A time series data prediction and anomaly detection method for mine equipment

Abstract

This invention provides a time-series data prediction and anomaly detection method for mining equipment, aiming to solve the problems of existing technologies such as loss of high-frequency information in multi-source heterogeneous data, separation of prediction and detection, reliance on labeled data, and high computing power. This method collects raw multi-frequency historical time-series data through heterogeneous sensors, obtains data of the same frequency through a unified feature layer scheme; inputs N historical frames of data into a DWLinear model, performs trend and noise component splitting and weighted fusion, and softmax normalization to output M future frames of data; after stitching, it outputs anomaly information through a mechanistic model, time-domain-frequency domain transformation (Fourier transform + wavelet transform), and isolated forest detection. This invention requires no manual annotation, the MSE of the DWLinear model is reduced by 17.82% compared to DLinear, it is lightweight and adaptable to edge devices, achieves "prediction-detection" collaboration, and covers multi-dimensional monitoring and various anomaly scenarios.

Description

Technical Field

[0001] This invention relates to the field of mining equipment condition monitoring technology, and in particular to a method for time-series data prediction and anomaly detection for mining equipment. Background Technology

[0002] Mining equipment (such as fans, conveyors, and pumps) is a core asset in mine production, and its operating status directly determines production efficiency and operational safety. The multi-dimensional time-series data generated during equipment operation, including vibration, temperature, pressure, and current / voltage, contains information about the equipment's health status. Predicting future operating conditions using time-series data can support preventative maintenance, and anomaly detection can identify early signs of failure to warn of potential risks. However, mining environments are characterized by high-dimensional data, nonlinearity, strong noise, and dynamic changes in operating conditions. Furthermore, equipment failures can easily lead to production interruptions, equipment damage, and even safety accidents. Therefore, stringent requirements are placed on prediction accuracy, detection reliability, and real-time performance.

[0003] Condition monitoring of mining equipment has become an essential requirement in the industry, but existing technologies struggle to adapt to complex scenarios. A publicly available method for predicting energy consumption in mining equipment, patent application number CN202510665895.9, achieves energy consumption prediction through multi-dimensional data collection, preprocessing, operating condition identification, transfer learning model construction, and edge-cloud collaboration. While this method improves accuracy to some extent, it has significant drawbacks: It relies on a large amount of labeled data: it requires labeling energy consumption data corresponding to more than 100,000 operation, process and environmental parameters, which is extremely costly in terms of human and financial resources; Limitations of single-dimensional prediction: It can only predict energy consumption in a single dimension and cannot cover the multi-dimensional health status of equipment; High computational requirements for the model: The cloud-based model uses a multi-layered Transformer teacher model with a large number of parameters. Training and inference require high-performance hardware, resulting in high hardware procurement and maintenance costs. Module separation leads to delayed response: the prediction and anomaly detection modules are independent, making it impossible to achieve "prediction-detection" collaboration and making it difficult to provide early warnings of future failures; Insufficient processing of multi-source heterogeneous data: The problem of high-frequency information loss caused by differences in sampling frequencies of different sensors has not been effectively solved; Poor adaptability: When new equipment lacks abnormal data, the model cannot be reused and is difficult to adapt to dynamic changes in operating conditions.

[0004] In summary, existing technologies are insufficient to meet the demands of mining scenarios for high-precision, low-cost, and highly adaptable time-series prediction and anomaly detection. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method for time series data prediction and anomaly detection for mining equipment.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for time-series data prediction and anomaly detection for mining equipment, comprising the following steps: Raw, multi-frequency, multi-dimensional historical time-series data of mining equipment are collected and acquired through multiple heterogeneous sensors. By adopting a unified feature layer scheme, the frequency of the sensor with the lowest sampling frequency is used as the reference output frequency. The corresponding sliding window value is set according to the ratio between the sampling frequency of multiple high sampling frequency sensors and the reference output frequency. Data features are extracted in each sliding window to obtain multi-dimensional time series feature data of the same frequency. By inputting historical N-frame multi-dimensional time series feature data from the same frequency multi-dimensional time series feature data at the same time interval into the DWLinear model for time series prediction, the future M-frame multi-dimensional time series feature data is output. A complete time-series dataset is formed by concatenating the historical N-frame multidimensional time-series feature data with the future M-frame multidimensional time-series feature data. Anomalies are detected in the time-series dataset using a mechanistic model, time-domain-frequency domain transformation, and unsupervised anomaly detection methods. Anomaly information is identified and output, wherein the anomaly information includes anomaly dimension, anomaly time period, and anomaly type.

[0007] In some possible embodiments, the original multi-frequency historical time-series data includes monitoring data for vibration, temperature, current / voltage, and pressure / flow dimensions, with different sampling frequencies for each dimension.

[0008] In some possible embodiments, the data features include general features and dimension-specific features; the general features are the mean and standard deviation; the dimension-specific features include: for vibration sensors: peak factor, FFT main frequency energy ratio and autocorrelation coefficient (1st hysteresis); for current sensors: RMS value, harmonic distortion rate and second-order difference extremum; for pressure sensors: first-order difference variance, static offset and moving mean change rate; for temperature sensors: raw data and one-dimensional slope.

[0009] In some possible embodiments, the DWLinear model uses one-dimensional convolutional layers. The trend component is obtained by smoothing the historical N frames of multidimensional temporal feature data. The historical N frames of multidimensional time-series feature data and trend components are combined. The noise component is obtained by subtraction. Then the trend components and the noise component After being processed by a linear layer, the corresponding output components are output. and The output components are fused using trainable weights. and The multidimensional temporal feature data of the future M frames is obtained after softmax normalization.

[0010] In some possible embodiments, the training method of the DWLinear model includes: using the Adam optimizer, adding an L2 regularization term and setting weight_decay to 1e-5, and using SmoothL1Loss as the loss function.

[0011] In some possible embodiments, the mechanism model includes three threshold detection methods: maximum and minimum threshold detection, trend threshold detection, and fluctuation threshold detection. The maximum and minimum threshold detection filters out abnormal value periods by setting a numerical threshold and the time that the value exceeds the threshold. The trend threshold detection filters out abnormal trend periods by setting a rate of change threshold. The fluctuation threshold detection filters out abnormal fluctuation periods by setting a standard deviation threshold.

[0012] In some possible embodiments, the time-domain to frequency-domain transformation includes Fourier transform and wavelet transform. The Fourier transform converts the time-series dataset into frequency-domain data by performing a fast Fourier transform, sets an amplitude threshold, filters out amplitude anomaly segments, and outputs the anomaly dimension and the anomaly category. The wavelet transform converts the time-series dataset into frequency-domain data by performing a continuous wavelet transform (CWT), sets an energy threshold, filters out energy anomaly segments, and outputs the anomaly dimension, the anomaly category, and the anomaly time period.

[0013] In some possible embodiments, the unsupervised anomaly detection employs the Isolation Forest algorithm, which includes a training phase and an inference phase. During the training phase, the time-series dataset is converted into sample features using a moving average (when N+M is too large). Samples and feature dimensions are randomly selected to build a binary search tree. This process is repeated multiple times to form an Isolation Forest containing 200 trees, with a default anomaly ratio of 0.01. During the inference phase, the time-series dataset is converted into sample features and then input into the Isolation Forest. An anomaly score is calculated using a level-order traversal. When the anomaly score exceeds a set threshold, the sample is identified as an anomaly, and the anomaly time period and the anomaly dimension are output.

[0014] In a second aspect, the present invention provides an electronic device, comprising: One or more processors; A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the time-series data prediction and anomaly detection method for mining equipment described above.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it can realize the time-series data prediction and anomaly detection method for mining equipment as described above.

[0016] The time-series data prediction and anomaly detection method for mining equipment according to the embodiments of the present invention has the following beneficial effects: This invention addresses the pain points of time-series data processing for mining equipment, bringing several significant benefits: First, it boasts low cost and high adaptability, eliminating the need for manual labeling of large amounts of data (the prediction task automatically divides training data and labels, and anomaly detection uses unsupervised or mechanistic models). The isolated forest algorithm adapts to scenarios with new equipment and no anomaly data, enabling reuse of similar equipment (for example, if a mine purchases 10 wind turbines of the same brand and model this month, with turbine 001 already installed and operational, and the other 9 turbines to be installed in one month, data can be collected on turbine 001 first to train the isolated forest algorithm. When the other 9 turbines are installed one month later, the isolated forest model trained on turbine 001 can be reused; however, once sufficient historical data is collected on the other 9 turbines, it is preferable to train each turbine individually to ensure inference accuracy), reducing model migration costs. Simultaneously, the DWLinear multidimensional time-series data prediction model overcomes the limitations of a single dimension, enabling simultaneous... This model predicts multi-dimensional monitoring data such as vibration, temperature, current / voltage, and pressure / flow rate, improving prediction accuracy without increasing computational load. By combining the Adam optimizer, L2 regularization (weight_decay=1e-5), and the SmoothL1 Loss function, it reduces the MSE by 17.82% compared to the DLinear model. The lightweight model can be directly deployed on edge mining equipment, eliminating reliance on high cloud computing power. Furthermore, by stitching together historical N frames and future M frames of predicted data to form a complete time-series set, it achieves collaborative optimization of "prediction-detection," enabling the detection of past anomalies and early warning of potential future faults, overcoming the response lag problem caused by traditional module separation. It also ensures data quality through a unified feature layer scheme of "sliding window + dimension-specific feature extraction," avoiding the loss of high-frequency information. More importantly, it integrates mechanistic models, time-domain-frequency transformation, and unsupervised anomaly detection methods, covering explicit known anomalies, latent frequency anomalies, and unknown novel faults, improving the comprehensiveness and reliability of detection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an example electronic device for the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 2The flowchart below shows the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 3 This is an overall framework diagram of the time-series data prediction and anomaly detection method for mining equipment according to the present invention; Figure 4 This is a schematic diagram of a multi-frequency data sliding window for a time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 5 This is a schematic diagram of multi-frequency data feature extraction and stitching for the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 6 This is a schematic diagram of the DWLinear model of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 7 This is a diagram showing the accuracy results of the sampling interval 1s_DLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 8 This is a diagram showing the accuracy of the sampling interval 1s_DWLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 9 This is a diagram showing the accuracy results of the sampling interval 10s_DLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 10 This is a diagram showing the accuracy results of the sampling interval 10s_DWLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 11 This is a graph showing the accuracy results of the sampling interval 30s_DLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 12 This is a diagram showing the accuracy results of the sampling interval 30s_DWLinear of the time-series data prediction and anomaly detection method for mining equipment of the present invention on the validation set; Figure 13 This is a schematic diagram of the threshold anomaly detection result of the time-series data prediction and anomaly detection method for mining equipment according to the present invention; Figure 14 This is a schematic diagram of the trend anomaly detection results of the time-series data prediction and anomaly detection method for mining equipment according to the present invention; Figure 15 This is a schematic diagram of the fluctuation anomaly detection results of the time-series data prediction and anomaly detection method for mining equipment according to the present invention; Figure 16This is a schematic diagram of the Fourier transform anomaly detection result of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 17 This is a schematic diagram of the wavelet transform anomaly detection result of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 18 This is a schematic diagram of time-series data conversion to sample features in the time-series data prediction and anomaly detection method for mining equipment according to the present invention; Figure 19 This is an example diagram illustrating the creation of an isolated forest for the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Detailed Implementation

[0018] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Figure 1 This is a schematic diagram of an example electronic device for implementing a time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 1 As shown, the electronic device 100 includes one or more processors 110, one or more storage devices 120, one or more input devices 130, one or more output devices 140, etc., and these components are interconnected via a bus system 150 and / or other forms of connection mechanisms. It should be noted that... Figure 1 The components and structures of the electronic devices shown are merely exemplary and not limiting; other components and structures may be used as needed.

[0020] The processor 110 may be a central processing unit (CPU), or may be a processing unit consisting of multiple processing cores, or other forms of processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 100 to perform desired functions.

[0021] Storage device 120 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, which a processor may execute to implement the client functions (implemented by the processor) in the embodiments of this disclosure described below, and / or other desired functions. Various applications and various data may also be stored in the computer-readable storage medium, such as various data used and / or generated by the applications.

[0022] The input device 130 may be a device used by a user to input commands, and may include one or more of a keyboard, mouse, microphone, and touch screen.

[0023] The output device 140 can output various information (such as images or sounds) to the outside (e.g., a user) and may include one or more of a display, a speaker, etc.

[0024] Figure 2 This is a flowchart of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 3 This is a diagram illustrating the overall framework of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. Figure 2 and Figure 3As shown, a time-series data prediction and anomaly detection method S200 for mining equipment includes steps S201 to S204. Step S201: Collect and acquire raw multi-frequency dimension historical time-series data of the mining equipment through multiple heterogeneous sensors; Step S202: By adopting a unified feature layer scheme, the frequency of the lowest sampling frequency sensor (usually a temperature sensor) is used as the reference output frequency. A corresponding sliding window value is set according to the ratio between the sampling frequencies of multiple high sampling frequency sensors and the reference output frequency. Data features are extracted within each sliding window to obtain multi-dimensional time-series feature data of the same frequency; Step S203: Step S204: By inputting historical N-frame multi-dimensional time series feature data from the same frequency and time interval into the DWLinear model for time series prediction, the future M-frame multi-dimensional time series feature data is output. Step S205: After concatenating the historical N-frame multi-dimensional time series feature data with the future M-frame multi-dimensional time series feature data to form a complete time series dataset, the time series dataset is subjected to anomaly detection using a mechanistic model, time-domain-frequency domain transformation, and unsupervised anomaly detection method. Anomaly information is identified and output. The anomaly information is divided into historical anomaly information and future anomaly information. The anomaly information includes anomaly dimension, anomaly time period, and anomaly type.

[0025] Specifically, multiple heterogeneous sensors include, but are not limited to, vibration sensors, temperature sensors, current sensors, and pressure sensors. Since multiple heterogeneous sensors collect different types of data, their required sampling frequencies are also different. The common sensor types and sampling frequencies used in mining equipment are shown in Table 1.

[0026] Table 1. Required sampling frequencies for different types of sensors.

[0027] Specifically, mining equipment includes, but is not limited to, fans, conveyors, and water pumps.

[0028] Specifically, the sliding window value is calculated based on the ratio between the sampling frequency of multiple high-sampling-frequency sensors and the reference output frequency. The formula is: Sliding window value = Target dimension sensor (vibration sensor, current sensor, and pressure sensor, etc.) sampling frequency / Reference frequency (temperature sensor sampling frequency). For example, if the temperature sensor sampling frequency is 0.1Hz (frame interval 10s) and the pressure sensor sampling frequency is 1Hz (frame interval 1s), then the sliding window value for the pressure dimension is 1 / 0.1 = 10. The sliding window values ​​for other high-sampling-frequency dimensions are calculated using the same logic. For details on multi-frequency data sliding window values, please refer to [link to relevant documentation]. Figure 4 The diagram shown is a multi-frequency data sliding window schematic of the time-series data prediction and anomaly detection method for mining equipment according to the present invention.

[0029] In some embodiments, the original multi-frequency historical time-series data includes, but is not limited to, monitoring data for vibration, temperature, current / voltage, and pressure / flow dimensions, with different sampling frequencies for each dimension.

[0030] In some embodiments, such as Figure 5 The diagram shown is a multi-frequency data feature extraction and stitching diagram of the time-series data prediction and anomaly detection method for mining equipment according to the present invention. The data features include general features and dimension-specific features. The general features are the mean and standard deviation. The dimension-specific features include: vibration sensor: peak factor, FFT main frequency energy ratio and autocorrelation coefficient (lag 1), current sensor: RMS value, harmonic distortion rate and second-order difference extremum, pressure sensor: first-order difference variance, static offset and moving mean change rate, temperature sensor: raw data and one-dimensional slope.

[0031] In some embodiments, the DWLinear model is an optimization of the DLinear model structure, and the DWLinear model uses one-dimensional convolutional layers. (Default one-dimensional convolution kernel size) 7) Smooth the historical N frames of multidimensional temporal feature data to obtain the trend component. The historical N frames of multi-dimensional time-series feature data and trend components are combined. The noise component is obtained by subtraction. Then, the trend components and noise components After being processed by a linear layer, the corresponding output components are output. and Through trainable weights (output components) and The trainable weight parameters are respectively and Weighted fusion output components and After softmax normalization (multiplying by each output component and then adding), the multidimensional temporal feature data of the future M frames (the fused data) is obtained. For details on the structure of the DWLinear model, please refer to [link / reference needed]. Figure 6 A schematic diagram of the structure of the DWLinear model for the time-series data prediction and anomaly detection method for mining equipment according to the present invention.

[0032] In some embodiments, the training method of the DWLinear model includes: using the Adam optimizer, adding an L2 regularization term and setting weight_decay to 1e-5 to effectively suppress overfitting, and using SmoothL1Loss as the loss function instead of MSELoss to reduce oscillations in the later stages of training.

[0033] Specifically, experiments show that, under otherwise identical conditions, DWLinear reduces MSE by 17.82% compared to DLinear. The larger the sampling interval, the greater the reduction in MSE. Please refer to [link to relevant documentation] for details. Figures 7-12 The following is a partial experimental result illustration.

[0034] In some embodiments, the mechanistic model includes three threshold detection methods: maximum and minimum threshold detection, trend threshold detection, and fluctuation threshold detection. Maximum and minimum threshold detection filters out periods of abnormal values ​​by setting numerical thresholds and specifying the duration of values ​​exceeding the threshold. For details, please refer to [link to relevant documentation]. Figure 13 The diagram shown illustrates the threshold anomaly detection results. Trend threshold detection filters out periods of trend anomalies by setting a rate of change threshold. For details, please refer to [link to relevant documentation]. Figure 14 The diagram shown illustrates the trend anomaly detection results. The fluctuation threshold detection filters out periods of abnormal fluctuation by setting a standard deviation threshold. For details, please refer to [link to relevant documentation]. Figure 15 The diagram shown illustrates the results of the fluctuation anomaly detection.

[0035] In some embodiments, the time-domain to frequency-domain transformation includes Fourier transform and wavelet transform. The Fourier transform converts the time-series dataset into frequency-domain data by performing a fast Fourier transform, sets an amplitude threshold, filters out amplitude anomaly segments, and outputs the anomaly dimension and anomaly category. For details, please refer to [link to relevant documentation]. Figure 16 The diagram shows the Fourier transform anomaly detection results. Wavelet transform converts the time-series dataset into frequency domain data by performing a continuous wavelet transform (CWT). An energy threshold is set to filter out energy anomaly segments, and the anomaly dimension, anomaly category, and anomaly time period are output. Wavelet transform anomaly detection relies on the time-frequency analysis capabilities of continuous wavelet transform (CWT), focusing on energy distribution anomalies of different frequency components. It is particularly adept at capturing transient anomalies in signals that change over time (such as sudden high-frequency disturbances or sudden increases in energy at specific frequencies), and is more suitable for detecting dynamic anomalies in non-stationary signals. For details, please refer to [link to relevant documentation]. Figure 17 The diagram shown illustrates the wavelet transform anomaly detection results.

[0036] In some embodiments, unsupervised anomaly detection employs the Isolation Forest algorithm (suitable for situations where new devices lack anomaly data and it is difficult to determine the anomaly threshold, or where certain special models of devices have unique detection mechanisms), including a training phase and an inference phase, wherein, as Figure 18As shown, during the training phase, the time-series dataset is converted into sample features using a moving average (i.e., converting each time-series segment within a sliding window into a sample). If the number of sensors is too large, resulting in high feature dimensionality, PCA principal component analysis can be performed to reduce the dimensionality of the sample features. A binary search tree is then built using randomly selected samples (without replacement) and feature dimensions. This process is repeated multiple times to create an isolated forest containing 200 trees. The default anomaly ratio is 0.01. For details, please refer to [link to relevant documentation]. Figure 19 The example diagram for creating an isolated forest is shown below. During the inference phase, the time-series dataset undergoes sample feature transformation and is then input into the isolated forest. Anomaly scores are calculated through hierarchical traversal. The formula for the anomaly score is: (d is the average tree depth), range of outlier scores When the anomaly score exceeds a set threshold (e.g., 0.5), it is judged as an anomalous sample; otherwise, it is a normal sample. The anomalous time period and anomalous dimension are output. In addition, the Isolation Forest algorithm cannot determine the specific category of the anomaly, but can only determine the anomalous time period and anomalous dimension (if a moving average was performed before, the time period needs to be calculated in reverse based on the number of consecutive samples). Isolation Forest is suitable for scenarios with a lack of anomalous data and high real-time requirements. The inference time can be effectively controlled by limiting the number of trees and the maximum tree depth.

[0037] The time-series data prediction and anomaly detection method for mining equipment according to embodiments of the present invention has the following advantages: This invention addresses the pain points of time-series data processing for mining equipment, bringing several significant benefits: First, it boasts low cost and high adaptability, eliminating the need for manual labeling of large amounts of data (the prediction task automatically divides training data and labels, and anomaly detection uses unsupervised or mechanistic models). The isolated forest algorithm adapts to scenarios with new equipment and no anomaly data, enabling reuse of similar equipment (for example, if a mine purchases 10 wind turbines of the same brand and model this month, with turbine 001 already installed and operational, and the other 9 turbines to be installed in one month, data can be collected on turbine 001 first to train the isolated forest algorithm. When the other 9 turbines are installed one month later, the isolated forest model trained on turbine 001 can be reused; however, once sufficient historical data is collected on the other 9 turbines, it is preferable to train each turbine individually to ensure inference accuracy), reducing model migration costs. Simultaneously, the DWLinear multidimensional time-series data prediction model overcomes the limitations of a single dimension, enabling simultaneous... This model predicts multi-dimensional monitoring data such as vibration, temperature, current / voltage, and pressure / flow rate, improving prediction accuracy without increasing computational load. By combining the Adam optimizer, L2 regularization (weight_decay=1e-5), and the SmoothL1 Loss function, it reduces the MSE by 17.82% compared to the DLinear model. The lightweight model can be directly deployed on edge mining equipment, eliminating reliance on high cloud computing power. Furthermore, by stitching together historical N frames and future M frames of predicted data to form a complete time-series set, it achieves collaborative optimization of "prediction-detection," enabling the detection of past anomalies and early warning of potential future faults, overcoming the response lag problem caused by traditional module separation. It also ensures data quality through a unified feature layer scheme of "sliding window + dimension-specific feature extraction," avoiding the loss of high-frequency information. More importantly, it integrates mechanistic models, time-domain-frequency transformation, and unsupervised anomaly detection methods, covering explicit known anomalies, latent frequency anomalies, and unknown novel faults, improving the comprehensiveness and reliability of detection.

[0038] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the method described above.

[0039] The computer-readable medium may be included in the apparatus, device, or system disclosed herein, or it may exist independently.

[0040] The computer-readable storage medium can be any tangible medium that contains or stores a program, and can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, optical fibers, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0041] The computer-readable storage medium may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof.

[0042] Parameter description: All parameter values ​​mentioned in this manual are not subjective assumptions, but rather a comprehensive result of the application scenario's security / efficiency requirements, industry standards and specifications, and industry practice experience thresholds. In actual applications, the parameters will be fine-tuned according to the relevant scenarios.

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for time-series data prediction and anomaly detection in mining equipment, characterized in that, Includes the following steps: Raw, multi-frequency, multi-dimensional historical time-series data of mining equipment are collected and acquired through multiple heterogeneous sensors. By adopting a unified feature layer scheme, the frequency of the sensor with the lowest sampling frequency is used as the reference output frequency. The corresponding sliding window value is set according to the ratio between the sampling frequency of multiple high sampling frequency sensors and the reference output frequency. Data features are extracted in each sliding window to obtain multi-dimensional time series feature data of the same frequency. By inputting historical N-frame multi-dimensional time series feature data from the same frequency multi-dimensional time series feature data at the same time interval into the DWLinear model for time series prediction, the future M-frame multi-dimensional time series feature data is output. A complete time-series dataset is formed by concatenating the historical N-frame multidimensional time-series feature data with the future M-frame multidimensional time-series feature data. Anomalies are detected in the time-series dataset using a mechanistic model, time-domain-frequency domain transformation, and unsupervised anomaly detection methods. Anomaly information is identified and output, wherein the anomaly information includes anomaly dimension, anomaly time period, and anomaly type.

2. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The original multi-frequency historical time-series data includes monitoring data for vibration, temperature, current / voltage, and pressure / flow, with different sampling frequencies for each dimension.

3. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The data features include general features and dimension-specific features; the general features are the mean and standard deviation; the dimension-specific features include: vibration sensor: peak factor, FFT main frequency energy ratio and autocorrelation coefficient (1st lag); current sensor: RMS value, harmonic distortion rate and second-order difference extremum; pressure sensor: first-order difference variance, static offset and moving mean change rate; temperature sensor: raw data and one-dimensional slope.

4. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The DWLinear model uses a one-dimensional convolutional layer. The trend component is obtained by smoothing the historical N frames of multidimensional temporal feature data. The historical N frames of multidimensional time-series feature data and trend components are combined. The noise component is obtained by subtraction. Then the trend components and the noise component After being processed by a linear layer, the corresponding output components are output. and The output components are fused using trainable weights. and The multidimensional temporal feature data of the future M frames is obtained after softmax normalization.

5. The method for time-series data prediction and anomaly detection of mining equipment according to claim 4, characterized in that, The training method for the DWLinear model includes: using the Adam optimizer, adding an L2 regularization term and setting weight_decay to 1e-5, and using SmoothL1Loss as the loss function.

6. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The mechanism model includes three threshold detection methods: maximum and minimum threshold detection, trend threshold detection, and fluctuation threshold detection. The maximum and minimum threshold detection filters out abnormal value periods by setting a numerical threshold and the time that the value exceeds the threshold. The trend threshold detection filters out abnormal trend periods by setting a rate of change threshold. The fluctuation threshold detection filters out abnormal fluctuation periods by setting a standard deviation threshold.

7. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The time-domain to frequency-domain transformation includes Fourier transform and wavelet transform. The Fourier transform converts the time-series dataset into frequency-domain data by performing a fast Fourier transform, sets an amplitude threshold, filters out amplitude anomaly segments, and outputs the anomaly dimension and the anomaly category. The wavelet transform converts the time-series dataset into frequency-domain data by performing a continuous wavelet transform (CWT), sets an energy threshold, filters out energy anomaly segments, and outputs the anomaly dimension, the anomaly category, and the anomaly time period.

8. The method for time-series data prediction and anomaly detection of mining equipment according to claim 1, characterized in that, The unsupervised anomaly detection employs the Isolation Forest algorithm, comprising a training phase and an inference phase. During the training phase, the time-series dataset is converted into sample features using a moving average (when N+M is too large). A binary search tree is then built by randomly selecting samples and feature dimensions. This process is repeated multiple times to form an Isolation Forest containing 200 trees, with a default anomaly ratio of 0.

01. During the inference phase, the time-series dataset is converted into sample features and then input into the Isolation Forest. An anomaly score is calculated using a level-order traversal. When the anomaly score exceeds a set threshold, the sample is identified as an anomaly, and the anomaly time period and anomaly dimension are output.

9. An electronic device, characterized in that, include: One or more processors; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the time-series data prediction and anomaly detection method for mining equipment according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it can implement the time-series data prediction and anomaly detection method for mining equipment according to any one of claims 1 to 8.

Images