Compression of multi-source heterogeneous time series data, motor fault prediction method and system

Abstract

The application discloses a compression method and system for multi-source heterogeneous time sequence data and a motor fault prediction method and system. Through the synergistic effect of multi-scale feature extraction, SE weighted fusion, mixed pooling and aggregator collapse, the intelligent compression and feature enhancement of time sequence data are realized. The architecture can simultaneously complete the reasonable reduction of sequence length and the effective improvement of feature depth, solve the technical problem of multi-source heterogeneous sensor data fusion in an industrial scene under the premise of retaining key fault features, and overcome the limitation of serious information loss of a traditional dimension reduction method.

Description

Technical Field

[0001] This invention relates to the field of compression processing of multi-source heterogeneous time-series data, specifically to compression of multi-source heterogeneous time-series data, and methods and systems for predicting motor faults. Background Technology

[0002] In the field of predictive maintenance of industrial equipment, existing technologies have three key shortcomings that directly limit the accuracy and practicality of fault prediction.

[0003] First, in terms of feature extraction, traditional methods mainly employ single-scale time-domain or frequency-domain analysis (such as FFT and wavelet transform), or simple statistical feature extraction (such as mean and variance). These methods struggle to effectively handle multi-source heterogeneous sensor data in industrial scenarios (such as high-frequency vibration and low-frequency temperature signals), leading to the loss or weakening of key fault features. Especially for complex operating conditions such as equipment motors, traditional dimensionality reduction methods (such as PCA) often destroy temporal correlations, failing to simultaneously meet the dual requirements of sequence length compression and feature depth preservation. Summary of the Invention

[0004] This invention provides a method and system for compressing multi-source heterogeneous time-series data and predicting motor faults, which solves the technical problem that existing feature extraction methods cannot simultaneously meet the dual requirements of sequence length compression and feature depth preservation.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0006] A method for compressing multi-source heterogeneous time-series data includes:

[0007] The preprocessed multi-source heterogeneous time series data is input into a trained data compression model to obtain compressed data; the multi-source heterogeneous time series data includes time series data from at least two different sources and in different formats;

[0008] The data compression model is configured as follows: multi-scale feature extraction is performed on the preprocessed multi-source heterogeneous time-series data, and the extracted multi-scale features are fused to obtain a first fused feature sequence; attention weights of the channels of the first fused feature sequence are calculated, and the first fused feature sequence is weighted and fused based on the attention weights of the channels to obtain a second fused feature sequence; hybrid pooling is performed on the second fused feature sequence to obtain a hybrid pooled feature sequence; and time dimension collapse processing is performed on the hybrid pooled feature sequence to obtain compressed data.

[0009] Preferably, the data compression model includes a multi-scale feature extraction module, which includes multiple parallel deep convolutional layers with different kernel sizes, a first activation layer, and a feature fusion layer. Each deep convolutional layer is configured to extract feature sequences of a corresponding scale. The first activation layer is configured to perform nonlinear operations on multiple feature sequences of different scales. The feature fusion layer is used to concatenate multiple feature sequences of different scales after nonlinear operations to obtain a first fused feature sequence.

[0010] and / or

[0011] The data compression model includes an SE attention module, which comprises a first average pooling layer, a fully connected nonlinear transformation layer, and a channel-wise dot product layer. The first average pooling layer is configured to perform average pooling on the input first fused feature sequence to obtain global information for each channel of the first fused feature sequence. The fully connected nonlinear transformation layer is configured to calculate the attention weight of each channel based on the global information of each channel. The channel-wise dot product layer is configured to calculate a weighted fusion of the first fused feature sequence based on the attention weight of each channel to obtain a second fused feature sequence.

[0012] and / or

[0013] The data compression model includes a hybrid pooling module, which comprises a second average pooling layer, a max pooling layer, and a channel concatenation layer. The second average pooling layer is configured to perform average pooling on the second fused feature sequence to obtain a first pooling sequence. The max pooling layer is configured to perform max pooling on the second fused feature sequence to obtain a second pooling sequence. The channel concatenation layer is configured to concatenate the first pooling sequence and the second pooling sequence to obtain a hybrid pooling feature sequence.

[0014] and / or

[0015] The data compression model includes an aggregation module, which includes a one-dimensional convolutional layer, a normalization layer, and a second activation layer. The one-dimensional convolutional layer is configured to extract a one-dimensional feature sequence from the hybrid pooling feature sequence. The normalization layer is configured to normalize the one-dimensional feature sequence. The second activation layer is configured to introduce a nonlinear transformation into the normalized one-dimensional feature sequence.

[0016] Preferably, the multi-source heterogeneous timing data is motor operating data, and the motor operating data includes any combination of the following:

[0017] The values ​​of the motor's three-phase current, vibration acceleration, speed, power factor, bearing temperature, and winding temperature are as follows:

[0018] Preferably, the preprocessing of the multi-source heterogeneous time-series data includes:

[0019] The multi-source heterogeneous time-series data is cleaned, and the data cleaning method includes any one or a combination of the following:

[0020] Method 1: Use an outlier detection algorithm to detect and remove outliers in the multi-source heterogeneous time-series data;

[0021] Method 2: Remove multi-source heterogeneous timing data for a preset time period after the motor starts and / or stops;

[0022] Method 3: Remove multi-source heterogeneous timing data where the motor's three-phase current is 0 but the motor speed is not 0;

[0023] Method 4: When the multi-source heterogeneous timing data includes the vibration acceleration value or the motor speed value of the motor, the vibration acceleration value or the motor speed value of the motor is low-pass filtered;

[0024] The preprocessing of the multi-source heterogeneous time series data includes: performing time alignment processing on the multi-source heterogeneous time series data;

[0025] The preprocessing of the multi-source heterogeneous time-series data includes: standardizing or normalizing the multi-source heterogeneous data.

[0026] Preferably, the preprocessing of the multi-source heterogeneous time-series data includes:

[0027] The multi-source heterogeneous time-series data is sliced ​​according to a preset time step to obtain multiple sliced ​​time-series data segments.

[0028] The process of obtaining multiple slice time-series data segments and inputting them into the trained data compression model includes:

[0029] The data is sequentially input into the data compression model to obtain local compressed data corresponding to each slice of time-series data. The local compressed data is then assembled to obtain the final compressed data.

[0030] A method for predicting motor faults includes the following steps:

[0031] The above-mentioned method for compressing multi-source heterogeneous time-series data is used to compress the collected motor operation data, and the compressed data is input into the trained fault prediction model to obtain the probability of possible faults in the motor within a predetermined future period.

[0032] Preferably, the fault prediction model is a GRU classifier;

[0033] and / or

[0034] The Adam optimization algorithm is used to jointly train the data compression model and the fault prediction model, and the training loss function is binary cross-entropy loss.

[0035] Preferably, the data compression model and the fault prediction model are jointly trained using the Adam optimization algorithm. The joint training includes the following steps:

[0036] Historical motor operating data and its corresponding service life are obtained, where the service life is the time point at which the motor fails due to a fault; the historical motor operating data is preprocessed to obtain multiple time-slice time-series data segments.

[0037] Based on the failure time point and the time node of each time slice, calculate the remaining service life of each time slice from the failure time point, and determine the training label of each time slice based on the remaining service life of each time slice. Construct training data based on each time slice and the training label to obtain multiple training data.

[0038] Multiple training data are sequentially input into the data compression model to obtain multiple compressed training data, and the obtained multiple compressed training data are sequentially input into the fault prediction model for training.

[0039] Preferably, determining the training labels for each time slice based on the remaining lifetime corresponding to each time slice includes:

[0040] The remaining lifespan corresponding to the time slice is compared with a preset threshold. If it is greater than the preset threshold, the time slice is labeled with a first training label. If it is less than or equal to the preset threshold, the time slice is labeled with a second training label.

[0041] Preferably, when the predicted probability of the motor failing within a predetermined future period is greater than a preset first probability threshold, an alarm signal is sent to the user.

[0042] When the predicted probability of a motor failure within a predetermined future period is less than or equal to a first probability threshold and greater than a preset second probability threshold, equipment status monitoring is strengthened.

[0043] A computer system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0044] The present invention has the following beneficial effects:

[0045] 1. The multi-source heterogeneous time-series data compression method of the present invention achieves intelligent compression and feature enhancement of time-series data through the synergistic effects of multi-scale feature extraction, SE (channel attention) weighted fusion, hybrid pooling, and aggregator collapse. This architecture can simultaneously achieve reasonable reduction of sequence length and effective improvement of feature depth, solving the technical challenge of multi-source heterogeneous sensor data fusion in industrial scenarios while preserving key fault features, and overcoming the limitation of severe information loss in traditional dimensionality reduction methods.

[0046] 2. In the preferred embodiment, this invention constructs an end-to-end joint learning architecture of temporal compression and GRU classifier. This architecture overcomes the limitations of traditional staged processing, enabling the feature compression module to dynamically optimize the feature extraction strategy according to the needs of the GRU classification task. The compressed high-density temporal features are input into the GRU network for deep temporal modeling, effectively capturing the long-term dependencies in equipment state evolution. This integrated design not only improves the model's ability to identify complex fault features but also significantly reduces the reliance on domain knowledge and manual feature engineering, enhancing the system's adaptability.

[0047] 3. In the preferred embodiment, this invention redefines the target paradigm of fault prediction, transforming traditional remaining useful life regression prediction into a binary classification problem based on time thresholds. By setting reasonable empirical period thresholds, the system clearly classifies equipment status into two categories: "safe operation" and "imminent failure," and outputs an intuitive probability assessment. This design enables the prediction results to directly correspond to differentiated maintenance decisions, including three response levels: immediate repair, enhanced monitoring, and normal maintenance, significantly simplifying the judgment process for maintenance personnel.

[0048] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0049] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0050] Figure 1 This is a flowchart of a motor fault prediction method in a preferred embodiment of the present invention. Detailed Implementation

[0051] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0052] Example 1:

[0053] As shown in Figure 1, this invention discloses a method for compressing multi-source heterogeneous time-series data, including:

[0054] The preprocessed multi-source heterogeneous time series data is input into a trained data compression model to obtain compressed data; the multi-source heterogeneous time series data includes time series data from at least two different sources and in different formats;

[0055] The data compression model is configured as follows: performing multi-scale feature extraction on the preprocessed multi-source heterogeneous time-series data, and fusing the extracted multi-scale features to obtain a fused feature sequence; calculating the attention weights of the channels of the first fused feature sequence, and performing weighted fusion on the first fused feature sequence based on the attention weights of the channels to obtain a fused feature sequence; performing hybrid pooling on the fused feature sequence to obtain a hybrid pooled feature sequence; and performing time dimension collapse processing on the hybrid pooled feature sequence to obtain compressed data.

[0056] In addition, in this embodiment, the present invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above method.

[0057] In this embodiment, the multi-source heterogeneous data preprocessing includes any of the following processing methods:

[0058] Data cleaning, data alignment, normalization, or standardization.

[0059] The specific data preprocessing method(s) to be used will be determined based on the actual application scenario and the final output target, and will not be limited here.

[0060] In this embodiment, the multi-scale feature extraction method can adopt a multi-branch parallel structure, feature pyramid structure, dilated convolution structure, skip connection structure or adaptive pooling, etc. The specific multi-scale feature extraction method is determined according to the specific application scenario and data characteristics, and is not limited here.

[0061] In this embodiment, the hybrid pooling method can be selected from multiple pooling methods such as max pooling, average pooling, and adaptive pooling for fusion. The specific fusion method can be weighted fusion, channel splicing, adaptive fusion, etc. The specific hybrid pooling method can be determined according to the specific application scenario and data characteristics, and is not limited here.

[0062] In this embodiment, the time dimension collapse processing can be performed using methods such as time-series pooling or aggregators.

[0063] This invention addresses the difficulty of data fusion from multi-source heterogeneous sensors by overcoming the information loss bottleneck of traditional feature extraction methods and achieving high-fidelity temporal feature compression.

[0064] Example 2:

[0065] Example 2 is an extended example of the example. The difference between Example 1 and Example 2 is that the specific steps of the compression method for multi-source heterogeneous time-series data are refined in combination with specific application scenarios.

[0066] In defining the prediction target, existing technologies mostly adopt regression prediction methods based on remaining useful life (RUL). This method has two inherent drawbacks: first, the degradation rate varies significantly among different devices, and RUL prediction with a fixed threshold lacks universality; second, the remaining time value output by the regression model is difficult to directly correspond to operation and maintenance decisions, requiring secondary manual interpretation, which increases the complexity of decision-making.

[0067] In terms of model architecture, traditional methods typically separate feature extraction and fault prediction into independent modules (e.g., manually extracting features before inputting them into a classifier). This staged processing leads to a disconnect between feature extraction and the final prediction target. Especially for time series models such as GRU, single-step input of features results in the loss of information about the evolution of equipment states. Furthermore, existing end-to-end models (such as pure LSTM) are inefficient at processing high-dimensional industrial time series data and lack targeted optimization for multi-scale features.

[0068] To address the aforementioned issues, this embodiment, building upon the content of Embodiment 1, creatively extracts a motor fault prediction method. This method aims to change the current disconnect between existing remaining service life prediction models and maintenance decisions, establishing an intelligent prediction mechanism directly oriented towards maintenance decisions. Secondly, it overcomes the mismatch between feature extraction and prediction targets caused by traditional staged processing by constructing an end-to-end adaptive learning framework. Through technological innovation, this solution strives to improve the accuracy of early fault warnings, simplify the maintenance decision-making process, and reduce the system's reliance on domain-specific knowledge, providing a more intelligent and practical predictive maintenance solution for industrial equipment.

[0069] Specifically, the motor fault prediction method in this embodiment includes the following steps:

[0070] 1. Acquisition of motor operating data

[0071] By installing smart meters, sensors, PT100 RTDs, frequency converters, and other instruments or equipment on the equipment motor, the operating parameters of the equipment, such as three-phase current (A phase, B phase, C phase), power factor, bearing temperature, vibration acceleration, winding temperature, and speed, can be collected or calculated.

[0072] 2. Data Preprocessing

[0073] 21. Data Cleaning

[0074] Outlier removal algorithms are used to eliminate outliers in the data. Specifically, these algorithms can be implemented using... Criteria, quartile algorithm, Z-score algorithm, etc., in this embodiment, combined with specific application scenarios and specific data characteristics, are adopted and used. The criteria exclude data that exceed the mean ± 3 standard deviations, such as those applicable to normally distributed parameters like three-phase current, vibration acceleration, and motor speed.

[0075] For example: Suppose the collected time series data is ,in Given the observed values ​​of the three-phase current parameters at time t, then the cleaned data... This can be formally expressed as:

[0076]

[0077] in, The mean of the sample; This represents the sample standard deviation.

[0078] Then, the same operation was performed on the vibration acceleration and motor speed.

[0079] Using the process constraint method, screen out those with 0 current but >0 rotation speed;

[0080] Data from the start-stop phase is filtered out, specifically data within 30 seconds of motor startup and after shutdown.

[0081] 22. Data Alignment

[0082] First, determine the target time axis. In this scheme, the highest sampling frequency of 1 second is selected as the reference time axis (t=0,1,2… seconds).

[0083] The effective value is calculated by averaging the squares of the instantaneous values ​​of the three-phase currents (phase A, phase B, and phase C) and then taking the square root. Representative values ​​are extracted from 1500 current sampling points for each phase. The calculation method for the effective value of the three-phase current is as follows:

[0084]

[0085] in, This indicates the sampling point for phase A current. The same calculation is then performed on the phase B and phase C currents.

[0086] The vibration acceleration is first low-pass filtered to effectively preserve the characteristics of the vibration signal and suppress high-frequency noise. The highest frequency of the filtered signal should be ≤2.0 kHz. Then, the effective value is calculated by alignment. The alignment method is the same as that for calculating the three-phase current, that is, the instantaneous value is squared, averaged, and then squared. Representative values ​​are extracted from 2000 vibration acceleration sampling points.

[0087] The motor speed data alignment method is consistent with the vibration acceleration.

[0088] 23. Data normalization

[0089] Standardize the data to facilitate subsequent algorithm processing. For example, convert data with different dimensions to the same dimension to make them comparable during model training.

[0090] 24. Data Slicing

[0091] For continuous time series historical data, the proposed method uses a window with a step size of T to segment the data and defines data labels based on the time window.

[0092] Specifically, the data tag acquisition process is as follows: acquire historical motor operating data and its corresponding service life, wherein the service life is the time point at which the motor fails due to a fault; preprocess the historical motor operating data to obtain multiple time-slice time-series data segments;

[0093] Based on the failure time point and the time node at the end of each time slice, calculate the remaining service life of each time slice from the failure time point:

[0094] Let EOL be the time of failure determined at each failure, and RUL be the remaining useful life. Then, EOL and RUL at time t have the following mathematical relationship:

[0095]

[0096] The training label for each time slice is determined based on the remaining lifetime corresponding to each time slice. Training data is constructed based on each time slice and the training label to obtain multiple training data.

[0097] The preferred scheme for determining the training labels for each time slice based on the remaining lifetime corresponding to each time slice is as follows:

[0098] The remaining lifespan corresponding to the time slice is compared with a preset threshold. If it is greater than the preset threshold, the time slice is labeled with a first training label. If it is less than or equal to the preset threshold, the time slice is labeled with a second training label.

[0099] For example, setting the time window threshold to If the remaining lifetime corresponding to a certain time slice Then mark it as D0 (indicating a safe operating period); if If the condition is close to failure, it is marked as D1 (indicating an imminent failure period). The label is represented in one-hot encoding form, such as D0 being [1, 0] and D1 being [0, 1].

[0100] 3. Data compression

[0101] 31. Construct a data compression model

[0102] Based on the operating data characteristics of the motor, in this embodiment, the data compression model includes a multi-scale feature extraction module, an SE attention module, a hybrid pooling module, and an aggregation module;

[0103] The multi-scale feature extraction module includes multiple parallel deep convolutional layers with different kernel sizes, a first activation layer, and a feature fusion layer. Each deep convolutional layer is configured to extract feature sequences of a corresponding scale. The first activation layer is configured to perform nonlinear operations on multiple feature sequences of different scales. The feature fusion layer is used to concatenate multiple feature sequences of different scales after nonlinear operations to obtain a first fused feature sequence.

[0104] In the preferred scheme, there are 3 deep convolutional layers with kernel sizes of 3, 5, and 7, and each kernel has 32 output channels.

[0105] The SE attention module includes a first average pooling layer, a fully connected nonlinear transformation layer, and a channel-wise dot product layer. The first average pooling layer is configured to perform average pooling on the input first fused feature sequence to obtain global information for each channel of the first fused feature sequence. The fully connected nonlinear transformation layer is configured to calculate the attention weight of each channel based on the global information of each channel. The channel-wise dot product layer is configured to calculate the weighted fusion of the first fused feature sequence based on the attention weight of each channel to obtain a second fused feature sequence.

[0106] The hybrid pooling module includes a second average pooling layer, a max pooling layer, and a channel concatenation layer. The second average pooling layer is configured to perform average pooling on the second fused feature sequence to obtain a first pooling sequence. The max pooling layer is configured to perform max pooling on the second fused feature sequence to obtain a second pooling sequence. The channel concatenation layer is configured to concatenate the first pooling sequence and the second pooling sequence to obtain a hybrid pooling feature sequence.

[0107] The aggregation module includes a one-dimensional convolutional layer, a normalization layer, and a second activation layer. The one-dimensional convolutional layer is configured to extract a one-dimensional feature sequence from the hybrid pooling feature sequence. The normalization layer is configured to normalize the one-dimensional feature sequence. The second activation layer is configured to introduce a nonlinear transformation into the normalized one-dimensional feature sequence to enhance the expressive power of the entire aggregation module.

[0108] In this embodiment, the first activation layer uses the ReLU activation function, and the second activation function uses the Swish activation function.

[0109] 4. Fault prediction

[0110] An improved version of LSTM, GRU, is used to construct a fault prediction model. This model can effectively predict the probability of a motor failure within a future window. The joint optimization process of the data compression model and the fault prediction model is as follows:

[0111] 41. GRU Input Layer Processing

[0112] The input layer is the network layer that feeds compressed data into the GRU. The input sample set can be represented as... ,in This indicates whether a malfunction has occurred, as indicated by the label information corresponding to the data in this window. Input data. The format is two-dimensional, namely [time step, number of features]. Here, the time step refers to the length of time that information from each feature can be passed to the next feature. In order to perform binary classification on the predicted value of equipment failure occurrence cycle, the time step needs to meet the minimum length of subsequent online data acquisition. The time step set in this scheme is 3 cycles.

[0113] 42. GRU Hidden Layer Processing

[0114] Hidden layers, situated between the input and output layers, are the core of model training and testing. The GRU unit computation process is formally expressed as follows:

[0115]

[0116] in This indicates the computation of the reset gate vector. and This represents the learnable weight matrix. This represents a learnable bias vector. This represents the hidden state sequence of the previous time step. This represents the sample vector at time t within the time window. Represents the sigmoid function; This indicates the computation and update of the gate vector. and This represents the learnable weight matrix. Represents a learnable bias vector; This indicates the computation of the candidate state sequence. and This represents the learnable weight matrix. This represents a learnable bias vector. This indicates element-wise multiplication. Indicates the activation function; This indicates the calculation of the final state.

[0117] 43. GRU Output Layer Processing

[0118] The output layer is a fully connected layer containing a feedforward neural network. In this scheme, the classifier output is defined as a two-element vector, where these features describe the probability that an observation belongs to one of two classes: Then, there are two units in the output layer, and they are used... Activation function. The output layer provides probability distributions for two classes (D0 and D1), formally represented as:

[0119]

[0120] in, Indicates the current window's hidden state at the last time step; Represents the learnable weight matrix; This represents a learnable bias. This represents the binary classification probability of the output, that is, the probability that the remaining lifetime exceeds a specified threshold. and the probability that the remaining lifetime does not exceed a specified threshold .

[0121] 44. Loss Calculation and Backpropagation

[0122] To jointly train the data compression model and the GRU classifier, the loss of the objective function is defined as the binary cross-entropy loss, which is specifically designed for solving two-class classification problems. The mathematical expression for the binary cross-entropy loss is as follows:

[0123]

[0124] The Adam optimization algorithm is used during training. It is an extension of the stochastic gradient descent algorithm and has the advantages of high computational efficiency, small memory requirements and high applicability to big data. It is widely used in deep learning models.

[0125] 5. Online prediction and maintenance decision-making

[0126] Real-time acquisition of the latest sensor data, followed by data alignment, window slicing, and data completion, constructs a sliding window of the same length as during training. The sliding window data is then input into a GRU classifier to obtain the probability that the current device may fail within the future target window (remaining lifetime less than a specified window size threshold). The preset maximum probability threshold for a fault is: The minimum probability threshold is Maintenance decisions are made based on probability.

[0127] when If so, arrange for repairs immediately;

[0128] when exist Within the scope, strengthen equipment status monitoring;

[0129] when Then only normal maintenance is required.

[0130] Example 3

[0131] Example 3 is an application example of Example 2, specifically including:

[0132] 1. Equipment operating parameter acquisition

[0133] By installing smart meters, sensors, PT100 RTDs, frequency converters, and other instruments or equipment on the main motor of the ball mill, operating parameters such as three-phase current, power factor, bearing temperature, vibration acceleration, winding temperature, and rotational speed were collected. The data covers approximately the past three years and amounts to 3.5TB.

[0134] The sampling frequency is as follows:

[0135] The three-phase current acquisition frequency is 1.5kHz, which means 1500 acquisitions per second.

[0136] The power factor sampling frequency is 1Hz, that is, it is sampled once per second;

[0137] The bearing temperature is sampled at a frequency of 1Hz, which means it is sampled once per second.

[0138] The vibration acceleration sampling frequency is 5kHz, which means 5000 samplings per second;

[0139] The motor speed is sampled at a frequency of 5kHz, which means 5000 samples are collected per second.

[0140] The winding temperature is sampled at a frequency of 1 Hz, which means it is sampled once per second.

[0141] 2. Data Preprocessing

[0142] 21. Data cleaning and alignment:

[0143] The data is cleaned and aligned according to the data cleaning and alignment method in Example 2. After cleaning and alignment, a dataset with a frequency of one data point per second is obtained. ( Indicates the number of time steps. (representing feature dimension), a real data point might be represented as 304.5 represents the A-phase current, 311 represents the B-phase current, and 299 represents the C-phase current. The units for the A-phase, B-phase, and C-phase currents are all in amperes (A), and this is under a 10kV voltage. 0.9 represents the power factor. 70 represents the bearing temperature in degrees Celsius. 3.5 represents the vibration acceleration in mm / s. 740 represents the rotational speed in r / min. 101 represents the winding temperature in degrees Celsius.

[0144] 22. Normalization

[0145] Convert data of different dimensions to the same dimension to make them comparable during model training. The normalized data will be between [0,1].

[0146] 23. Data Slicing

[0147] We collected time interval data from motor startup to failure multiple times during past production (the failure period may vary each time). We denote EOL as the time of failure determined for each failure and RUL as the remaining service life. Then, the following mathematical relationship exists between EOL and RUL at time point t:

[0148]

[0149] Taking an EOL of 600 production cycles as an example (8 hours per cycle, 3 cycles / day), the entire lifecycle data of the motor from start-up to failure is divided into multiple time windows, each window being 3 cycles long (i.e., 1 day). At the end of each window, the Remaining Useful Life (RUL) is calculated. If the RUL exceeds a set threshold... (For example, 21 cycles, or 7 days), then the window is marked as category D0, indicating that the motor will not fail within the next 1 cycle; otherwise, it is marked as D1, indicating that a failure is about to occur. The labels use one-hot encoding: D0 is represented as [1, 0], and D1 is represented as [0, 1], corresponding to the binary classification judgment of "failure" and "normal" respectively.

[0150] Therefore, within the time interval from start-up to failure of multiple motors, we will obtain multiple training units composed of time window data and label information. Each training unit consists of a time window of data and its corresponding label information. We denote a time window of data as a matrix. The label data is denoted as D0 or D1.

[0151] 3. Data compression

[0152] The multi-scale feature fusion compression process is as follows:

[0153] 1) Input time window data The data is divided into blocks, into 1200 72×8 blocks, each block... .

[0154] 2) Convolution operations are performed using 3 sets of parallel convolution kernels, formally represented as follows:

[0155] For convolution calculation with a kernel size of 3:

[0156]

[0157] in, This represents a one-dimensional convolution operation. For convolution kernel weights, For bias, This represents the ReLU activation function.

[0158] For convolution calculation with a kernel size of 5:

[0159]

[0160] in, For convolution kernel weights, For bias.

[0161] For convolution calculation with a kernel size of 7:

[0162]

[0163] in, For convolution kernel weights, For bias.

[0164] The three channel features are concatenated, and the formal representation is as follows:

[0165]

[0166] 3) Channel weighting is performed after the SE attention layer. The calculation process is as follows:

[0167] The average pooling operation is formally represented as follows:

[0168]

[0169] Nonlinear transformation operation, output channel weights The formal representation is as follows:

[0170]

[0171] in, and Represents the weight matrix. and Indicates bias.

[0172] Channel-weighted calculation, formally expressed as follows:

[0173]

[0174] Among them weighted features .

[0175] 4) Multi-scale pooling is performed after passing through a multi-scale pooling layer, as follows:

[0176] Max pooling, formally represented as follows:

[0177]

[0178] in This represents the max pooling function.

[0179] Average pooling, formally represented as follows:

[0180]

[0181] in This represents the average pooling function.

[0182] Channel splicing, formally represented as follows:

[0183]

[0184] 5) The time dimension collapses after the aggregator is executed, and the process is as follows:

[0185] 1D convolution computation, formalized as follows:

[0186]

[0187] The 1D convolution kernel size is 72, and the number of output channels is 128. This represents a flattening function, which flattens multiple tensors into one-dimensional vectors. Represents the aggregator weight matrix. This indicates the aggregator bias.

[0188] The execution layer is normalized and activated using the Swish function, formally expressed as follows:

[0189]

[0190] in Represents the normalization function. This represents the Swish activation function.

[0191] After the above data compression operations, 1200 blocks can be obtained. Time window data.

[0192] 4. Joint optimization of data compression model and GRU classifier

[0193] In the hidden layers constructed using this method, layers with 100 and 50 units were set sequentially. Additionally, Dropout was applied after each GRU layer to reduce overfitting of the neural network's training data, thereby improving the network's feature extraction capabilities.

[0194] In this approach, we collected data from the motor's start-up to failure interval five times, totaling approximately 3000 = 5 × 600 cycles (one cycle is approximately 8 hours, and 3000 cycles are approximately 1000 days). Each training unit is defined as a time window of one day, resulting in 1000 training units over 1000 days. Furthermore, the model is trained for 100 rounds.

[0195] In one round of training, 1000 training units are used for model training, which corresponds to 1000 time windows of data. Within each window, data compression is first performed, compressing the original 86400×8 training data into 1200×128 high-dimensional data. Next, at time step t, the 1200 128-dimensional data points are sequentially fed into the GRU classifier, implementing GRU forward propagation, including calculating the reset gate vector, the update gate vector, the candidate state vector, and the final hidden state vector, until the last time step. Then, the final hidden state is used to predict the probability of the device malfunctioning in the next 7 days. After forward propagation is complete, the loss is calculated based on the real label data and the model-predicted label data corresponding to the time window. The derivative of the loss is then calculated, and the gradient is backpropagated to update the model parameters.

[0196] After completing 100 rounds of model training, we output the final optimized model, which can be used for online prediction of equipment failures in real-world environments.

[0197] 5. Online prediction and maintenance decision-making

[0198] 51. Collect the latest sensor data in real time, then perform data alignment, window slicing, and data compression to construct a sliding window test sample of the same length as during training (3 cycles, i.e., 1 day). ;

[0199] 52. Input sliding window test sample The GRU classifier is used to obtain the probability P(D1) that the current device may fail within a future target window (remaining lifetime is less than a specified window size threshold);

[0200] 53. The maximum probability threshold for a fault is preset to 90%, and the minimum probability threshold is preset to 60%. Maintenance decisions are made based on the probability levels.

[0201] If the probability of failure P(D1) > 90%, maintenance should be arranged immediately.

[0202] When the probability of failure P(D1) is within the range of [60%, 90%], strengthen equipment status monitoring;

[0203] If the probability of failure P(D1) < 60%, then only normal maintenance is required.

[0204] In summary, this invention proposes three core innovative technologies in the field of equipment motor fault prediction, which significantly improves the intelligence level of predictive maintenance.

[0205] The first innovation lies in the design of a multi-scale time-series feature compression model. Through the synergistic effects of multi-scale feature extraction, SE (channel attention) weighted fusion, hybrid pooling, and aggregator collapse, intelligent compression and feature enhancement of time-series data are achieved. This architecture can simultaneously achieve reasonable reduction of sequence length and effective enhancement of feature depth. While preserving key fault features, it solves the technical challenge of fusing multi-source heterogeneous sensor data in industrial scenarios and overcomes the limitation of severe information loss in traditional dimensionality reduction methods.

[0206] The second innovation is the construction of an end-to-end multi-source heterogeneous time-series data compression and GRU joint learning architecture. This architecture breaks through the limitations of traditional staged processing, enabling the feature compression module to dynamically optimize the feature extraction strategy according to the needs of the GRU classification task. The compressed high-density time-series features are input into the GRU network for deep time-series modeling, effectively capturing the long-term dependencies in equipment state evolution. This integrated design not only improves the model's ability to identify complex fault features, but also significantly reduces the reliance on domain knowledge and manual feature engineering, enhancing the system's adaptability.

[0207] The third innovation redefines the target paradigm for fault prediction, transforming traditional remaining useful life regression prediction into a binary classification problem based on time thresholds. By setting reasonable empirical period thresholds, the system clearly classifies equipment status into two categories: "safe operation" and "imminent failure," and outputs an intuitive probability assessment. This design allows the prediction results to directly correspond to differentiated maintenance decisions, including three response levels: immediate repair, enhanced monitoring, and normal maintenance, significantly simplifying the judgment process for maintenance personnel.

[0208] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A compression method of multi-source heterogeneous time series data, characterized in that, The method comprises the following steps: inputting preprocessed multi-source heterogeneous time series data into a trained data compression model to obtain compressed data; the multi-source heterogeneous time series data comprises at least two time series data of different sources and different formats; the data compression model is configured to perform multi-scale feature extraction on the preprocessed multi-source heterogeneous time series data, and fuse the extracted multi-scale features to obtain a first fused feature sequence; calculate the attention weight of the channel of the first fused feature sequence, and perform weighted fusion on the first fused feature sequence based on the attention weight of the channel to obtain a second fused feature sequence; perform mixed pooling on the second fused feature sequence to obtain a mixed pooled feature sequence, and perform time dimension collapse processing on the mixed pooled feature sequence to obtain compressed data; the data compression model comprises a multi-scale feature extraction module, the multi-scale feature extraction module comprises a plurality of parallel deep convolution layers with different convolution kernel sizes, a first activation layer and a feature fusion layer, each deep convolution layer is configured to extract a feature sequence of a corresponding scale size, the first activation layer is configured to perform nonlinear operation on a plurality of feature sequences of different scales, and the feature fusion layer is used for splicing the plurality of feature sequences of different scales after nonlinear operation to obtain a first fused feature sequence; the data compression model comprises an SE attention module, the SE attention module comprises a first average pooling layer, a fully connected nonlinear transformation layer and a channel-by-channel dot product layer, the first average pooling layer is configured to perform average pooling processing on the input first fused feature sequence to obtain global information of each channel of the first fused feature sequence; the fully connected nonlinear transformation layer is configured to calculate the attention weight of each channel based on the global information of each channel; the channel-by-channel dot product layer is configured to calculate the weighted fusion of the first fused feature sequence based on the attention weight of each channel to obtain a second fused feature sequence; the data compression model comprises a mixed pooling module, the mixed pooling module comprises a second average pooling layer, a maximum pooling layer and a channel splicing layer, the second average pooling layer is configured to perform average pooling on the second fused feature sequence to obtain a first pooled sequence, the maximum pooling layer is configured to perform maximum pooling on the second fused feature sequence to obtain a second pooled sequence; and the channel splicing layer is configured to splice the first pooled sequence and the second pooled sequence in the channel to obtain a mixed pooled feature sequence; the data compression model comprises an aggregation module, the aggregation module comprises a one-dimensional convolution layer, a normalization layer and a second activation layer, the one-dimensional convolution layer is configured to extract a one-dimensional feature sequence of the mixed pooled feature sequence, the normalization layer is configured to normalize the one-dimensional feature sequence, and the second activation layer is configured to introduce nonlinear transformation again to the normalized one-dimensional feature sequence.

2. The compression method of multi-source heterogeneous time series data according to claim 1, characterized in that, The multi-source heterogeneous time series data is motor operation data, and the motor operation data comprises any combination of the following: The motor three-phase current value, the motor vibration acceleration value, the motor rotating speed value, the motor power factor value, the motor bearing temperature value, and the motor winding temperature value.

3. The compression method of multi-source heterogeneous time series data according to claim 2, characterized in that, The preprocessing of the multi-source heterogeneous time series data comprises: The data cleaning of the multi-source heterogeneous time series data comprises any one or a combination of the following modes: Mode 1: detecting abnormal values in the multi-source heterogeneous time series data by using an abnormal value detection algorithm and eliminating the abnormal values; Mode 2: eliminating the multi-source heterogeneous time series data in a preset time period after the motor starts and / or stops; Mode 3: eliminating the multi-source heterogeneous time series data in which the motor three-phase current value is 0 but the motor rotating speed value is not 0; Mode 4: when the multi-source heterogeneous time series data comprises the motor vibration acceleration value or the motor rotating speed value, performing low-pass filtering on the motor vibration acceleration value or the motor rotating speed value; The preprocessing of the multi-source heterogeneous time series data comprises time alignment processing of the multi-source heterogeneous time series data; The preprocessing of the multi-source heterogeneous time series data comprises standardization or normalization processing of the multi-source heterogeneous time series data.

4. The compression method of multi-source heterogeneous time series data according to claim 1, characterized in that, The preprocessing of the multi-source heterogeneous time series data comprises: slicing the multi-source heterogeneous time series data according to a preset time step to obtain a plurality of sliced time series data segments; The obtained plurality of sliced time series data segments are input into the trained data compression model, comprising: sequentially inputting into the data compression model in sequence to obtain local compression data corresponding to each sliced time series data segment, assembling the local compression data to obtain final compression data.

5. A method of motor fault prediction, characterized by, comprising the following steps: using the compression method of the multi-source heterogeneous time series data according to any one of claims 1-4 to compress the collected motor operation data, and inputting the compressed data into the trained fault prediction model to obtain a probability of possible faults of the motor in a future predetermined period.

6. The motor fault prediction method of claim 5, wherein, The fault prediction model is a GRU classifier. and / or The data compression model and the fault prediction model are jointly trained by using an Adam optimization algorithm, and a training loss function is binary cross-entropy loss.

7. The motor fault prediction method of claim 5, wherein The data compression model and the fault prediction model are jointly trained by using an Adam optimization algorithm, and the joint training comprises the following steps: obtaining historical motor operation data and corresponding service life thereof, the service life being a failure time point of the motor due to a fault; preprocessing the historical motor operation data to obtain a plurality of time-sliced time series data segments; calculating a remaining service life of each time slice from the failure time point according to the failure time point and a time node at the end of each time slice, and determining a training label of each time slice according to the remaining service life corresponding to each time slice, constructing training data according to each time slice and the training label to obtain a plurality of training data; sequentially inputting the plurality of training data into the data compression model to obtain a plurality of compressed training data, and sequentially inputting the obtained plurality of compressed training data into the fault prediction model for training.

8. The motor fault prediction method of claim 7, wherein, The determination of the training label of each time slice according to the remaining service life corresponding to each time slice comprises: The remaining service life corresponding to the time slice is compared with a preset threshold value, when greater than the preset threshold value, the time slice is labeled using a first training label, and when less than or equal to the preset threshold value, the time slice is labeled using a second training label.

9. The motor fault prediction method of claim 8, wherein, When the predicted probability of the motor possibly failing in a future predetermined period is greater than a preset first probability threshold value, an alarm signal is sent to a user; When the predicted probability of the motor possibly failing in a future predetermined period is less than or equal to the first probability threshold value and greater than a preset second probability threshold value, the equipment state monitoring is strengthened.

10. A computer system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the steps of the method of any one of claims 1 to 9 when executing the computer program.

Images