Specifically, the motion direction vector of the real-time state point needs to be calculated first. This vector can be obtained from the coordinate difference between the current real-time state point and the previous real-time state point in the state space, and is used to characterize the evolution direction of the device state in the state space. Then, for each typical fault cloud route, at the point corresponding to the shortest distance, the tangent direction vector of the parameterized curve is calculated using the first derivative of the parameterized curve at that location, and is used to characterize the typical fault evolution direction.

[0054] Then, the angle or direction consistency index between the motion direction vector of the real-time state point and the tangent direction vector is calculated. For example, the cosine similarity between the two vectors can be calculated. When the cosine similarity is greater than a preset direction threshold, or the angle is less than a preset angle threshold, the motion direction of the real-time state point is determined to be consistent with the advancement direction of the typical fault cloud path, and the second state matching result of the path is recorded as satisfied; otherwise, it is recorded as not satisfied. The above direction matching judgment is performed for all typical fault cloud paths, and the corresponding second state matching result is generated.

[0055] Finally, the first state matching result and the second state matching result are comprehensively judged to determine whether the preset matching state is met, and the corresponding fault cloud route is generated. Specifically, for each typical fault cloud route, the first state matching result and the second state matching result are checked simultaneously. If the first state matching result and the second state matching result corresponding to the route are both satisfied, that is, the real-time state point matches the typical fault cloud route in distance (i.e., the shortest vertical distance is less than the typical width threshold) and in direction (i.e., the motion direction vector is consistent with the tangent direction), then it is determined that the real-time state point has entered the typical fault cloud route, and the route is output as the currently matched fault cloud route.

[0056] Based on this assessment, an early warning signal is issued, containing information such as the fault type, possible fault location, and fault severity. Simultaneously, proactive intervention suggestions are generated tailored to the corresponding fault cloud path. These suggestions, based on the fault type and equipment operating status, provide specific maintenance measures or operational recommendations, such as suggesting equipment shutdown for inspection, replacement of specific components, or adjustment of equipment operating parameters. The early warning signal and intervention suggestions are then communicated to equipment maintenance personnel via system interface or notification to ensure timely action and prevent further escalation of the fault.

[0057] Furthermore, distance matching status judgment is performed between the real-time status point and the multiple typical fault cloud routes to generate a first status matching result. Step P41 in this embodiment of the application also includes: P41-1: Calculate the shortest vertical distance from the real-time state point to the parameterized centerline of the multiple typical fault cloud routes; P41-2: Calculate the progress of the real-time state point on the multiple typical fault cloud routes, and extract the typical width of the multiple typical fault cloud routes at the progress point; P41-3: Compare whether the shortest vertical distance is less than the typical width of the multiple typical fault cloud routes at the progress point, and generate the first state matching result.

[0058] Specifically, the process of determining the status by distance matching between real-time status points and multiple typical fault cloud routes can be further refined.

[0059] First, the shortest vertical distance from the real-time state point to the parameterized centerline of multiple typical fault cloud routes is calculated. Since each typical fault cloud route is represented as a continuous parameterized curve in a low-dimensional state space and described by curve parameters, such as the normalized parameter s∈[0,1], for the real-time state point generated at the current moment, the shortest distance calculation needs to be performed for each typical fault cloud route. Specifically, the coordinates of the real-time state point in the state space can be used as a fixed point, and the curve position point with the smallest distance to that point can be searched on the parameterized centerline of the corresponding typical fault cloud route. For each typical fault cloud route, the shortest vertical distance from the real-time state point to the centerline is calculated using a numerical optimization method based on the mathematical expression of its parameterized centerline. For example, by minimizing the Euclidean distance between the real-time state point and each point on the parameterized centerline, the closest point is found, and the distance between that point and the real-time state point is calculated, thus obtaining the shortest vertical distance.

[0060] Next, the progress of the real-time state point along multiple typical fault cloud paths is calculated, and the typical width of each path at that progress point is extracted. Specifically, the previous steps have determined the shortest distance point on the parameterized center line corresponding to the real-time state point. This point corresponds to a unique curve parameter value on the parameterized curve. This curve parameter value serves as the progress of the real-time state point along the typical fault cloud path, characterizing the relative position of the device's current state on the fault evolution path. Then, based on the progress, the typical width at the corresponding position is extracted from the width distribution data of the typical fault cloud paths. The typical width is obtained from the previous steps based on the vertical distance from the trajectory point to the center line, reflecting the allowable fluctuation range of the fault evolution trajectory in the state space during this progress stage.

[0061] Finally, the shortest vertical distance is compared with the typical width of multiple typical fault cloud routes at the advancement progress point to generate the first state matching result. Specifically, for each typical fault cloud route, the calculated shortest vertical distance is compared with the typical width at the corresponding advancement progress point. If the shortest vertical distance is less than the typical width, it means that the real-time state point is spatially close to the typical fault cloud route, and the matching status between the real-time state point and the typical fault cloud route is recorded as a match. Conversely, if the shortest vertical distance is greater than or equal to the typical width, the matching status is recorded as a mismatch. The matching statuses of all typical fault cloud routes are summarized to generate the first state matching result. This result records in detail the matching status between the real-time state point and each typical fault cloud route in terms of distance, which can provide important distance matching basis for subsequent fault prediction, ensuring the accuracy and reliability of fault prediction.

[0062] Furthermore, the real-time status point is matched with the routes of the multiple typical fault cloud clusters to determine the direction and generate a second status matching result. Step P42 in this embodiment of the application also includes: P42-1: Collect multiple continuous real-time state points at multiple consecutive moments and calculate the motion direction vector; P42-2: Calculate the tangent direction of the motion direction vector at the current progress point of the multiple typical fault cloud routes; P42-3: Determine whether the motion direction vector is consistent with the tangent direction of the multiple typical fault cloud routes and generate the second state matching result.

[0063] Optionally, the process of determining the direction matching status between real-time status points and multiple typical fault cloud routes can be further refined.

[0064] First, multiple consecutive real-time state points are collected to calculate the device's motion direction vector in the state space. Specifically, a real-time data acquisition system acquires the device's operational data at consecutive time points and inputs this data into a depth autoencoder to generate a series of consecutive real-time state points. These state points reflect the device's operational state at different times. Next, the motion direction vector of the device's state is obtained by calculating the difference between adjacent real-time state points. For example, if there are three consecutive state points, the motion direction vector can be obtained by calculating the difference between the later and earlier state points. To improve accuracy, the average of multiple consecutive difference vectors can be taken as the final motion direction vector to more accurately reflect the changing trend of the device's state.

[0065] Next, the tangent direction of the motion direction vector at the current progress point of multiple typical faulty cloud cluster routes is calculated. Specifically, firstly, based on the position of the real-time state point, its current progress point on each typical faulty cloud cluster route is determined. Then, using this position point as a reference, the derivative of the parameterized centerline of the typical faulty cloud cluster route is calculated to obtain the tangent direction vector at that position point. This tangent direction vector reflects the theoretical evolution direction of the typical faulty cloud cluster route in the current progress stage and is used to describe the directional trend from healthy cloud clusters to the corresponding faulty cloud clusters.

[0066] Finally, the system determines whether the motion direction vector matches the tangent direction of multiple typical fault cloud routes, generating a second-state matching result. In practice, for each typical fault cloud route, the angle between its motion direction vector and tangent direction vector is compared. If the two vectors are essentially aligned (i.e., the angle is small, for example, less than 15 degrees), the motion direction is considered to match the tangent direction, and the matching status of the real-time state point with that typical fault cloud route is recorded as a match. Conversely, if the angle is large, the matching status is recorded as a mismatch. The matching statuses of all typical fault cloud routes are summarized to generate the second-state matching result. This result details the matching status of the real-time state point with each typical fault cloud route in terms of direction, providing crucial directional matching information for subsequent fault prediction and ensuring the accuracy and reliability of fault prediction.

[0067] Furthermore, in generating proactive intervention suggestions adapted to the corresponding faulty cloud path, step P40 of this embodiment also includes: P44: A pre-set intervention strategy knowledge base is bound to the multiple typical fault cloud routes. Each typical fault cloud route is associated with a strategy set, and the strategies are defined in segments according to the route progress. P45: When it is determined that the coal mine equipment has entered the corresponding fault cloud route, based on the calculated real-time progress, the intervention measures corresponding to the real-time progress are retrieved from the strategy set associated with the corresponding fault cloud route, and the active intervention suggestion is generated.

[0068] In one possible embodiment of this application, in order to generate proactive intervention suggestions adapted to the corresponding faulty cloud cluster routes, it is first necessary to pre-define an intervention strategy knowledge base bound to multiple typical faulty cloud cluster routes. This intervention strategy knowledge base is established during the system deployment or initialization phase and organized and stored using typical faulty cloud cluster routes as an index.

[0069] Specifically, for each typical fault cloud path, a detailed set of intervention strategies is defined during the system initialization phase based on historical data and expert experience. These strategy sets are stored in a knowledge base, with each set containing multiple intervention measures defined in segments according to the progress of the fault cloud path. For example, a typical fault cloud path can be divided into three phases: initial, intermediate, and late stages, each with corresponding intervention measures. Initial stage interventions might include simple checks and adjustments; intermediate stage interventions might require maintenance or replacement of some components; and late stage interventions might require a complete shutdown for overhaul. These intervention measures are documented in detail, including operating procedures, required tools, and expected results, to facilitate rapid execution in practical applications.

[0070] Next, when it is determined that the coal mine equipment has entered the corresponding fault cloud path, the system first reads the progress parameter of the real-time status point on the corresponding typical fault cloud path. This progress parameter has been calculated during distance matching or direction matching and is used to characterize the relative position of the equipment's current state on the fault evolution path. Subsequently, using the typical fault cloud path as an index, the system retrieves the set of strategies bound to that path from the intervention strategy knowledge base, and locates the target progress interval containing the real-time progress within this strategy set according to predefined progress segmentation rules. Then, the corresponding intervention strategy entry is extracted from the target progress interval as the intervention measure output for the current equipment state. For example, if the real-time progress is in the middle stage of the fault cloud path, the system will retrieve the intervention measures for the middle stage, compile them into a detailed proactive intervention suggestion, and convey this intervention suggestion to the equipment maintenance personnel through the system interface or notification so that they can take timely measures to prevent the fault from worsening.

[0071] Through this process, as the equipment status progresses along a typical fault cloud path, the system can automatically match and output intervention measures appropriate to the specific degradation stage the equipment is currently in, thereby improving the efficiency and reliability of equipment maintenance and providing strong support for fault prevention and management of coal mine equipment.

[0072] In summary, the embodiments of this application have at least the following technical effects: This application utilizes a deep autoencoder to extract features and map states from historical multi-source time-series operational data, accurately constructing health and fault data clouds, establishing typical fault paths, and achieving early warning and accurate prediction of faults. It also collects current equipment operating data in real time and maps it to the state space, enabling real-time monitoring of equipment operating status and timely detection of abnormal changes. Combined with a pre-built intervention strategy knowledge base, it generates proactive intervention suggestions adapted to the fault type based on the location and progress of real-time state points, helping maintenance personnel to quickly take measures to prevent further deterioration of faults. This significantly improves the operational reliability of coal mine equipment, reduces production interruptions caused by sudden faults, thereby lowering equipment maintenance costs, reducing unnecessary equipment downtime and maintenance workload, and enhancing the overall safety of coal mine production.

[0073] The technology achieves the effect of continuously judging the real-time status by constructing typical fault cloud routes, enabling forward-looking prediction and proactive intervention of the fault evolution trend of coal mine equipment.

[0074] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0075] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0076] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for predicting coal mine equipment faults based on deep learning, characterized in that, include: Collect historical multi-source time-series operating data of coal mine equipment, including normal status data and fault data of different known fault types; The trained deep autoencoder is used to map the historical multi-source time-series running data into points in the state space, constructing a healthy data cloud and multiple sets of fault data clouds. The connection paths of the discrete state points of the healthy data cloud and the multiple sets of fault data clouds are analyzed, and multiple typical fault cloud routes are established. Real-time data collection of the current operating data of the coal mining equipment is mapped to points in the state space to generate real-time state points; Determine the matching status between the real-time status point and the multiple typical fault cloud routes. If the preset matching status is met, determine that the coal mine equipment has entered the corresponding fault cloud route, issue an early warning signal, and generate an active intervention suggestion adapted to the corresponding fault cloud route.

2. The method for predicting coal mine equipment faults based on deep learning as described in claim 1, characterized in that, The construction steps of the depth autoencoder include: Construct a deep network consisting of a symmetric encoder and decoder; Using normal state data from historical multi-source time-series running data as input, the deep network is pre-trained in an unsupervised manner with the primary objective of minimizing reconstruction error; On the pre-trained model, fault data of various known fault types are used. The second objective is to minimize the intra-class distance of labeled samples in the state space and maximize the inter-class distance. Supervised fine-tuning is performed, and the convergence objective is to obtain the deep autoencoder by achieving a separation clarity of different state data in the low-dimensional state space that is greater than a preset threshold.

3. The coal mine equipment fault prediction method based on deep learning as described in claim 2, characterized in that, The encoder contains three convolutional layers for extracting features layer by layer, and the decoder contains three deconvolutional layers for reconstructing the input.

4. The coal mine equipment fault prediction method based on deep learning as described in claim 2, characterized in that, Analyze the connection paths of discrete state points in the health data cloud and the multiple fault data cloud sets, and establish multiple typical fault cloud routes, including: The health data cloud clusters are connected to the cloud cluster data in the multiple fault data cloud cluster sets to establish multiple health-fault transition cloud cluster sets. For the multiple health-failure transition clouds, a continuous sequence of state points starting from the window marked as normal and continuing until the window where the failure is confirmed constitutes multiple sets of initial degradation trajectories. The dynamic time warping algorithm is used to align the initial degradation trajectories in multiple initial degradation trajectory sets. Based on the aligned trajectories, the center lines of the same fault type are defined to form typical fault cloud routes from healthy cloud clusters to fault cloud clusters of the corresponding fault type. The multiple typical fault cloud routes are generated, and each typical fault cloud route carries a fault type label.

5. The coal mine equipment fault prediction method based on deep learning as described in claim 4, characterized in that, The aligned trajectories are defined by their center lines of the same fault type as typical fault cloud routes pointing from healthy cloud clusters to fault cloud clusters of the corresponding fault type, including: After aligning multiple initial degenerate trajectory sets, the centerlines corresponding to cubic B-spline curves are fitted to obtain multiple parameterized curves. Based on the correspondence, the vertical distance from other degenerate trajectory points in multiple initial degenerate trajectory sets to the corresponding parameterized curves is calculated, and the width distribution at different progress levels is statistically obtained. The multiple typical fault cloud routes are generated using the multiple parameterized curves and their corresponding width distributions.

6. The method for predicting coal mine equipment faults based on deep learning as described in claim 1, characterized in that, Determine the matching status of the real-time status point to the multiple typical fault cloud routes. If a preset matching status is met, determine that the coal mine equipment has entered the corresponding fault cloud route, including: The distance matching status of the real-time status point and the multiple typical fault cloud routes is determined to generate a first status matching result. The real-time status points are matched with the routes of the multiple typical fault cloud clusters to determine the direction and status, and a second status matching result is generated. By combining the first state matching result and the second state matching result, it is determined whether the preset matching state is met, and a corresponding fault cloud route is generated.

7. The coal mine equipment fault prediction method based on deep learning as described in claim 6, characterized in that, The distance matching status is determined between the real-time status point and the multiple typical fault cloud routes to generate a first status matching result, including: Calculate the shortest vertical distance from the real-time status point to the parameterized centerline of the multiple typical fault cloud routes; Calculate the progress of the real-time status point on the multiple typical fault cloud path, and extract the typical width of the multiple typical fault cloud path at the progress point; The first state matching result is generated by comparing whether the shortest vertical distance is less than the typical width of the multiple typical fault cloud routes at the advancement progress.

8. The coal mine equipment fault prediction method based on deep learning as described in claim 6, characterized in that, The real-time status points are matched with the routes of the multiple typical fault cloud clusters to determine their direction and generate a second status matching result, including: Collect multiple consecutive real-time state points at multiple consecutive moments and calculate the motion direction vector; Calculate the tangential direction of the motion direction vector at the current progress point of the multiple typical fault cloud routes; Determine whether the motion direction vector is consistent with the tangent direction of the multiple typical fault cloud routes, and generate the second state matching result.

9. The coal mine equipment fault prediction method based on deep learning as described in claim 6, characterized in that, The preset matching state is that the shortest vertical distance is less than the typical width of the typical fault cloud path at the advancement progress and the motion direction vector is consistent with the tangent direction.

10. The coal mine equipment fault prediction method based on deep learning as described in claim 1, characterized in that, Generate proactive intervention recommendations adapted to the corresponding faulty cloud path, including: A pre-defined intervention strategy knowledge base is provided, which is bound to the multiple typical fault cloud cluster routes. Each typical fault cloud cluster route is associated with a set of strategies, and the strategies are defined in segments according to the route progress. When it is determined that the coal mining equipment has entered the corresponding fault cloud route, the intervention measures corresponding to the real-time progress are retrieved from the strategy set associated with the corresponding fault cloud route based on the calculated real-time progress, and the active intervention suggestion is generated.

Images