A fault prediction method and system for a high-voltage motor

Abstract

The application discloses a kind of high-voltage motor fault prediction method and system, to solve the problem that prior art cannot carry out high sensitivity early warning to weak, progressive early fault under the premise of not relying on fault sample.The method comprises: synchronously collecting three-phase current, three-axis vibration and stator temperature signal;Based on historical normal data, train multivariate LSTM-VAE prediction model;Real-time calculation multidimensional residual vector;Residual covariance matrix is calculated in sliding window;Generate anomaly score by structural similarity with reference covariance matrix;When anomaly score is continuously over threshold, trigger early warning.The system comprises multi-source sensor array, synchronous acquisition module, prediction model, residual calculation module, covariance analysis module and early warning decision module.The application realizes early fault high robustness early warning when signal amplitude is not over limit by monitoring multivariate residual covariance structure evolution, significantly improves predictive maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and intelligent operation and maintenance of industrial equipment, specifically to a fault prediction method and system for high-voltage motors. Background Technology

[0002] Against the backdrop of rapid development in industrial automation and intelligent manufacturing, high-voltage motors, as core power equipment in industries such as power, metallurgy, and petrochemicals, directly impact the safety and efficiency of the entire production system through their operational reliability. High-voltage motors typically operate under high loads, strong electromagnetic interference, and complex conditions for extended periods. Sudden failures can not only cause significant economic losses but also potentially trigger safety accidents. Therefore, effective condition monitoring and early fault warning for high-voltage motors have become a key technological requirement in the field of intelligent operation and maintenance of industrial equipment.

[0003] Online monitoring technology based on multi-source sensor information is widely used in the health management of high-voltage motors. Current mainstream solutions often employ independent channel threshold alarm mechanisms, such as setting fixed upper limits for current RMS values, vibration velocity, or winding temperature to determine if these limits are exceeded. However, such methods can only identify obviously abnormal faults and cannot capture early signs where signal amplitudes are still within the normal range but the internal state has already undergone slight degradation. Existing technologies generally neglect the dynamic coupling relationship changes exhibited by multi-source sensor signals in the fault initiation stage. Under normal operating conditions of a high-voltage motor, there is a stable physical correlation between variables such as three-phase current, vibration, and temperature; however, when early faults (such as slight winding asymmetry or micro-damage to the bearing surface) begin to develop, this coupling structure often becomes disturbed first, before any single signal amplitude anomaly. Due to the lack of modeling capabilities for the evolution of multivariate residual covariance structures, existing systems struggle to achieve proactive perception of such subcritical states, leading to delayed or even missed warnings, severely limiting the actual effectiveness of predictive maintenance. Therefore, there is an urgent need for an early fault warning method that does not rely on fault samples and can identify early fault degradation in coupling relationships from multivariate temporal dynamics in an unsupervised manner. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a fault prediction method and system for high-voltage motors, effectively solving the problems mentioned in the background section.

[0005] To achieve the above objectives, the present invention provides a fault prediction method for high-voltage motors, comprising the following steps:

[0006] Step 1: Acquire multi-source time series signals during the operation of the high-voltage motor: The multi-source sensor array deployed in the high-voltage motor is collected through the synchronous acquisition module to synchronously acquire three-phase current signals, bearing housing triaxial vibration signals and stator winding temperature signals;

[0007] Step 2: Train the multivariate time series prediction model: Based on the multi-source time series signals collected during the historical stable operation phase of the equipment, construct a multivariate LSTM-VAE prediction model and train it. Use the multi-source time series signals within a predetermined time period in the past as the input of the model, and output the multivariate signal prediction values ​​for several future time steps. The training objective is to minimize the mean square error loss function.

[0008] Step 3: Real-time calculation of multidimensional residual vector: During the runtime phase, the currently measured multi-source time series signal is compared with the model's multivariate signal prediction values ​​at predetermined time intervals to generate a residual vector of dimension d. The residual vector Before being stored in the sliding window, the residual vector is zero-mean processed to obtain a centered residual vector. ;

[0009] Step 4: Sliding calculation of the sample covariance matrix: Store the most recent predetermined number of residual vectors into a sliding window of fixed length, and calculate the sample covariance matrix based on the data within this window at predetermined intervals. ;

[0010] Simultaneously, based on data collected during historical stable operation phases, the covariance matrix calculation process is repeated to obtain M covariance matrices under normal operating conditions. The arithmetic mean of these matrices is then used as the reference covariance matrix. ;

[0011] Step 5: Generate anomaly scores: Compare the structural similarity between the current sample covariance matrix and the reference covariance matrix, and calculate the anomaly scores using the normalized trace inner product formula;

[0012] Step 6: Trigger early fault warning: When the abnormal score continuously exceeds the preset threshold and the duration is ≥ the predetermined time period, the early warning decision module outputs an early fault warning signal to the central monitoring platform.

[0013] To optimize the above technical solution, the specific measures also include:

[0014] Furthermore, in step 1, the multi-source sensor array includes a three-phase current transformer installed in the motor junction box, a triaxial acceleration vibration sensor fixed in the bearing housing, and a platinum resistance temperature sensor embedded in the stator slot. All sensor signals are synchronously triggered by the same high-precision clock source to ensure that the multi-channel data are aligned on the time axis; the sampling frequency is not lower than the preset frequency threshold, and the time alignment accuracy does not exceed the predetermined time threshold.

[0015] Furthermore, in step 2, the encoder of the multivariate LSTM-VAE prediction model consists of two layers of long short-term memory network units stacked together, with the number of hidden units in each layer being a preset value and the latent space dimension set to a predetermined value; the decoder adopts a symmetrical structure and generates 7-dimensional signal prediction values ​​for future time moments through reverse temporal expansion. During model training, the Adam optimizer is used, with the initial learning rate set to a preset learning rate, and it decays to a predetermined proportion of the original value when the verification loss does not improve for a consecutive predetermined number of rounds.

[0016] Furthermore, in step 3, the residual vector The calculation method is as follows, where each component corresponds to the prediction bias of a sensing channel.

[0017] ,

[0018] in, These are measured values ​​of multi-source time series signals. The value is the predicted value of the multivariate signal from the model, and k is the time step index; The dimension is 7; the residual vector is zero-mean processed before being stored in the sliding window to eliminate systematic bias.

[0019] Furthermore, in step 4, the sample covariance matrix The calculation method is as follows:

[0020] ,

[0021] Where t is the index of the current time step, N is the length of the sliding window, and i is the index of the centered residual vector within the sliding window. The centered residual vector after zero-mean processing. , It is the mean vector of all residual vectors within the sliding window; a 7×7 symmetric positive semidefinite matrix fully represents the covariance structure of the current multivariate residuals.

[0022] Furthermore, in step 4, the reference covariance matrix This was obtained by repeating the computation process in step 4 on a historical normal dataset.

[0023] = ,

[0024] in, is the covariance matrix under normal operating conditions, and m is the index of the reference covariance matrix.

[0025] Further, in step 5, the outlier score is calculated. Specifically:

[0026] ,

[0027] in, Represents the trace of a matrix. Let Frobenius norm be the matrix. For reference covariance matrix.

[0028] Furthermore, in step 6, the early warning decision module adopts a dual threshold confirmation mechanism. In addition to the abnormal score threshold determination, an auxiliary criterion is also set. That is, when the abnormal score continuously exceeds the preset threshold and the duration is ≥ the predetermined time period, and the rate of change of the largest eigenvalue of the covariance matrix exceeds the predetermined rate of change threshold, the early warning decision module outputs an early fault warning signal to the central monitoring platform.

[0029] Furthermore, after the warning is triggered, the system automatically records all original signals, residual vectors, and covariance matrix evolution trajectories within a predetermined time period before and after the trigger time, and generates a diagnostic report containing a covariance structure heatmap for experts to remotely analyze the fault mode.

[0030] This invention also provides a fault prediction system for high-voltage motors, used to execute the aforementioned fault prediction method for high-voltage motors, comprising: a multi-source sensor array for synchronously acquiring three-phase current signals, bearing housing triaxial vibration signals, and stator winding temperature signals of the high-voltage motor; a synchronous acquisition module connected to the multi-source sensor array for time-aligned data acquisition; an edge computing unit deployed with a multivariable LSTM-VAE prediction model for training the model based on multi-source time series signals from historical stable operating phases, using the multi-source time series signals as input to the model, and outputting multivariable signal prediction values; a residual calculation module located in the edge computing unit, connected to the synchronous acquisition module via a high-speed serial bus, for comparing the measured signals with the multivariable signal prediction values ​​to generate a d-dimensional residual vector; a covariance analysis module for calculating the sample covariance matrix and the reference covariance matrix based on the residual vector within a sliding window; an early warning decision module for comparing the current covariance matrix with the reference covariance matrix to calculate anomaly scores, and outputting an early fault warning signal when preset conditions are met; and a central monitoring platform for receiving the warning signal and triggering further diagnostic processes.

[0031] The beneficial effects of this invention are as follows: This invention provides unsupervised early warning, requiring only historical data from normal equipment operation to complete model training and reference benchmark calibration. It eliminates the need for scarce fault samples, solving the generalization problem caused by a lack of early-stage fault samples. This allows for rapid deployment during the initial equipment commissioning phase or after major overhauls. Simultaneously, by monitoring the structural evolution of the multivariate residual covariance matrix, this invention can accurately capture subtle perturbations in the dynamic coupling relationships between variables, exhibiting higher sensitivity than traditional threshold methods and univariate detection schemes. Multivariate joint modeling effectively suppresses random fluctuations such as electromagnetic interference and environmental noise from single sensors. As a high-order statistical feature, the covariance matrix is ​​insensitive to local anomalies, triggering an early warning only when a systematic shift occurs in the multi-channel correlation pattern, significantly reducing the false alarm rate and significantly improving robustness and anti-interference capabilities. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating the fault prediction method for high-voltage motors of the present invention.

[0033] Figure 2 This is a schematic diagram of the fault prediction system for high-voltage motors of the present invention. Detailed Implementation

[0034] The invention will now be described in further detail with reference to the accompanying drawings.

[0035] It should be noted that the terms such as "upper", "lower", "left", "right", "front", and "back" used in the invention are only for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0036] The embodiments described in this invention are merely some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0037] like Figure 1 , 2 As shown, this embodiment is applied to an early fault warning scenario for a 10kV / 2500kW high-voltage asynchronous motor in the main exhaust fan drive system of a large steel enterprise. This motor operates under conditions of high dust, strong electromagnetic interference, and variable load (70%-100% rated load) for extended periods, making it prone to non-open-circuit faults such as early bearing micro-pitting and stator insulation aging. Traditional threshold-based alarm monitoring systems have repeatedly failed to detect faults, especially in the early stages of degradation such as bearing micro-pitting and slight stator winding asymmetry. To address this issue, a fault prediction method and system for high-voltage motors as described in this invention are deployed. This method is based on residual covariance analysis using multivariate time-series prediction.

[0038] First, the system architecture consists of a multi-source sensor array, a synchronous data acquisition module, an edge computing unit, a multivariate time-series prediction model, a residual calculation module, a covariance analysis module, an early warning decision module, and a central monitoring platform, all integrated into an industrial IoT edge smart gateway. The multi-source sensor array includes three types of physical entities: a three-phase current transformer (model: LEM ITN 900-S) installed inside the motor junction box, with a rated primary current of 3000A and an output of ±10V analog signal; a triaxial MEMS acceleration and vibration sensor (model: PCB 352C33) fixed on the bearing housings of the non-drive and drive ends of the motor, with a range of ±50g and a frequency response range of 0.5Hz–10kHz; and a Pt100 platinum resistance temperature sensor (compliant with IEC 60751 Class A standard) embedded near the end in the stator slot, with a four-wire lead-out to eliminate the influence of wire resistance. All sensor signals are synchronously sampled and triggered by the same high-precision GPS timing clock source (PPS pulse accuracy ±50ns), ensuring that the multi-channel data are strictly aligned on the time axis, with the time alignment error controlled within 0.8 milliseconds.

[0039] The synchronous data acquisition module uses an industrial-grade data acquisition card (such as the NI PXIe-6368 multifunction data acquisition card). In this embodiment, it is configured with a 1kHz sampling frequency, an independent anti-aliasing filter for each channel (cutoff frequency 450Hz), and interconnects with the edge computing unit via the PXIExpress backplane bus. The edge computing unit is built on the NVIDIA Jetson AGX Orin embedded AI platform, equipped with an ARM Cortex-A78AE CPU and an Ampere architecture GPU, 32GB of LPDDR5 memory and 1TB of NVMe SSD storage, running the Ubuntu 20.04 LTS operating system and ROS 2 Foxy middleware. This unit is directly connected to the acquisition module via a PCIeGen4 x8 interface, and the measured data transmission latency is 0.32 milliseconds, meeting the real-time requirements.

[0040] The multivariate time series prediction model was implemented using the deep learning framework PyTorch 1.12 and deployed in a GPU-accelerated environment on an edge computing unit. The model employs an LSTM-VAE (Long Short-Term Memory Network-Variational Autoencoder) architecture, where the encoder consists of two stacked LSTM layers with 128 hidden units per layer and an input dimension of 7 (corresponding to three-phase current). , , triaxial vibration , , The stator temperature (T) is used, and the time window length is 5000 sampling points (i.e., 5 seconds of historical data). The encoder output mean μ and logarithmic variance are... A 32-dimensional latent variable z is obtained through reparameterized sampling. The decoder adopts a symmetrical structure, also composed of two LSTM layers, and generates 7-dimensional signal predictions for the next 1 to 3 time steps (i.e., 100ms, 200ms, 300ms) through reverse temporal unfolding. During the model training phase, continuous 72 hours of normal operating data (sampling rate 1kHz) from the initial stage of equipment commissioning are used to construct input-output pairs in a sliding window manner: the input is a 7-dimensional sequence of the past 5 seconds (i.e., 5000 time points), and the output is the true value of the next time point. The loss function uses mean squared error (MSE), and the Adam optimizer is used. =0.9, =0.999), with an initial learning rate of 0.001. When the validation set loss shows no improvement for 10 consecutive rounds, the learning rate is reduced to 0.5 of the original value until convergence.

[0041] The residual calculation module and the covariance analysis module are integrated into the same software process, running on the CPU core of the edge computing unit's server. The residual calculation module's operations are implemented in C++, accelerated using the SIMD instruction set (AVX2), with a single calculation taking less than 0.1 milliseconds, ensuring sufficient real-time responsiveness. The residual vectors are stored directly as raw floating-point arrays in a circular buffer without any statistical extraction or manual feature engineering. This buffer has a fixed length of 600, corresponding to a 1-minute (600×100ms) observation window. The covariance analysis module triggers covariance matrix calculation every 5 seconds (i.e., every 50 residual vector update cycles). Internally, it maintains a 600×7 double-precision floating-point matrix storing the most recent 600 residual vectors.

[0042] The early warning decision module is responsible for anomaly detection and signal output. Implemented in Python 3.8, it integrates an OPC UA client (based on the FreeOpcUa library) for interface with the SCADA system (monitoring and data acquisition system). The entire system connects to the factory's industrial ring network via Gigabit Ethernet (IEEE 802.3ab), supporting Modbus / TCP, Profinet, and OPC UA protocols. Early warning signals can directly drive the SAP PM module to generate maintenance work orders, which are then pushed to maintenance personnel's mobile terminals via the MQTT protocol.

[0043] In one implementation, the system workflow is as follows: After system startup, it first enters the offline calibration phase: During the initial commissioning of the equipment or during stable operation after a major overhaul, the synchronous acquisition module continuously acquires multi-source raw time series signals at a frequency of 1kHz for no less than 72 hours to form a historical normal dataset. The dataset was divided into a training set (80%), a validation set (10%), and a test set (10%). The multivariate LSTM-VAE prediction model was trained end-to-end on the training set, with the objective function being the mean squared error (MSE) loss.

[0044]

[0045] Where T is the total number of time steps (in this embodiment, T=5000, corresponding to 5 seconds of historical data), t represents the time step index, and 1≤t≤T. These are measured values ​​of multi-source time series signals. This represents the predicted value of the multivariate signal from the model.

[0046] After training, the model is solidified and deployed to the inference engine of the edge computing unit.

[0047] Once in the online operation phase, the residual calculation module executes a prediction and residual calculation loop every 100 milliseconds. Specifically, it packages the 7-dimensional time series (5000 sampling points) within the current 5-second window into a tensor and transmits it to the edge computing unit via a high-speed serial bus. The multivariate time series prediction model receives this input and outputs the 7-dimensional prediction value for the next time step. At the same time, the synchronous acquisition module obtains the measured value of k at the current time. The two are compared in the residual calculation module to generate a 7-dimensional residual vector. Before being stored in the sliding window, the residual vector is first subjected to zero-mean normalization: the mean of the most recent 600 residual vectors is calculated. The mean value is then subtracted from the current residual to obtain the centralized residual. This is to eliminate systematic offsets caused by sensor drift or environmental temperature drift.

[0048] The centered residual vectors are pushed into a circular buffer (FIFO queue) of length 600, corresponding to a 60-second observation window. The covariance analysis module triggers covariance matrix calculation every 5 seconds (i.e., after every 50 residual vector updates), using an unbiased estimation formula to calculate its sample covariance matrix. : , where i is the global time step index, with a value range of t-N+1≤i≤t, used to traverse the N centered residual vectors within the sliding window, where the denominator 599=600-1. This 7×7 symmetric positive semi-definite matrix fully characterizes the second-order statistical coupling characteristics between the current multivariate residuals. For example, the covariance term between the current and vibration channels reflects their synchronicity in prediction bias. Meanwhile, the system has constructed a reference covariance matrix during the offline calibration phase. In historical normal datasets Repeat the above covariance calculation process to obtain the covariance matrix under normal conditions. (where m is the index of the reference covariance matrix), traverse all centered residual vectors of the reference set to generate M reference set sample covariance matrices, forming a set of reference covariance matrices. To eliminate the impact of random fluctuations on the benchmark under normal operating conditions, the reference covariance matrix set is statistically averaged: the arithmetic mean of the covariance matrices of the M reference set samples is calculated as the final offline calibration reference covariance matrix. The expression is: = .in, It is a 7×7 symmetric positive semi-definite matrix. This reference covariance matrix remains unchanged throughout the entire lifespan of the equipment, unless major repairs or replacement of core components occur.

[0049] To obtain the current covariance matrix With reference matrix Then, the early warning decision module calculates the anomaly score. This embodiment uses the trace inner product formula normalized to the Frobenius norm:

[0050] ,

[0051] in Represents the trace of a matrix. The Frobenius norm is defined as the square root of the sum of the squares of all elements of the matrix. The denominator is normalized by multiplying the Frobenius norms of the two matrices to achieve trace inner product normalization. This index measures the structural similarity between the two matrices, with a value range of [0,1]. The closer it is to 1, the greater the deviation of the covariance structure.

[0052] To improve the reliability of early warnings, the system adopts a dual-threshold confirmation mechanism. The primary criterion is: when AnomalyScore(t) > 0.35 (preset threshold) and this state is maintained continuously for more than 10 minutes (i.e., 600 5-second cycles); the secondary criterion is: the relative rate of change per minute of the largest eigenvalue of the covariance matrix exceeds 5%, that is, the current largest eigenvalue and the largest eigenvalue 60 seconds ago satisfy the following: ,in, The largest eigenvalue of the covariance matrix calculated at the current time t. The maximum eigenvalue of the covariance matrix calculated 60 seconds prior (i.e., at time t-60s); this criterion is used to eliminate instantaneous interference caused by sudden changes in operating conditions, ensuring the reliability of the early warning signal. Only when both conditions are met simultaneously does the early warning decision module determine it as an early fault bud and send an "early fault warning" signal to the central monitoring platform via the OPC UA protocol. Simultaneously, it triggers the diagnostic report generation process: automatically recording all original signals, residual vector sequences, and covariance matrix evolution trajectories within 30 minutes before and after the trigger time, and using the Matplotlib library to generate a covariance structure heatmap, marking significantly changed matrix elements (such as the position corresponding to (1,4)). and (covariance), providing experts with remote analysis of failure modes.

[0053] Furthermore, this embodiment also implements an adaptive operating condition function. The system pre-establishes multiple normal operating condition data subsets covering different load levels (30%, 50%, 75%, 100% rated load), speed ranges (950–990 rpm), and ambient temperatures (10–40℃), and calibrates the corresponding reference covariance matrix for each subset. This forms a reference matrix library. During operation, the edge computing unit reads the load rate and speed signals provided by the DCS system in real time, and automatically switches to the best-matching reference matrix through the nearest neighbor matching algorithm. This effectively suppresses the natural drift of the covariance structure caused by fluctuations in operating conditions, ensuring the stability of the early warning logic under changing operating conditions.

[0054] In actual operation, the system successfully provided an early warning of bearing failure 18 months after the motor was put into operation. At that time, the three-phase current RMS was 1850A (rated 1920A), the vibration velocity RMS was 2.8mm / s (below the 4.5mm / s alarm limit specified in ISO 10814), the stator temperature was 78℃ (ambient temperature 35℃), and all individual signals were within limits. However, the system detected a covariance structure anomaly: current-vertical vibration (… - The correlation coefficient significantly increased from 0.12 in the normal state to 0.41, causing the AnomalyScore to remain above 0.38 for 12 minutes, with the maximum eigenvalue change rate reaching 6.2% / min, satisfying the dual threshold condition. After the warning was triggered, maintenance personnel arranged for a shutdown inspection and found micro-pitting pits with a diameter of approximately 0.3 mm in the raceway of the non-drive end bearing, verifying the high sensitivity of this method in capturing subcritical states.

[0055] In summary, this embodiment achieves unsupervised and highly robust early warning of early faults in high-voltage motors by constructing high-precision synchronous acquisition hardware, deploying an LSTM-VAE prediction model, calculating the residual covariance matrix in real time, and introducing a structural similarity metric mechanism. This effectively solves the technical problem of high false negative rate in traditional methods during the weak degradation stage.

[0056] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method of failure prediction of a high voltage electric machine, characterized in that, Includes the following steps: Step 1: Acquire multi-source time series signals during the operation of the high-voltage motor: Data is acquired from the multi-source sensor array deployed on the high-voltage motor through the synchronous acquisition module, and the three-phase current signal, bearing housing triaxial vibration signal and stator winding temperature signal are acquired synchronously. Step 2: Train the multivariate time series prediction model: Based on the multi-source time series signals collected during the historical stable operation phase of the equipment, construct a multivariate LSTM-VAE prediction model and train it. Use the multi-source time series signals within a predetermined time period in the past as the input of the model, and output the multivariate signal prediction values ​​for several future time steps. The training objective is to minimize the mean square error loss function. Step 3: Real-time calculation of multidimensional residual vector: During the operation phase, the currently measured multi-source time series signal is compared with the model's multivariate signal prediction values ​​at predetermined time intervals to generate a residual vector of dimension d. The residual vector Before being stored in the sliding window, the residual vector is zero-mean processed to obtain a centered residual vector. ; Step 4: Sliding calculation of the sample covariance matrix: Store the most recent predetermined number of residual vectors into a sliding window of fixed length, and calculate the sample covariance matrix based on the data within this window at predetermined intervals. ; Simultaneously, based on data collected during historical stable operation phases, the covariance matrix calculation process is repeated to obtain M covariance matrices under normal operating conditions. The arithmetic mean of these matrices is then used as the reference covariance matrix. ; Step 5: Generate anomaly scores: Compare the structural similarity between the current sample covariance matrix and the reference covariance matrix, and calculate the anomaly scores using the normalized trace inner product formula; Step 6: Trigger early fault warning: When the abnormal score continuously exceeds the preset threshold and the duration is ≥ the predetermined time period, the early warning decision module outputs an early fault warning signal to the central monitoring platform.

2. The fault prediction method according to claim 1, characterized in that, In step 1, the multi-source sensor array includes a three-phase current transformer installed in the motor junction box, a triaxial acceleration vibration sensor fixed in the bearing housing, and a platinum resistance temperature sensor embedded in the stator slot. All sensor signals are synchronously triggered by the same high-precision clock source; the sampling frequency is not lower than a preset frequency threshold, and the time alignment accuracy does not exceed a predetermined time threshold.

3. The fault prediction method according to claim 1, characterized in that, In step 2, the encoder of the multivariate LSTM-VAE prediction model consists of two layers of long short-term memory network units stacked together, with the number of hidden units in each layer being a preset value and the latent space dimension being set to a predetermined value. The decoder adopts a symmetrical structure and generates 7-dimensional signal prediction values ​​for future time moments through reverse temporal expansion. The Adam optimizer is used during model training. The initial learning rate is set to a preset learning rate and decays to a predetermined proportion of the original value when the verification loss does not improve for a predetermined number of consecutive rounds.

4. The fault prediction method according to claim 1, characterized in that, Residual vector in step 3 The calculation method is as follows, where each component corresponds to the prediction bias of a sensing channel. ， in, These are measured values ​​of multi-source time series signals. The value is the predicted value of the multivariate signal from the model, and k is the time step index; The dimension is 7.

5. The fault prediction method according to claim 1, characterized in that, Sample covariance matrix in step 4 The calculation method is as follows: ， Where t is the index of the current time step, N is the length of the sliding window, and i is the index of the centered residual vector within the sliding window. The centered residual vector after zero-mean processing. , It is the mean vector of all residual vectors within the sliding window.

6. The fault prediction method according to claim 5, characterized in that, In step 4, refer to the covariance matrix This was obtained by repeating the computation process in step 4 on a historical normal dataset. = ， in, Let be the covariance matrix under normal operating conditions, and m be the index of the covariance matrix under normal operating conditions.

7. The fault prediction method according to claim 1, characterized in that, In step 5, the outlier score is calculated. Specifically: ， in, Represents the trace of a matrix. Let Frobenius norm be the matrix. For reference covariance matrix.

8. The fault prediction method according to claim 1, characterized in that, In step 6, the early warning decision module adopts a dual threshold confirmation mechanism. In addition to the abnormal score threshold judgment, an auxiliary criterion is also set. That is, when the abnormal score continuously exceeds the preset threshold and the duration is ≥ the predetermined time period, and the rate of change of the largest eigenvalue of the covariance matrix exceeds the predetermined rate of change threshold, the early warning decision module outputs an early fault warning signal to the central monitoring platform.

9. The fault prediction method according to claim 1, characterized in that, After the warning is triggered, the system automatically records all original signals, residual vectors and covariance matrix evolution trajectories within a predetermined time period before and after the trigger time, and generates a diagnostic report containing a covariance structure heatmap for experts to remotely analyze the fault mode.

10. A fault prediction system for a high-voltage motor, used to execute the fault prediction method as described in any one of claims 1 to 9, characterized in that, include: A multi-source sensor array is used to synchronously acquire the three-phase current signal, the three-axis vibration signal of the bearing housing, and the temperature signal of the stator winding of the high-voltage motor. The synchronous acquisition module is connected to the multi-source sensor array and is used to acquire time-aligned data. The edge computing unit is equipped with a multivariable LSTM-VAE prediction model, which is used to train the model based on multi-source time series signals from historical stable operation phases. The multi-source time series signals are used as input to the model, and the predicted values ​​of the multivariable signals are output. The residual calculation module is located in the edge computing unit and is connected to the synchronous acquisition module through a high-speed serial bus. It is used to compare the measured signals with the predicted values ​​of the multivariable signals to generate a d-dimensional residual vector. The covariance analysis module is used to calculate the sample covariance matrix and the reference covariance matrix based on the residual vector within the sliding window. The early warning decision module is used to compare the structural similarity between the current covariance matrix and the reference covariance matrix to calculate the anomaly score, and output an early fault warning signal when the preset conditions are met. The central monitoring platform is used to receive early warning signals and trigger further diagnostic processes.

Images