A method and system for diagnosing faults in a multi-port energy router

Abstract

The present application belongs to the technical field of power electronic equipment fault diagnosis, and discloses a kind of multi-port energy router fault diagnosis method and system.The method obtains the load side voltage, input side voltage, input current and output current of the monitored DC-DC interface module, after pre-processing and synchronization windowing, constructs time series data matrix, periodic characteristic matrix and physical perception attention vector;Through the time series feature branch and periodic feature branch of shared diagnostic model, respectively extract features, adjust the time series feature weight using the physical perception attention vector, and fuse the two types of features;Through the unified fault classifier, the classification probability of each preset state category is obtained, and the diagnosis result is determined accordingly.The system is used to implement the above method.The present application can identify open-circuit fault of power switch device and load side voltage sensor fault in the same diagnosis process, reducing the confusion of different fault types.

Description

Technical Field

[0001] This invention belongs to the field of power electronic equipment condition monitoring and fault diagnosis technology, and more specifically, relates to a fault diagnosis method and system for a multi-port energy router. Background Technology

[0002] Multi-port energy routers are power electronic conversion devices used to realize energy exchange and power dispatch between multiple energy ports. They can be applied to scenarios such as port integrated energy systems, DC microgrids, renewable energy consumption, and the access of various types of loads. A multi-port energy router typically includes a common DC bus, multiple port power conversion modules, port voltage and current sampling units, an energy management controller, and communication interfaces. Different ports can be connected to photovoltaic systems, energy storage systems, shore power, DC loads, electric ship charging equipment, or electric heavy truck charging equipment, etc.

[0003] In multi-port energy routers, DC-DC interface modules are used to connect the common DC bus to the corresponding energy ports to achieve voltage matching and energy transfer between different DC voltage levels. Dual Active Bridge (DAB) type isolated bidirectional DC-DC modules typically include a primary-side full-bridge, a secondary-side full-bridge, a high-frequency transformer, a series inductor, input and output-side filter capacitors, and port voltage and current sensors, enabling electrical isolation and bidirectional power transfer.

[0004] During the operation of a multi-port energy router, the power switching devices in the DC-DC interface module may experience open-circuit faults or drive abnormalities. Port voltage and current sensors may also experience faults such as open circuits, biasing, gain drift, or jamming. These faults can cause port power imbalances, fluctuations in the common DC bus voltage, or port outages, and in severe cases, may even affect the normal operation of other ports. Especially under closed-loop control conditions, open-circuit faults in power switching devices will reduce effective power transmission capability; sensor faults may distort feedback signals and cause incorrect controller adjustments. Therefore, both types of faults may manifest as abnormal output voltage, current, or port power, and their external characteristics share certain similarities.

[0005] Existing fault protection methods are mostly based on overcurrent, overvoltage, undervoltage, fuse operation, or fixed thresholds for judgment. They typically only implement protection after a fault has caused a significant exceedance of electrical limits, limiting early fault identification and further determination of fault ports and fault types. On the other hand, existing fault diagnosis technologies usually detect open-circuit faults in the switching devices or sensor faults of the power converter. For example, open-circuit faults can be identified based on the transmission power characteristics under different operating states of the converter, or sensor anomalies can be judged by constructing state models and residuals.

[0006] However, during the operation of multi-port energy routers, open-circuit faults in power switching devices and sensor faults can occur separately, and the two types of faults exhibit certain similarities in their abnormal manifestations of external electrical quantities such as output voltage, current, and port power. If separate diagnostic algorithms are configured for each type of fault, or if judgments are made based solely on a single electrical quantity or a single fault characteristic, it not only increases the complexity of the diagnostic system but also easily leads to fault type confusion, making it difficult to accurately determine the port where the fault occurs, the fault type, and whether the anomaly originates from the sensor feedback link. Therefore, there is still room for improvement in how to use a single diagnostic algorithm to identify both open-circuit faults in power switching devices and sensor faults, and to improve the reliability of fault diagnosis under different operating conditions. Summary of the Invention

[0007] In view of the fact that the external electrical quantity changes of open circuit faults in power switching devices and sensor faults in the prior art are somewhat similar, and that using independent diagnostic algorithms can easily increase the complexity of the diagnostic system and may cause confusion of fault types, this invention provides a fault diagnosis method and system for multi-port energy routers.

[0008] This invention constructs a multi-channel timing data matrix, an input current cycle feature matrix, and a physical sensing attention vector, and processes the above information using the same shared diagnostic model to achieve unified identification of open-circuit faults in power switching devices and load-side voltage sensor faults.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] In a first aspect, the present invention provides a method for diagnosing faults in a multi-port energy router, comprising: The load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router are acquired to form multi-channel timing data; The multi-channel time-series data is preprocessed and synchronized windowed to obtain multiple diagnostic windows; For each diagnostic window, a timing data matrix is ​​constructed based on the multi-channel timing data, a periodic feature matrix is ​​constructed based on the waveform of the input current in multiple switching cycles, and a physical sensing attention vector is constructed based on the multi-channel timing data. The time-series data matrix is ​​input into the time-series feature branch of the shared diagnostic model to extract time-series features, and the physical perception attention vector is encoded as an attention bias to adjust the feature weights of the time-series features. The periodic feature matrix is ​​input into the periodic feature branch of the shared diagnostic model to extract periodic features, and the time-series features adjusted by feature weights are fused with the periodic features to obtain fused features; The fused features are input into the unified fault classifier of the shared diagnostic model to obtain the classification probability of the monitored DC-DC interface module in each preset state category, and the diagnostic result of the monitored DC-DC interface module is determined based on the classification probability. The preset state categories include healthy state, open circuit fault of power switching device, and load-side voltage sensor fault.

[0011] Optionally, the monitored DC-DC interface module is a dual active bridge isolated bidirectional DC-DC interface module, including an input port, a primary full bridge, a series inductor, a high-frequency transformer, a secondary full bridge, and an output port.

[0012] The input voltage and the input current are collected at the input port, and the load voltage and the output current are collected at the output port.

[0013] Optionally, the preprocessing includes low-pass filtering of the load-side voltage, input-side voltage, input current, and output current.

[0014] The synchronous windowing process includes synchronously segmenting the load-side voltage, input-side voltage, input current, and output current after low-pass filtering according to a unified time base, and grouping data belonging to the same sampling period into the same diagnostic window.

[0015] The load-side voltage, input-side voltage, input current, and output current within each diagnostic window are normalized, and the normalized data are arranged according to the measurement channel to form the timing data matrix.

[0016] The physical perception attention vector is constructed based on the load-side voltage, input-side voltage, input current, and output current corresponding to each diagnostic window.

[0017] Optionally, constructing the periodic feature matrix includes: The start time of each switching cycle is determined based on the switching reference signal of the monitored DC-DC interface module, and the input current in the diagnostic window is divided into multiple periodic signal segments according to each start time, with each periodic signal segment corresponding to a switching cycle. Based on the positive and negative polarities of the switch reference signal in each of the periodic signal segments, the input current sampling values ​​in each of the periodic signal segments are divided into positive half-cycle sampling values ​​and negative half-cycle sampling values. The positive half-cycle energy, negative half-cycle energy, positive half-cycle peak value, and negative half-cycle peak value are determined based on the positive half-cycle sampling value and the negative half-cycle sampling value, respectively, and the period average value is determined based on the input current sampling value within the same period signal segment. The half-cycle energy ratio is determined based on the positive half-cycle energy and the negative half-cycle energy, and the positive half-cycle peak value ratio is determined based on the positive half-cycle peak value and the negative half-cycle peak value. According to the order of each periodic signal segment in the diagnostic window, the half-cycle energy ratio, positive and negative half-cycle peak ratio and period mean corresponding to each periodic signal segment are respectively composed into a periodic feature sequence, and the periodic feature matrix is ​​constructed based on the periodic feature sequence.

[0018] Optionally, constructing the physical perception attention vector includes: The window input power is determined based on the input side voltage and input current at the same sampling time within the diagnostic window, and the window output power is determined based on the load side voltage and output current at the same sampling time. The absolute value of the difference between the window input power and the window output power is determined as the power imbalance, and the average DC value of the input current, the standard deviation of the input voltage, and the standard deviation of the load voltage within the diagnostic window are determined respectively. The above physical quantities are dimensionless according to the preset normalization parameters corresponding to the DC average of the input current, the power imbalance, the standard deviation of the input voltage, and the standard deviation of the load voltage. The dimensionless input current DC mean, power imbalance, input voltage standard deviation, and load voltage standard deviation are used to form the physical perception attention vector.

[0019] Optionally, the temporal feature branch includes a local pattern learning layer, a multi-scale convolutional module, and a physical awareness attention module.

[0020] The local pattern learning layer is used to perform local smoothing on the temporal data matrix. The multi-scale convolution module is used to extract temporal features at different time scales from the locally smoothed temporal data matrix. The physical perception attention module is used to encode the physical perception attention vector into the attention bias and adjust the feature weights of the temporal features according to the attention bias.

[0021] The periodic feature branch includes a periodic feature projection module and a periodic attention module.

[0022] The periodic feature projection module is used to map the periodic features corresponding to each switching period in the periodic feature matrix to a preset feature space, and the periodic attention module is used to perform attention weighting on the periodic features corresponding to different switching periods after mapping.

[0023] The shared diagnostic model also includes a feature fusion module, which is used to perform pooling processing on the time-series features adjusted by feature weights and the periodic features weighted by attention, and to concatenate the pooling results to obtain the fused features.

[0024] Optionally, the preset state categories include health status, primary-side power switch device open circuit fault, secondary-side power switch device open circuit fault, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault.

[0025] The unified fault classifier calculates the classification probability of the monitored DC-DC interface module in each of the preset state categories based on the fusion features, and determines the preset state category with the highest classification probability as the state category of the corresponding diagnostic window.

[0026] The number of occurrences of each preset status category is counted based on the status categories of multiple consecutive diagnostic windows, and the preset status category with the highest number of occurrences is determined as the final diagnostic result.

[0027] When outputting fault categories, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault are grouped into load-side voltage sensor faults.

[0028] Secondly, the present invention provides a multi-port energy router fault diagnosis system, including a data acquisition module, a preprocessing and windowing module, an auxiliary feature construction module, a shared diagnostic model, and a diagnostic result determination module.

[0029] The data acquisition module is used to acquire the load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router, forming multi-channel timing data.

[0030] The preprocessing and windowing module is used to preprocess and synchronously window the multi-channel time series data to obtain multiple diagnostic windows, and to construct a time series data matrix based on the multi-channel time series data corresponding to each diagnostic window.

[0031] The auxiliary feature construction module is used to construct a periodic feature matrix based on the waveform of the input current in each diagnostic window over multiple switching cycles, and to construct a physical sensing attention vector based on the multi-channel timing data corresponding to each diagnostic window.

[0032] The shared diagnostic model includes a time-series feature branch, a periodic feature branch, a feature fusion module, and a unified fault classifier.

[0033] The temporal feature branch is used to extract temporal features from the temporal data matrix and encode the physical perception attention vector as an attention bias to adjust the feature weights of the temporal features.

[0034] The periodic feature branch is used to extract periodic features from the periodic feature matrix.

[0035] The feature fusion module is used to fuse the time-series features adjusted by feature weights and the periodic features to obtain fused features.

[0036] The unified fault classifier is used to calculate the classification probability of the monitored DC-DC interface module being in each preset state category based on the fusion features.

[0037] The diagnostic result determination module is used to determine the diagnostic result of the monitored DC-DC interface module based on the classification probability.

[0038] The preset state categories include healthy state, open circuit fault of power switching device, and load-side voltage sensor fault.

[0039] Thirdly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the above-mentioned multi-port energy router fault diagnosis method.

[0040] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described multi-port energy router fault diagnosis method.

[0041] Compared with the prior art, the present invention has at least the following beneficial effects.

[0042] 1. This invention processes the time-series data matrix, periodic feature matrix, and physical perception attention vector using the same shared diagnostic model, enabling the identification of open-circuit faults in power switching devices and load-side voltage sensor faults within the same diagnostic process, without the need to configure separate diagnostic models for the two types of faults.

[0043] 2. This invention constructs a periodic feature matrix based on the waveform of the input current over multiple switching cycles, and uses the half-cycle energy ratio, the positive and negative half-cycle peak ratio, and the period mean to reflect the periodic structure and changes of the input current, which is helpful in identifying half-cycle current anomalies caused by open circuits in power switching devices.

[0044] 3. The present invention constructs a physical perception attention vector based on the DC mean of the input current, the power imbalance, the standard deviation of the input voltage, and the standard deviation of the load voltage, and encodes the physical perception attention vector as an attention bias, so that the shared diagnostic model can adjust the feature weights of the timing features in combination with the changes of the corresponding electrical quantities.

[0045] 4. This invention extracts features of multi-channel time-series data at different time scales through time-series feature branches, extracts correlation features between multiple switching cycles through periodic feature branches, and fuses the two types of features, which is beneficial for distinguishing different fault categories with similar external electrical quantity changes.

[0046] 5. This invention uses a unified fault classifier to output the classification probabilities corresponding to health status, primary-side power switch device open circuit fault, secondary-side power switch device open circuit fault, and various load-side voltage sensor faults. It can output specific fault subtypes and fault categories according to application needs.

[0047] 6. The present invention determines the final diagnostic result based on the state categories of multiple consecutive diagnostic windows, which can reduce the possibility of misjudgment caused by transient disturbances or local sampling noise affecting a single diagnostic window, and improve the stability of the diagnostic results. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of a four-port isolated DC-DC circuit topology for a multi-port energy router according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the main circuit topology of a DAB-type DC-DC module according to an embodiment of the present invention; Figure 3 This is a flowchart illustrating a multi-port energy router fault diagnosis method provided according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the online diagnostic waveform for health status under voltage sag disturbance provided in accordance with an embodiment of the present invention; Figure 5 This is a schematic diagram of the diagnostic waveform and fault flag bits when the primary-side switch S1 experiences an open-circuit fault, according to an embodiment of the present invention. Figure 6 This is a schematic diagram of the diagnostic waveform and fault flag bits when the secondary switch Q1 experiences an open-circuit fault, according to an embodiment of the present invention. Figure 7 This is a schematic diagram of the diagnostic waveform and fault flag bits when the load-side voltage sensor experiences an open-circuit fault, according to an embodiment of the present invention. Figure 8 This is a schematic diagram of a diagnostic waveform under load fluctuation conditions provided in accordance with an embodiment of the present invention; Figure 9 This is a schematic diagram of the diagnostic waveform under input voltage fluctuation conditions provided in accordance with an embodiment of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0050] Example 1 This embodiment provides a fault diagnosis method for a multi-port energy router. The DC-DC interface module in the multi-port energy router is used as a port-level monitoring unit, and the same shared diagnostic model is used to identify open circuit faults in power switching devices and load-side voltage sensor faults.

[0051] Among them, open circuit faults of power switching devices include open circuit faults of primary-side power switching devices and open circuit faults of secondary-side power switching devices; load-side voltage sensor faults include open circuit faults of load-side voltage sensors, bias faults, gain drift faults and jamming faults.

[0052] The use of the same shared diagnostic model to identify open-circuit faults in power switching devices and load-side voltage sensor faults in this embodiment means that, without switching the diagnostic model, the same shared diagnostic model is used to identify each of the above-mentioned single faults, without requiring the open-circuit faults in power switching devices and the load-side voltage sensor faults to occur simultaneously.

[0053] like Figure 1 As shown, this embodiment is applicable to multi-port energy routers with a common DC bus. Figure 1 A four-port isolated power conversion topology is shown, wherein port 1 is connected to a battery energy storage unit, port 2 is connected to a DC load, port 3 is connected to a photovoltaic array, and port 4 is connected to a common DC bus.

[0054] Ports 1 and 2 are located on one side of the high-frequency transformer, corresponding to the energy storage branch and the DC load branch, respectively; ports 3 and 4 are located on the other side of the high-frequency transformer, corresponding to the photovoltaic access branch and the common DC bus branch, respectively. Each port achieves energy transmission, electrical isolation, and voltage matching through its corresponding power conversion branch.

[0055] Figure 1 This is used to illustrate the system-level port connection relationships and energy conversion relationships of a multi-port energy router. Figure 2 This describes the circuit structure of a DAB-type DC-DC interface module used to illustrate one of the monitored isolated DC power channels.

[0056] The first embodiment referred to in this example The monitored DC-DC interface module refers to the module connected to the first... Each monitored port corresponds to an interface unit capable of acquiring input and output voltage and current and performing DC-DC power conversion. When a multi-port energy router adopts a multi-port integrated magnetic topology, an equivalent power channel with independent sampling and control capabilities can also be used as the monitored DC-DC interface module. Therefore, Figure 1 and Figure 2 These represent system-level application locations and port-level diagnostic objects, respectively. It is not required that the power switching devices in the two diagrams correspond one-to-one.

[0057] like Figure 2 As shown, the monitored DC-DC interface module can be a dual active bridge isolated bidirectional DC-DC interface module, or DAB type DC-DC interface module for short.

[0058] The DAB-type DC-DC interface module includes an input port, an input-side filter capacitor, a primary-side full bridge, a series inductor, a high-frequency transformer, a secondary-side full bridge, an output-side filter capacitor, and an output port.

[0059] The primary-side full bridge includes power switching devices S1, S2, S3, and S4; the secondary-side full bridge includes power switching devices Q1, Q2, Q3, and Q4.

[0060] The primary-side full-bridge and the secondary-side full-bridge are connected via a series inductor and a high-frequency transformer. By adjusting the phase shift relationship between the primary-side and secondary-side full-bridges, the power transmission direction and magnitude of the DAB-type DC-DC interface module can be adjusted.

[0061] During fault diagnosis, each monitored DC-DC interface module is treated as a port-level monitoring unit. Module identifiers can be configured for each monitored DC-DC interface module, and diagnostic windows, diagnostic features, and diagnostic results can be generated for each module.

[0062] like Figure 3 As shown, the multi-port energy router fault diagnosis method in this embodiment specifically includes the following steps.

[0063] Step 1: Obtain the load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router to form multi-channel timing data.

[0064] Specifically, for the first Each monitored DC-DC interface module synchronously acquires the load-side voltage according to a unified time base. Input side voltage Input current and output current .

[0065] Input side voltage and input current The load-side voltage is collected at the input port of the monitored DC-DC interface module. and output current Data is collected from the output port of the monitored DC-DC interface module.

[0066] For including Construct a four-channel time series matrix from the multi-channel time series data at each sampling time:

[0067] In the formula, Indicates the first Multi-channel timing data corresponding to each monitored DC-DC interface module. This indicates the number of sampling points contained in each measurement channel. Sample values ​​with the same sequence number in different measurement channels correspond to the same sampling time.

[0068] Step 2: Preprocess and synchronize windowing the multi-channel time-series data to obtain multiple diagnostic windows.

[0069] Specifically, the load-side voltage, input-side voltage, input current, and output current acquired synchronously are low-pass filtered to reduce the impact of sampling noise on subsequent feature extraction.

[0070] In one specific implementation, a fourth-order Butterworth low-pass filter is used to filter the data from each measurement channel. The cutoff frequency of the low-pass filter is determined based on the switching frequency of the monitored DC-DC interface module, the sampling frequency, and the range of signal frequencies to be retained.

[0071] Furthermore, the load-side voltage, input-side voltage, input current, and output current after low-pass filtering are synchronously segmented according to a unified time reference, and the data from the four measurement channels belonging to the same sampling period are grouped into the same diagnostic window to maintain the time correspondence between the measurement channels within the same diagnostic window.

[0072] For the data of the time-series characteristic branch of the input shared diagnostic model, the load-side voltage, input-side voltage, input current and output current in each diagnostic window are normalized by maximum value to minimum value, and the normalized data are arranged according to the measurement channel to form a time-series data matrix.

[0073] For the data used to construct the periodic feature matrix and the physical perception attention vector, a low-pass filter is applied to retain the actual amplitude and dimensions of the data in order to calculate the periodic characteristics of the input current, the window input power, the window output power, the DC mean of the input current, and the standard deviation of the voltage.

[0074] For example, in this embodiment, the switching frequency is 100kHz, and the sampling rate for the load-side voltage, input-side voltage, input current, and output current is 1MHz. The diagnostic window length is 500μs, and each diagnostic window includes 500 synchronous sampling points, covering 50 switching cycles.

[0075] The sliding step size of the diagnostic window is 250 μs, and the overlap ratio between adjacent diagnostic windows is 50%. By using overlapping diagnostic windows, the diagnostic results can be continuously updated, and the periodic variation characteristics of the input current over multiple switching cycles can be preserved.

[0076] Step 3: Determine the start time of each switching cycle based on the switching reference signal of the monitored DC-DC interface module, and divide the input current in the diagnostic window into multiple periodic signal segments according to each start time, with each periodic signal segment corresponding to one switching cycle.

[0077] Specifically, in single-phase-shift modulation mode, the input current sampling value in each period signal segment is divided into positive half-cycle sampling value and negative half-cycle sampling value according to the positive and negative polarities of the primary-side full-bridge output voltage or the corresponding switch reference signal.

[0078] For the The monitored DC-DC interface module is the One switching cycle, with Represents the set of indices for positive half-cycle sampling points, in order to This represents the set of indices for sampling points in the negative half-cycle.

[0079] No. The positive and negative half-cycle energies corresponding to each switching cycle are expressed as follows:

[0080]

[0081] In the formula, Indicates the first Energy corresponding to the positive half-cycle of each switching cycle Indicates the first The energy of the negative half-cycle corresponding to each switching cycle This indicates the sampling time interval between adjacent sampling moments.

[0082] No. The peak values ​​of the positive and negative half-cycles corresponding to each switching cycle are respectively expressed as:

[0083]

[0084] In the formula, This represents the maximum absolute value of the input current during the positive half-cycle. This represents the maximum absolute value of the input current during the negative half-cycle.

[0085] No. The average period value corresponding to each switching cycle is expressed as:

[0086] In the formula, This indicates the number of sampling points within a single switching cycle.

[0087] The sampling rate in this embodiment is 1MHz, and the switching frequency is 100kHz. Therefore, each switching cycle includes 10 sampling points. Sampling time interval .

[0088] The half-cycle energy ratio is determined based on the positive and negative half-cycle energies, and the positive and negative half-cycle peak ratio is determined based on the positive and negative half-cycle peak values.

[0089]

[0090] In the formula, Indicates the first Half-cycle energy ratio corresponding to each switching cycle Indicates the first The ratio of positive and negative half-cycle peak values ​​for each switching cycle. This indicates a positive number used to avoid a denominator of zero.

[0091] The above processing is performed on each switching cycle within the diagnostic window to obtain the half-cycle energy ratio, positive and negative half-cycle peak ratio, and cycle average for each switching cycle.

[0092] According to the order of each switching cycle within the diagnostic window, the half-cycle energy ratio, positive and negative half-cycle peak ratio and cycle mean corresponding to each switching cycle are respectively composed into a cycle feature sequence, and a cycle feature matrix is ​​constructed based on the cycle feature sequence.

[0093] In this embodiment, a diagnostic window comprises 50 switching cycles; therefore, the periodic feature matrix comprises 50 periodic features arranged in chronological order. The size of the periodic feature matrix is ​​expressed as follows: ,in, This indicates the number of periodic feature channels.

[0094] The periodic features include at least the half-cycle energy ratio, the ratio of positive to negative half-cycle peak values, and the period mean. In engineering applications, the number of periodic feature channels can be determined based on the input configuration of the actual training model. .

[0095] When an open-circuit fault occurs in either the primary or secondary power switching device, the corresponding half-cycle current pulse will weaken or be absent, causing the half-cycle energy ratio and the positive-negative half-cycle peak ratio to deviate from their values ​​under healthy conditions. The period average is used to reflect the change in the operating point of the input current within the corresponding switching cycle.

[0096] Step 4: Construct a physical perception attention vector based on the load-side voltage, input-side voltage, input current, and output current corresponding to each diagnostic window.

[0097] Specifically, the window input power is determined based on the input side voltage and input current at the same sampling time within the diagnostic window, and the window output power is determined based on the load side voltage and output current at the same sampling time.

[0098] Window input power and window output power are expressed as follows:

[0099]

[0100] In the formula, Indicates the first The window input power of each monitored DC-DC interface module within the corresponding diagnostic window. This indicates the output power within the corresponding diagnostic window. The window output power can also be referred to as the measured output power.

[0101] The absolute value of the difference between the window input power and the window output power is defined as the power imbalance:

[0102] In the formula, Indicates the first The power imbalance of each monitored DC-DC interface module within the corresponding diagnostic window.

[0103] Furthermore, the average DC value of the input current, the standard deviation of the input voltage, and the standard deviation of the load voltage within the diagnostic window are determined respectively.

[0104] The DC mean of the input current is used to reflect the average operating point of the input current within the diagnostic window; the standard deviation of the input side voltage is used to reflect the degree of fluctuation of the input side voltage within the diagnostic window; the standard deviation of the load side voltage is used to reflect the degree of fluctuation of the load side voltage within the diagnostic window.

[0105] The above physical quantities are dimensionless according to the preset normalization parameters corresponding to the DC mean of the input current, the power imbalance, the standard deviation of the input voltage, and the standard deviation of the load voltage.

[0106] Each preset normalization parameter is predetermined based on the value range of the corresponding physical quantity in the training samples, and the preset normalization parameters determined in the training phase are used in the online diagnosis process of the shared diagnostic model.

[0107] The dimensionless input current DC mean, power imbalance, input voltage standard deviation, and load voltage standard deviation are used to form a physical perception attention vector.

[0108] For ease of representation, each symbol in the physical perception attention vector represents the corresponding physical quantity after dimensionless processing. Therefore, the physical perception attention vector is represented as:

[0109] In the formula, Indicates the first The physical perception attention vector of each monitored DC-DC interface module within the corresponding diagnostic window. This represents the average DC value of the input current. This represents the standard deviation of the input voltage. This represents the standard deviation of the load-side voltage.

[0110] Step 5: Input the time series data matrix into the time series feature branch of the shared diagnostic model to extract time series features; encode the physical perception attention vector into an attention bias, and adjust the feature weights of the time series features according to the attention bias.

[0111] The periodic feature matrix is ​​input into the periodic feature branch of the shared diagnostic model to extract periodic features. The time-series features adjusted by feature weights are then fused with the periodic features to obtain fused features.

[0112] This embodiment uses an LCNN-AAM-PINN hybrid diagnostic model as a shared diagnostic model. The shared diagnostic model includes a temporal feature branch, a periodic feature branch, a feature fusion module, and a unified fault classifier.

[0113] The temporal feature branch includes a local pattern learning layer, a multi-scale convolution module, and a physical awareness attention module.

[0114] The local pattern learning layer is used to perform local smoothing on the time series data matrix. In this embodiment, the size of the time series data matrix input to the local pattern learning layer is... The local pattern learning layer uses a moving average with a window length of 7, while maintaining the output size. .

[0115] The multi-scale convolution module includes three multi-scale one-dimensional convolution blocks arranged in sequence, with the number of output channels of the three multi-scale one-dimensional convolution blocks being 64, 128 and 256 respectively.

[0116] Each multi-scale one-dimensional convolutional block includes parallel large convolutional kernel branches and small convolutional kernel branches. The large convolutional kernel branch uses a one-dimensional convolutional kernel of size 11 to extract the envelope variation features of the waveform; the small convolutional kernel branch uses a one-dimensional convolutional kernel of size 3 to extract the edge and local variation features of the switching pulse.

[0117] After processing with three multi-scale one-dimensional convolutional blocks, a size of [size missing] is obtained. The time-series characteristic map.

[0118] The physical perception attention module is set after the multi-scale convolution module. It is used to encode the physical perception attention vector into an attention bias and introduce the attention bias into the attention weight calculation process of the temporal features to adjust the feature weights of the temporal features.

[0119] In this embodiment, the physical perception attention module adopts a four-head attention structure, which is used for the size of Attention weighting is applied to the temporal feature maps.

[0120] The periodic feature branch includes a periodic feature projection module and a periodic attention module.

[0121] The periodic feature projection module is used to map the periodic features corresponding to each switching period in the periodic feature matrix to a preset feature space. In this embodiment, the periodic feature projection module maps the periodic features of size [missing information]. The periodic characteristic matrix is ​​projected as size Its periodic characteristics.

[0122] The periodic attention module is used to perform attention weighting on the periodic features corresponding to different mapping switching cycles. In this embodiment, the periodic attention module adopts a four-head attention structure.

[0123] The feature fusion module performs global pooling on the time-series features adjusted by feature weights and the periodic features weighted by attention, respectively.

[0124] In this embodiment, the temporal features are processed by global pooling to form a 256-dimensional temporal pooled feature, and the periodic features are processed by global pooling to form a 64-dimensional periodic pooled feature. The temporal pooled feature and the periodic pooled feature are concatenated to obtain a 320-dimensional fused feature.

[0125] The unified fault classifier employs a hidden layer comprising 128 neurons and outputs the classification probability of each preset state category based on the fusion features.

[0126] Through the above processing, the shared diagnostic model can utilize the time-series variation characteristics of load-side voltage, input-side voltage, input current and output current, the periodic variation characteristics of input current in multiple switching cycles, as well as the DC average value of input current, power imbalance and voltage fluctuation characteristics to diagnose faults in the monitored DC-DC interface module.

[0127] Step 6: Input the fused features into the unified fault classifier of the shared diagnostic model to obtain the classification probability of the monitored DC-DC interface module in each preset state category.

[0128] In this embodiment, the preset state categories include healthy state, primary-side power switch device open circuit fault, secondary-side power switch device open circuit fault, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault.

[0129] The unified fault classifier outputs the classification probability of each preset state category through the Softmax function, and determines the state category of the corresponding diagnostic window based on each classification probability:

[0130] In the formula, Indicates the first The category score corresponding to each diagnostic window This represents the classification probability vector corresponding to the respective diagnostic window. This indicates that the corresponding diagnostic window belongs to the first... The classification probability of a preset state category. This indicates the status category of the corresponding diagnostic window.

[0131] The seven categories of probabilities correspond to the following in order: healthy state, open circuit fault of primary power switching device, open circuit fault of secondary power switching device, open circuit fault of load-side voltage sensor, bias fault of load-side voltage sensor, gain drift fault of load-side voltage sensor, and stuck fault of load-side voltage sensor.

[0132] Step 7: Obtain the status categories of multiple diagnostic windows in chronological order, and count the number of times each preset status category appears in multiple consecutive diagnostic windows.

[0133] The most frequently occurring preset state category is determined as the final diagnostic result to reduce the possibility of misjudgment caused by transient disturbances or local sampling noise affecting a single diagnostic window.

[0134] Optionally, a uniform confidence threshold can be set. When the maximum classification probability of the corresponding diagnostic window reaches the uniform confidence threshold, the preset state category corresponding to the maximum classification probability is determined as the state category of the diagnostic window; when the maximum classification probability is lower than the uniform confidence threshold, the diagnostic window is marked as pending confirmation, and the diagnostic results are updated using subsequent diagnostic windows.

[0135] When a specific fault subtype needs to be output, the preset state category that occurs most frequently is directly output; when a major fault category needs to be output, the load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault are grouped into load-side voltage sensor faults.

[0136] Diagnostic results may include module identifier, port identifier, fault occurrence time, fault type, fault side, and fault confidence level.

[0137] The diagnostic results can be sent to the energy management controller of the multi-port energy router. Based on the diagnostic results, the energy management controller performs actions such as faulty port isolation, faulty bridge arm drive blocking, port derating, backup sensor switching, sensor signal reconstruction, or port power reallocation.

[0138] Before implementing the above fault diagnosis method, the shared diagnostic model is trained in advance using training samples.

[0139] The training samples required for the shared diagnostic model can be generated using the DAB simulation model.

[0140] The DAB simulation model includes a DC power supply, a primary-side full-bridge, a series inductor, a high-frequency transformer, a secondary-side full-bridge, a voltage outer-loop PI controller, a sensor fault injection unit, and a training data export unit.

[0141] In the simulation model, the DC power supply voltage is set to 750 V, the turns ratio of the high-frequency transformer is set to 75:40, the primary-side full bridge includes power switching devices S1 to S4, and the secondary-side full bridge includes power switching devices Q1 to Q4.

[0142] The proportional gain of the voltage outer loop PI controller Set as Integral coefficient Set to 0.01. The fixed simulation step size of the simulation model is set to... This corresponds to a sampling rate of 1 MHz.

[0143] An open-circuit fault in a power switch device is injected by forcibly setting the drive signal of the corresponding power switch device to zero.

[0144] Open-circuit fault samples of primary-side power switching devices S1 to S4 are generated by setting the drive signals of the power switching devices S1 to S4 in the primary-side full-bridge to zero respectively; open-circuit fault samples of secondary-side power switching devices are generated by setting the drive signals of the power switching devices Q1 to Q4 in the secondary-side full-bridge to zero respectively.

[0145] Forced zeroing of the drive signal is used to simulate open-circuit faults caused by drive failure, broken bond wires, or damage to power switching devices.

[0146] The sensor fault verification object in this embodiment is the load-side voltage sensor. By setting zeroing, addition, multiplication, and sample-and-hold circuits in the output link of the load-side voltage sensor, open-circuit faults, bias faults, gain drift faults, and jamming faults of the load-side voltage sensor are simulated respectively.

[0147] The DAB simulation model synchronously exports the load-side voltage, input-side voltage, input current, and output current, and generates corresponding sample labels based on the simulation operation status and fault injection status.

[0148] The training data generation process includes synchronous sampling, data preprocessing, diagnostic window construction, periodic feature and physical perception attention vector construction, sample labeling, and dataset partitioning.

[0149] First, load-side voltage, input-side voltage, input current, and output current are synchronously acquired from the DAB simulation model to form multi-channel timing data.

[0150] Then, low-pass filtering and synchronous windowing are performed on the multi-channel time-series data. The data used for time-series feature extraction is normalized and a time-series data matrix is ​​formed. A periodic feature matrix is ​​constructed based on the waveform of the input current in multiple switching cycles. A physical sensing attention vector is constructed based on the DC mean of the input current, the power imbalance, the standard deviation of the input voltage, and the standard deviation of the load voltage.

[0151] Subsequently, a single-category label was set according to the simulation running status and fault injection status corresponding to each sample.

[0152] The training samples are categorized as follows: health status, primary-side power switch open circuit fault, secondary-side power switch open circuit fault, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault.

[0153] The training samples in this embodiment do not include composite fault samples where both power switching device open-circuit fault and load-side voltage sensor fault occur simultaneously.

[0154] The processed samples were divided into training, validation, and test sets in an 8:1:1 ratio. The training set was used to update the model parameters of the shared diagnostic model, the validation set was used to select model parameters and determine when to stop training, and the test set was used to evaluate the shared diagnostic model after training.

[0155] This embodiment uses a seven-class weighted cross-entropy as the classification loss:

[0156] In the formula, Represents classification loss. This indicates the number of samples in a training batch. Indicates the first The category weights corresponding to each preset state category Indicates the first The training sample corresponds to the first... Labels for each preset state category, This represents the corresponding classification probability output by the shared diagnostic model.

[0157] In this embodiment, the category weights corresponding to health status, primary-side power switch device open circuit fault, secondary-side power switch device open circuit fault, and four types of load-side voltage sensor faults are set to 4.0, 1.0, 1.0, 4.0, 4.0, 4.0, and 4.0, respectively.

[0158] The physical perception attention vector influences the feature weights of temporal features through the physical perception attention module, and participates in the model parameter updates of the temporal feature branch, periodic feature branch, physical perception attention module, feature fusion module, and unified fault classifier through backpropagation of the seven-class weighted cross-entropy classification loss.

[0159] In one specific implementation, the AdamW optimizer is used to train the shared diagnostic model, with an initial learning rate set to... The weight decay coefficient is set to The batch size was set to 56, the maximum number of training rounds was set to 100, the early stopping patience value was set to 15, and the gradient clipping threshold was set to 1.0.

[0160] During training, the ReduceLROnPlateau learning rate decay strategy was adopted, with the decay factor set to 0.5 and the patience value set to 5.

[0161] Gaussian noise and amplitude scaling can also be used to augment the training samples to expand the training sample size.

[0162] The model structure parameters and training parameters described above are a specific set of parameters used in this embodiment. In other embodiments, the corresponding parameters can be adjusted according to the sampling frequency, diagnostic window length, number of training samples, and computing power of the model deployment device.

[0163] Figures 4 to 9 The waveforms of input current, load-side voltage, and output current under different operating conditions are shown. Figures 5 to 7 The changes in the fault flag bits before and after the corresponding fault occurred are also shown.

[0164] like Figure 4 As shown, under voltage sag disturbances, the input current, load-side voltage, and output current undergo corresponding dynamic changes, and the shared diagnostic model still identifies the monitored DC-DC interface module as healthy. Figure 4 The "health status 250 μs" indicates that the health status diagnosis result is obtained 250 μs after the voltage drop occurs.

[0165] like Figure 5 As shown, after an open-circuit fault occurs in the primary-side power switching device S1, the periodic waveform of the input current changes, and the load-side voltage and output current generate corresponding dynamic responses as the power transmission state changes. The fault flag changes from 0 to 1 to indicate an open-circuit fault in the primary-side power switching device. Figure 5 The fault diagnosis time shown is 250 μs.

[0166] like Figure 6 As shown, after an open-circuit fault occurs in the secondary-side power switch Q1, the input current, load-side voltage, and output current change accordingly. The fault flag changes from 0 to 2, indicating an open-circuit fault in the secondary-side power switch. Figure 6 The fault diagnosis time shown is 250 μs.

[0167] like Figure 7 As shown, after an open-circuit fault occurs in the load-side voltage sensor, the measured value output by the load-side voltage sensor changes, and the input current and output current change accordingly with the closed-loop control process. The fault flag changes from 0 to 3 to indicate a load-side voltage sensor fault. Figure 7 The fault diagnosis time shown is 500 μs.

[0168] Figure 7 The fault flag bit 3 is used to indicate the major category of load-side voltage sensor faults. The shared diagnostic model still distinguishes between open-circuit faults, bias faults, gain drift faults, and jamming faults of the load-side voltage sensor by using the classification probabilities corresponding to the seven preset state categories.

[0169] like Figure 8As shown, when the load changes, the input current, load-side voltage, and output current change dynamically accordingly. Figure 8 Used to display multi-channel waveforms under load fluctuation conditions.

[0170] like Figure 9 As shown, when the input voltage fluctuates, the input current, load-side voltage, and output current undergo corresponding dynamic changes. Figure 9 Used to display multi-channel waveforms under input voltage fluctuation conditions.

[0171] By jointly utilizing multi-channel timing features, input current cycle features, and physical perception attention vectors, the shared diagnostic model can distinguish between healthy states, open-circuit faults in power switching devices, load-side voltage sensor faults, and non-fault transients caused by load fluctuations and input voltage fluctuations.

[0172] Example 2 This embodiment provides a multi-port energy router fault diagnosis system to implement the multi-port energy router fault diagnosis method described in Embodiment 1.

[0173] The multi-port energy router fault diagnosis system includes a data acquisition module, a preprocessing and windowing module, an auxiliary feature construction module, a shared diagnostic model, and a diagnostic result determination module.

[0174] The data acquisition module is used to acquire the load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router, forming multi-channel timing data.

[0175] The preprocessing and windowing module is used to perform low-pass filtering and synchronous windowing processing on multi-channel time-series data to obtain multiple diagnostic windows, and to construct a time-series data matrix based on the multi-channel time-series data corresponding to each diagnostic window.

[0176] In this embodiment, the preprocessing and windowing module uses a fourth-order Butterworth low-pass filter to filter the data of each measurement channel and normalizes the data used for time series feature extraction.

[0177] For example, at a sampling rate of 1 MHz, the diagnostic window length is 500 μs, the sliding step size is 250 μs, and each diagnostic window includes 500 synchronous sampling points and 50 switching cycles.

[0178] The auxiliary feature construction module is used to construct a periodic feature matrix based on the waveform of the input current in each diagnostic window over multiple switching cycles, and to construct a physical sensing attention vector based on the multi-channel timing data corresponding to each diagnostic window.

[0179] Specifically, the auxiliary feature construction module determines the start time of each switching cycle based on the switching reference signal, and divides the input current sampling value in each switching cycle into positive half-cycle sampling value and negative half-cycle sampling value.

[0180] The auxiliary feature construction module determines the positive half-cycle energy, negative half-cycle energy, positive half-cycle peak value, and negative half-cycle peak value based on the positive half-cycle sampling value and the negative half-cycle sampling value, and determines the cycle average value based on the input current sampling value within the corresponding switching cycle, thereby determining the half-cycle energy ratio and the positive and negative half-cycle peak value ratio.

[0181] The auxiliary feature construction module constructs a periodic feature matrix based on the order of each switching cycle within the diagnostic window, according to the half-cycle energy ratio, the positive and negative half-cycle peak ratio, and the cycle mean.

[0182] The auxiliary feature construction module is also used to determine the DC mean of the input current, the power imbalance, the standard deviation of the input voltage and the standard deviation of the load voltage corresponding to each diagnostic window, and to perform dimensionless processing on the above physical quantities to form a physical perception attention vector.

[0183] The shared diagnostic model includes a time-series feature branch, a periodic feature branch, a feature fusion module, and a unified fault classifier.

[0184] The temporal feature branch includes a local pattern learning layer, a multi-scale convolution module, and a physical awareness attention module.

[0185] The local pattern learning layer is used to perform local smoothing on the temporal data matrix; the multi-scale convolution module is used to extract temporal features at different time scales from the locally smoothed temporal data matrix; the physical perception attention module is used to encode the physical perception attention vector into an attention bias and adjust the feature weights of the temporal features according to the attention bias.

[0186] The periodic feature branch includes a periodic feature projection module and a periodic attention module.

[0187] The periodic feature projection module is used to map the periodic features corresponding to each switching cycle in the periodic feature matrix to a preset feature space; the periodic attention module is used to perform attention weighting on the periodic features corresponding to different switching cycles after mapping.

[0188] The feature fusion module is used to perform pooling processing on the time-series features adjusted by feature weights and the periodic features weighted by attention, and then concatenate the pooling results to obtain fused features.

[0189] The unified fault classifier is used to calculate the classification probability of the monitored DC-DC interface module in each preset state category based on the fusion features.

[0190] The diagnostic result determination module is used to determine the preset state category with the highest classification probability as the state category of the corresponding diagnostic window, and to determine the final diagnostic result based on the state categories of multiple consecutive diagnostic windows.

[0191] Furthermore, the multi-port energy router fault diagnosis system can communicate with the energy management controller. The energy management controller receives the diagnostic results and performs corresponding control operations based on the module identifier, port identifier, fault category, and fault confidence level.

[0192] Example 3 This embodiment provides an electronic device, which includes a processor and a memory.

[0193] The memory stores a computer program, which, when executed by the processor, implements the multi-port energy router fault diagnosis method described in Embodiment 1.

[0194] Example 4 This embodiment provides a computer-readable storage medium in which a computer program is stored.

[0195] When the computer program is executed by the processor, it implements the multi-port energy router fault diagnosis method described in Embodiment 1.

[0196] Example 5 This embodiment uses a multi-port energy router that interconnects a photovoltaic array, a battery energy storage unit, a DC load, and a common DC bus as the application object.

[0197] The multi-port energy router adopts Figure 1 The diagram shows a four-port isolated power conversion topology. Port 1 is connected to a battery energy storage unit, port 2 is connected to a DC load, port 3 is connected to a photovoltaic array, and port 4 is connected to a common DC bus.

[0198] During normal operation, the energy management controller coordinates photovoltaic power access, energy storage charging and discharging, load power supply, and power exchange on the common DC bus based on the output power of the photovoltaic array, the power demand of the DC load, the state of charge of the battery energy storage unit, and the power command of the common DC bus.

[0199] The fault diagnosis system uses each monitored DC-DC interface module as a port-level monitoring unit to synchronously acquire the load-side voltage, input-side voltage, input current, and output current of the corresponding monitored DC-DC interface module.

[0200] The fault diagnosis system preprocesses and synchronously windowes multi-channel time-series data, and constructs a time-series data matrix, a periodic feature matrix, and a physical perception attention vector, respectively.

[0201] Subsequently, the time-series data matrix, periodic feature matrix, and physical perception attention vector are input into the shared diagnostic model to obtain the classification probability of the monitored DC-DC interface module in each preset state category, and the final diagnostic result is determined based on the state category of multiple consecutive diagnostic windows.

[0202] When an open-circuit fault occurs in the power switching device of the DC-DC interface module corresponding to the energy storage port, the half-cycle pulse of the input current, the effective power transmitted at the port, and the dynamic response on the output side change accordingly. The shared diagnostic model identifies either an open-circuit fault in the primary-side power switching device or an open-circuit fault in the secondary-side power switching device based on timing characteristics, periodic characteristics, and physical perception attention vectors.

[0203] When the load-side voltage sensor experiences an open-circuit fault, bias fault, gain drift fault, or jamming fault, the load-side voltage measurement and the closed-loop operating status of the monitored DC-DC interface module will change accordingly. The shared diagnostic model identifies the corresponding load-side voltage sensor fault subtype based on these changes.

[0204] After receiving the diagnostic results, the energy management controller performs corresponding fault-tolerant control operations based on the fault category.

[0205] For open-circuit faults in power switching devices, the drive signal of the faulty bridge arm can be blocked, and the faulty port can be isolated or derated. For load-side voltage sensor faults, the system can switch to a backup sensor, use the observer's estimated value, or enter a preset fault-tolerant control mode.

[0206] In this embodiment, the sensor fault verification object is the load-side voltage sensor. For input-side voltage sensors, input current sensors, or output current sensors, the same diagnostic framework can be used for extension after collecting corresponding sensor fault samples and retraining or calibrating the shared diagnostic model.

[0207] The above embodiments mainly use DAB-type isolated bidirectional DC-DC interface modules as the monitored objects. For three-phase DAB, multi-active bridge, or other DC-DC interface modules with periodic switching current characteristics, the construction method of the periodic characteristics can be adjusted according to the switching modulation method and input current periodic structure of the corresponding topology, and the shared diagnostic model can be retrained or calibrated using training samples collected under the corresponding topology.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A fault diagnosis method for a multi-port energy router, characterized in that, include: The load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router are acquired to form multi-channel timing data; The multi-channel time-series data is preprocessed and synchronized windowed to obtain multiple diagnostic windows; For each diagnostic window, a timing data matrix is ​​constructed based on the multi-channel timing data, a periodic feature matrix is ​​constructed based on the waveform of the input current in multiple switching cycles, and a physical sensing attention vector is constructed based on the multi-channel timing data. The time-series data matrix is ​​input into the time-series feature branch of the shared diagnostic model to extract time-series features, and the physical perception attention vector is encoded as an attention bias to adjust the feature weights of the time-series features. The periodic feature matrix is ​​input into the periodic feature branch of the shared diagnostic model to extract periodic features, and the time-series features adjusted by feature weights are fused with the periodic features to obtain fused features; The fused features are input into the unified fault classifier of the shared diagnostic model to obtain the classification probability of the monitored DC-DC interface module in each preset state category, and the diagnostic result of the monitored DC-DC interface module is determined based on the classification probability. The preset state categories include healthy state, open circuit fault of power switching device, and load-side voltage sensor fault.

2. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: The monitored DC-DC interface module is a dual active bridge isolated bidirectional DC-DC interface module, including an input port, a primary full bridge, a series inductor, a high-frequency transformer, a secondary full bridge, and an output port; The input voltage and the input current are collected at the input port, and the load voltage and the output current are collected at the output port.

3. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: The preprocessing includes low-pass filtering of the load-side voltage, input-side voltage, input current, and output current; The synchronous windowing process includes synchronously segmenting the load-side voltage, input-side voltage, input current, and output current after low-pass filtering according to a unified time base, and grouping data belonging to the same sampling period into the same diagnostic window; The load-side voltage, input-side voltage, input current, and output current within each diagnostic window are normalized, and the normalized data are arranged according to the measurement channel to form the timing data matrix. The physical perception attention vector is constructed based on the load-side voltage, input-side voltage, input current, and output current corresponding to each diagnostic window.

4. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: Constructing the periodic feature matrix includes: The start time of each switching cycle is determined based on the switching reference signal of the monitored DC-DC interface module, and the input current in the diagnostic window is divided into multiple periodic signal segments according to each start time, with each periodic signal segment corresponding to a switching cycle. Based on the positive and negative polarities of the switch reference signal in each of the periodic signal segments, the input current sampling values ​​in each of the periodic signal segments are divided into positive half-cycle sampling values ​​and negative half-cycle sampling values. The positive half-cycle energy, negative half-cycle energy, positive half-cycle peak value, and negative half-cycle peak value are determined based on the positive half-cycle sampling value and the negative half-cycle sampling value, respectively, and the period average value is determined based on the input current sampling value within the same period signal segment. The half-cycle energy ratio is determined based on the positive half-cycle energy and the negative half-cycle energy, and the positive half-cycle peak value ratio is determined based on the positive half-cycle peak value and the negative half-cycle peak value. According to the order of each periodic signal segment in the diagnostic window, the half-cycle energy ratio, positive and negative half-cycle peak ratio and period mean corresponding to each periodic signal segment are respectively composed into a periodic feature sequence, and the periodic feature matrix is ​​constructed based on the periodic feature sequence.

5. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: Constructing the physical perception attention vector includes: The window input power is determined based on the input side voltage and input current at the same sampling time within the diagnostic window, and the window output power is determined based on the load side voltage and output current at the same sampling time. The absolute value of the difference between the window input power and the window output power is determined as the power imbalance, and the average DC value of the input current, the standard deviation of the input voltage, and the standard deviation of the load voltage within the diagnostic window are determined respectively. The above physical quantities are dimensionless according to the preset normalization parameters corresponding to the DC average of the input current, the power imbalance, the standard deviation of the input voltage, and the standard deviation of the load voltage. The dimensionless input current DC mean, power imbalance, input voltage standard deviation, and load voltage standard deviation are used to form the physical perception attention vector.

6. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: The temporal feature branch includes a local pattern learning layer, a multi-scale convolutional module, and a physical awareness attention module; The local pattern learning layer is used to perform local smoothing on the temporal data matrix. The multi-scale convolution module is used to extract temporal features at different time scales from the locally smoothed temporal data matrix. The physical perception attention module is used to encode the physical perception attention vector into the attention bias and adjust the feature weights of the temporal features according to the attention bias. The periodic feature branch includes a periodic feature projection module and a periodic attention module; The periodic feature projection module is used to map the periodic features corresponding to each switching period in the periodic feature matrix to a preset feature space, and the periodic attention module is used to perform attention weighting on the periodic features corresponding to different switching periods after mapping. The shared diagnostic model also includes a feature fusion module, which is used to perform pooling processing on the time-series features adjusted by feature weights and the periodic features weighted by attention, and to concatenate the pooling results to obtain the fused features.

7. The method for fault diagnosis of a multi-port energy router according to claim 1, characterized in that: The preset state categories include health status, primary-side power switch device open circuit fault, secondary-side power switch device open circuit fault, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault. The unified fault classifier calculates the classification probability of the monitored DC-DC interface module in each of the preset state categories based on the fusion features, and determines the preset state category with the highest classification probability as the state category of the corresponding diagnostic window. The number of occurrences of each preset status category is counted based on the status categories of multiple consecutive diagnostic windows, and the preset status category with the highest number of occurrences is determined as the final diagnostic result. When outputting fault categories, load-side voltage sensor open circuit fault, load-side voltage sensor bias fault, load-side voltage sensor gain drift fault, and load-side voltage sensor stuck fault are grouped into load-side voltage sensor faults.

8. A multi-port energy router fault diagnosis system, used to implement the multi-port energy router fault diagnosis method according to any one of claims 1-7, characterized in that, include: The module includes a data acquisition module, a preprocessing and windowing module, an auxiliary feature construction module, a shared diagnostic model, and a diagnostic result determination module. The data acquisition module is used to acquire the load-side voltage, input-side voltage, input current, and output current of the monitored DC-DC interface module in the multi-port energy router, forming multi-channel timing data; The preprocessing and windowing module is used to preprocess and synchronously window the multi-channel time series data to obtain multiple diagnostic windows, and to construct a time series data matrix based on the multi-channel time series data corresponding to each diagnostic window. The auxiliary feature construction module is used to construct a periodic feature matrix based on the waveform of the input current in each diagnostic window over multiple switching cycles, and to construct a physical sensing attention vector based on the multi-channel timing data corresponding to each diagnostic window. The shared diagnostic model includes a temporal feature branch, a periodic feature branch, a feature fusion module, and a unified fault classifier; The temporal feature branch is used to extract temporal features from the temporal data matrix and encode the physical perception attention vector as an attention bias to adjust the feature weights of the temporal features; The periodic feature branch is used to extract periodic features from the periodic feature matrix; The feature fusion module is used to fuse the time-series features adjusted by feature weights and the periodic features to obtain fused features; The unified fault classifier is used to calculate the classification probability of the monitored DC-DC interface module being in each preset state category based on the fusion features. The diagnostic result determination module is used to determine the diagnostic result of the monitored DC-DC interface module based on the classification probability. The preset state categories include healthy state, open circuit fault of power switching device, and load-side voltage sensor fault.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the multi-port energy router fault diagnosis method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a multi-port energy router fault diagnosis method according to any one of claims 1 to 7.

Images