An online monitoring and early warning method for bus capacitance of an electric drive system

Abstract

The application relates to an online monitoring and early warning method for a bus capacitor of an electric drive system, and belongs to the field of electric drive systems. The electric drive system sequentially comprises an isolation circuit, a full-wave bridge rectifier, a bus capacitor group, an inverter, an actuator and a capacitor online testing device; the capacitor online testing device comprises a relay, a load resistor, a voltage acquisition module, a current acquisition module and a main control computing unit; when the electric drive system is idle, the load resistor is connected to build a test path, current and voltage signals are acquired; the test path is disconnected, a degradation representation quantity of multi-index information fusion is built through time-frequency domain feature extraction, and an equal-length time interval sequence degradation curve is built; a TCN-BRU-GRU combined prediction model learns and trains trend characteristics according to the degradation curve. The application solves the problems of waveform fitting of non-equal-interval time sampling, local and global data feature capturing, realizes trend prediction of bus capacitor performance degradation, gives early warning before the bus capacitor fails, and guarantees the safe and reliable operation of the electric drive system.

Description

Technical Field

[0001] This invention relates to the field of electric drive systems, and more specifically to an online monitoring and early warning method for the bus capacitance of an electric drive system. Background Technology

[0002] Electric drive systems are a crucial component of current new energy technologies, responsible for energy conversion and output. They transform direct current (DC) into alternating current (AC) of fixed frequency and voltage, or frequency and voltage regulation, to drive motors for motion control or energy conversion. The bus capacitor in an electric drive system absorbs power pulsations to conserve and recover energy, providing a smooth and stable output for the downstream inverter, increasing output voltage stability, reducing power loss, and improving energy efficiency. Connecting the rectification and inversion stages in the intermediate circuit, the bus capacitor provides voltage regulation, filtering, and reactive power to the motor load, playing a vital role in the system. However, with prolonged use, the electrical parameters of the bus capacitor degrade due to material defects, manufacturing processes, design flaws, and environmental factors such as temperature, humidity, and vibration.

[0003] Existing technologies for monitoring and early warning of bus capacitors in electric drive systems include the following: One approach involves removing the bus capacitor from the equipment and constructing a relationship between electrical and non-electrical parameters, i.e., a data conversion method. For example, the correlation between non-electrical physical quantities such as the weight, internal pressure, and internal temperature of aluminum electrolytic capacitors and the degree of capacitor performance degradation. Another approach utilizes structural changes in the capacitor to identify its fault state. Methods such as X-ray imaging to assess damage and acoustic detection to detect microcracks in capacitors are used to evaluate capacitor performance. However, due to the uncertainty of structural changes, it is difficult to establish a relationship between electrical parameters and structural changes, and this approach typically requires expensive measuring instruments and is not very practical. A third approach involves online acquisition and analysis of the electrical parameter signals of the bus capacitor. This involves using filtering algorithms such as median filtering, Kalman filtering, and wavelet transform denoising to filter the sampled signals, and then using parameter estimation algorithms to obtain the capacitor parameters. However, in harsh electromagnetic interference (EMI) environments, the noise from voltage and current measurements leads to low accuracy in estimating RSR or C.

[0004] The equivalent model of a capacitor includes the effective capacitance C, leakage resistance Rleak, equivalent resistance ESR, and equivalent inductance ESL, such as... Figure 1As shown, the equivalent capacitance model consists of the equivalent resistance (ESR) and equivalent inductance (ESL) connected in series, followed by the leakage resistance (Rleak) and capacitance (C) connected in parallel. The equivalent inductance (ESL) is composed of the lead inductance of the capacitor and the equivalent inductance between the capacitor terminals connected in series. The equivalent resistance (ESR) consists of the lead resistance of the capacitor and the equivalent resistance between the capacitor terminals, and its value depends on the capacitor's operating temperature, operating frequency, and the resistance of its lead wires. The leakage resistance (Rleak) depends on the capacitor's leakage characteristics. The leakage resistance (Rleak), equivalent resistance (ESR), and equivalent inductance (ESL) components in the equivalent capacitance model negatively impact the capacitor's capacitive reactance. While these factors cannot be eliminated by material processing, they need to be controlled within acceptable limits.

[0005] The mechanisms of capacitor degradation include slow electrolyte evaporation, dielectric decay and expansion, oxide film damage, and electrode corrosion. The degradation process of electrolytic capacitors is essentially an irreversible degradation of the electrochemical system. Gradual evaporation of water in the electrolyte leads to an increase in the equivalent series resistance (ESR), significantly reducing high-frequency filtering capability. Dielectric decay is due to the accumulation of lattice defects in the dielectric layer's microstructure under long-term electric field, resulting in a decrease in the effective dielectric constant. Electrode corrosion is caused by the formation of an oxide layer on the electrode surface through electrochemical processes, leading to increased contact resistance.

[0006] After analysis and selection, this invention characterizes the degradation of capacitors in inverters from the following aspects:

[0007] (1) Correlation of ripple amplitude: When the capacitor degrades, the equivalent series resistance (ESR) increases and the capacitance decreases, which leads to an increase in the peak-to-peak value of the bus voltage ripple.

[0008] (2) High-frequency harmonic characteristics: Changes were found in the high-frequency ripple energy distribution of the bus voltage. If the proportion of high-frequency components increased, it indicated that the capacitor ESR was deteriorating.

[0009] (3) Current fluctuation response: Monitor the transient response of the bus current when the load changes suddenly. If the energy storage capacity of the capacitor decreases, the transient peak amplitude of the bus current will increase and the recovery time will be prolonged. Summary of the Invention

[0010] To address the aforementioned technical problems, this invention proposes an online monitoring and early warning method for bus capacitors in an electric drive system, belonging to the field of electric drive systems. The electric drive system sequentially includes an isolation circuit, a full-wave bridge rectifier, a bus capacitor bank, an inverter, an actuator, and an online capacitor testing device. The online capacitor testing device includes a relay, a load resistor, a voltage acquisition module, a current acquisition module, and a main control computing unit. When the electric drive system is idle, a test path is established by connecting the load resistor to acquire current and voltage signals. The test path is then disconnected, and a degradation characterization quantity based on multi-index information fusion is constructed through time-frequency domain feature extraction, resulting in a degradation curve with equal time intervals. A TCN-BRU-GRU combined prediction model is used to learn and train trend features on the degradation curve. This invention solves the problems of waveform fitting and local and global data feature capture in non-equal time interval sampling, achieving trend prediction of bus capacitor performance degradation and providing early warning before bus capacitor failure, ensuring the safe and reliable operation of the electric drive system.

[0011] The electric drive system sequentially includes an isolation circuit, a full-wave bridge rectifier, a bus capacitor bank, an inverter, an actuator, and an online capacitor testing device. In the electric drive system topology, unidirectional AC power is input to the isolation circuit. After opto-isolation conversion by the isolation circuit, the AC power is converted to DC power by the full-wave bridge rectifier as the bus voltage. The bus capacitor bank is connected in parallel between the bus voltages, and then the actuator is driven by a three-phase two-level inverter. A PMSM motor (permanent magnet synchronous motor) or a BLDC motor (brushless DC motor) is used as the actuator to actuate the actuator. The inverter includes six switching transistors, which are connected via the electric drive... The controller outputs pulses for phase control; the online capacitance testing device includes a relay, a load resistor, a voltage acquisition module, a current acquisition module, and a main control computing unit; the normally open contact of the relay is connected in series with the load resistor and the current acquisition module, and the relay and the current acquisition module are respectively connected to the positive and negative ends of the bus voltage of the electric drive system; the two ends of the voltage acquisition module are respectively connected to the two ends of the load resistor to acquire the voltage signal of the load resistor, and simultaneously acquire the bus voltage; the current acquisition module is used to acquire the current signal of the load resistor, and simultaneously acquire the bus current; the acquisition of voltage signal and current signal is initiated by the main control computing unit.

[0012] Furthermore, the bus capacitor bank includes 2-5 bus capacitors.

[0013] An online monitoring and early warning method for bus capacitors in an electric drive system, specifically including the following steps:

[0014] When the electric drive system is in an idle state, including after initialization and during shutdown, the electric drive controller energizes the relay coil, connects the load resistor to form a test circuit, and starts the voltage and current acquisition modules to collect current and voltage signals as raw data. After the test is completed, the test circuit is disconnected, and the raw data is used to extract signal features in the time and frequency domain to construct a degradation characterization quantity based on multi-index information fusion. A degradation curve based on an equal time interval sequence is constructed through numerical fitting. The time series combined prediction model (TCN-BRU-GRU) learns and trains trend features on the degradation curve to achieve early warning.

[0015] Step 1: Connect the test circuit and data acquisition;

[0016] Step 2, feature extraction in the time and frequency domains;

[0017] Step 3, construct time series data; the time series data includes the mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak Three characteristic parameters;

[0018] Step 4: Train and predict time series data using the TCN-BRA-GRU combined prediction model; the TCN-BRA-GRU combined prediction model is a comprehensive time series prediction model.

[0019] The mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak The time series data reconstructed from samples with equal time intervals of the three feature parameters is used as a single capacitor aging assessment index. The TCN-BRA-GRU combined prediction model is used to train and predict the trend and changes of the time series data.

[0020] Step 5: Compare the predicted values ​​of the capacitor aging assessment indicators with the failure alarm threshold to make a comprehensive judgment on whether to trigger an alarm.

[0021] Furthermore, in step 1, the connection test circuit and data acquisition process are as follows:

[0022] Step 1.1, based on the bus voltage U of the electric drive system L Select the load resistor R connected with the rated current I. L This ensures that the test circuit current operates at 80% of the rated current. ;

[0023] Step 1.2: The electric drive controller sets all six switching devices of the inverter to be off, so that the inverter and the back-end actuator are in an isolated state.

[0024] Step 1.3: Start the load resistor R through the main control computing unit. L The voltage signal and current signal are acquired, and the bus voltage U is acquired synchronously. L waveform and bus current I L The waveform is used as the raw data.

[0025] Furthermore, in step 1.3, the bus voltage U L The waveform is acquired through a high-voltage differential sampling circuit; the bus current I L The waveform is conditioned and acquired using a linear Hall effect sensor.

[0026] Bus voltage U L waveform and bus current I L The sampling rate for waveform acquisition is set to 30ks / s, and the program sampling time is 2 seconds. Within the 2-second sampling period, the load resistor is connected and disconnected once, with both the connection and disconnection times being 1 second.

[0027] Furthermore, in step 1, the time window control for accessing the test circuit and data acquisition is as follows: during the inverter output stage, the access relay control signal is at a low level; the sampling of bus voltage and bus current is in a non-operating state, i.e., an idle state.

[0028] First, the relay control pulse changes from low to high, the relay contacts are closed, the test circuit is formed, and the sampling of bus voltage and bus current is valid, with a sampling time window of 1 second;

[0029] Then, the relay control pulse returns to a low level, and the test circuit is disconnected.

[0030] Furthermore, in step 2, the feature extraction process in the time domain and frequency domain is as follows:

[0031] The time-domain and frequency-domain characteristics include the bus voltage U L ripple mean V ripple-mean High-frequency energy ratio K and bus current I L Peak current I peak Three characteristic parameters; mean ripple V ripple-mean Used to reflect the average fluctuation range;

[0032] The mean ripple value V ripple-mean The solution process is as follows:

[0033] First, calculate the DC component: by obtaining the bus voltage U. L The mean value is the DC component U. L-mean ;

[0034] Then, extract the ripple component: bus voltage U L Subtract DC component UL-mean The alternating current ripple V is obtained. ripple ;

[0035] Finally, calculate the AC ripple V. ripple The mean of the absolute values ​​V ripple-mean ;

[0036] The process of solving for the high-frequency energy proportion K is as follows:

[0037] First, the bus voltage U collected in step 1.3 is... L The waveform data is subjected to Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal;

[0038] Then, the energy spectrum is calculated: the square of the modulus of the complex number is calculated point by point from the result of the FFT transformation, which is used to represent the energy at the corresponding frequency at each point: energy = (abs(fft_result) / n) 2 , where n represents the data length, and fft_result is the set of complex numbers obtained by FFT;

[0039] Next, process the single-sided spectrum: Considering the symmetry of the real signal, retain the first half of the spectrum, (n / 2)+1 points, convert and retain the effective frequency band, with a sampling rate Fs of 30kHz. According to the Nyquist sampling theorem, the maximum effective frequency band is 15kHz. The frequency corresponding to the x-th point in the (n / 2)+1 FFT calculation results is equal to... Fs = 30K;

[0040] Next, determine the dividing point index: find the frequency corresponding to 10kHz as the dividing point index;

[0041] Finally, calculate the high-frequency energy ratio K: accumulate the low-frequency and high-frequency energy respectively, and calculate the high-frequency energy ratio K; the low frequency is ≤10kHz, and the high frequency is >10kHz;

[0042] The peak current I peak The solution method is as follows:

[0043] Traverse the bus current I in step 1.3 L The waveform data are sorted from largest to smallest, and the 20 largest values ​​are averaged to obtain the peak bus voltage I. peak .

[0044] Furthermore, in step 3, the time series data is constructed as follows:

[0045] Ripple mean V ripple-mean High-frequency energy ratio K and peak current I peakThe sampling time window for the three characteristic parameters is during the initialization of the electric drive system or during the idle state during operation. The data samples collected during the long-term operation are the original time series samples. The time intervals of the original time series samples are randomly distributed. The new time series sample data, i.e., the time series data, is obtained by reconstructing samples with equal time intervals.

[0046] Step 3.1: Based on the data characteristics of the original time series samples and the time step to be predicted, clarify the time interval of the reconstructed samples to generate new samples with equal time intervals. The time interval of the new samples is equal to the time step to be predicted.

[0047] Step 3.2: Reconstruct the original time series samples using the time intervals determined in Step 3.1: The sampling points are represented as... , ,in Indicates the sampling time of the sampling point. This represents the signal value collected at each sampling point; the original time series sample consisting of n sampling points is... After clarifying the time interval of the reconstructed samples, respectively in arrive , arrive … arrive A smooth interpolation curve is generated for the interval using cubic spline interpolation. arrive The time interval is 24 hours. arrive The time interval is 8 hours, interpolated according to a 1-hour time interval. arrive This will reconstruct 22 interpolation samples. arrive Six interpolation samples will be reconstructed from these samples;

[0048] Step 3.3: After interpolation, perform min-max normalization to adjust the values ​​of the time series data to be between 0 and 1; the time series data is... N represents the length of the reconstructed data;

[0049] Ripple mean V ripple-mean High-frequency energy ratio K and peak current I peak The original time series sample data of the three feature parameters were reconstructed at equal time intervals to obtain the time series data of the three feature parameters respectively; the time series data of the three feature parameters were used as aging assessment indicators and modeled and predicted using the TCN-BRA-GRU combined prediction model respectively.

[0050] Furthermore, in step 4, the training and prediction of the time series data are as follows:

[0051] The mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak The time series data reconstructed from the time interval sequences of the three feature parameters is used as a single capacitor aging assessment index. The TCN-BRA-GRU combined prediction model is used to train and predict the trend and changes of the time series data.

[0052] Step 4.1, Obtaining the training and test sets:

[0053] A sliding window with a width of 99 is used to perform sliding sampling on the time series data to obtain the total sample set. Then, 80% of the total sample set is selected as the training set and 20% of the samples are selected as the test set.

[0054] Capacitor aging assessment index: mean ripple V ripple-mean Time series data is It contains n original data points. After sampling through a sliding window, the number of samples is n-99+1. The first sample is... The second sample is ...The last sample is ;

[0055] Capacitor aging assessment indicators include high-frequency energy percentage K and peak current I. peak Time series data partitioning method and capacitor aging assessment index V ripple-mean same;

[0056] Step 4.2: Train and predict time series data using the TCN-BRA-GRU combined prediction model:

[0057] Each sample contains 99 data points. The first 96 data points serve as the input to the TCN-BRA-GRU combined prediction model, representing the trend characteristics of the time series. The last 3 data points serve as the output of the TCN-BRA-GRU combined prediction model, representing 3 follow-up prediction values ​​learned based on the time series characteristics of the first 96 data points.

[0058] Furthermore, in step 4, the TCN-BRA-GRU combined prediction model is a lightweight prediction model with excellent fitting ability and high efficiency. It combines three modules: Temporal Convolutional Network (TCN), Bi-level routing attention (BRA), and Gate Recurrent Unit (GRU). Through a dynamic routing mechanism, it efficiently models the sparse, long-distance spatial dependencies of time series. The combination of the dynamic routing mechanism with TCN's feature capture and dependency expansion, and GRU's sequence dynamic capture, provides an innovative approach to solving long sequence modeling and complex dependencies between variables. In the TCN-BRA-GRU combined prediction model, the TCN layer is used to extract basic features of the time series, the BRA layer establishes long dependencies, captures long-term dependency characteristics in the time series, and extracts the overall trend information of data changes. The GRU layer is used to ensure that the TCN-BRA-GRU combined prediction model understands the dynamic changes of the time series.

[0059] Implementation process of TCN-BRA-GRU combined prediction model: Input the aging assessment index data of individual capacitors into the TCN-BRA-GRU combined prediction model respectively; First, divide the aging assessment index data into training set and test set in a ratio of 8:2. Input the training set into the TCN-BRA-GRU combined prediction model in batches to iterate and solve the model parameters. After the TCN-BRA-GRU combined model converges, verify the generalization ability of the model through the test set. Finally, apply the TCN-BRA-GRU combined model for forward prediction.

[0060] The TCN-BRA-GRU combined prediction model consists of an input layer, a TCN layer, a BRA layer, a GRU layer, a fully connected layer, and an output layer.

[0061] (1) Input layer: The input consists of X_01 to X_96 from a single sample data;

[0062] (2) TCN layer: The TCN layer is a residual topology architecture. The TCN layer contains a residual operation and two combined convolutions. Each combined convolution includes a one-dimensional extended causal convolution, weight normalization, ReLU activation and Dropout operation. The kernel size of the TCN layer is set to 3. The Dropout coefficient is set to 0.2. The Dropout setting can randomly select some neurons to be deactivated to prevent overfitting and accelerate the convergence speed of the TCN layer. The expansion coefficient is set to (1, 2, 4). The filter is 64. The output of the first combined convolution is used as the input of the second combined convolution. The output of the second combined convolution and the input of the first combined convolution are used as the final output of the TCN layer after performing a residual operation. The two inputs of the residual operation have different dimensions. The input of the first combined convolution is passed through a 1×1 convolution layer for dimensionality reduction before being input into the residual operation.

[0063] (3) BRA layer: For time-series data, it implements a dual attention mechanism of region-to-region routing and token-to-token attention. It calculates the relevance based on the aggregated information of time-series data and dynamically selects the source region with the maximum overlap for each target region. The proportion of the source region with the maximum overlap is configurable and is configured to 0.5. Then, it performs intensive token-level attention calculation in the relevant regions selected by coarse-grained routing. While maintaining the ability to capture global long-distance dependencies, it significantly reduces the computational complexity.

[0064] The process by which the BRA layer extracts information about the overall trend of changes in time series data is as follows:

[0065] First, region partitioning and input projection are performed. The time series of length LEN is divided into S non-overlapping regions, each containing 12 time steps (tokens). Region M = LEN / S. Taking LEN = 96 and S = 12, we get M = 8. The input sample sequence is then... Remodeling , where R represents the sample set, B is the BATCH_SIZE value for training the BRA layer, used to define the batch of training data each time, and BATCH_SIZE is the higher-order dimension of the training data, set to 64;

[0066] Then, the mean pooling method is used to aggregate the representation values ​​for each region. calculate; There are 8 time series samples;

[0067] Next, establish region-level attention and calculate the region query key separately. Region key Attention score ;in ; ; ; for The weight matrix, for The weight matrix, for The dimension;

[0068] Next, select the four most relevant source regions with the highest correlation coefficients for each target region;

[0069] Finally, we move to fine-grained attention, collecting all tokens from the four most relevant regions and calculating the attention between tokens in the target region and tokens in the relevant regions. Only four relevant regions are needed, with eight tokens in each region, for a total of 32 relevant tokens, to replace all token patterns in full attention, thus improving the solution speed.

[0070] (4) GRU layer: The GRU layer has long short-term memory function. It uses the update gate to control whether the information of the current time step needs to be updated, and the reset gate to control whether the information of the previous time step needs to be forgotten. The candidate hidden state represents the generation of new memory content, and the hidden state determines whether to retain the old information or adopt the new information through the update gate. The 32 token states of the BRA layer are input to the GRU layer. The GRU layer contains 50 hidden layers. The three intermediate variables of the GRU layer's gated recurrent unit, update gate, reset gate and hidden state, are used to learn the correlation between the 32 tokens of the input time series and establish feature mapping.

[0071] (5) Fully connected layer: The feature maps extracted by the GRU layer are transformed into the output; the fully connected layer enables the TCN-BRA-GRU combined prediction model to perform more complex integration and transformation of the features extracted by the GRU layer, thereby improving the expressive power and performance of the TCN-BRA-GRU combined prediction model.

[0072] (6) Output layer: The output layer consists of three nodes, which are X_97 to X_99 in a single sample data. As the output of the TCN-BRA-GRU combined prediction model, it means that three predicted values ​​following the time series are obtained based on X_01 to X_96 of the input layer.

[0073] The TCN-BRA-GRU combined prediction model uses the Adam optimizer to optimize parameters and reduce loss. Adam is an adaptive optimization algorithm that performs first-order gradient optimization on a stochastic objective function, combining the advantages of the AdaGrad optimizer and the RMSProp optimizer. The loss function L is the mean absolute error (MAE).

[0074]

[0075] Where: K is the total number of samples in the training set; Given the predicted value for the i-th sample, These are the predicted values ​​actually calculated by the TCN-BRA-GRU combined prediction model.

[0076] Furthermore, in step 5, the method for comprehensively determining whether to trigger an alarm is as follows:

[0077] Capacitor degradation is a lengthy process. The acquisition and prediction of time-series data characterizing the parameters require a sufficient dataset, with at least 1000 samples. The TCN-BRA-GRU combined prediction model demonstrates better accuracy and usability in fitting the parameter degradation trend. This is achieved by individually pairing the ripple mean V... ripple-mean High-frequency energy ratio K and peak current I peak After predicting the time series data, the predicted values ​​are calculated separately; the predicted values ​​are compared with the failure alarm threshold to make a comprehensive judgment on whether to trigger an alarm.

[0078] Step 5.1, select the ripple mean V ripple-mean Failure alarm threshold TH_V ripple-mean It is 1.5 times the original value;

[0079] Step 5.2, calculate the ripple mean V. ripple-mean The predicted value and the failure alarm threshold TH_V ripple-mean For comparison, if it exceeds the failure alarm threshold TH_V ripple-mean If so, an alarm will be triggered; otherwise, the high-frequency energy ratio K value will be further determined.

[0080] Step 5.3: Select the failure alarm threshold TH_K of the high-frequency energy ratio K as 1.4 times the original value;

[0081] Step 5.4: Compare the predicted value of the high-frequency energy proportion K with the failure alarm threshold TH_K. If it is greater than the failure alarm threshold TH_K, an alarm is triggered; otherwise, the peak current value I is further determined. peak ;

[0082] Step 5.5, select the peak current value I peak Failure alarm threshold TH-I peak Five times the original value;

[0083] Step 5.6, set the peak current value I peak Predicted value and failure alarm threshold TH-I peak For comparison, if it exceeds the failure alarm threshold TH-I peak If an alarm is triggered, an alert will be issued; otherwise, the data will be collected again and compared once more.

[0084] The technical effects of this invention are as follows: The TCN-BRA-GRU combined prediction model adopted in this invention is a lightweight prediction model with excellent fitting ability and high efficiency. It combines three modules: Temporal Convolutional Network (TCN), Two-Layer Routing Attention (BRA), and Gated Recurrent Unit (GRU). Through a dynamic routing mechanism, it efficiently models sparse, long-distance spatial dependencies in time series. The combination of the dynamic routing mechanism with TCN's feature capture and dilated dependency and GRU's sequence dynamic capture provides an innovative approach to solving long sequence modeling and complex dependencies between variables. This invention uses artificial intelligence algorithms to solve waveform fitting for non-equidistant time sampling, capture local and global data features, and achieve trend prediction of bus capacitor performance degradation. It provides early warning before bus capacitor failure, ensuring the safe and reliable operation of the electric drive system. Attached Figure Description

[0085] Figure 1 This is the equivalent model diagram of a capacitor;

[0086] Figure 2 This diagram illustrates the composition of an electric drive system and an online capacitor testing scheme.

[0087] Figure 3 This is a flowchart of the online prediction process for capacitance.

[0088] Figure 4 This is a schematic diagram of the capacitance test and data acquisition time window;

[0089] Figure 5 Flowchart of applying the TCN-BRA-GRU combined prediction model to a single capacitor aging assessment index;

[0090] Figure 6 Diagram of the TCN-BRA-GRU combined prediction model architecture;

[0091] Figure 7 This is a flowchart for the comprehensive judgment of capacitor failure. Detailed Implementation

[0092] An online monitoring and early warning method for bus capacitors in an electric drive system is disclosed. The electric drive system sequentially includes an isolation circuit, a full-wave bridge rectifier, a bus capacitor bank, an inverter, an actuator, and an online capacitor testing device. The online capacitor testing device includes a relay, a load resistor, a voltage acquisition module, a current acquisition module, and a main control computing unit. When the electric drive system is idle, a test path is established by connecting the load resistor to acquire current and voltage signals. The test path is then disconnected, and a degradation characterization quantity with multi-index information fusion is constructed through time-frequency domain feature extraction, resulting in a degradation curve with equal time intervals. A TCN-BRU-GRU combined prediction model is used to learn and train trend features on the degradation curve. This invention solves the problems of waveform fitting and local and global data feature capture in non-equal time interval sampling, realizes trend prediction of bus capacitor performance degradation, provides early warning before bus capacitor failure, and ensures the safe and reliable operation of the electric drive system.

[0093] The electric drive system sequentially includes an isolation circuit, a full-wave bridge rectifier, a bus capacitor bank, an inverter, an actuator, and an online capacitor testing device; the bus capacitor bank includes 2-5 bus capacitors; for example... Figure 2 As shown, in the electric drive system topology, a unidirectional AC input is isolated. After opto-isolation conversion by the isolation circuit, the AC power is converted into DC power by a full-wave bridge rectifier as the bus voltage. A bus capacitor bank is connected in parallel between the bus voltages, and then a three-phase two-level inverter drives the actuator. A PMSM motor (permanent magnet synchronous motor) or a BLDC motor (brushless DC motor) is used as the actuator to actuate the actuator. The inverter includes six switching transistors, which are phase-controlled by pulses output by the electric drive controller. The online capacitor testing device includes a relay, a load resistor, a voltage acquisition module, a current acquisition module, and a main control computing unit. The normally open contact of the relay is connected in series with the load resistor and the current acquisition module. The relay and the current acquisition module are respectively connected to the positive and negative ends of the electric drive system bus voltage. The two ends of the voltage acquisition module are respectively connected to the two ends of the load resistor to acquire the voltage signal of the load resistor and the bus voltage. The current acquisition module is used to acquire the current signal of the load resistor and the bus current. The acquisition of voltage and current signals is initiated by the main control computing unit.

[0094] An online monitoring and early warning method for bus capacitance in an electric drive system, such as... Figure 3 As shown, the specific steps include the following:

[0095] When the electric drive system is in an idle state, including after initialization and during shutdown, the electric drive controller energizes the relay coil, connects the load resistor to form a test circuit, and starts the voltage and current acquisition modules to collect current and voltage signals as raw data. After the test is completed, the test circuit is disconnected, and the raw data is used to extract signal features in the time and frequency domain to construct a degradation characterization quantity based on multi-index information fusion. A degradation curve based on an equal time interval sequence is constructed through numerical fitting. The time series combined prediction model (TCN-BRU-GRU) learns and trains trend features on the degradation curve to achieve early warning.

[0096] Step 1: Connect the test circuit and data acquisition;

[0097] Step 2, feature extraction in the time and frequency domains;

[0098] Step 3, construct time series data; the time series data includes the mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak Three characteristic parameters;

[0099] Step 4: Train and predict time series data using the TCN-BRA-GRU combined prediction model; the TCN-BRA-GRU combined prediction model is a comprehensive time series prediction model.

[0100] The mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak The time series data reconstructed from samples with equal time intervals of the three feature parameters is used as a single capacitor aging assessment index. The TCN-BRA-GRU combined prediction model is used to train and predict the trend and changes of the time series data.

[0101] Step 5: Compare the predicted values ​​of the capacitor aging assessment indicators with the failure alarm threshold to make a comprehensive judgment on whether to trigger an alarm.

[0102] In step 1, the connection test circuit and data acquisition process are as follows:

[0103] Step 1.1, based on the bus voltage U of the electric drive system L Select the load resistor R connected with the rated current I. L This ensures that the test circuit current operates at 80% of the rated current. ;

[0104] Step 1.2: The electric drive controller sets all six switching devices of the inverter to be off, so that the inverter and the back-end actuator are in an isolated state.

[0105] Step 1.3: Start the load resistor R through the main control computing unit. L The voltage signal and current signal are acquired, and the bus voltage U is acquired synchronously. L waveform and bus current I L The waveform is used as the raw data.

[0106] In step 1.3, the bus voltage U L The waveform is acquired through a high-voltage differential sampling circuit; the bus current I L The waveform is conditioned and acquired using a linear Hall effect sensor.

[0107] Bus voltage U L waveform and bus current I L The sampling rate for waveform acquisition is set to 30ks / s, and the program sampling time is 2 seconds. Within the 2-second sampling period, the load resistor is connected and disconnected once, with both the connection and disconnection times being 1 second.

[0108] In step 1, the time window control for accessing the test circuit and data acquisition is as follows: Figure 4 As shown, during the inverter output stage, the relay control signal is at a low level; the sampling of bus voltage and bus current is in an inactive state, i.e., an idle state.

[0109] First, the relay control pulse changes from low to high, the relay contacts are closed, the test circuit is formed, and the sampling of bus voltage and bus current is valid, with a sampling time window of 1 second;

[0110] Then, the relay control pulse returns to a low level, and the test circuit is disconnected.

[0111] In step 2, the feature extraction process in the time domain and frequency domain is as follows:

[0112] The time-domain and frequency-domain characteristics include the bus voltage U L ripple mean V ripple-mean High-frequency energy ratio K and bus current I L Peak current I peak Three characteristic parameters; mean ripple V ripple-mean Used to reflect the average fluctuation range;

[0113] The mean ripple value V ripple-mean The solution process is as follows:

[0114] First, calculate the DC component: by obtaining the bus voltage U. L The mean value is the DC component U. L-mean ;

[0115] Then, extract the ripple component: bus voltage U LSubtract DC component U L-mean The alternating current ripple V is obtained. ripple ;

[0116] Finally, calculate the AC ripple V. ripple The mean of the absolute values ​​V ripple-mean ;

[0117] The process of solving for the high-frequency energy proportion K is as follows:

[0118] First, the bus voltage U collected in step 1.3 is... L The waveform data is subjected to Fast Fourier Transform (FFT) to convert the time-domain signal into a frequency-domain signal;

[0119] Then, the energy spectrum is calculated: the square of the modulus of the complex number is calculated point by point from the result of the FFT transformation, which is used to represent the energy at the corresponding frequency at each point: energy = (abs(fft_result) / n) 2 , where n represents the data length, and fft_result is the set of complex numbers obtained by FFT;

[0120] Next, process the single-sided spectrum: Considering the symmetry of the real signal, retain the first half of the spectrum, (n / 2)+1 points, convert and retain the effective frequency band, with a sampling rate Fs of 30kHz. According to the Nyquist sampling theorem, the maximum effective frequency band is 15kHz. The frequency corresponding to the x-th point in the (n / 2)+1 FFT calculation results is equal to... Fs = 30K;

[0121] Next, determine the dividing point index: find the frequency corresponding to 10kHz as the dividing point index;

[0122] Finally, calculate the high-frequency energy ratio K: accumulate the low-frequency and high-frequency energy respectively, and calculate the high-frequency energy ratio K; the low frequency is ≤10kHz, and the high frequency is >10kHz;

[0123] The peak current I peak The solution method is as follows:

[0124] Traverse the bus current I in step 1.3 L The waveform data are sorted from largest to smallest, and the 20 largest values ​​are averaged to obtain the peak bus voltage I. peak .

[0125] In step 3, the time series data is constructed as follows:

[0126] Ripple mean V ripple-mean High-frequency energy ratio K and peak current I peakThe sampling time window for the three characteristic parameters is during the initialization of the electric drive system or during the idle state during operation. The data samples collected during the long-term operation are the original time series samples. The time intervals of the original time series samples are randomly distributed. The new time series sample data, i.e., the time series data, is obtained by reconstructing samples with equal time intervals.

[0127] Step 3.1: Based on the data characteristics of the original time series samples and the time step to be predicted, clarify the time interval of the reconstructed samples to generate new samples with equal time intervals. The time interval of the new samples is equal to the time step to be predicted.

[0128] Step 3.2: Reconstruct the original time series samples using the time intervals determined in Step 3.1: The sampling points are represented as... , ,in Indicates the sampling time of the sampling point. This represents the signal value collected at each sampling point; the original time series sample consisting of n sampling points is... After clarifying the time interval of the reconstructed samples, respectively in arrive , arrive … arrive A smooth interpolation curve is generated for the interval using cubic spline interpolation. arrive The time interval is 24 hours. arrive The time interval is 8 hours, interpolated according to a 1-hour time interval. arrive This will reconstruct 22 interpolation samples. arrive Six interpolation samples will be reconstructed from these samples;

[0129] Step 3.3: After interpolation, perform min-max normalization to adjust the values ​​of the time series data to be between 0 and 1; the time series data is... N represents the length of the reconstructed data;

[0130] Ripple mean V ripple-mean High-frequency energy ratio K and peak current I peak The original time series sample data of the three feature parameters were reconstructed at equal time intervals to obtain the time series data of the three feature parameters respectively; the time series data of the three feature parameters were used as aging assessment indicators and modeled and predicted using the TCN-BRA-GRU combined prediction model respectively.

[0131] In step 4, the training and prediction of the time series data are as follows:

[0132] The mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak The time series data reconstructed from the time interval sequences of the three feature parameters is used as a single capacitor aging assessment index. The TCN-BRA-GRU combined prediction model is used to train and predict the trend and changes of the time series data.

[0133] Step 4.1, Obtaining the training and test sets:

[0134] A sliding window with a width of 99 is used to perform sliding sampling on the time series data to obtain the total sample set. Then, 80% of the total sample set is selected as the training set and 20% of the samples are selected as the test set.

[0135] Capacitor aging assessment index: mean ripple V ripple-mean Time series data is It contains n original data points. After sampling through a sliding window, the number of samples is n-99+1. The first sample is... The second sample is ...The last sample is ;

[0136] Capacitor aging assessment indicators include high-frequency energy percentage K and peak current I. peak Time series data partitioning method and capacitor aging assessment index V ripple-mean same;

[0137] Step 4.2: Train and predict time series data using the TCN-BRA-GRU combined prediction model:

[0138] Each sample contains 99 data points. The first 96 data points serve as the input to the TCN-BRA-GRU combined prediction model, representing the trend characteristics of the time series. The last 3 data points serve as the output of the TCN-BRA-GRU combined prediction model, representing 3 follow-up prediction values ​​learned based on the time series characteristics of the first 96 data points.

[0139] In step 4, the TCN-BRA-GRU combined prediction model is a lightweight prediction model with excellent fitting ability and high efficiency. It combines three modules: Temporal Convolutional Network (TCN), Bi-level routing attention (BRA), and Gate Recurrent Unit (GRU). Through a dynamic routing mechanism, it efficiently models the sparse, long-distance spatial dependencies of time series. The combination of the dynamic routing mechanism with TCN's feature capture and dependency expansion, and GRU's sequence dynamic capture, provides an innovative approach to solving long sequence modeling and complex dependencies between variables. In the TCN-BRA-GRU combined prediction model, the TCN layer is used to extract basic features of the time series, the BRA layer establishes long dependencies, captures long-term dependency characteristics in the time series, and extracts the overall trend information of data changes. The GRU layer is used to ensure that the TCN-BRA-GRU combined prediction model understands the dynamic changes of the time series.

[0140] The implementation process of the TCN-BRA-GRU combined prediction model is as follows: Figure 5 As shown; the aging assessment index data of individual capacitors are input into the TCN-BRA-GRU combined prediction model respectively; firstly, the aging assessment index data are divided into training set and test set in a ratio of 8:2. The training set is input into the TCN-BRA-GRU combined prediction model in batches to iterate and solve the model parameters. After the TCN-BRA-GRU combined model converges, the generalization ability of the model is verified by the test set. Finally, the TCN-BRA-GRU combined model is applied for forward prediction.

[0141] The TCN-BRA-GRU combined prediction model consists of an input layer, a TCN layer, a BRA layer, a GRU layer, a fully connected layer, and an output layer, as follows: Figure 6 As shown;

[0142] (1) Input layer: The input consists of X_01 to X_96 from a single sample data;

[0143] (2) TCN layer: The TCN layer is a residual topology architecture. The TCN layer contains a residual operation and two combined convolutions. Each combined convolution includes a one-dimensional extended causal convolution, weight normalization, ReLU activation and Dropout operation. The kernel size of the TCN layer is set to 3. The Dropout coefficient is set to 0.2. The Dropout setting can randomly select some neurons to be deactivated to prevent overfitting and accelerate the convergence speed of the TCN layer. The expansion coefficient is set to (1, 2, 4). The filter is 64. The output of the first combined convolution is used as the input of the second combined convolution. The output of the second combined convolution and the input of the first combined convolution are used as the final output of the TCN layer after performing a residual operation. The two inputs of the residual operation have different dimensions. The input of the first combined convolution is passed through a 1×1 convolution layer for dimensionality reduction before being input into the residual operation.

[0144] (3) BRA layer: For time-series data, it implements a dual attention mechanism of region-to-region routing and token-to-token attention. It calculates the relevance based on the aggregated information of time-series data and dynamically selects the source region with the maximum overlap for each target region. The proportion of the source region with the maximum overlap is configurable and is configured to 0.5. Then, it performs intensive token-level attention calculation in the relevant regions selected by coarse-grained routing. While maintaining the ability to capture global long-distance dependencies, it significantly reduces the computational complexity.

[0145] The process by which the BRA layer extracts information about the overall trend of changes in time series data is as follows:

[0146] First, region partitioning and input projection are performed. The time series of length LEN is divided into S non-overlapping regions, each containing 12 time steps (tokens). Region M = LEN / S. Taking LEN = 96 and S = 12, we get M = 8. The input sample sequence is then... Remodeling , where R represents the sample set, B is the BATCH_SIZE value for training the BRA layer, used to define the batch of training data each time, and BATCH_SIZE is the higher-order dimension of the training data, set to 64;

[0147] Then, the mean pooling method is used to aggregate the representation values ​​for each region. calculate; There are 8 time series samples;

[0148] Next, establish region-level attention and calculate the region query key separately. Region key Attention score ;in ; ; ; for The weight matrix, for The weight matrix, for The dimension;

[0149] Next, select the four most relevant source regions with the highest correlation coefficients for each target region;

[0150] Finally, we move to fine-grained attention, collecting all tokens from the four most relevant regions and calculating the attention between tokens in the target region and tokens in the relevant regions. Only four relevant regions are needed, with eight tokens in each region, for a total of 32 relevant tokens, to replace all token patterns in full attention, thus improving the solution speed.

[0151] (4) GRU layer: The GRU layer has long short-term memory function. It uses the update gate to control whether the information of the current time step needs to be updated, and the reset gate to control whether the information of the previous time step needs to be forgotten. The candidate hidden state represents the generation of new memory content, and the hidden state determines whether to retain the old information or adopt the new information through the update gate. The 32 token states of the BRA layer are input to the GRU layer. The GRU layer contains 50 hidden layers. The three intermediate variables of the GRU layer's gated recurrent unit, update gate, reset gate and hidden state, are used to learn the correlation between the 32 tokens of the input time series and establish feature mapping.

[0152] (5) Fully connected layer: The feature maps extracted by the GRU layer are transformed into the output; the fully connected layer enables the TCN-BRA-GRU combined prediction model to perform more complex integration and transformation of the features extracted by the GRU layer, thereby improving the expressive power and performance of the TCN-BRA-GRU combined prediction model.

[0153] (6) Output layer: The output layer consists of three nodes, which are X_97 to X_99 in a single sample data. As the output of the TCN-BRA-GRU combined prediction model, it means that three predicted values ​​following the time series are obtained based on X_01 to X_96 of the input layer.

[0154] The TCN-BRA-GRU combined prediction model uses the Adam optimizer to optimize parameters and reduce loss. Adam is an adaptive optimization algorithm that performs first-order gradient optimization on a stochastic objective function, combining the advantages of the AdaGrad optimizer and the RMSProp optimizer. The loss function L is the mean absolute error (MAE).

[0155]

[0156] Where: K is the total number of samples in the training set; Given the predicted value for the i-th sample, These are the predicted values ​​actually calculated by the TCN-BRA-GRU combined prediction model.

[0157] In step 5, the method for comprehensively determining whether to trigger an alarm is as follows:

[0158] Capacitor degradation is a lengthy process. The acquisition and prediction of time-series data characterizing the parameters require a sufficient dataset, with at least 1000 samples. The TCN-BRA-GRU combined prediction model demonstrates better accuracy and usability in fitting the parameter degradation trend. This is achieved by individually pairing the ripple mean V... ripple-mean High-frequency energy ratio K and peak current I peak After forecasting the time series data, the predicted values ​​are calculated separately; the predicted values ​​are then compared with the failure alarm threshold to comprehensively determine whether an alarm should be triggered. Figure 7 As shown;

[0159] Step 5.1, select the ripple mean V ripple-mean Failure alarm threshold TH_V ripple-mean It is 1.5 times the original value;

[0160] Step 5.2, calculate the ripple mean V. ripple-mean The predicted value and the failure alarm threshold TH_V ripple-mean For comparison, if it exceeds the failure alarm threshold TH_V ripple-mean If so, an alarm will be triggered; otherwise, the high-frequency energy ratio K value will be further determined.

[0161] Step 5.3: Select the failure alarm threshold TH_K of the high-frequency energy ratio K as 1.4 times the original value;

[0162] Step 5.4: Compare the predicted value of the high-frequency energy proportion K with the failure alarm threshold TH_K. If it is greater than the failure alarm threshold TH_K, an alarm is triggered; otherwise, the peak current value I is further determined. peak ;

[0163] Step 5.5, select the peak current value I peak Failure alarm threshold TH-I peak Five times the original value;

[0164] Step 5.6, set the peak current value I peak Predicted value and failure alarm threshold TH-I peak For comparison, if it exceeds the failure alarm threshold TH-I peakIf an alarm is triggered, an alert will be issued; otherwise, the data will be collected again and compared once more.

Claims

1. A method for online monitoring and early warning of bus capacitance in an electric drive system, characterized in that, Includes the following steps: Step 1: Connect the test circuit and data acquisition; Step 1.1, based on the bus voltage U of the electric drive system L Select the load resistor R connected with the rated current I. L This ensures that the test circuit current operates at 80% of the rated current, that is, ; Step 1.2: The electric drive controller sets all six switching devices of the inverter to be off, so that the inverter and the back-end actuator are in an isolated state. Step 1.3: Start the load resistor R through the main control computing unit. L The voltage signal and current signal are acquired, and the bus voltage U is acquired synchronously. L waveform and bus current I L The waveform is used as the raw data; Step 2, feature extraction in the time and frequency domains; The time-domain and frequency-domain characteristics include the bus voltage U L ripple mean V ripple-mean High-frequency energy ratio K and bus current I L Peak current I peak Three characteristic parameters; mean ripple V ripple-mean Used to reflect the average fluctuation range; Step 3, construct time series data; the time series data includes the mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak Three characteristic parameters; Ripple mean V ripple-mean High-frequency energy ratio K and peak current I peak The sampling time window for the three characteristic parameters is during the initialization of the electric drive system or during the idle state during operation. The data samples collected during the long-term operation are the original time series samples. The time intervals of the original time series samples are randomly distributed. The new time series sample data, i.e., the time series data, is obtained by reconstructing samples with equal time intervals. Step 4: Train and predict time series data using the TCN-BRA-GRU combined prediction model; the TCN-BRA-GRU combined prediction model is a comprehensive time series prediction model. The mean ripple V ripple-mean High-frequency energy ratio K and peak current I peak The time series data reconstructed from samples with equal time intervals of the three feature parameters is used as a single capacitor aging assessment index. The TCN-BRA-GRU combined prediction model is used to train and predict the trend and changes of the time series data. Step 5: Compare the predicted values ​​of the capacitor aging assessment indicators with the failure alarm threshold to make a comprehensive judgment on whether to trigger an alarm.

2. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 1.3, the bus voltage U L The waveform is acquired through a high-voltage differential sampling circuit; the bus current I L The waveform is conditioned and acquired using a linear Hall effect sensor. Bus voltage U L waveform and bus current I L The sampling rate for waveform acquisition is set to 30ks / s, and the program sampling time is 2 seconds. Within the 2-second sampling period, the load resistor is connected and disconnected once, with both the connection and disconnection times being 1 second.

3. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 1, the time window for accessing the test circuit and data acquisition is controlled during the inverter output stage, and the access relay control signal is at a low level; the sampling of bus voltage and bus current is in a non-working state, i.e., an idle state. First, the relay control pulse changes from low to high, the relay contacts are closed, the test circuit is formed, and the sampling of bus voltage and bus current is valid, with a sampling time window of 1 second; Then, the relay control pulse returns to a low level, and the test circuit is disconnected.

4. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 2, the feature extraction process in the time domain and frequency domain is as follows: The mean ripple value V ripple-mean The solution process is as follows: First, calculate the DC component: by obtaining the bus voltage U. L The mean value is the DC component U. L-mean ; Then, extract the ripple component: bus voltage U L Subtract DC component U L-mean The alternating current ripple V is obtained. ripple ; Finally, calculate the AC ripple V. ripple The mean of the absolute values ​​V ripple-mean ; The process of solving for the high-frequency energy proportion K is as follows: First, the bus voltage U collected in step 1.3 is... L The waveform data is subjected to FFT transformation to convert the time-domain signal into a frequency-domain signal; Then, the energy spectrum is calculated: the square of the modulus of the complex number is calculated point by point from the result of the FFT transformation, which is used to represent the energy at the corresponding frequency at each point: energy = (abs(fft_result) / n) 2 , where n represents the data length, and fft_result is the set of complex numbers obtained by FFT; Next, process the single-sided spectrum: Considering the symmetry of the real signal, retain the first half of the spectrum, (n / 2)+1 points, convert and retain the effective frequency band, with a sampling rate Fs of 30kHz. According to the Nyquist sampling theorem, the maximum effective frequency band is 15kHz. The frequency corresponding to the x-th point in the (n / 2)+1 FFT calculation results is equal to... Fs = 30K; Next, determine the dividing point index: find the frequency corresponding to 10kHz as the dividing point index; Finally, calculate the high-frequency energy ratio K: accumulate the low-frequency and high-frequency energy respectively, and calculate the high-frequency energy ratio K; the low frequency is ≤10kHz, and the high frequency is >10kHz; The peak current I peak The solution method is as follows: Traverse the bus current I in step 1.3 L The waveform data are sorted from largest to smallest, and the 20 largest values ​​are averaged to obtain the peak bus voltage I. peak .

5. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 3, the time series data is constructed as follows: Step 3.1: Based on the data characteristics of the original time series samples and the time step to be predicted, clarify the time interval of the reconstructed samples to generate new samples with equal time intervals. The time interval of the new samples is equal to the time step to be predicted. Step 3.2: Reconstruct the original time series samples using the time intervals determined in Step 3.1: The sampling points are represented as... , ,in Indicates the sampling time of the sampling point. This represents the signal value collected at each sampling point; the original time series sample consisting of n sampling points is... After clarifying the time interval of the reconstructed samples, respectively in arrive , arrive … arrive A smooth interpolation curve is generated for the interval using cubic spline interpolation. arrive The time interval is 24 hours. arrive The time interval is 8 hours, interpolated according to a 1-hour time interval. arrive This will reconstruct 22 interpolation samples. arrive Six interpolation samples will be reconstructed from these samples; Step 3.3: After interpolation, perform min-max normalization to adjust the values ​​of the time series data to be between 0 and 1; the time series data is... N represents the length of the reconstructed data.

6. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 4, the training and prediction of the time series data are as follows: Step 4.1, Obtaining the training and test sets: A sliding window with a width of 99 is used to perform sliding sampling on the time series data to obtain the total sample set. Then, 80% of the total sample set is selected as the training set and 20% of the samples are selected as the test set. Capacitor aging assessment index: mean ripple V ripple-mean Time series data is It contains n original data points. After sampling through a sliding window, the number of samples is n-99+1. The first sample is... The second sample is ...The last sample is ; Capacitor aging assessment indicators include high-frequency energy percentage K and peak current I. peak Time series data partitioning method and capacitor aging assessment index V ripple-mean same; Step 4.2: Train and predict time series data using the TCN-BRA-GRU combined prediction model: Each sample contains 99 data points. The first 96 data points serve as the input to the TCN-BRA-GRU combined prediction model, representing the trend characteristics of the time series. The last 3 data points serve as the output of the TCN-BRA-GRU combined prediction model, representing 3 follow-up prediction values ​​learned based on the time series characteristics of the first 96 data points.

7. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 4, the TCN-BRA-GRU combined prediction model is a lightweight prediction model with excellent fitting ability and high efficiency. It combines three modules: Temporal Convolutional Network (TCN), Two-Layer Routing Attention (BRA), and Gated Recurrent Unit (GRU). Through a dynamic routing mechanism, it efficiently models the sparse, long-distance spatial dependencies of time series. The dynamic routing mechanism is combined with the feature capture and dependency expansion of TCN and the sequence dynamic capture of GRU. In the TCN-BRA-GRU combined prediction model, the TCN layer is used to extract the basic features of the time series, the BRA layer establishes long dependencies, captures the long-term dependency characteristics in the time series, and extracts the overall trend information of data changes. The GRU layer is used to help the TCN-BRA-GRU combined prediction model understand the dynamic changes of the time series. Individual capacitor aging assessment index data are input into the TCN-BRA-GRU combined prediction model. First, the aging assessment index data are divided into training set and test set in an 8:2 ratio. The training set is input into the TCN-BRA-GRU combined prediction model in batches to iterate and solve the model parameters. After the TCN-BRA-GRU combined model converges, the generalization ability of the model is verified through the test set. Finally, the TCN-BRA-GRU combined model is applied for forward prediction. The TCN-BRA-GRU combined prediction model consists of an input layer, a TCN layer, a BRA layer, a GRU layer, a fully connected layer, and an output layer. (1) Input layer: The input consists of X_01 to X_96 from a single sample data; (2) TCN layer: The TCN layer is a residual topology architecture. The TCN layer contains a residual operation and two layers of combined convolution. Each layer of combined convolution includes one-dimensional extended causal convolution, weight normalization, ReLU activation and Dropout operation in sequence. The kernel size of the TCN layer is set to 3. The Dropout coefficient is set to 0.

2. The Dropout setting can randomly select some neurons to be deactivated to prevent overfitting during training and accelerate the convergence speed of the TCN layer. The dilation coefficient is set to (1, 2, 4); the filter is 64; the output of the first combined convolution is used as the input of the second combined convolution. The output of the second combined convolution and the input of the first combined convolution are used as the final output of the TCN layer after residual operation. The two inputs of the residual operation have different dimensions. The input of the first combined convolution is passed through a 1×1 convolutional layer for dimensionality reduction before being input into the residual operation. (3) BRA layer: For time-series data, it implements a dual attention mechanism of region-to-region routing and token-to-token attention. It calculates the relevance based on the aggregated information of time-series data and dynamically selects the source region with the maximum overlap for each target region. The proportion of the source region with the maximum overlap is configurable and is set to 0.

5. Then, it performs intensive token-level attention calculation in the relevant regions selected by coarse-grained routing. While maintaining the ability to capture global long-distance dependencies, it significantly reduces the computational complexity. The process by which the BRA layer extracts information about the overall trend of changes in time series data is as follows: First, region partitioning and input projection are performed. The time series of length LEN is divided into S non-overlapping regions, each containing 12 time steps. Region M = LEN / S. Taking LEN = 96 and S = 12, we get M = 8. The input sample sequence is then... Remodeling , where R represents the sample set, B is the BATCH_SIZE value for training the BRA layer, used to define the batch of training data each time, and BATCH_SIZE is the higher-order dimension of the training data, set to 64; Then, average pooling is used to aggregate the characterization values ​​for each region. calculate, There are 8 time series samples; Next, establish region-level attention and calculate the region query key separately. Region key Attention score ;in ; ; ; for The weight matrix, for The weight matrix, for The dimension; Next, select the four most relevant source regions with the highest correlation coefficients for each target region; Finally, we move to fine-grained attention, collecting all tokens from the four most relevant regions and calculating the attention between tokens in the target region and tokens in the relevant regions. Only four relevant regions are needed, with eight tokens in each region, for a total of 32 relevant tokens, to replace all token patterns in full attention, thus improving the solution speed. (4) GRU layer: The GRU layer has long short-term memory function. It uses the update gate to control whether the information of the current time step needs to be updated, and the reset gate to control whether the information of the previous time step needs to be forgotten. The candidate hidden state represents the generation of new memory content, and the hidden state determines whether to retain the old information or adopt the new information through the update gate. The 32 token states of the BRA layer are input to the GRU layer. The GRU layer contains 50 hidden layers. The three intermediate variables of the GRU layer's gated recurrent unit, update gate, reset gate and hidden state, are used to learn the correlation between the 32 tokens of the input time series and establish feature mapping. (5) Fully connected layer: The feature maps extracted by the GRU layer are transformed into the output; the fully connected layer enables the TCN-BRA-GRU combined prediction model to perform more complex integration and transformation of the features extracted by the GRU layer, thereby improving the expressive power and performance of the TCN-BRA-GRU combined prediction model. (6) Output layer: The output layer consists of three nodes, which are X_97 to X_99 in a single sample data. As the output of the TCN-BRA-GRU combined prediction model, it means that three predicted values ​​following the time series are obtained based on X_01 to X_96 of the input layer. The TCN-BRA-GRU combined prediction model uses the Adam optimizer to optimize parameters to reduce loss. Adam is an adaptive optimization algorithm that performs first-order gradient optimization on a stochastic objective function, combining the advantages of the AdaGrad optimizer and the RMSProp optimizer. The loss function L is chosen to be the mean absolute error; ； Where: K is the total number of samples in the training set; Given the predicted value for the i-th sample, These are the predicted values ​​actually calculated by the TCN-BRA-GRU combined prediction model.

8. The method for online monitoring and early warning of bus capacitance in an electric drive system according to claim 1, characterized in that, In step 5, the method for comprehensively determining whether to trigger an alarm is as follows: Step 5.1, select the ripple mean V ripple-mean Failure alarm threshold TH_V ripple-mean It is 1.5 times the original value; Step 5.2, calculate the ripple mean V. ripple-mean The predicted value and the failure alarm threshold TH_V ripple-mean For comparison, if it exceeds the failure alarm threshold TH_V ripple-mean If so, an alarm will be triggered; otherwise, the high-frequency energy ratio K value will be further determined. Step 5.3: Select the failure alarm threshold TH_K of the high-frequency energy ratio K as 1.4 times the original value; Step 5.4: Compare the predicted value of the high-frequency energy proportion K with the failure alarm threshold TH_K. If it is greater than the failure alarm threshold TH_K, an alarm is triggered; otherwise, the peak current value I is further determined. peak ; Step 5.5, select the peak current value I peak Failure alarm threshold TH-I peak Five times the original value; Step 5.6, set the peak current value I peak Predicted value and failure alarm threshold TH-I peak For comparison, if it exceeds the failure alarm threshold TH-I peak If an alarm is triggered, an alert will be issued; otherwise, the data will be collected again and compared once more.

Images