Equipment fault prediction method based on industrial big data

Abstract

The invention discloses an equipment fault prediction method based on industrial big data, and belongs to the technical field of equipment fault prediction, and the method comprises the following steps: S1, obtaining historical operation data and real-time monitoring data of industrial equipment, and generating a standardized data set; s2, analyzing the quantitative proportion of fault samples and normal samples in the standardized data set to form a balanced data set; s3, dividing the balanced data set into a training set and a test set, training a plurality of preset base models in parallel through the training set, and evaluating the weight of each trained base model through the test set; s4, constructing an integrated prediction model according to the weight of the trained base model, and outputting a fault probability and a type identifier; and S5, when the fault probability exceeds a preset early warning threshold value, generating an alarm signal and a maintenance suggestion. By generating the simulated fault data, the problem of data imbalance is effectively solved, and the prediction sensitivity of key fault modes such as rare faults is improved.

Description

Technical Field

[0001] This invention relates to the field of equipment failure prediction technology, and in particular to a method for equipment failure prediction based on industrial big data. Background Technology

[0002] Equipment failure prediction refers to the technical field of building predictive models based on historical and real-time monitoring data generated during the operation of industrial equipment, using data mining, machine learning, and statistical analysis to identify potential equipment failure states, assess the probability and type of failure, and provide support for maintenance decisions. Its core objective is to reduce unplanned downtime, optimize the allocation of maintenance resources, and improve the reliability and safety of production systems through early warning and equipment health management.

[0003] The working principle of existing technologies typically includes: collecting operating parameters through sensor arrays (such as temperature, vibration, and current sensors) deployed in industrial equipment and transmitting them to a central database for storage; performing basic preprocessing (such as missing value imputation and normalization) on historical and real-time data to generate a standardized dataset; directly using this dataset to train a fault prediction model (such as support vector machine or neural network), and triggering an alarm when the model outputs a fault probability exceeding a threshold; the entire process relies on manual intervention for data quality control and model adjustment, and often adopts a single model architecture.

[0004] The shortcomings of existing technologies are mainly reflected in the following aspects: First, there is a serious imbalance problem at the data level, with a very small number of fault samples (especially rare faults), resulting in large training bias in the model and insufficient sensitivity in predicting key fault modes; second, the model lacks a dynamic optimization mechanism, and the weight allocation is static, making it unable to adapt to real-time data fluctuations. Therefore, there is an urgent need to provide a method for equipment fault prediction based on industrial big data to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the prior art, namely, the serious imbalance at the data level, the extremely small number of fault samples (especially rare faults), which leads to large model training bias and insufficient sensitivity in predicting key fault modes; and the lack of dynamic optimization mechanism at the model level, with static weight allocation, which cannot adapt to real-time data fluctuations. This invention provides a method for predicting equipment faults based on industrial big data.

[0006] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a method for predicting equipment failure based on industrial big data, comprising the following steps:

[0007] S1. Acquire historical operating data and real-time monitoring data of industrial equipment according to a preset sampling period, and preprocess the historical operating data and real-time monitoring data to generate a standardized dataset;

[0008] S2. Analyze the ratio of fault samples to normal samples in the standardized dataset. When the number of fault samples is lower than a preset threshold, extract the feature distribution pattern of the fault samples and generate simulated fault data based on the normal samples to form a balanced dataset.

[0009] S3. Divide the balanced dataset into a training set and a test set, train multiple preset base models in parallel using the training set, and evaluate the weights of each trained base model using the test set.

[0010] S4. Construct an integrated prediction model based on the weights of the trained base model, input the real-time monitoring data into the integrated prediction model, and output the fault probability and type identifier;

[0011] S5. When the fault probability exceeds the preset warning threshold, an alarm signal and maintenance suggestion are generated, and the real-time monitoring data, the fault probability and the type identifier are archived to the preset historical operation database.

[0012] The present invention is further configured such that: the historical operating data and the real-time monitoring data in step S1 are collected in real time by a sensor group deployed on the industrial equipment to obtain operating parameters such as temperature, vibration and current, and transmitted to a preset central database.

[0013] The present invention is further configured such that the generation step of the standardized dataset in step S1 is as follows:

[0014] S11. Receive the historical operating data and the real-time monitoring data collected in real time by the sensor group deployed on the industrial equipment, perform timestamp synchronization and format consistency verification on the historical operating data and the real-time monitoring data, and generate an initial multi-source data stream;

[0015] S12. Detect and process missing values ​​and outliers in the initial multi-source data stream to generate an intermediate processing data stream, and normalize the intermediate processing data stream to generate a standardized dataset.

[0016] The present invention is further configured such that: the criterion for distinguishing between the fault samples and the normal samples in step S2 is: based on the historical operating data, a dynamic threshold boundary is set for the key parameters in the historical operating parameters; when the real-time monitoring data in the standardized dataset exceeds the dynamic threshold boundary for multiple consecutive preset sampling periods, or when the matching degree with the feature waveform in the preset fault mode library is higher than the preset similarity threshold, the data within the preset sampling period is marked as a fault sample.

[0017] Real-time monitoring data whose preset sampling period does not exceed the dynamic threshold boundary or whose matching degree with the feature waveform in the preset fault mode library is lower than the preset similarity threshold and is within the stable operating range of industrial equipment are classified as normal samples.

[0018] The present invention is further configured such that the formation steps of the balanced dataset in step S2 are as follows:

[0019] S21. Analyze the ratio of the number of fault samples to the number of normal samples in the standardized dataset. When the number of fault samples is lower than a preset threshold, extract the feature distribution pattern of the fault samples in the key historical operating parameters, and generate the distribution profile of the fault samples according to the feature distribution pattern. Based on the distribution profile, perform spatial transformation on the normal samples to generate simulated fault data points.

[0020] S22. The simulated fault data points are fused with the original normal samples to form a preliminary balanced data pool. The preliminary balanced data pool is subjected to stratified sampling and sequence rearrangement. The rearranged preliminary balanced data pool is divided according to a preset time window based on the time axis to generate a balanced dataset.

[0021] The ratio of faulty samples to normal samples in the balanced dataset reaches a preset threshold, and the sample type and source timestamp of each data item in the balanced dataset are marked.

[0022] The present invention is further configured such that the step of generating the simulated fault data points in step S21 is as follows:

[0023] S211. Extract historical operating data of temperature, vibration and current operating parameters from the fault sample, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form a first statistical characteristic set, and generate a first time series variation feature set by analyzing the long-term trend slope and periodic fluctuation law of each operating parameter. Merge the first statistical characteristic set and the first time series variation feature set to construct the feature distribution pattern of the fault sample. Based on the discrete data points in the feature distribution pattern, use a preset algorithm to fit the spatial boundary and generate the distribution profile of the fault sample.

[0024] S212. Based on the real-time monitoring data of the operating parameters of temperature, vibration and current in the normal sample, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form a second statistical characteristic set. By analyzing the short-term fluctuation period and long-term degradation slope of each operating parameter, generate a second time-series change feature set. Merge the second statistical characteristic set and the second time-series change feature set to generate the characteristic distribution pattern of the normal sample. Then, perform density estimation on the characteristic distribution pattern to generate the distribution profile of the normal sample.

[0025] S213. The distribution contours of the faulty samples and the distribution contours of the normal samples are superimposed and aligned in a preset feature space to generate a superimposed distribution contour map. The adjacent point pairs of the distribution contours of the faulty samples and the distribution contours of the normal samples in the superimposed distribution contour map are identified. The Euclidean distance of each adjacent point pair is calculated. The contour line segments formed by the set of adjacent point pairs whose Euclidean distance is less than a preset matching threshold are selected as candidate contour line segments.

[0026] S214. The contour segments of the faulty samples in the candidate contour segments are marked as the first contour segment, and the contour segments of the normal samples are marked as the second contour segment. The first contour segment and the second contour segment are segmented according to a preset interpolation ratio, and the first spatial coordinates and the second spatial coordinates of each segmentation point generated during the segmentation are extracted in the preset feature space.

[0027] S215. Map the second spatial coordinates to the feature subspace where the first spatial coordinates are located, and assign the transformed second spatial coordinates to the random disturbance factor derived from the first statistical characteristic set of the fault sample to generate simulated fault data points.

[0028] The present invention is further configured such that the division rules for the training set and the test set in step S3 are as follows: First, the balanced dataset is sorted in chronological order according to the source timestamp of each data item in the balanced dataset; the faulty samples and the normal samples are stratified according to the sample type label; the faulty samples and the normal samples are randomly selected from each stratum according to a preset ratio to form a training set candidate pool and a test set candidate pool; time series overlap detection is performed on the training set candidate pool, and data items that have time window overlap with the test set candidate pool are removed to generate the training set; a preset sample distribution balance check is performed on the test set candidate pool to generate the test set.

[0029] The preset base model includes a time series prediction model and a classification model;

[0030] The specific content of step S3, which involves training multiple preset base models in parallel using the training set and evaluating the weights of each trained base model using the test set, is as follows:

[0031] S31. The training set is input into the time series prediction model and the classification model for parallel training. The time series prediction model analyzes the timestamp sequence of each data item in the training set and the periodic fluctuation pattern of the operating parameters to establish a prediction function for the evolution trajectory of the equipment state and generate a time series prediction result set. The classification model extracts the feature difference distribution contours of the fault samples and the normal samples in the training set, constructs a classification decision function in the preset feature space, and generates a classification result set. The internal parameters of the trained time series prediction model and the classification model are optimized respectively to generate an optimized base model set.

[0032] S32. Input the test set into the optimized base model set, obtain the first prediction result set of each base model for the test set, compare the first prediction result set with the true label of the test set to generate deviation distribution data, calculate the initial accuracy and false alarm rate of each base model based on the deviation distribution data, and convert the initial accuracy and false alarm rate into initial weight coefficients according to a preset weight initialization rule, calculate the number of accurate identifications of fault samples and the number of false alarms of normal samples for each base model based on the initial weight coefficients, generate a model performance parameter set, and generate a weight correction factor by fusing the number of accurate identifications and the number of false alarms through a dynamic weighting algorithm according to the model performance parameter set, and adjust the initial weight coefficients according to the weight correction factor to generate the weights of each base model.

[0033] The present invention is further configured such that the construction step of the integrated prediction model in step S4 is as follows:

[0034] S401. Based on the weights of each base model and the model performance parameter set, a dynamic weight allocation mechanism is constructed. The dynamic weight allocation mechanism adjusts the contribution ratio of each base model according to the characteristic fluctuation amplitude of the real-time monitoring data, generates a weight configuration scheme, and performs weighted fusion on the optimized base model set according to the weight configuration scheme to construct a primary integrated model framework.

[0035] S402. Input the real-time monitoring data into the primary integrated model framework, obtain the independent prediction results of each base model for the real-time monitoring data, perform probability fusion calculation based on the independent prediction results and the weight configuration scheme, generate a preliminary fault probability distribution, calibrate the confidence level of the preliminary fault probability distribution, and correct the preset probability threshold in combination with the feature distribution pattern of the fault sample to generate the calibrated fault probability and type identifier.

[0036] S403. Bind the calibrated fault probability and type identifier with the real-time monitoring data to generate a second prediction result set, verify the second prediction result set, and generate an integrated prediction model.

[0037] The present invention is further configured such that: the specific content of step S5 is as follows: real-time monitoring of the fault probability output by the integrated prediction model; when the fault probability exceeds a preset warning threshold, an alarm signal generation process is automatically triggered; targeted maintenance suggestions are generated by matching the type identifier with a pre-stored fault handling knowledge base; a preset data archiving thread is simultaneously started to timestamp and bind the real-time monitoring data, the fault probability and the type identifier within the current time window, and store them in a specific partition of the historical operation database after compression.

[0038] The beneficial effects of this invention are as follows:

[0039] 1. This invention extracts the feature distribution patterns of fault samples and generates simulated fault data based on normal samples to form a balanced dataset, effectively solving the problem of data imbalance and significantly improving the prediction sensitivity and recognition accuracy of key fault patterns such as rare faults.

[0040] 2. This invention dynamically evaluates the base model weights using a test set and constructs an integrated prediction model, thereby achieving adaptive adjustment of model weights. This overcomes the limitations of static weight allocation mechanisms and significantly enhances the model's adaptability to real-time data fluctuations and its predictive robustness. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is a flowchart illustrating the steps involved in forming the balanced dataset according to the present invention.

[0043] Figure 3 This is a flowchart illustrating the specific content of step S3 of the present invention. Detailed Implementation

[0044] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0045] Please see Figure 1 - Figure 3 The equipment failure prediction method based on industrial big data includes the following steps:

[0046] S1. Acquire historical operating data and real-time monitoring data of industrial equipment according to the preset sampling period, and preprocess the historical operating data and real-time monitoring data to generate a standardized dataset;

[0047] S2. Analyze the ratio of fault samples to normal samples in the standardized dataset. When the number of fault samples is lower than a preset threshold, extract the feature distribution pattern of the fault samples and generate simulated fault data based on normal samples to form a balanced dataset.

[0048] S3. Divide the balanced dataset into a training set and a test set. Train multiple pre-defined base models in parallel using the training set, and evaluate the weights of each trained base model using the test set.

[0049] S4. Construct an integrated prediction model based on the weights of the trained base model, input real-time monitoring data into the integrated prediction model, and output the fault probability and type identifier.

[0050] S5. When the probability of a fault exceeds the preset warning threshold, an alarm signal and maintenance suggestions are generated, and the real-time monitoring data, fault probability and type identifier are archived to the preset historical operation database.

[0051] Steps S1-S5 ensure data consistency by generating a standardized dataset, address sample imbalance by using a balanced dataset, optimize model robustness by training base models in parallel and evaluating weights, construct an integrated prediction model to improve fault identification accuracy, and achieve a closed loop of real-time early warning and data archiving, significantly improving prediction reliability and reducing unplanned downtime.

[0052] One embodiment of the present invention is as follows: the historical operating data and real-time monitoring data in step S1 are collected in real time by a sensor group deployed on industrial equipment to obtain operating parameters such as temperature, vibration and current, and transmitted to a preset central database.

[0053] The sensor group includes temperature sensors, vibration sensors, and current sensors, used to collect multi-dimensional operating parameters of industrial equipment in real time.

[0054] Preferably, the steps for generating the standardized dataset in step S1 are as follows:

[0055] S11. Receive historical operating data and real-time monitoring data collected in real time by sensor groups deployed on industrial equipment, perform timestamp synchronization and format consistency verification on historical operating data and real-time monitoring data, and generate an initial multi-source data stream.

[0056] S12. Detect and process missing and outlier values ​​in the initial multi-source data stream to generate an intermediate processing data stream. Normalize the intermediate processing data stream to generate a standardized dataset.

[0057] The specific content of detecting and processing missing values ​​and outliers in the initial multi-source data stream in step S12 is as follows: using a preset linear interpolation algorithm or time series prediction method, the missing values ​​are dynamically filled according to the numerical trend of adjacent data points; using statistical methods such as Z-score calculation or interquartile range method to identify data points that exceed the reasonable range, and using mean replacement or regression model-based correction strategies to correct them, so as to ensure the integrity and reliability of the data stream.

[0058] This embodiment collects multi-dimensional operating parameters in real time using temperature, vibration, and current sensors. After timestamp synchronization, format verification, missing value processing, and normalization, a high-quality standardized dataset is generated to ensure data consistency and integrity, providing a reliable foundation for fault prediction.

[0059] One embodiment of the present invention is as follows: the criterion for distinguishing fault samples from normal samples in step S2 is: based on historical operating data, a dynamic threshold boundary is set for key parameters (such as vibration acceleration, temperature gradient, and current harmonic distortion) in historical operating parameters. When the real-time monitoring data in the standardized dataset exceeds the dynamic threshold boundary for multiple consecutive preset sampling periods, or when the matching degree with the characteristic waveforms (such as specific frequency components of bearing wear or current leakage mode of insulation aging) in the preset fault mode library is higher than the preset similarity threshold, the data within this preset sampling period is marked as a fault sample.

[0060] Setting dynamic threshold boundaries for key parameters: Based on the parameter distribution characteristics under normal equipment conditions in historical operating data, the moving average and standard deviation of each parameter (such as vibration acceleration, temperature gradient, and current harmonic distortion) are calculated using a sliding time window statistical method. The basic threshold boundary is "mean ± 3 times standard deviation", and the equipment aging factor and operating load coefficient are introduced for dynamic correction. Among them, the vibration acceleration threshold needs to be adaptively adjusted in combination with the equipment speed spectrum characteristics, the temperature gradient threshold is compensated in relation to the rate of change of ambient temperature, and the current harmonic distortion threshold is dynamically calibrated according to the grid voltage fluctuation, forming an elastic threshold boundary that is updated in real time with the equipment operating status.

[0061] The optimal preset sampling period is as follows: for rapidly changing vibration parameters, high-frequency sampling of 10ms-100ms is used to capture impact characteristics; for slowly changing temperature parameters, medium-frequency sampling of 1s-10s is used to track thermal inertia effects; and for current parameters, sampling is set to integer multiples of the power frequency period of 20ms-200ms, taking into account the power frequency characteristics. Dynamic adjustments are made for different fault types: bearing wear monitoring needs to cover harmonics more than 32 times the shaft rotation frequency, so the sampling period should not be greater than 1 / 64 of the rotation period; insulation aging monitoring needs to capture the cumulative effect of leakage current, so the sampling period can be extended to 1-5 minutes, and the complete capture of fault characteristics is achieved through multi-rate sampling fusion.

[0062] Preset Fault Mode Library: The preset fault mode library is constructed by integrating historical equipment fault records, simulation data, and physical models. It includes a characteristic waveform template library of typical faults (such as the impact response waveform of bearing inner ring faults and the stepped leakage current waveform of insulation aging), fault evolution trajectory models (such as the mapping relationship between wear depth and vibration harmonic energy), and fault association rules (such as the composite fault mode of sudden temperature rise accompanied by current harmonic growth). Each fault mode in the library is labeled with the fault cause, severity level, and corresponding maintenance strategy. It is continuously iterated and optimized through actual operation data to form a dynamic knowledge base with self-learning capabilities.

[0063] The preset similarity threshold adopts a multi-level judgment mechanism. For waveform matching degree calculation, dynamic time warping algorithm (DTW) or cosine similarity algorithm is used. Differentiated thresholds are set according to fault type: sudden faults (such as bearing breakage) require a similarity threshold ≥0.85 to ensure alarm timeliness, while progressive faults (such as insulation aging) are set with a threshold ≥0.70 to reduce the risk of missed alarms. At the same time, a confidence calibration mechanism is introduced. When the data quality indicators (such as signal-to-noise ratio >20dB) meet the standard, the threshold strictness is automatically increased, and the threshold level is dynamically adjusted in combination with the historical false alarm rate to achieve a balance between accuracy and robustness.

[0064] Real-time monitoring data whose preset sampling period does not exceed the dynamic threshold boundary or whose matching degree with the feature waveform in the preset fault mode library is lower than the preset similarity threshold and is within the stable operating range of industrial equipment are classified as normal samples.

[0065] Preferably, the steps for forming the balanced dataset in step S2 are as follows:

[0066] S21. Analyze the ratio of fault samples to normal samples in the standardized dataset. When the number of fault samples is lower than the preset threshold, extract the feature distribution pattern of fault samples in key historical operating parameters, and generate the distribution profile of fault samples based on the feature distribution pattern. Based on the distribution profile, perform spatial transformation on normal samples to generate simulated fault data points.

[0067] Preferably, the step of generating simulated fault data points in step S21 is as follows:

[0068] S211. Extract historical operating data of temperature, vibration and current operating parameters from the fault sample, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form the first statistical characteristic set, and analyze the long-term trend slope and periodic fluctuation law of each operating parameter through dynamic time warping algorithm to generate the first time series change feature set. Integrate the first statistical characteristic set and the first time series change feature set to construct the feature distribution pattern of the fault sample. Based on the discrete data points in the feature distribution pattern, use a preset algorithm (the preset algorithm is the convex hull algorithm) to fit the spatial boundary and generate the distribution contour of the fault sample.

[0069] The specific steps for calculating the mean, variance, and extreme values ​​of each operating parameter in the time domain to form the first statistical characteristic set are as follows: Based on the historical operating data sequences of temperature, vibration, and current operating parameters in the fault sample, the historical operating parameter subsequences are divided according to time windows. The arithmetic mean of each subsequence is calculated as the mean parameter to measure the central tendency of the data. The average squared difference between each subsequence data point and the mean parameter is calculated as the variance parameter to characterize the degree of data dispersion. The maximum and minimum values ​​in each subsequence are extracted as extreme value parameters to reflect the range of parameter fluctuations. The mean, variance, and extreme values ​​of each operating parameter are integrated into a multi-dimensional vector according to parameter type to form the first statistical characteristic set. The first statistical characteristic set serves as the basic data source for subsequent time series feature analysis.

[0070] The steps for generating the first time-series variation feature set are as follows: Based on the parameter distribution provided by the first statistical characteristic set, the time series of each operating parameter are aligned using a dynamic time warping algorithm, the trend slope between the series is calculated to capture the long-term change direction, and the periodic fluctuation pattern of the series is analyzed to identify the inherent oscillation mode; the weighted average of the trend slope and the covariance matrix of the fluctuation period are extracted to generate the first time-series variation feature set characterizing the dynamic evolution of the parameters. The first time-series variation feature set and the first statistical characteristic set together constitute the complete input for fault feature analysis.

[0071] The steps for constructing the feature distribution pattern of fault samples are as follows: First, the first set of statistical characteristics and the first set of temporal variation features are integrated. The mean, variance, and extreme values ​​in the statistical parameters are concatenated with the trend slope and fluctuation period in the temporal parameters using multi-dimensional vectors. Principal component analysis is used to reduce dimensionality and remove redundant information, retaining the principal components with the highest contribution as feature vectors. Based on these feature vectors, a scatter plot of the fault samples is constructed. A continuous probability distribution surface is generated using kernel density estimation, forming the feature distribution pattern of the fault samples. This feature distribution pattern intuitively reflects the clustering regions and diffusion patterns of fault parameters in the feature space.

[0072] S212. Based on real-time monitoring data of operating parameters such as temperature, vibration and current in normal samples, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form a second statistical characteristic set. Analyze the short-term fluctuation period and long-term degradation slope of each operating parameter through sliding window Fourier transform to generate a second time-series change feature set. Merge the second statistical characteristic set and the second time-series change feature set to generate the characteristic distribution pattern of normal samples. Use a Gaussian mixture model to estimate the density of the characteristic distribution pattern and generate the distribution profile of normal samples.

[0073] S213. Overlay and align the distribution contours of faulty samples and normal samples in a preset feature space to generate an overlay distribution contour map. Identify adjacent point pairs of the distribution contours of faulty samples and normal samples in the overlay distribution contour map, calculate the Euclidean distance of each adjacent point pair, and select the contour line segments formed by the set of adjacent point pairs with Euclidean distance less than a preset matching threshold as candidate contour line segments.

[0074] Preset matching threshold: Dynamically set based on the superposition of the distribution contours of fault samples and normal samples. By calculating the percentile of the Euclidean distance distribution of all adjacent point pairs in historical operation data, the 15th percentile is taken as the lower limit of the threshold to ensure sensitivity. At the same time, the upper limit of the threshold is adjusted in combination with the severity coefficient of the fault type to avoid overfitting. For high-risk fault modes, the threshold is appropriately relaxed to the 20th percentile to reduce the false negative rate, while for common faults, it is strictly controlled within the 10th percentile to improve accuracy.

[0075] S214. Mark the contour segments of faulty samples in the candidate contour segments as the first contour segment and the contour segments of normal samples as the second contour segment. Segment the first contour segment and the second contour segment according to the preset interpolation ratio, and extract the first spatial coordinates and the second spatial coordinates of each segmentation point generated during segmentation in the preset feature space.

[0076] Preset interpolation ratio: The preset interpolation ratio is adaptively determined based on the length and curvature of the candidate contour segments. For contour segments whose length is greater than 20% of the average segment length, a uniform interpolation ratio of 0.25 is used to ensure segmentation uniformity. For high curvature segments, a dynamic interpolation ratio of 0.15 is used to preserve contour details. The selection of the interpolation ratio must also consider the distribution overlap between faulty samples and normal samples. When the overlap is higher than 60%, the ratio is adjusted to 0.30 to enhance the diversity of generated data. When the overlap is lower than 30%, the ratio is reduced to 0.10 to maintain structural stability.

[0077] S215. The second spatial coordinates are mapped to the feature subspace where the first spatial coordinates are located by the affine transformation algorithm, and the transformed second spatial coordinates are assigned to the random disturbance factor derived from the first statistical characteristic set of the fault samples to generate simulated fault data points that conform to the fault evolution law.

[0078] The process of deriving the random disturbance factor is as follows: the ratio of the extreme values ​​of the variances of each operating parameter is extracted from the first set of statistical characteristics as the basic disturbance coefficient, the variation exponent of the mean of the operating parameters is calculated as the amplitude adjustment factor, and a random number sequence conforming to a Gaussian distribution is generated through Monte Carlo simulation; the random number sequence is convolved with the basic disturbance coefficient and then superimposed with the amplitude adjustment factor to generate a random disturbance factor sequence. After affine transformation, the random disturbance factor is injected into spatial coordinates to simulate the natural fluctuation characteristics of fault data, ensuring that the generated data retains the original statistical regularity and has reasonable randomness.

[0079] S22. The simulated fault data points are fused with the original normal samples to form a preliminary balanced data pool. The preliminary balanced data pool is subjected to stratified sampling and sequence rearrangement. Based on the time axis, the rearranged preliminary balanced data pool is divided according to a preset time window to generate a time-series coherent balanced dataset.

[0080] When the ratio of faulty samples to normal samples in the balanced dataset reaches a preset threshold, the sample type (faulty sample or normal sample) and source timestamp of each data item in the balanced dataset are marked.

[0081] Preferably, during the generation of the balanced dataset, the quality of the balanced dataset is continuously optimized by dynamically adjusting the generation parameters of the simulated fault data, so as to ensure that the training data can effectively cover various states of equipment operation, especially rare but critical fault modes.

[0082] Example: Taking industrial bearing fault prediction as an example, based on historical vibration acceleration data, a sliding window is used to calculate the moving average and standard deviation. The dynamic threshold boundary is set as "mean ± 3 times standard deviation" and a bearing aging factor is introduced for correction. When the real-time monitoring data exceeds the threshold for three consecutive sampling periods (10ms per period) or the DTW matching degree with the inner ring wear waveform in the fault mode library is ≥0.85, the data in that period is marked as a fault sample. If the number of fault samples is insufficient, the first statistical characteristic set (mean, variance, extreme value) and the first time series change feature set (trend slope, fluctuation period) of its vibration parameters are extracted, and the fault distribution profile is generated by the convex hull algorithm. The normal samples generate distribution profiles through a Gaussian mixture model. After superposition, the Euclidean distance between the profiles is calculated. Candidate line segments with a distance less than the 15th quantile are selected. The coordinate points are divided according to the 0.25 interpolation ratio. The simulated fault data points are generated by superimposing a random disturbance factor through affine transformation. Finally, the data is merged with the normal samples to form a balanced dataset with a fault sample ratio of 30%.

[0083] This example accurately captures bearing fault features through dynamic thresholding and multi-pattern matching, and generates physically reasonable simulation data by using contour overlay and coordinate transformation, effectively solving the problem of model bias caused by sample scarcity. Adaptive interpolation and perturbation mechanisms enhance data diversity and significantly improve the prediction sensitivity of sudden wear faults (reducing the false negative rate by 20%). At the same time, threshold calibration balances the risk of false alarms, enabling the model to maintain more than 95% classification robustness in imbalanced data scenarios.

[0084] One embodiment of the present invention is as follows: the rules for dividing the training set and the test set in step S3 are as follows: First, the balanced dataset is sorted in chronological order according to the source timestamp of each data item in the balanced dataset. Based on the sample type label, fault samples and normal samples are stratified separately. Fault samples and normal samples are randomly selected from each stratum according to a preset ratio to form the training set candidate pool and the test set candidate pool. Time series overlap detection is performed on the training set candidate pool, and data items that have time window overlap with the test set candidate pool are removed to generate the training set. The sample distribution balance of the test set candidate pool is checked by a preset method to ensure that it covers representative data of all fault types and normal states to generate the test set.

[0085] Preset ratio: Based on the initial ratio of faulty samples to normal samples in the balanced dataset, the ratio of training set to test set is set to 9:1 when the proportion of faulty samples is less than 5% to enhance small sample learning; when the proportion is between 5% and 20%, an 8:2 ratio is used to balance generalization ability; and when the proportion is greater than 20%, a 7:3 standard ratio is used.

[0086] The pre-defined base models include time series prediction models and classification models;

[0087] The specific content of step S3, which involves training multiple pre-defined base models in parallel using the training set and evaluating the weights of each trained base model using the test set, is as follows:

[0088] S31. Input the training set into the time series prediction model and the classification model for parallel training. The time series prediction model analyzes the timestamp sequence of each data item in the training set and the periodic fluctuation pattern of the operating parameters to establish a prediction function for the equipment state evolution trajectory and generate a time series prediction result set. The classification model extracts the feature difference distribution contours between fault samples and normal samples in the training set, constructs a classification decision function in the preset feature space, and generates a classification result set. Optimize the internal parameters of the trained time series prediction model and the classification model respectively to generate an optimized base model set.

[0089] The steps for establishing the prediction function for the equipment state evolution trajectory are as follows: Based on the data sequence sorted by timestamp in the training set, extract the gradient change pattern of the frequency domain energy distribution of the vibration signal and the temperature sequence as the temporal feature vector, map the feature vector to the transition probability matrix of the hidden Markov chain through the state space model, derive the state transition equation by maximum likelihood estimation, and finally generate a prediction function that can describe the state degradation path by combining the physical constraints of the equipment.

[0090] The classification decision function in the preset feature space is defined as follows: the multidimensional running parameters in the training set are mapped to the high-dimensional feature space through the kernel function, and the Mahalanobis distance between the center points of fault samples and normal samples is calculated in this space. The hyperplane decision boundary is constructed by the distance ratio. The output of the decision function is the signed distance from the sample to the hyperplane. Positive values ​​are classified as fault samples and negative values ​​are classified as normal samples. The threshold is dynamically calibrated by the principle of maximizing the area under the ROC curve.

[0091] The specific content of optimizing the internal parameters of the trained time series prediction model and classification model is as follows: the time series prediction model uses the backpropagation algorithm to adjust the weight coefficients of the state transition matrix, with the goal of minimizing the root mean square of the rolling prediction error; the classification model updates the Lagrange multipliers of the support vectors through the sequence minimization optimization algorithm, with the sum of the interval loss function and the regularization term as the optimization index; an early stopping mechanism is introduced in the optimization process to prevent overfitting.

[0092] S32. Input the test set into the optimized base model set, obtain the first prediction result set of each base model on the test set, compare the first prediction result set with the true label of the test set, generate deviation distribution data, calculate the initial accuracy and false alarm rate of each base model based on the deviation distribution data, and convert the initial accuracy and false alarm rate into initial weight coefficients according to the preset weight initialization rules. Based on the initial weight coefficients, calculate the number of accurate identifications of fault samples and the number of false alarms of normal samples for each base model, generate a model performance parameter set, and generate a weight correction factor by fusing the number of accurate identifications and the number of false alarms through a dynamic weighting algorithm according to the model performance parameter set. Adjust the initial weight coefficients according to the weight correction factor to generate the weights of each base model, bind the final weights to the first prediction result set of the base model, generate a weight evaluation report, and feed it back to the sampling strategy optimization process of the training set.

[0093] The calculation steps for the initial accuracy and false alarm rate of each base model are as follows: The initial accuracy is calculated as the percentage of correctly predicted samples in the test set out of the total number of samples, which is the sum of true positives and true negatives divided by the total number of samples multiplied by 100%; the false alarm rate is calculated as the percentage of normal samples that are incorrectly predicted as faults out of the total number of normal samples, which is the false positive divided by the sum of true negatives and false positives multiplied by 100%; the calculation results are rounded to three decimal places to ensure accuracy.

[0094] The preset weight initialization rule is as follows: the basic weight coefficient is set according to the difference between the initial accuracy and the false alarm rate. When the accuracy is higher than 95% and the false alarm rate is lower than 5%, a basic weight of 0.7 or higher is assigned. When the accuracy is in the range of 80% to 95%, the weight is calculated by linear interpolation. At the same time, a model complexity penalty factor is introduced to moderately decay the weight of the multilayer network model.

[0095] The calculation steps for the number of accurately identified faulty samples and the number of false positives for normal samples are as follows: The number of accurately identified samples is obtained directly from the true positive count of the confusion matrix, which is the actual number of faulty samples correctly predicted by the model; the number of false positives is obtained from the false positive count of the confusion matrix, which is the number of normal samples that the model incorrectly identifies as faulty; both types of numbers need to be averaged after three independent verifications.

[0096] The specific content of generating a weight correction factor by fusing the number of accurate identifications and the number of false alarms through a dynamic weighted algorithm is as follows: the number of accurate identifications is used as the positive weight factor, and the number of false alarms is used as the negative weight factor to construct a weighted summation formula; where the positive weight factor is multiplied by the accuracy confidence coefficient, the negative weight factor is multiplied by the false alarm severity coefficient, and the final correction factor is the log-normalized value of the ratio of the weighted sum to the total sample size.

[0097] The specific calculation formula for adjusting the initial weight coefficients based on the weight correction factor is: the adjusted weight is equal to the initial weight multiplied by e raised to the power of the correction factor, and then divided by the sum of the adjusted weights of all base models.

[0098] Taking the fault prediction of a wind turbine gearbox as an example, the proportion of fault samples in its balanced dataset is 4%. Based on the preset ratio, the training set and test set are divided into 9:1. The time-series prediction model analyzes the frequency domain energy distribution of gearbox vibration signals and the bearing temperature gradient change pattern, establishes a state transition equation to generate a prediction result set, and the classification model maps the operating parameters to a high-dimensional feature space through a kernel function to construct the Mahalanobis distance decision boundary to generate a classification result set. After internal parameter optimization, the model achieves an initial accuracy of 93.5% and a false alarm rate of 3.8% on the test set. The initial weight coefficient is calculated to be 0.74 according to the weight initialization rule. The correction factor is generated by fusing the number of accurately identified fault samples and the number of false alarms of normal samples through a dynamic weighting algorithm. Finally, the weight is adjusted to 0.78 and bound to the prediction result set to generate a weight evaluation report, which is fed back to the training set sampling strategy optimization process.

[0099] This example significantly improves the predictive ability for low-proportion fault samples (4%) through dynamic proportional partitioning and weight optimization mechanisms. The model accuracy reaches 93.5% and the false alarm rate is controlled at 3.8%, effectively solving the problem of missed alarms caused by sample imbalance. The weight correction factor enhances the model's adaptability to real-time data fluctuations, providing a highly reliable basis for equipment maintenance decisions and reducing the risk of unplanned downtime by more than 20%.

[0100] One embodiment of the present invention is as follows: the construction step of the integrated prediction model in step S4 is as follows:

[0101] S401. Based on the weights and model performance parameter sets of each base model, a dynamic weight allocation mechanism is constructed. The dynamic weight allocation mechanism adjusts the contribution ratio of each base model according to the characteristic fluctuation amplitude of real-time monitoring data, generates a weight configuration scheme, and performs weighted fusion on the optimized base model set according to the weight configuration scheme to construct a primary integrated model framework.

[0102] The construction steps of the dynamic weight allocation mechanism are as follows: Based on the weights of each base model and the model performance parameter set, the characteristic fluctuation amplitude of the data is monitored in real time and quantified by calculating the variance and rate of change of the running parameters. A dynamic adjustment factor is generated according to the fluctuation amplitude. The dynamic adjustment factor is combined with the initial weights of the base models, and the contribution ratio of each model under the current data conditions is calculated by the weighting algorithm to form a weight configuration scheme, ensuring that the model weights adapt to the data characteristics.

[0103] Weight Configuration Scheme: The weight configuration scheme is output by a dynamic weight allocation mechanism, which clarifies the weight allocation ratio of each base model in the integration. The weight configuration scheme is dynamically adjusted based on the fluctuation range of real-time data. For example, the weight of the time series prediction model is increased during the high fluctuation phase, and the weight of the classification model is enhanced during the stable phase. The weight configuration scheme is directly used to perform weighted fusion on the optimized set of base models to build a primary integration model framework.

[0104] S402. Input the real-time monitoring data into the primary integrated model framework, obtain the independent prediction results of each base model for the real-time monitoring data, perform probability fusion calculation based on the independent prediction results and weight configuration scheme, generate a preliminary fault probability distribution, calibrate the confidence of the preliminary fault probability distribution, and correct the preset probability threshold by combining the feature distribution pattern of the fault samples, and generate the calibrated fault probability and type identifier.

[0105] Preliminary Fault Probability Distribution: Real-time monitoring data is input into the primary ensemble model framework to obtain the independent prediction results of each base model; based on the weight configuration scheme, the independent prediction results are probabilistically fused and calculated, and a preliminary fault probability distribution is generated using a weighted average or probability superposition method; this distribution represents the comprehensive probability level of fault occurrence and provides a basic input for subsequent calibration;

[0106] Preset probability threshold: The preset probability threshold serves as the initial boundary for fault determination. The initial value is set by analyzing the characteristic distribution patterns of historical fault samples. During the confidence calibration stage, the probability threshold is dynamically corrected by combining the distribution characteristics of real-time data and the fault evolution law to ensure that the probability threshold can both capture early fault signals and control the risk of false alarms.

[0107] S403. Bind the calibrated fault probability and type identifier with the real-time monitoring data to generate a second prediction result set. Perform logical consistency verification on the second prediction result set to ensure that the matching relationship between the fault type identifier and the fault probability value conforms to the preset fault evolution law, and finally generate an integrated prediction model.

[0108] This embodiment uses a dynamic weight allocation mechanism to enable the model to adapt to data fluctuations, probability fusion and confidence calibration to improve fault prediction accuracy, and logical consistency verification to ensure that the output conforms to the fault evolution law, thus significantly enhancing the robustness and reliability of the integrated model.

[0109] One embodiment of the present invention is as follows: Step S5 specifically involves: real-time monitoring of the fault probability output by the integrated prediction model; when the fault probability exceeds a preset warning threshold, an alarm signal generation process is automatically triggered; targeted maintenance suggestions are generated by matching the pre-stored fault handling knowledge base with the type identifier; a preset data archiving thread is simultaneously started to timestamp and bind the real-time monitoring data, fault probability, and type identifier within the current time window, compress them, and store them in a specific partition of the historical operation database; based on the fault probability evolution pattern of this alarm event and the handling effect records of similar events in the historical operation database, the sensitivity parameters of the warning threshold are dynamically optimized and the maintenance strategy of the fault handling knowledge base is updated, forming a closed-loop management link from real-time warning, decision support to knowledge iteration.

[0110] This embodiment monitors the probability of failure in real time and automatically triggers alarms. It generates accurate maintenance suggestions by combining type identifier matching knowledge base, synchronizes archived data, and dynamically optimizes thresholds and strategies based on evolutionary patterns to form closed-loop management, which significantly improves fault response speed, maintenance accuracy and system self-optimization capabilities.

[0111] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting equipment failures based on industrial big data, characterized in that: Includes the following steps: S1. Acquire historical operating data and real-time monitoring data of industrial equipment according to a preset sampling period, and preprocess the historical operating data and real-time monitoring data to generate a standardized dataset; S2. Analyze the ratio of fault samples to normal samples in the standardized dataset. When the number of fault samples is lower than a preset threshold, extract the feature distribution pattern of the fault samples and generate simulated fault data based on the normal samples to form a balanced dataset. S3. Divide the balanced dataset into a training set and a test set, train multiple preset base models in parallel using the training set, and evaluate the weights of each trained base model using the test set. S4. Construct an integrated prediction model based on the weights of the trained base model, input the real-time monitoring data into the integrated prediction model, and output the fault probability and type identifier; S5. When the fault probability exceeds the preset warning threshold, an alarm signal and maintenance suggestion are generated, and the real-time monitoring data, the fault probability and the type identifier are archived to the preset historical operation database.

2. The equipment fault prediction method based on industrial big data according to claim 1, characterized in that: The historical operating data and the real-time monitoring data in step S1 are obtained by real-time acquisition of operating parameters such as temperature, vibration and current through sensor groups deployed on industrial equipment, and transmitted to a preset central database.

3. The equipment fault prediction method based on industrial big data according to claim 2, characterized in that: The steps for generating the standardized dataset in step S1 are as follows: S11. Receive the historical operating data and the real-time monitoring data collected in real time by the sensor group deployed on the industrial equipment, perform timestamp synchronization and format consistency verification on the historical operating data and the real-time monitoring data, and generate an initial multi-source data stream; S12. Detect and process missing values ​​and outliers in the initial multi-source data stream to generate an intermediate processing data stream, and normalize the intermediate processing data stream to generate a standardized dataset.

4. The equipment fault prediction method based on industrial big data according to claim 3, characterized in that: The criteria for distinguishing between fault samples and normal samples in step S2 are as follows: Based on the historical operating data, a dynamic threshold boundary is set for the key parameters in the historical operating parameters. When the real-time monitoring data in the standardized dataset exceeds the dynamic threshold boundary for multiple consecutive preset sampling periods, or when the matching degree with the feature waveform in the preset fault mode library is higher than the preset similarity threshold, the data within this preset sampling period is marked as a fault sample. Real-time monitoring data whose preset sampling period does not exceed the dynamic threshold boundary or whose matching degree with the feature waveform in the preset fault mode library is lower than the preset similarity threshold and is within the stable operating range of industrial equipment are classified as normal samples.

5. The equipment fault prediction method based on industrial big data according to claim 4, characterized in that: The steps for forming the balanced dataset in step S2 are as follows: S21. Analyze the ratio of the number of fault samples to the number of normal samples in the standardized dataset. When the number of fault samples is lower than a preset threshold, extract the feature distribution pattern of the fault samples in the key historical operating parameters, and generate the distribution profile of the fault samples according to the feature distribution pattern. Based on the distribution profile, perform spatial transformation on the normal samples to generate simulated fault data points. S22. The simulated fault data points are fused with the original normal samples to form a preliminary balanced data pool. The preliminary balanced data pool is subjected to stratified sampling and sequence rearrangement. The rearranged preliminary balanced data pool is divided according to a preset time window based on the time axis to generate a balanced dataset. The ratio of faulty samples to normal samples in the balanced dataset reaches a preset threshold, and the sample type and source timestamp of each data item in the balanced dataset are marked.

6. The equipment fault prediction method based on industrial big data according to claim 5, characterized in that: The step of generating the simulated fault data points in step S21 is as follows: S211. Extract historical operating data of temperature, vibration and current operating parameters from the fault sample, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form a first statistical characteristic set, and generate a first time series variation feature set by analyzing the long-term trend slope and periodic fluctuation law of each operating parameter. Merge the first statistical characteristic set and the first time series variation feature set to construct the feature distribution pattern of the fault sample. Based on the discrete data points in the feature distribution pattern, use a preset algorithm to fit the spatial boundary and generate the distribution profile of the fault sample. S212. Based on the real-time monitoring data of the operating parameters of temperature, vibration and current in the normal sample, calculate the mean, variance and extreme values ​​of each operating parameter in the time domain to form a second statistical characteristic set. By analyzing the short-term fluctuation period and long-term degradation slope of each operating parameter, generate a second time-series change feature set. Merge the second statistical characteristic set and the second time-series change feature set to generate the characteristic distribution pattern of the normal sample. Then, perform density estimation on the characteristic distribution pattern to generate the distribution profile of the normal sample. S213. The distribution contours of the faulty samples and the distribution contours of the normal samples are superimposed and aligned in a preset feature space to generate a superimposed distribution contour map. The adjacent point pairs of the distribution contours of the faulty samples and the distribution contours of the normal samples in the superimposed distribution contour map are identified. The Euclidean distance of each adjacent point pair is calculated. The contour line segments formed by the set of adjacent point pairs whose Euclidean distance is less than a preset matching threshold are selected as candidate contour line segments. S214. The contour segments of the faulty samples in the candidate contour segments are marked as the first contour segment, and the contour segments of the normal samples are marked as the second contour segment. The first contour segment and the second contour segment are segmented according to a preset interpolation ratio, and the first spatial coordinates and the second spatial coordinates of each segmentation point generated during the segmentation are extracted in the preset feature space. S215. Map the second spatial coordinates to the feature subspace where the first spatial coordinates are located, and assign the transformed second spatial coordinates to the random disturbance factor derived from the first statistical characteristic set of the fault sample to generate simulated fault data points.

7. The equipment fault prediction method based on industrial big data according to claim 6, characterized in that: The rules for dividing the training set and the test set in step S3 are as follows: First, the balanced dataset is sorted chronologically according to the source timestamp of each data item in the balanced dataset. Based on the sample type label, the faulty samples and the normal samples are stratified separately. Faulty samples and normal samples are randomly selected from each stratum according to a preset ratio to form a training set candidate pool and a test set candidate pool. Time series overlap detection is performed on the training set candidate pool, and data items that have time window overlap with the test set candidate pool are removed to generate the training set. The sample distribution balance of the test set candidate pool is checked according to a preset ratio to generate the test set. The preset base model includes a time series prediction model and a classification model; The specific content of step S3, which involves training multiple preset base models in parallel using the training set and evaluating the weights of each trained base model using the test set, is as follows: S31. The training set is input into the time series prediction model and the classification model for parallel training. The time series prediction model analyzes the timestamp sequence of each data item in the training set and the periodic fluctuation pattern of the operating parameters to establish a prediction function for the evolution trajectory of the equipment state and generate a time series prediction result set. The classification model extracts the feature difference distribution contours of the fault samples and the normal samples in the training set, constructs a classification decision function in the preset feature space, and generates a classification result set. The internal parameters of the trained time series prediction model and the classification model are optimized respectively to generate an optimized base model set. S32. Input the test set into the optimized base model set, obtain the first prediction result set of each base model for the test set, compare the first prediction result set with the true label of the test set to generate deviation distribution data, calculate the initial accuracy and false alarm rate of each base model based on the deviation distribution data, and convert the initial accuracy and false alarm rate into initial weight coefficients according to a preset weight initialization rule, calculate the number of accurate identifications of fault samples and the number of false alarms of normal samples for each base model based on the initial weight coefficients, generate a model performance parameter set, and generate a weight correction factor by fusing the number of accurate identifications and the number of false alarms through a dynamic weighting algorithm according to the model performance parameter set, and adjust the initial weight coefficients according to the weight correction factor to generate the weights of each base model.

8. The equipment fault prediction method based on industrial big data according to claim 7, characterized in that: The steps for constructing the integrated prediction model in step S4 are as follows: S401. Based on the weights of each base model and the model performance parameter set, a dynamic weight allocation mechanism is constructed. The dynamic weight allocation mechanism adjusts the contribution ratio of each base model according to the characteristic fluctuation amplitude of the real-time monitoring data, generates a weight configuration scheme, and performs weighted fusion on the optimized base model set according to the weight configuration scheme to construct a primary integrated model framework. S402. Input the real-time monitoring data into the primary integrated model framework, obtain the independent prediction results of each base model for the real-time monitoring data, perform probability fusion calculation based on the independent prediction results and the weight configuration scheme, generate a preliminary fault probability distribution, calibrate the confidence level of the preliminary fault probability distribution, and correct the preset probability threshold in combination with the feature distribution pattern of the fault sample to generate the calibrated fault probability and type identifier. S403. Bind the calibrated fault probability and type identifier with the real-time monitoring data to generate a second prediction result set, verify the second prediction result set, and generate an integrated prediction model.

9. The equipment fault prediction method based on industrial big data according to claim 8, characterized in that: The specific content of step S5 is as follows: real-time monitoring of the fault probability output by the integrated prediction model; when the fault probability exceeds the preset warning threshold, an alarm signal generation process is automatically triggered; and targeted maintenance suggestions are generated by matching the pre-stored fault handling knowledge base with the type identifier. A preset data archiving thread is started synchronously, and the real-time monitoring data, the fault probability and type identifier within the current time window are timestamped, compressed and stored in a specific partition of the historical operation database.

Images