Industrial time series data processing method, system and equipment

Abstract

The invention discloses an industrial time series data processing method, system and device, and relates to the technical field of industrial time series data processing, and the method comprises the steps: carrying out the time-space alignment and fusion of multi-source heterogeneous sensor time series data, and then carrying out the hybrid adaptive noise reduction processing, and obtaining the time series data after noise reduction; carrying out data segmentation by adopting a self-adaptive window function to obtain an initial data sample; identifying an equipment working condition corresponding to the initial data sample, and performing standardization processing on the initial data sample based on the equipment working condition to obtain a standardized data sample; classifying and identifying the standardized data sample to obtain a preset rare fault category sample and other category samples; and enhancing a preset rare fault category sample based on the generative adversarial network, and combining the enhanced fault sample with other categories of samples to obtain a balanced training data set. The technical problem that in the prior art, industrial time sequence data cannot meet the training requirement of a high-precision prediction model is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial time series data processing, in particular to an industrial time series data processing method, system and device. BACKGROUND

[0003] In the traditional industrial equipment predictive maintenance practice, the data processing flow is usually rough, and the following technical problems and challenges exist universally: 1. Single data segmentation scale: The traditional method is to use a fixed-length sliding window to segment continuous time series data to generate samples for model input. However, the running state of the equipment is dynamically changing, which includes both long-term and slow performance degradation process and short-term and sudden fault impact. The fixed window cannot consider both features of different time scales: if the window is too large, the instantaneous fault impact signal may be smoothed out, resulting in feature ambiguity; if the window is too small, the long-term degradation trend of the equipment cannot be effectively captured, causing information loss.

[0004] 2. Difficulty in multi-source data fusion: Modern industrial equipment is usually equipped with multiple types of sensors, such as accelerometers for measuring vibration, thermocouples for measuring equipment surface temperature, optical encoders for measuring main shaft speed, and pressure sensors for measuring load, etc. These sensor data sources are diverse, with different physical meanings, dimensions, sampling frequencies, and data formats. For example, the sampling rate of vibration signals can reach tens of kilohertz, while the temperature signal may only be collected once per second. Simply concatenating these data or using conventional resampling methods (such as linear interpolation or nearest neighbor interpolation) often introduces noise and causes data distortion, making it difficult to achieve true information fusion and fully reflect the comprehensive state of the equipment at a certain time.

[0005] 3. Serious industrial noise interference: The industrial production site environment is harsh, with a lot of electromagnetic interference, mechanical vibration and production process noise. These noises will superimpose on the original sensor signals, seriously affecting the signal-to-noise ratio of the data. Traditional filters, such as mean filtering, median filtering or Gaussian filtering, can filter out some random noise, but they are not effective for complex, non-Gaussian background noise and periodic interference from adjacent equipment or power systems. More importantly, these simple filtering methods may damage early weak, non-periodic fault impact signals while filtering out noise, resulting in the loss of critical fault information.

[0006] 4. Extremely scarce fault samples: In actual production, the equipment is in normal operation most of the time, and the occurrence of faults, especially early faults of specific types, is a small probability event. This leads to a large number of normal samples and extremely few fault samples (especially different types of fault samples) in the collected data set. This serious data imbalance problem will cause the machine learning model to tend to "ignore" the minority class of fault samples during training, resulting in a very low recognition rate of rare faults, poor generalization ability, and difficulty in effectively warning in actual application.

[0007] 5. Difficulty in separating the influence of variable working conditions: Large industrial equipment, such as wind turbines, numerical control machine tools, and rolling mills, usually has dynamic changes in operating conditions (such as speed, load, feed rate, etc.). Changes in working conditions will directly cause significant shifts in the statistical characteristics of sensor data (such as vibration amplitude, equipment temperature). If all data is uniformly standardized without distinction (for example, global normalization), the data fluctuations caused by changes in working conditions will be mistakenly considered by the model as changes in the health status of the equipment, resulting in a large number of false alarms or fault misjudgments, seriously affecting the reliability of the predictive maintenance system. SUMMARY

[0008] The purpose of the present application is to provide an industrial time series data processing method, system and device to solve at least one of the above technical problems.

[0009] In a first aspect, an embodiment of the present application provides an industrial time series data processing method, comprising: performing spatio-temporal alignment and fusion on multi-source heterogeneous sensor time series data from at least one industrial equipment to obtain multi-dimensional feature vector time series data; performing mixed adaptive noise reduction processing on the multi-dimensional feature vector time series data to obtain noise-reduced time series data; based on the statistical characteristics of the noise-reduced time series data, using an adaptive window function to segment the data to obtain initial data samples; identifying the equipment working conditions corresponding to the initial data samples, and performing standardization processing on the initial data samples based on the equipment working conditions to obtain standardized data samples; performing classification identification on the standardized data samples to obtain preset rare fault class samples and other class samples except the preset rare fault class samples; based on a generative adversarial network, enhancing the preset rare fault class samples to generate enhanced fault samples, and merging the enhanced fault samples with the other class samples to obtain an balanced training data set.

[0010] Optionally, the multi-source heterogeneous sensor time series data includes high-frequency vibration sensor data stream, low-frequency temperature sensor data stream, and speed sensor data stream.

[0011] Optionally, spatiotemporal alignment and fusion are performed on time-series data from multi-source heterogeneous sensors from at least one industrial device to obtain multi-dimensional feature vector time-series data. This includes: using a network time protocol server as a reference, performing timestamp correction on the multi-source heterogeneous sensor time-series data to obtain timestamp-corrected sampled data; the timestamp-corrected sampled data includes high-sampling-frequency data and low-sampling-frequency data; downsampling and aligning the high-sampling-frequency data to the low-sampling-frequency data using a weighted average method based on timestamp proximity to obtain time-aligned multi-channel sensor data; and vectorizing the time-aligned multi-channel sensor data to obtain multi-dimensional feature vector time-series data at the same timestamp.

[0012] Optionally, the multidimensional feature vector time-series data is subjected to hybrid adaptive denoising processing to obtain denoised time-series data, including: performing discrete wavelet transform on the high-frequency signal components in the multidimensional feature vector time-series data, and using a soft thresholding method in the wavelet domain to remove Gaussian white noise to obtain preliminary denoised data; inputting the preliminary denoised data into a trained deep autoencoder for filtering out non-Gaussian background noise and periodic interference to obtain denoised time-series data; wherein the deep autoencoder is trained based on reconstructed healthy device signals, and the training objective is to minimize the difference between the reconstructed signal and the original healthy signal.

[0013] Optionally, based on the statistical characteristics of the time-series data after denoising, an adaptive window function is used to segment the data to obtain initial data samples, including: continuously calculating the information entropy of the time-series data after denoising based on a sliding preset benchmark window; if the information entropy is lower than a preset threshold, then segmenting the time-series data after denoising based on a first preset analysis window; if the information entropy is higher than or equal to the preset threshold, then segmenting the time-series data after denoising based on a second preset analysis window to obtain initial data samples; wherein, the first preset analysis window is larger than the second preset analysis window.

[0014] Optionally, identifying the equipment operating conditions corresponding to the initial data sample and standardizing the initial data sample based on the equipment operating conditions to obtain standardized data samples includes: dividing the historical operating condition data of the at least one industrial device into multiple operating condition zones based on an unsupervised clustering algorithm; determining the operating condition zone to which each sample in the initial data sample belongs; calling the statistical parameters corresponding to each operating condition zone to perform Z-score standardization on the initial data sample to obtain standardized data samples; the statistical parameters include the mean and standard deviation.

[0015] Optionally, classifying and identifying the standardized data samples includes: classifying and identifying the standardized data samples based on a pre-trained classifier.

[0016] Optionally, the preset rare fault category samples are enhanced based on the generative adversarial network to generate enhanced fault samples, including: using the preset rare fault category samples as seed samples and inputting them into the generator of the generative adversarial network to generate enhanced fault samples that are similar in feature distribution but have differences.

[0017] Secondly, embodiments of the present invention also provide an industrial time-series data processing system for implementing the industrial time-series data processing method provided in the embodiments of the present invention; the system includes: an alignment and fusion module, a noise reduction module, a segmentation module, a standardization module, a classification and recognition module, and an enhancement module; wherein, the alignment and fusion module is used to perform spatiotemporal alignment and fusion on time-series data from multi-source heterogeneous sensors from at least one industrial device to obtain multi-dimensional feature vector time-series data; the noise reduction module is used to perform hybrid adaptive noise reduction processing on the multi-dimensional feature vector time-series data to obtain noise-reduced time-series data; the segmentation module is used to, based on the statistical characteristics of the noise-reduced time-series data, adopt... The data is segmented using an adaptive window function to obtain initial data samples. The standardization module is used to identify the equipment operating conditions corresponding to the initial data samples and to standardize the initial data samples based on the equipment operating conditions to obtain standardized data samples. The classification and recognition module is used to classify and recognize the standardized data samples to obtain preset rare fault category samples and other category samples besides the preset rare fault category samples. The enhancement module is used to enhance the preset rare fault category samples based on a generative adversarial network to generate enhanced fault samples, and to merge the enhanced fault samples with the other category samples to obtain a balanced training dataset.

[0018] Thirdly, embodiments of the present invention also provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0019] This invention provides a method, system, and device for processing industrial time-series data. First, it performs spatiotemporal alignment and fusion of collected multi-source heterogeneous sensor data to ensure data consistency. Second, it uses a hybrid adaptive denoising module, combining wavelet transform and deep autoencoder, to denoise the fused data, improving the signal-to-noise ratio. Third, it employs an adaptive window function based on information entropy to segment the denoised data to capture features at different scales. Subsequently, considering the impact of varying equipment operating conditions, it dynamically standardizes the segmented data to remove operating condition interference. Finally, it utilizes a generative adversarial network to augment rare fault samples, addressing the data imbalance problem. This invention, through multi-stage serial processing, significantly improves the signal-to-noise ratio, spatiotemporal consistency, and sample balance of the data, providing a solid data foundation for subsequent high-precision prediction model training. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a method for processing industrial time-series data provided in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the working principle of a generative adversarial network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an industrial time-series data processing system provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1 Figure 1 This is a flowchart of a method for processing industrial time-series data according to an embodiment of the present invention. Figure 1 As shown, the method specifically includes the following steps: Step S102: Spatiotemporal alignment and fusion of time-series data from multi-source heterogeneous sensors from at least one industrial device are performed to obtain multi-dimensional feature vector time-series data.

[0024] Step S104: Perform hybrid adaptive denoising processing on the multidimensional feature vector time series data to obtain the denoised time series data.

[0025] Step S106: Based on the statistical characteristics of the time series data after noise reduction, an adaptive window function is used to segment the data to obtain the initial data sample.

[0026] Step S108: Identify the equipment operating conditions corresponding to the initial data samples, and standardize the initial data samples based on the equipment operating conditions to obtain standardized data samples.

[0027] Step S110: Classify and identify the standardized data samples to obtain preset rare fault category samples and other category samples besides the preset rare fault category samples.

[0028] Step S112: Enhance the preset rare fault category samples based on the generative adversarial network to generate enhanced fault samples, and merge the enhanced fault samples with other category samples to obtain a balanced training dataset.

[0029] Specifically, step S102 aims to address the inconsistency in sampling rate and time reference of time-series data from multiple sensors, providing a unified spatiotemporal reference for subsequent data fusion and analysis.

[0030] First, data is acquired through a data acquisition interface. This interface supports various communication protocols commonly used in industrial settings, such as OPC-UA, Modbus, and TCP / IP, enabling flexible access to sensors and data acquisition devices from different manufacturers and of different types. In this embodiment of the invention, the multi-source heterogeneous sensor time-series data includes at least a high-frequency vibration sensor data stream (e.g., an accelerometer with a sampling rate of 25.6 kHz), a low-frequency temperature sensor data stream (e.g., a thermocouple with a sampling rate of 1 Hz), and a speed sensor data stream.

[0031] Specifically, step S102 further includes the following steps: Step S1021: Using the network time protocol server as a reference, perform timestamp correction on the time series data of multi-source heterogeneous sensors to obtain timestamp-corrected sampling data; the timestamp-corrected sampling data includes high sampling frequency data and low sampling frequency data.

[0032] Specifically, in industrial settings, the local clocks of different data acquisition units may deviate, necessitating unification to a common time base. This embodiment of the invention employs a high-precision Network Time Protocol (NTP) server as the time synchronization center for the entire system. While acquiring data, all data acquisition units attach an NTP-synchronized timestamp to each data point or data packet. After the data enters the processing system, the system recalibrates the timestamps of all data points to ensure the consistency of the time base.

[0033] Step S1022: The high sampling frequency data is downsampled and aligned to the low sampling frequency data using a weighted average method based on the proximity of timestamps, to obtain time-aligned multi-channel sensor data.

[0034] Specifically, after unifying the time base, the issue of different sampling rates needs to be addressed. Directly upsampling low-frequency data through interpolation introduces spurious information, while directly downsampling high-frequency data results in the loss of detail. This invention employs a more faithful alignment method: using the timestamp sequence of the sensor with the lowest sampling frequency (e.g., a temperature sensor) as a reference, all higher-frequency data are aligned to these timestamps. The specific alignment method uses a weighted average method. For example, to align vibration data to the timestamp t of a certain temperature data point... temp Above, the system will search for a value containing t. temp Time window [t] temp −Δt,t temp All vibration data points within [+Δt]. Then, a weighted average is calculated for the values ​​of these vibration data points, as expressed by the formula: In the formula, v aligned This represents the aligned vibration data, v(t) i ) is the value of the i-th vibration data point within the time window, and its weight w i Based on its timestamp t i With target timestamp t temp The degree of proximity can be determined, for example, by using a Gaussian function to calculate the weights: Here, σ is a hyperparameter that controls the rate of weight decay. This method ensures that the aligned data points better reflect the main trends within the time period, rather than simply taking a certain instantaneous value, thereby reducing information distortion.

[0035] Step S1023: Vectorize the time-aligned multi-channel sensor data to obtain multi-dimensional feature vector time-series data under the same timestamp.

[0036] Specifically, after time alignment, the data from all sensors are on a common time axis. At this point, all sensor data values ​​at the same timestamp t can be combined into a multidimensional feature vector. For example, at time t, the fused multidimensional feature vector time series data X(t) can be represented as: X(t) = [v aligned (t),T(t),p(t),…] Among them, v aligned X(t) is the aligned vibration value, T(t) is the temperature value, and p(t) is the rotational speed value. This multidimensional feature vector time series data X(t) completely describes the overall state of the device at that moment and serves as the input for all subsequent processing steps.

[0037] Specifically, step S104 further includes the following steps: Step S1041: Perform discrete wavelet transform on the high-frequency signal components in the multidimensional feature vector time series data, and remove Gaussian white noise in the wavelet domain using the soft thresholding method to obtain preliminary denoised data. Step S1042: Input the preliminary denoising data into the trained deep autoencoder to filter out non-Gaussian background noise and periodic interference, and obtain the denoised time series data. The deep autoencoder is trained based on the reconstructed healthy device signal, and the training objective is to minimize the difference between the reconstructed signal and the original healthy signal.

[0038] Specifically, to filter out strong noise from the industrial site in the multidimensional feature vector time series data X(t) while preserving the weak signal features related to early faults to the maximum extent, this invention adopts a hybrid noise reduction strategy that combines wavelet transform and deep autoencoder.

[0039] First, the high-frequency signal components (e.g., aligned vibration signals v) in the multidimensional feature vector time series data are analyzed. aligned (t) Perform Discrete Wavelet Transform (DWT). DWT can decompose a signal into multiple scales in the time and frequency domains to obtain wavelet coefficients in different frequency bands. Typically, the energy of Gaussian white noise is evenly distributed across the wavelet coefficients in each frequency band, while the energy of the useful signal is mainly concentrated in a few wavelet coefficients with larger amplitudes. Therefore, a threshold can be set to process the wavelet coefficients. This embodiment of the invention uses the soft thresholding method, the formula of which is: In the formula, c j These are the original wavelet coefficients, ĉ jThese are the processed wavelet coefficients, and λ is the threshold. This method sets wavelet coefficients with amplitudes less than the threshold λ to zero and shrinks wavelet coefficients with amplitudes greater than the threshold λ by λ towards zero. After processing all wavelet coefficients, the signal is reconstructed using inverse discrete wavelet transform (IDWT). This step effectively filters out most of the Gaussian white noise.

[0040] However, industrial noise is often complex, containing non-Gaussian noise and periodic interference (such as 50Hz power frequency interference from the power grid). Traditional wavelet thresholding methods have limited effectiveness in processing such noise. Therefore, this invention also introduces a deep autoencoder (DAE) for deep noise reduction. A DAE is an unsupervised neural network consisting of an encoder and a decoder. Its core idea is that the encoder compresses the input data into a low-dimensional latent representation, while the decoder attempts to perfectly reconstruct the original input from this compressed representation.

[0041] The DAE in this embodiment of the invention is pre-trained on a large amount of data from devices in a "healthy" state. The training objective is to minimize the difference (e.g., mean square error) between the reconstructed signal and the original healthy signal. In this way, the DAE learns the intrinsic structure and patterns of the healthy signal. When a signal containing complex noise and fault impulses is input into this trained DAE, since the DAE is only good at reconstructing the healthy signal patterns it is familiar with, it effectively filters out components that do not conform to the healthy pattern (i.e., complex background noise and periodic interference), while retaining the main body of the healthy signal. More importantly, although early fault impulse signals are also abnormal, their morphology is different from random noise. The DAE can also retain these non-periodic impulse components to a certain extent because they deviate greatly from the learned normal pattern and will not be completely smoothed out during the reconstruction process. Finally, the signal after wavelet transform and DAE two-stage processing has a significantly improved signal-to-noise ratio, laying the foundation for subsequent feature extraction and analysis.

[0042] Specifically, step S106 further includes the following steps: Step S1061: Based on a sliding preset reference window, continuously calculate the information entropy of the time series data after noise reduction.

[0043] Step S1062: If the information entropy is lower than the preset threshold, then the time series data after noise reduction is segmented based on the first preset analysis window.

[0044] Step S1063: If the information entropy is higher than or equal to a preset threshold, the time series data after noise reduction is segmented based on the second preset analysis window to obtain an initial data sample; wherein, the first preset analysis window is larger than the second preset analysis window.

[0045] Specifically, step S106 aims to address the problem that traditional fixed-window segmentation methods cannot simultaneously take into account both long-term trends and instantaneous features.

[0046] This invention uses information entropy as an indicator to measure the stationarity or complexity of data. For a discrete time-series data set, its amplitude is first quantized and binned, and then its information entropy is calculated using the following formula: Where H(X) represents information entropy, n is the number of quantization bins, and p(x) is the number of bins. i ) represents the probability that a data point falls into the i-th bin. The higher the information entropy, the greater the uncertainty, the more volatile the data, or the richer the information it contains; conversely, the lower the information entropy, the more stable and regular the data.

[0047] The specific process is as follows: 1. Calculate the sliding information entropy: On the time series data after noise reduction, set a sliding preset baseline window (e.g., the window length is 1024 data points). The preset baseline window slides over the entire time series with a fixed step size, and the information entropy of the data within the preset baseline window is calculated in real time, thus obtaining a time series of information entropy.

[0048] 2. Threshold Judgment: Set an information entropy threshold. This information entropy threshold can be obtained by statistically analyzing the information entropy of a large amount of data under healthy operating conditions (e.g., taking the mean plus three standard deviations). When the information entropy calculated by the sliding preset benchmark window is consistently lower than the information entropy threshold, the device is judged to be currently operating stably.

[0049] 3. Large Window Segmentation: During stable equipment operation, a large analysis window (i.e., the first preset analysis window mentioned above, for example, 4096 data points) is used for data segmentation. The purpose of using a large window is to capture long-term, slow performance degradation trends of the equipment, such as the slow increase in vibration energy caused by bearing wear.

[0050] 4. Small Window Segmentation: When the calculated information entropy suddenly exceeds the information entropy threshold, it is determined that the data may contain a sudden impact signal or a state change (e.g., a bearing crack causing a momentary impact). At this time, it immediately and automatically switches to a smaller analysis window (i.e., the second preset analysis window mentioned above, for example, 512 data points). Using a small window allows for sampling of the fault area at a higher time resolution, thereby accurately capturing the complete waveform of the instantaneous fault characteristics and avoiding smoothing.

[0051] Through this adaptive window switching mechanism, the present invention cleverly achieves a balance between features at different time scales. The final output segmented data samples contain both macroscopic degradation information and microscopic fault details.

[0052] Specifically, step S108 further includes the following steps: Step S1081: Based on the unsupervised clustering algorithm, the historical operating condition data of at least one industrial device is divided into multiple operating condition zones.

[0053] Step S1082: Determine the operating zone to which each sample in the initial data sample belongs.

[0054] Step S1083: Call the statistical parameters corresponding to each working condition zone to perform Z-score standardization on the initial data sample to obtain standardized data samples; the statistical parameters include mean and standard deviation.

[0055] Specifically, step S108 aims to eliminate the impact of changes in equipment operating conditions (such as speed and load) on sensor data, enabling the model to focus on changes in the health status of the equipment itself.

[0056] First, the current operating condition needs to be identified. This embodiment of the invention employs an unsupervised clustering algorithm (such as the K-Means algorithm) to automatically classify operating conditions. Sensor data directly related to the equipment's operating condition, such as rotational speed and load, are received in parallel. After collecting a large amount of historical operating condition data, the K-Means algorithm is run using this data as input. The algorithm divides the data points into k clusters, with each cluster center representing a typical stable operating condition zone. For example, clustering may yield three operating condition zones: Condition 1 (high speed - heavy load), Condition 2 (medium speed - medium load), and Condition 3 (low speed - light load).

[0057] Then, before standardizing the initial data sample, perform the following operations: Operating condition zone determination: Obtain the operating condition data (average speed and load) of each data sample in the initial data sample within the corresponding time period, and determine which operating condition zone cluster center the data point is closest to, thereby determining the operating condition zone to which the data sample belongs.

[0058] Standardization by operating condition: Pre-calculate and store statistical parameters of various sensor data (such as vibration and temperature) during healthy equipment operation in each operating condition zone, mainly the mean μ. k and standard deviation σ k (Where k is the index of the operating zone). Then, call the μ corresponding to the operating zone k to which the data sample belongs. k and σ k The data sample is standardized using the Z-score formula: Where X is the initial data sample, X std It is a standardized data sample.

[0059] In this way, regardless of the equipment's operating conditions, as long as its health status remains unchanged, its standardized data will fall within a similar distribution range. Only when the equipment's health status deteriorates will its data deviate from the normal range for that operating condition, thus exhibiting obvious outliers after standardization. This effectively eliminates the interference from changes in operating conditions.

[0060] Specifically, step S110 includes: classifying and identifying standardized data samples based on a pre-trained classifier.

[0061] For example, all samples in the standardized data sample are initially classified based on a pre-trained classifier, the number of samples in each category is counted, and the fault categories with a number of samples far less than normal samples are identified as preset rare fault categories.

[0062] Specifically, step S112 includes: using a preset rare fault category sample as a seed sample and inputting it into the generator of the generative adversarial network to generate enhanced fault samples that are similar in feature distribution but have differences.

[0063] Figure 2 This is a schematic diagram illustrating the working principle of a generative adversarial network according to an embodiment of the present invention. Figure 2 As shown, the following data augmentation process is performed on each preset rare fault category sample: 1. Generative Adversarial Network (GAN) Model Training: A GAN consists of a generator (G) and a discriminator (D). In this embodiment of the invention, a specific GAN is trained for each preset rare fault category. During training, all real samples of that preset rare fault category are fed into the GAN.

[0064] 2. Sample Generation: The generator G's task is to learn the data distribution of real fault samples. It receives a random noise vector as input and attempts to output a virtual sample that is as similar as possible to a real fault sample. The discriminator D's task is to receive a sample (which may be real or a virtual sample generated by G) and determine its authenticity.

[0065] 3. Adversarial Training: G and D engage in a zero-sum game. G's goal is to generate samples sufficient to "deceive" D, while D's goal is to identify the fake samples generated by G as accurately as possible. Through repeated adversarial training, both G's generation ability and D's discrimination ability will continuously improve. Ultimately, when the training reaches Nash equilibrium, G will be able to generate virtual fault samples that are highly consistent with the feature distribution of real fault samples and are difficult for the human eye or simple classifiers to distinguish.

[0066] 4. Dataset Expansion: Using the trained generator G, a large number of high-quality and diverse virtual fault samples can be generated from random noise. These newly generated samples are then mixed with the original dataset, thereby significantly increasing the number of samples in rare fault categories, forming a final training dataset with a relatively balanced class distribution, i.e., a balanced training dataset.

[0067] This balanced training dataset is then output to subsequent model training modules (e.g., CNNs for fault diagnosis or LSTM models for RUL prediction), which can significantly improve the model's recognition accuracy and robustness when facing real rare faults.

[0068] In summary, the industrial time-series data processing method provided by this invention constructs a complete, advanced, and automated industrial time-series data processing workflow, systematically solving many pain points in existing technologies and providing a solid data foundation for achieving high-precision predictive maintenance. Compared with existing technologies, this invention has the following significant advantages: (1) Improved feature fidelity: By adopting an adaptive window function segmentation method based on information entropy, this invention can intelligently select the scale of the analysis window according to the dynamic characteristics of the data itself. This enables the processed data samples to accurately capture the instantaneous impact characteristics when the fault occurs at high resolution, and to fully display the long-term degradation trend during the stable operation of the equipment at a long scale, with less information loss and higher feature fidelity.

[0069] (2) Ensuring data consistency and comprehensiveness: The spatiotemporal alignment and fusion method proposed in this invention effectively solves the alignment problem of sensor data from different physical meanings and sampling rates in the temporal and spatial dimensions. The multidimensional feature vector formed after fusion can more comprehensively and accurately describe the overall operating status of the device at any given time, providing high-quality input for subsequent multivariate analysis models.

[0070] (3) Significantly improves signal-to-noise ratio: This invention innovatively combines wavelet threshold denoising with deep autoencoder (DAE) denoising. This hybrid method takes into account both Gaussian white noise and complex non-Gaussian noise filtering capabilities. More importantly, through the reconstruction learning of healthy signals, DAE can effectively filter out noise while preserving the non-periodic impact characteristics of early weak fault signals to the maximum extent, providing the possibility of achieving early warning of faults.

[0071] (4) Enhanced model focus on equipment status: By introducing a data standardization method that considers varying operating conditions, this invention successfully removes data drift caused by changes in equipment operating conditions (such as speed and load) from the data. This allows subsequent machine learning models to focus more on learning data patterns caused by real changes in equipment health status, thereby significantly reducing false alarm rates and improving the model's generalization ability and diagnostic accuracy.

[0072] (5) Improve the model's ability to identify rare faults: To address the common problem of insufficient fault samples in industrial applications, this invention utilizes Generative Adversarial Networks (GANs) for data augmentation, generating a large number of high-quality virtual fault samples, effectively solving the class imbalance problem in the training data. This significantly enhances the sufficiency of model training, enabling the final trained model to have higher recognition accuracy and robustness for various faults, especially rare faults.

[0073] In summary, this invention provides an end-to-end, automated industrial time-series data processing solution that systematically solves several key technical bottlenecks in the data processing workflow. It provides a solid, reliable, and high-quality data foundation for subsequent predictive maintenance models and has extremely high engineering application value.

[0074] Example 2 Figure 3 This is a schematic diagram of an industrial time-series data processing system provided according to an embodiment of the present invention. Figure 3 As shown, the system includes: alignment and fusion module 10, noise reduction module 20, segmentation module 30, standardization module 40, classification and recognition module 50, and enhancement module 60.

[0075] Specifically, the alignment and fusion module 10 is used to perform spatiotemporal alignment and fusion of time-series data from multi-source heterogeneous sensors from at least one industrial device to obtain multi-dimensional feature vector time-series data. The noise reduction module 20 is used to perform hybrid adaptive noise reduction processing on multidimensional feature vector time series data to obtain noise-reduced time series data. The segmentation module 30 is used to segment the data based on the statistical characteristics of the time series data after noise reduction, using an adaptive window function to obtain the initial data samples. The standardization module 40 is used to identify the equipment operating conditions corresponding to the initial data samples and to standardize the initial data samples based on the equipment operating conditions to obtain standardized data samples. The classification and recognition module 50 is used to classify and recognize standardized data samples to obtain preset rare fault category samples and other category samples besides the preset rare fault category samples; The enhancement module 60 is used to enhance the preset rare fault category samples based on the generative adversarial network, generate enhanced fault samples, and merge the enhanced fault samples with other category samples to obtain a balanced training dataset.

[0076] Specifically, the alignment and fusion module 10 is also used for: Using a network time protocol server as a reference, time-stamp correction is performed on the time-series data of multi-source heterogeneous sensors to obtain timestamp-corrected sampled data; the timestamp-corrected sampled data includes high-sampling-frequency data and low-sampling-frequency data. High-sampling-frequency data is downsampled and aligned to low-sampling-frequency data using a weighted average method based on the proximity of timestamps, resulting in time-aligned multi-channel sensor data. The time-aligned multi-channel sensor data is vectorized to obtain multi-dimensional feature vector time-series data under the same timestamp.

[0077] Specifically, the noise reduction module 20 is also used for: Discrete wavelet transform is performed on the high-frequency signal components in the multidimensional feature vector time series data, and Gaussian white noise is removed in the wavelet domain using the soft thresholding method to obtain preliminary denoised data; The initial denoised data is input into a trained deep autoencoder to filter out non-Gaussian background noise and periodic interference, resulting in denoised time-series data. The deep autoencoder is trained based on the reconstructed healthy device signal, and the training objective is to minimize the difference between the reconstructed signal and the original healthy signal.

[0078] Specifically, segmentation module 30 is also used for: Based on a sliding preset baseline window, the information entropy of the time series data after noise reduction is continuously calculated; If the information entropy is lower than the preset threshold, the time series data after noise reduction is segmented based on the first preset analysis window. If the information entropy is higher than or equal to the preset threshold, the time series data after noise reduction is segmented based on the second preset analysis window to obtain the initial data sample. The first preset analysis window is larger than the second preset analysis window.

[0079] Specifically, standardized module 40 is also used for: The historical operating data of at least one industrial device is divided into multiple operating zones based on an unsupervised clustering algorithm. Determine the operating zone to which each sample in the initial data sample belongs; The statistical parameters corresponding to each operating condition zone are called to perform Z-score standardization on the initial data sample to obtain standardized data samples; the statistical parameters include the mean and standard deviation.

[0080] Specifically, the classification and recognition module 50 is also used to classify and recognize standardized data samples based on a pre-trained classifier.

[0081] Specifically, the enhancement module 60 is also used to input preset rare fault category samples as seed samples into the generator of the generative adversarial network to generate enhanced fault samples that are similar in feature distribution but have differences.

[0082] The present invention also provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and running on the processor, wherein the processor executes the computer program to implement the method provided in the embodiments of the present invention.

[0083] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0084] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for processing industrial time series data, characterized by, The method comprises the following steps: spatiotemporal alignment and fusion of multi-source heterogeneous sensor time series data from at least one industrial equipment to obtain multi-dimensional feature vector time series data; hybrid adaptive noise reduction processing of the multi-dimensional feature vector time series data to obtain noise-reduced time series data; segmentation of the noise-reduced time series data based on statistical characteristics thereof using an adaptive window function to obtain initial data samples; identification of the equipment working conditions corresponding to the initial data samples and standardization processing of the initial data samples based on the equipment working conditions to obtain standardized data samples; classification and identification of the standardized data samples to obtain preset rare fault category samples and other category samples except the preset rare fault category samples; enhancement of the preset rare fault category samples based on a generative adversarial network to generate enhanced fault samples, and merging of the enhanced fault samples with the other category samples to obtain an equalized training data set.

2. The method of claim 1, wherein: The multi-source heterogeneous sensor time series data comprises high-frequency vibration sensor data stream, low-frequency temperature sensor data stream, and rotating speed sensor data stream.

3. The method of claim 1, wherein: The spatiotemporal alignment and fusion of multi-source heterogeneous sensor time series data from at least one industrial equipment to obtain multi-dimensional feature vector time series data comprises the following steps: timestamp correction of the multi-source heterogeneous sensor time series data based on a network time protocol server to obtain timestamp-corrected sampling data; the timestamp-corrected sampling data comprises high-sampling-frequency data and low-sampling-frequency data; down-sampling alignment of the high-sampling-frequency data to the low-sampling-frequency data by a weighted average method based on timestamp proximity to determine weights to obtain time-aligned multi-channel sensor data; vectorization of the time-aligned multi-channel sensor data to obtain multi-dimensional feature vector time series data under the same timestamp.

4. The method of claim 1, wherein: The hybrid adaptive noise reduction processing of the multi-dimensional feature vector time series data to obtain noise-reduced time series data comprises the following steps: discrete wavelet transform of high-frequency signal components in the multi-dimensional feature vector time series data, and removal of Gaussian white noise in the wavelet domain by a soft threshold method to obtain preliminary noise-reduced data; inputting of the preliminary noise-reduced data into a trained deep autoencoder for filtering of non-Gaussian background noise and periodic interference to obtain noise-reduced time series data; wherein the deep autoencoder is trained based on reconstructed healthy equipment signals, and the training target is to minimize the difference between the reconstructed signals and the original healthy signals.

5. The method of claim 1, wherein: The segmentation of the noise-reduced time series data based on statistical characteristics thereof using an adaptive window function to obtain initial data samples comprises the following steps: continuous calculation of information entropy of the noise-reduced time series data based on a sliding preset reference window; if the information entropy is lower than a preset threshold, data segmentation of the noise-reduced time series data based on a first preset analysis window; if the information entropy is higher than or equal to the preset threshold, data segmentation of the noise-reduced time series data based on a second preset analysis window to obtain initial data samples; The first preset analysis window is greater than the second preset analysis window.

6. The method of claim 1, wherein: The initial data samples are standardized based on the equipment working condition to obtain standardized data samples, including: The historical working condition data of the at least one industrial equipment is divided based on an unsupervised clustering algorithm to obtain multiple working condition areas; It is judged that each sample in the initial data sample belongs to a working condition area; The statistical parameters corresponding to each working condition area are called to perform Z-score standardization processing on the initial data samples to obtain standardized data samples; the statistical parameters include mean and standard deviation.

7. The method of claim 1, wherein: The standardized data samples are classified and identified, including: based on a pre-trained classifier, the standardized data samples are classified and identified.

8. The method of claim 1, wherein: The preset rare fault category samples are enhanced based on a generative adversarial network to generate enhanced fault samples, including: the preset rare fault category samples are input into the generator of the generative adversarial network as seed samples to generate enhanced fault samples similar in feature distribution but with differences.

9. An industrial time series data processing system, characterized by, The processing method of the industrial time series data is used to realize any one of claims 1-8; the system comprises: an alignment fusion module, a noise reduction module, a segmentation module, a standardization module, a classification identification module and an enhancement module; wherein, The alignment fusion module is configured to perform spatio-temporal alignment and fusion on multi-source heterogeneous sensor time series data from at least one industrial equipment to obtain multi-dimensional feature vector time series data; The noise reduction module is configured to perform mixed adaptive noise reduction processing on the multi-dimensional feature vector time series data to obtain noise-reduced time series data; The segmentation module is configured to perform data segmentation using an adaptive window function based on the statistical characteristics of the noise-reduced time series data to obtain initial data samples; The standardization module is configured to identify the equipment working condition corresponding to the initial data samples and perform standardization processing on the initial data samples based on the equipment working condition to obtain standardized data samples; The classification identification module is configured to classify and identify the standardized data samples to obtain preset rare fault category samples and other category samples except the preset rare fault category samples; The enhancement module is configured to enhance the preset rare fault category samples based on a generative adversarial network to generate enhanced fault samples, and combine the enhanced fault samples with the other category samples to obtain a balanced training data set.

10. An electronic device, comprising: It includes: A memory, a processor and a computer program stored on the memory and running on the processor, the processor implements the method of any one of claims 1-8 when executing the computer program.

Images