A fault diagnosis method for low-voltage flexible direct current regulator based on artificial intelligence

Abstract

The present invention discloses an artificial intelligence-based fault diagnosis method for a low-voltage flexible direct current regulator, which relates to the field of intelligent power technology. The method comprises the following steps: collecting multimodal operating data through a distributed sensor array, and performing constraint verification and cleaning on the multimodal data based on the laws of electrical physics; inputting the multimodal operating data into a multimodal fusion network, and making the voltage-current characteristics conform to Ohm's law according to a cross-modal attention mechanism, and correcting the feature weights through a differentiable constraint loss; applying constrained perturbations to the feature space using a fast gradient sign method and a projected gradient descent method to obtain an adversarial feature set; constructing a training set, adopting an adversarial training framework, and using a meta-learning mechanism for dynamic adversarial defense; inputting real-time multimodal data into a fault classification model, outputting a fault probability distribution, setting a fault threshold, and comparing and issuing an alarm. The present invention realizes intelligent fault diagnosis for complex working conditions by combining adversarial training with a meta-learning mechanism.

Description

Technical Field

[0001] The present invention relates to the field of intelligent power technology, and in particular to an artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method. Background Art

[0002] Low-voltage flexible DC voltage regulators (LVDCs), core components of smart grids, face complex operating conditions that can lead to faults such as short circuits, overheating, and insulation aging, posing a threat to grid security. Traditional diagnostic methods face limitations. Existing detection methods based on rule matching, threshold analysis, or manual judgment struggle to achieve high-precision, real-time diagnosis in complex electromagnetic environments. A single sensor model further compromises the dimensionality of condition monitoring data. Furthermore, feature conflicts between heterogeneous data sources (such as voltage / current / temperature and humidity) exacerbate the challenges of multimodal fusion. Traditional machine learning and single-source deep learning models are sensitive to feature distribution biases, lack robustness to environmental interference, and lack adversarial defense mechanisms. This leads to a significant drop in the generalization performance of fault classification models under dynamic operating conditions. There is an urgent need to develop an intelligent diagnostic framework that integrates multimodal sensor data, possesses adaptive feature learning, and robust decision-making capabilities. By deeply mining the correlations between multidimensional fault features, this framework can overcome the dynamic diagnostic bottleneck in complex power scenarios and improve grid system reliability. Summary of the Invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] Therefore, the present invention provides an artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method to solve the problem of insufficient generalization ability of traditional fault classification models under complex working conditions.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In the first aspect, the present invention provides a low-voltage flexible direct current regulator fault diagnosis method based on artificial intelligence, which includes: collecting multimodal operating data through a distributed sensor array, the multimodal operating data including the voltage, current, temperature and humidity of the regulator, and constraining and cleaning the multimodal data based on the laws of electrical physics; inputting the multimodal operating data into a multimodal fusion network, extracting high-frequency ripple features, temperature and humidity features and time domain statistics, making the voltage-current features conform to Ohm's law according to the cross-modal attention mechanism, correcting the feature weights through differentiable constraint loss, and generating a multimodal feature vector; using the fast gradient sign method and the projected gradient descent method to impose constrained perturbations in the feature space to obtain an adversarial feature set; mixing the multimodal feature vector with the adversarial feature set to construct a training set, adopting an adversarial training framework and using a meta-learning mechanism for dynamic adversarial defense to obtain a fault classification model; inputting real-time multimodal data into the fault classification model, outputting a fault probability distribution, setting a fault threshold and comparing it for alarm.

[0007] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method of the present invention, wherein: the constraint verification and cleaning of multimodal data based on the laws of power physics includes the following steps:

[0008] Based on Kirchhoff's law and the principle of power conservation, the voltage and current time series data are checked for consistency;

[0009] The temperature data is dynamically calibrated by using heat conduction equation constraints and combining the regulator heat sink parameters.

[0010] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method of the present invention, wherein: the multimodal fusion network includes:

[0011] Extract high-frequency ripple features and time-domain statistics from voltage and current data using a one-dimensional convolutional neural network.

[0012] From the temperature data, a long short-term memory network is used to capture the dynamic thermal response pattern;

[0013] Combining humidity data with regulator heat dissipation parameters, a heat dissipation efficiency correction factor is generated.

[0014] As an optimal solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method described in the present invention, the cross-modal attention mechanism refers to the use of a physical constraint attention gating unit to realize the Ohm's law hard constraint of the voltage-current characteristics.

[0015] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method described in the present invention, the applying of constrained disturbance refers to generating adversarial disturbance in the feature space, and the disturbance range is limited by physical constraints.

[0016] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method described in the present invention, the adversarial training framework refers to the use of a dual-channel adversarial training architecture, the main channel optimizes the fault classifier through cross-entropy loss, the adversarial channel updates the disturbance generator through the maximum-minimum optimization objective, and the dynamic balance factor decays exponentially with the training rounds.

[0017] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method of the present invention, the meta-learning mechanism includes the following steps:

[0018] Extract fault scenarios from the historical fault database as the support set and new faults as the query set;

[0019] Rapid adaptation is achieved through inner loop updates, combined with outer loop optimization.

[0020] As a preferred solution of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method described in the present invention, the fault threshold refers to the use of a dynamic threshold adjustment strategy combined with a heat dissipation efficiency correction coefficient to achieve working condition adaptive alarm triggering.

[0021] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method as described in the first aspect of the present invention is implemented.

[0022] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method as described in the first aspect of the present invention.

[0023] The beneficial effects of the present invention are: the use of power electronic devices to replace traditional mechanical transmission components reduces noise and vibration, and reduces the impact on the environment. Efficient energy utilization helps to reduce energy waste and carbon emissions. The equipment has a small size and weight, which is convenient for installation and transportation, and its control algorithm and mechanical structure also have good adaptability. Through built-in sensors and artificial intelligence control algorithms, real-time monitoring and intelligent fault diagnosis of the equipment are achieved, improving the maintenance efficiency and operational stability of the equipment. The low-voltage flexible direct current voltage regulator adopts advanced intelligent control and real-time monitoring technology in its core technology, and has the characteristics of high efficiency, high protection, and multi-functional integration. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 This is a schematic diagram of the overall architecture of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method in Example 1.

[0026] Figure 2 This is a data collection and preprocessing flow chart of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method in Example 1.

[0027] Figure 3 This is a multimodal fusion network structure diagram of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method in Example 1. DETAILED DESCRIPTION

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0029] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0030] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0031] Example 1, reference Figures 1 to 3 , which is the first embodiment of the present invention, provides a low-voltage flexible direct current regulator fault diagnosis method based on artificial intelligence, comprising the following steps:

[0032] S1: Collect multimodal operating data through a distributed sensor array, where the multimodal operating data includes voltage, current, temperature, and humidity of the voltage regulator, and perform constraint verification and cleaning on the multimodal data based on the laws of electrical physics.

[0033] Specifically, the following steps are included:

[0034] S1.1: Conduct distributed sensor array deployment.

[0035] Specifically, high-frequency voltage sensors are used to collect voltage data. They have a range of ±500 VDC, a sampling rate of 2 MHz (meeting 50 kHz ripple capture), and an accuracy of ±0.1% FS. They are deployed at the positive and negative poles of the DC bus (two redundant sensors are configured for each phase).

[0036] Current data is collected using a Rogowski coil ammeter with a DC-200kHz bandwidth, a linearity error of <0.05%, and a temperature drift of ±10ppm / °C. It is deployed at the IGBT module input / output terminals.

[0037] Fiber optic temperature sensors are used to collect temperature data. The measurement range is -40°C to +150°C, the accuracy is ±0.3°C, and the response time is <10ms. They are deployed on the power device substrate (configured with three-point temperature measurement).

[0038] Humidity data is collected using a MEMS humidity sensor with a measurement range of 0-100% RH and an accuracy of ±1.5% RH. The sensor is deployed inside the sealed cavity of the device (two sensors are arranged diagonally).

[0039] IEEE 1588 Precision Time Protocol (PTP) is used for data synchronization to achieve multi-channel data time alignment (error < 1μs).

[0040] S1.2: Perform constraint checks on multimodal data based on the laws of electrical physics.

[0041] Specifically, Kirchhoff's law is used to perform current calibration and calculate the impedance deviation rate for each sampling point. The expression is:

[0042] ;

[0043] in, is the impedance deviation rate, is the voltage data, is the current data, is the calibration resistance, is the time variable.

[0044] When the impedance deviation rate is greater than 5% and lasts for more than 10 μs, it is marked as abnormal data.

[0045] Constraints are imposed by the three-dimensional heat conduction equation, which is expressed as:

[0046] ;

[0047] in, is the temperature field, is the partial derivative identifier, is the thermal diffusion coefficient, which is determined by the heat sink material, is the second-order derivative term in space, describing the conduction rate of temperature in three-dimensional space. is the heat loss term, is the volumetric heat capacity of the material, is the temperature, , , are the three orthogonal directions that constitute the three-dimensional rectangular coordinate system.

[0048] S1.3: Identify and process abnormal data.

[0049] Specifically, for current and voltage data, if five consecutive sampling points exceed the range of the average value ±3 times the historical standard deviation, or the impedance deviation rate is greater than 5% and lasts for more than 10μs, it is judged as abnormal and replaced with the sliding average of the previous 1ms.

[0050] For temperature data, if If the duration is greater than 10ms, it is considered abnormal, marked and switched to the backup sensor.

[0051] For humidity data, if the difference between adjacent sensors is greater than 5%RH, it is judged as abnormal and the Kalman filter is used to fuse the humidity data.

[0052] Preferably, by deploying high-frequency voltage sensors, Rogowski coil ammeters, fiber optic temperature sensors and redundant MEMS humidity sensor arrays, combined with the IEEE 1588 protocol to achieve microsecond-level multi-channel data synchronization, a high-precision multimodal data acquisition system is constructed; based on the impedance deviation rate verification of Kirchhoff's law and the dynamic constraint of the three-dimensional heat conduction equation, the circuit physical consistency verification of voltage and current and the matching of the temperature field evolution law are realized; through the sliding window statistical threshold, thermal change rate threshold and redundant sensor difference detection, a multi-dimensional anomaly discrimination mechanism is constructed, and intelligent repair strategies such as sliding mean replacement and Kalman filter fusion are adopted to improve the accuracy, integrity and real-time performance of data acquisition in complex power environments, providing a highly reliable data foundation for power network status monitoring and fault warning.

[0053] S2: Input the multimodal operating data into the multimodal fusion network to extract high-frequency ripple features, temperature and humidity features, and time-domain statistics. Based on the cross-modal attention mechanism, the voltage-current features are made to conform to Ohm's law. The feature weights are corrected through the differentiable constrained loss to generate a multimodal feature vector.

[0054] Specifically, the following steps are included:

[0055] S2.1: Use 1D-CNN network to extract high-frequency ripple features from voltage and current data.

[0056] Specifically, the one-dimensional convolution is set to dual input channels of voltage and current, the number of output channels is 32, the convolution kernel size is 5, and the convolution stride is 2.

[0057] Select the ReLU activation function, use the pooling layer for dimensionality reduction, extract higher-level features through the second layer of convolution, and normalize to prevent overfitting.

[0058] The high-frequency ripple characteristic expression is:

[0059] ;

[0060] in, is the high frequency ripple characteristic, is the convolution kernel index, To get the maximum value identifier, For the The time series output of the convolution kernel, is the Sigmoid activation function, is the weight coefficient of the pooling layer, which is automatically optimized through back propagation. Output features for the pooling layer.

[0061] Extract the voltage high-frequency ripple characteristics and current high frequency ripple characteristics .

[0062] S2.2: Extract time-domain statistics of voltage and current data.

[0063] It should be noted that the time domain statistics include the waveform factor, kurtosis, pulse factor and margin factor of the voltage and current data.

[0064] The form factor is calculated by the ratio of the peak value to the root mean square value of the voltage and current data, reflecting the relationship between the peak value and the effective value. A larger form factor indicates the presence of a larger mutation or pulse.

[0065] By calculating the mean and standard deviation of the voltage and current data, and calculating the fourth power deviation of each sample point relative to the mean, the mean is calculated and normalized to obtain the kurtosis. The larger the kurtosis, the more likely there are larger pulses or abnormal data points in the voltage and current data.

[0066] By calculating the peak and average absolute values of the voltage and current data and calculating the ratio, the pulse factor is obtained to reflect whether the voltage and current data have prominent pulse components. A larger pulse factor usually indicates that the voltage and current data have stronger short-term impacts.

[0067] The envelope root mean square value of the voltage and current data is calculated by Hilbert transform, and the margin factor is calculated in combination with the peak value. It measures the ratio of the maximum peak value of the voltage and current data to the envelope RMS. A high margin factor indicates that the voltage and current data may contain high-amplitude transient events.

[0068] S2.3: Use BiLSTM to extract thermal dynamic features from temperature data.

[0069] Specifically, a bidirectional LSTM (BiLSTM) network was constructed with an input feature dimension of 1 and a hidden layer dimension of 32. A bidirectional LSTM was used. The attention mechanism was used, with the query vector (Query) and key vector (Key) dimensions set to 64 and a normalization coefficient of 0.1.

[0070] The thermal dynamic characteristic expression is:

[0071] ;

[0072] in, is the thermal dynamic characteristic, is the humidity change rate, is the natural base, Starts the device in timestamp.

[0073] S2.4: Voltage-current feature correction using cross-modal attention mechanism.

[0074] Specifically, based on the high-frequency ripple characteristics of current and voltage, the current and voltage query vectors, key vectors, and value vectors are generated through linear changes.

[0075] The attention score of voltage to current is calculated by scaling the dot product attention, taking the calibration resistance as prior knowledge and constraining the attention output. The expression is:

[0076] ;

[0077] in, To constrain the attention output, is the current and voltage query vector, is the key vector, is a value vector, is the feature dimension, is the normalization function.

[0078] The original current and voltage features and the constrained attention output are fused through a gating mechanism, and the expression is:

[0079] ;

[0080] ;

[0081] in, is the gating coefficient, is the gate weight matrix, obtained through training, is the gate bias term, is the final fusion feature, is element-wise multiplication.

[0082] The final corrected voltage-current characteristic is , whose implicit relationship satisfies Ohm's law and preserves high-frequency ripple details.

[0083] S2.5: Correct feature weights via differentiable constrained loss.

[0084] Specifically, the physical constraint loss function expression is:

[0085] ;

[0086] in, is the physical constraint loss function, is the number of sampling points in the time window, is the index of the number of sampling points, For the The voltage data of the samples, For the Current data of samples.

[0087] The square term enforces Ohm's law constraints, and the exponential term balances the degrees of freedom and physical consistency.

[0088] Optimally, by combining deep learning with physical constraints, multimodal data fusion is achieved through high-frequency ripple feature extraction, time-domain statistical analysis, BiLSTM thermal dynamic modeling, cross-modal attention correction of voltage-current features, and differentiable constrained loss optimization. This solution not only accurately extracts device operating characteristics and improves prediction accuracy, but also ensures that the characteristics conform to Ohm's law through physical constraints, enhancing the method's reliability and generalization capabilities. Furthermore, end-to-end optimization reduces computational complexity and makes inference more efficient.

[0089] S3: Use the fast gradient sign method and projected gradient descent method to impose constrained perturbations in the feature space to obtain an adversarial feature set.

[0090] Specifically, the following steps are included:

[0091] S3.1: Generate initial perturbations using the Fast Gradient Sign Method (FGSM).

[0092] Specifically, the loss gradient of the final fusion feature is calculated by backpropagation, the perturbation direction is generated and the perturbation is applied using the fast gradient sign method to generate the initial perturbation sample.

[0093] S3.2: Iteratively optimize the perturbation using projected gradient descent (PGD).

[0094] It should be understood that the initial perturbation samples generated by FGSM are too simple to simulate complex adversarial attacks. Therefore, we further use the projected gradient descent (PGD) method to optimize and make the perturbation more deceptive while limiting the range.

[0095] Specifically, based on the initial perturbation generated by FGSM, multiple iterative updates are performed, each time adjusting the perturbation direction based on the current high-frequency ripple characteristics, temperature and humidity characteristics, and the gradient of time-domain statistics to obtain the current perturbation characteristics. During each iteration, the loss of the current perturbation characteristic input is calculated, and the loss gradient is calculated. Based on the loss gradient, the perturbation amount is updated to move in the direction of increasing error.

[0096] To ensure that the perturbation does not destroy the overall physical properties of the feature, it is necessary to limit the perturbation to a reasonable range after each update to ensure that it does not exceed the limited range. Projection methods are used to project the perturbation back to the specified limited range, limiting the maximum value of the perturbation or ensuring that the perturbation does not exceed the statistical deviation.

[0097] The above process is continued until the loss value of the current perturbation feature converges or the maximum number of iterations is reached.

[0098] S3.2: Apply perturbations to the final fusion features to generate adversarial feature sets.

[0099] Preferably, adversarial perturbations are applied to the feature space through FGSM and PGD to enhance the robustness and defense capabilities of the fault classification model. FGSM generates initial perturbations, improving the fault classification model's ability to cope with small perturbations, while PGD further optimizes the perturbations to make them more aggressive, simulating complex adversarial attack scenarios. To maintain physical consistency, the PGD process limits the perturbations to a certain extent, ensuring that the physical meaning of the features remains unchanged. By applying perturbations to different feature modes, an adversarial feature set is generated, which can be used for fault classification model training, robustness assessment, and safety analysis, ensuring that the fault classification model maintains stable and efficient prediction capabilities even in complex environments.

[0100] S4: Mix the multimodal feature vectors with the adversarial feature set to construct a training set, adopt the adversarial training framework and use the meta-learning mechanism for dynamic adversarial defense to obtain a fault classification model.

[0101] Specifically, the following steps are included:

[0102] S4.1: Mix multimodal feature vectors and adversarial feature sets to construct training sets.

[0103] Specifically, the multimodal feature vector and the adversarial feature set are mixed in a 3:1 ratio to ensure that the fault classification model can learn both normal features and adapt to adversarial interference.

[0104] S4.2: Build an adversarial training framework.

[0105] Specifically, cross entropy loss is used as the loss function, adversarial perturbation loss is added, and a joint training objective is set.

[0106] The loss function expression is:

[0107] ;

[0108] in, is the loss function, is the training set, is the predicted probability distribution.

[0109] The dynamic adversarial training process uses an alternating training strategy: In the first phase, training is performed on raw data, with the adversarial example generator fixed and only the multimodal feature vectors used to update the fault classification model parameters. In the second phase, adversarial training is performed, with the fault classification model fixed and new adversarial examples generated to update the training set. Alternating between the first and second phases gradually improves the robustness of the fault classification model.

[0110] S4.3: Dynamic adversarial defense using meta-learning mechanisms.

[0111] Specifically, a batch of data is randomly sampled from the training set as the support set, and the newly generated adversarial perturbation is applied to the support set to obtain the query set.

[0112] The meta-loss function expression is:

[0113] ;

[0114] in, is the meta-loss function, are the fault classification model parameters, For the support set, The cloud learning weight is set to 1.2 to enhance the generalization ability of the fault classification model to unknown attacks. For the query set.

[0115] The fault classification model parameters are updated by calculating the meta-gradient, which is expressed as:

[0116] ;

[0117] ;

[0118] in, is the meta-gradient, To calculate the gradient of the meta-loss for the fault classification model parameters, it represents the optimization direction of the fault classification model. is the meta-learning rate (set to 0.001), which controls the update step size of the fault classification model parameters and affects the speed at which the fault classification model learns new tasks.

[0119] Furthermore, the perturbation statistics of the input features (such as feature norm deviation) are monitored in real time. If a perturbation is detected (the deviation exceeds a threshold of 0.1), the meta-learning optimizer is activated to quickly adjust the fault classification model parameters.

[0120] Optimally, the multimodal feature vectors and adversarial feature sets are mixed in a 3:1 ratio, ensuring that the fault classification model can learn both normal operating conditions and adapt to adversarial perturbations. A dynamic adversarial training strategy is employed, alternating between normal and adversarial training, enabling the fault classification model to maintain high classification performance even in the face of unseen attacks. Finally, a meta-learning mechanism is introduced to optimize the fault classification model using the support set and adversarial query set. This allows the model to quickly adapt to new attacks and adjust parameters in real time when anomalies in input features are detected, improving generalization capabilities.

[0121] S5: Input real-time multimodal data into the fault classification model, output the fault probability distribution, set the fault threshold and compare it to issue an alarm.

[0122] Specifically, the following steps are included:

[0123] S5.1: Set the fault threshold.

[0124] Specifically, the fault threshold expression is:

[0125] ;

[0126] in, For the Failure threshold for class failure, is the fault category index, For the The historical average failure probability of this type of failure, For the The standard deviation of the class fault, It is an experience factor, and its value is between 2 and 3.

[0127] S5.2: Compare the fault probability distribution with the fault threshold and issue an alarm.

[0128] It should be understood that if the predicted failure probability is greater than the failure threshold, it is determined that a failure has occurred and an alarm is triggered; otherwise, it is determined that no failure has occurred and no alarm is triggered.

[0129] Alarms include fault type, probability of occurrence, trigger timestamp, and associated devices and regions. Local alarms are generated and remote notifications are sent to maintenance personnel. Alarm information is logged for later analysis and inspection.

[0130] Ideally, by setting fault thresholds based on historical data and standard deviations and inputting real-time multimodal data into a fault classification model, the probability of a fault occurring can be accurately determined and an alarm triggered. This approach effectively reduces false positives and missed positives, ensuring accurate fault detection and real-time response. Furthermore, through a combination of local alarms and remote notifications, fault information is promptly communicated to maintenance personnel, and detailed alarm logs are recorded for subsequent analysis and optimization. Adaptive threshold adjustment, intelligent decision-making mechanisms, and fault prediction capabilities help improve equipment stability and reliability, reduce unexpected failures, and enhance maintenance efficiency.

[0131] This embodiment also provides a computer device, which is suitable for the low-voltage flexible direct current regulator fault diagnosis method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the low-voltage flexible direct current regulator fault diagnosis method based on artificial intelligence proposed in the above embodiment.

[0132] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0133] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0134] In summary, the present invention reduces noise and vibration and reduces the impact on the environment by: using power electronic devices to replace traditional mechanical transmission components. Efficient energy utilization helps to reduce energy waste and carbon emissions. The equipment has a small size and weight, which is convenient for installation and transportation, and its control algorithm and mechanical structure also have good adaptability. Through built-in sensors and artificial intelligence control algorithms, real-time monitoring and intelligent fault diagnosis of the equipment are achieved, improving the maintenance efficiency and operational stability of the equipment. The low-voltage flexible direct current voltage regulator adopts advanced intelligent control and real-time monitoring technology in its core technology, and has the characteristics of high efficiency, high protection, and multi-functional integration.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A fault diagnosis method for a low-voltage flexible direct current regulator based on artificial intelligence, characterized by: include, Collect multimodal operating data, including voltage, current, temperature, and humidity of the low-voltage flexible DC voltage regulator, through a distributed sensor array, and perform constraint verification and cleaning on the multimodal data based on the laws of electrical physics; Multimodal operating data is fed into a multimodal fusion network to extract high-frequency ripple features, temperature and humidity features, and time-domain statistics. The voltage-current features are then made to conform to Ohm's law using a cross-modal attention mechanism. Feature weights are then modified using a differentiable constraint loss to generate a multimodal feature vector. The multimodal fusion network comprises: Extract high-frequency ripple features and time-domain statistics from voltage and current data using a one-dimensional convolutional neural network. Extract thermal dynamic features from temperature data using a bidirectional long short-term memory network; Combine humidity data and low-voltage flexible DC voltage regulator heat dissipation parameters to generate heat dissipation efficiency correction coefficient; The cross-modal attention mechanism refers to the use of physical constraint attention gating units to implement Ohm's law hard constraints on voltage-current characteristics; The fast gradient sign method and projected gradient descent method are used to impose constrained perturbations in the feature space to obtain an adversarial feature set; The constrained perturbation is to generate adversarial perturbations in the feature space, and the perturbation range is limited by physical constraints; The multimodal feature vectors are mixed with the adversarial feature set to construct a training set. The adversarial training framework and meta-learning mechanism are used for dynamic adversarial defense to obtain a fault classification model. Input real-time multimodal data into the fault classification model, output the fault probability distribution, set the fault threshold and compare it to issue an alarm.

2. The artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to claim 1, characterized in that: The constraint checking and cleaning of multimodal data based on the laws of electrical physics includes the following steps: Based on Kirchhoff's law and the principle of power conservation, the voltage and current time series data are checked for consistency; The temperature data is dynamically calibrated by using the heat conduction equation constraint and combining the low-voltage flexible DC voltage regulator radiator parameters.

3. The artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to claim 1, characterized in that: The adversarial training framework adopts a dual-channel adversarial training architecture. The main channel optimizes the fault classifier through cross-entropy loss, and the adversarial channel updates the perturbation generator through the maximum-minimum optimization objective. The dynamic balance factor decays exponentially with the training rounds.

4. The artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to claim 1, characterized in that: The meta-learning mechanism includes the following steps: Extract fault scenarios from the historical fault database as the support set and new faults as the query set; Rapid adaptation is achieved through inner loop updates, combined with outer loop optimization.

5. The artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to claim 1, characterized in that: The fault threshold refers to the use of a dynamic threshold adjustment strategy combined with a heat dissipation efficiency correction coefficient to achieve working condition adaptive alarm triggering.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to any one of claims 1 to 5 are implemented.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence-based low-voltage flexible direct current regulator fault diagnosis method according to any one of claims 1 to 5 are implemented.

Images