Fault early warning system and method based on multi-data fusion and causal reasoning mechanism

Abstract

The invention discloses a fault early warning system and method based on multi-data fusion and a causal reasoning mechanism, and relates to the technical field of power system fault diagnosis. Comprising the steps that a data acquisition module synchronously acquires and preprocesses multi-source sensor data; the data fusion module fuses data through adaptive Kalman filtering to generate a state evaluation value; the fault prediction module screens features based on a causal graph and inputs LSTM to output a fault risk value; the dynamic threshold adjustment module dynamically calculates a threshold in combination with a power grid load; the early warning module compares the risk value with a threshold value and generates early warning; the data acquisition and fusion module is deployed at an edge end, and other modules are deployed at a cloud end. The state estimation precision is improved through multi-data fusion, feature redundancy is reduced through causal reasoning, the dynamic threshold adapts to working condition changes, and the fault prediction accuracy and the system reliability are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of power system fault diagnosis technology, specifically to a fault early warning system and method based on multi-data fusion and causal reasoning mechanisms. Background Technology

[0002] The global energy structure is rapidly transforming towards cleaner and lower-carbon energy sources. High-voltage cascaded energy storage converter systems, as a core supporting technology for new energy power generation grid integration, smart grid flexible regulation, and electric vehicle charging infrastructure, play a crucial role in renewable energy consumption and grid peak shaving and valley filling through modular design, enabling efficient conversion and flexible configuration of high-voltage power. In existing technologies, CN119324565A discloses a fault early warning method and system for high-voltage cascaded energy storage converter systems. This scheme achieves monitoring and early warning of system operating status through steps such as data acquisition, preprocessing, historical fault classification and feature extraction, dual-algorithm prediction, fault matching and identification, and early warning response system maintenance and updates.

[0003] However, the fault development of high-voltage cascaded energy storage converter systems is both insidious and rapid, making it difficult for existing early warning methods to capture early fault characteristics in a timely manner. Summary of the Invention

[0004] The purpose of this invention is to provide a fault early warning system and method based on multi-data fusion and causal reasoning mechanisms, which can solve the technical problem of insufficient accuracy in fault prediction of high-voltage cascaded energy storage converter systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a fault early warning system based on multi-data fusion and causal reasoning mechanisms for fault prediction in high-voltage cascaded energy storage converter systems, the system comprising: The data acquisition module is used to simultaneously acquire data from multiple sensor sources, perform data preprocessing, and output purification data. The data fusion module is used to fuse the purification data through adaptive Kalman filtering to generate system status assessment values; The fault prediction module is used to filter out features directly related to the target fault from the system state evaluation value based on the pre-constructed causal graph, and input the filtered features into the pre-trained LSTM model to output the fault risk value. The dynamic threshold adjustment module is used to receive the fault risk value and, in combination with the power grid load and environmental parameters, dynamically calculate and output the fault judgment threshold. The early warning module is used to compare the fault risk value with the fault determination threshold, and generate and display early warning information when the fault risk value exceeds the fault determination threshold; The data acquisition module and the data fusion module are deployed at the edge computing end of the high-voltage cascaded energy storage converter system, while the fault prediction module, the dynamic threshold adjustment module, and the early warning module are deployed in the cloud.

[0006] Furthermore, the data acquisition module includes: Sensor networks include voltage sensors, current sensors, temperature sensors, and vibration sensors; A synchronous acquisition unit, wherein the synchronous acquisition unit uses GPS timestamps to synchronize multi-source sensor data; The data preprocessing unit uses a sliding window statistical method to identify and remove outliers in the original data, and uses linear interpolation to fill in missing data values ​​to generate cleaned data.

[0007] Furthermore, the data fusion module includes: The state vector construction unit is used to construct a system state vector with voltage, current, temperature and vibration amplitude as components based on the purification data. The Kalman filter unit is used to filter and fuse the system state vector and output the system state evaluation value. An adaptive optimization unit is used to dynamically adjust the process noise covariance matrix and the observation noise covariance matrix during the Kalman filtering process.

[0008] Furthermore, the adaptive optimization unit calculates the root mean square value of the filtered residual through a sliding window, and adjusts the process noise covariance matrix and the observation noise covariance matrix according to the root mean square value.

[0009] Furthermore, the fault prediction module includes: The causal graph construction unit constructs a causal graph reflecting the causal relationships between variables based on the system state evaluation value; The feature filtering unit is used to filter out features directly related to the target fault from the real-time input system state evaluation values ​​based on the cause-effect graph. The LSTM prediction unit is used to predict the fault risk value sequence for the next hour based on the selected features and output the fault risk value.

[0010] Furthermore, the causal graph construction unit uses the PC algorithm or the NOTEARS algorithm to construct the initial causal graph.

[0011] Secondly, the present invention also provides a fault early warning method based on multi-data fusion and causal reasoning mechanisms. The method, using the aforementioned system, includes the following steps: S1. Through the data acquisition module, multi-source sensor data of the high-voltage cascaded energy storage converter system are collected in real time and synchronously, and the multi-source sensor data is preprocessed to obtain purified data. S2. The purification data is fused using an adaptive Kalman filter algorithm through the data fusion module to obtain the system status evaluation value. S3. Through the fault prediction module, the pre-constructed causal graph is used to filter features directly related to the target fault from the system state evaluation value, and the filtered features are input into the pre-trained LSTM model for prediction to obtain the fault risk value. S4. The fault risk value is received through the dynamic threshold adjustment module, and the fault judgment threshold is dynamically calculated in combination with the power grid load and environmental parameters. S5. The early warning module compares the fault risk value with the fault judgment threshold. When the fault risk value exceeds the fault judgment threshold, an early warning message is generated and displayed.

[0012] Furthermore, the specific process of step S2 is as follows: S21. Based on the purification data, construct a system state vector with voltage, current, temperature and vibration amplitude as components; S22. Perform Kalman filtering iteration on the system state vector and output the fused state evaluation value; S23. During the Kalman filter iteration process, the process noise covariance matrix and the observation noise covariance matrix are dynamically adjusted based on the statistical characteristics of the filter residuals.

[0013] Furthermore, the specific process of S23 is as follows: S231. Calculate the root mean square value of the filtered residual within the set sliding window; S232. Based on the root mean square value, the process noise covariance matrix and the observation noise covariance matrix are adjusted proportionally or exponentially.

[0014] Furthermore, the specific process of S3 is as follows: S31. Based on the real-time acquired system state evaluation value, extract the feature variables within the time window; S32. Based on the initial causal graph, select direct causal parent node variables that are directly related to the target fault from the feature variables.

[0015] Compared with the prior art, the present invention has the following beneficial technical effects: This invention discloses a fault early warning system based on multi-data fusion and causal reasoning mechanisms. The system synchronously acquires data from multiple sensors via a data acquisition module, performs preprocessing, and outputs cleaned data to eliminate time asynchrony and data loss issues, ensuring the integrity and reliability of the input data. The data fusion module uses adaptive Kalman filtering to fuse the cleaned data, generating a system state assessment value. It dynamically adjusts the noise covariance matrix to adapt to environmental changes, effectively suppressing noise interference and improving the accuracy of state estimation. Based on a pre-constructed causal graph, the fault prediction module filters features directly related to the target fault from the system state assessment value and inputs these features into a pre-trained LSTM model to output the fault. The system employs a risk value mechanism to reduce interference from irrelevant features, thereby improving prediction efficiency and interpretability. A dynamic threshold adjustment module receives the fault risk value and dynamically calculates the fault determination threshold based on grid load and environmental parameters, allowing the threshold to adaptively adjust with operating conditions and reducing false alarms and missed alarms. An early warning module compares the fault risk value with the fault determination threshold and generates and displays early warning information when the threshold is exceeded, achieving timely and effective fault warnings. Furthermore, the data acquisition and data fusion modules are deployed at the edge computing end, while the fault prediction, dynamic threshold adjustment, and early warning modules are deployed in the cloud, forming an edge-cloud collaborative architecture. This ensures real-time data processing while fully utilizing cloud computing resources, significantly improving the overall performance and reliability of the system. In summary, this invention effectively solves the problem of insufficient fault prediction accuracy in high-voltage cascaded energy storage converter systems through multi-data fusion, causal reasoning, and dynamic threshold mechanisms, providing strong support for the safe and stable operation of the power grid. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the fault early warning system structure based on multi-data fusion and causal reasoning mechanism in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the fault early warning system based on multi-data fusion and causal reasoning mechanism in an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the fault early warning method based on multi-data fusion and causal reasoning mechanism in an embodiment of the present invention. Detailed Implementation

[0019] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0020] See Figure 1 and Figure 2This invention provides a fault early warning system based on multi-data fusion and causal reasoning mechanisms for fault prediction in high-voltage cascaded energy storage converter systems. The system includes: The data acquisition module is used to simultaneously acquire data from multiple sensor sources, perform data preprocessing, and output purification data. The data fusion module is used to fuse the purification data through adaptive Kalman filtering to generate system status assessment values; In a more specific embodiment of the present invention, both the data acquisition module and the data fusion module are deployed at the edge computing end of the high-voltage cascaded energy storage converter system. The data fusion module uses voltage (U), current (I), temperature (T) and vibration amplitude (A) as state vector components, eliminates the inherent noise and measurement deviation of each sensor through recursive filtering, and outputs the optimal estimate of each state quantity to form the system state evaluation value.

[0021] The fault prediction module is used to filter out features directly related to the target fault from the system state evaluation value based on the pre-constructed causal graph, and input the filtered features into the pre-trained LSTM model to output the fault risk value. In a more specific embodiment of the present invention, the fault prediction module is deployed in the cloud and includes two working stages: offline modeling and online inference. In the offline modeling stage, historical operating data (including normal and typical fault samples) is used to learn the causal relationships between variables through PC or NOTEARS algorithms, constructing a causal graph reflecting these relationships, and identifying the set of direct causal parent nodes corresponding to each target fault (such as IGBT overheating failure, DC support capacitor capacitance decay, and cooling fan jamming). In the online inference stage, system state evaluation values ​​uploaded from the edge are received. Based on the causal graph, features directly related to the target fault, such as temperature change rate, current peak-to-peak value, and voltage ripple coefficient, are selected from the real-time input system state evaluation values. Then, based on the current target fault type to be predicted, its direct parent nodes are located in the causal graph, and only the features corresponding to these nodes are input into the pre-trained LSTM model. The LSTM model models the evolution of health status in time series form, outputting a fault risk value sequence (in units of probability values ​​within the [0,1] interval) for the next hour.

[0022] The dynamic threshold adjustment module is used to receive the fault risk value and, in combination with the power grid load and environmental parameters, dynamically calculate and output the fault judgment threshold. In a more specific embodiment of the present invention, the dynamic threshold adjustment module is deployed in the cloud and dynamically adjusts the system's fault determination threshold using an adaptive algorithm based on the power grid's operating conditions. The dynamic threshold adjustment module establishes a baseline model of normal operating conditions by analyzing historical system data, compares the actual operating conditions with the baseline model in real time, and automatically adjusts the threshold to adapt to complex environmental changes. It receives the fault risk value and the system's normal operating baseline output by the fault prediction module (based on statistical analysis of long-term historical normal data to establish a multi-dimensional system normal operating baseline). The threshold D is dynamically updated based on the deviation between the actual operating status and the baseline.

[0023] in, As the baseline threshold, The deviation value is the mean relative deviation between the previous feature and the baseline. This is the adjustment coefficient. The output dynamically calculates and outputs the fault determination threshold.

[0024] The early warning module is used to compare the fault risk value with the fault determination threshold, and generate and display early warning information when the fault risk value exceeds the fault determination threshold; In a more specific embodiment of the present invention, the early warning module presents fault information through an intuitive user interface when a fault occurs, displaying the fault risk in the form of dynamic charts, heat maps, etc., to help operators quickly locate the problem. It also automatically formulates a troubleshooting plan based on the fault, improving fault resolution efficiency. The early warning module also supports intelligent interactive functions, adaptively adjusting the interface layout and data display method according to user operation records to optimize the fault troubleshooting process.

[0025] The data acquisition module and the data fusion module are deployed at the edge computing end of the high-voltage cascaded energy storage converter system, while the fault prediction module, the dynamic threshold adjustment module, and the early warning module are deployed in the cloud. Data interaction between the system modules is achieved through standardized interfaces, ensuring the system's modularity and scalability. The system adopts a layered computing architecture, deploying the data acquisition module and the data fusion module at the edge to complete edge data processing; and performing advanced analysis in the cloud, deploying complex fault analysis tasks such as the fault prediction module, the dynamic threshold adjustment module, and the fault early warning module, achieving low-latency real-time monitoring and fault early warning.

[0026] In the operation of a high-voltage cascaded energy storage converter system, when an abnormal increase in IGBT junction temperature occurs, the data acquisition module synchronously captures millisecond-level waveforms from the power module temperature sensor and the DC-side current sensor via GPS. The data fusion module uses an adaptive Kalman filter to suppress the phase lag introduced by the thermal inertia of the temperature sensor and outputs a corrected real-time junction temperature estimate. The fault prediction module identifies the "current slope dI / dt" and "junction temperature change rate dT / dt" as the direct parent nodes of IGBT overheating failure based on the causal graph. These are input into an LSTM model, and a dynamic threshold adjustment module dynamically calculates and outputs a fault judgment threshold based on grid load and environmental parameters. An early warning module compares the fault risk value with the fault judgment threshold. When the fault risk value exceeds the fault judgment threshold, an early warning message is generated and displayed. Employing a hierarchical computing architecture, and combining sensor data fusion technology and causal reasoning mechanisms, real-time monitoring and intelligent analysis of multi-source sensor data are achieved. This eliminates the need to transmit all sensor data to a central server for analysis, significantly reducing communication latency and computational burden, and improving the timeliness of fault warnings.

[0027] In a more specific embodiment provided by the present invention, the data acquisition module includes: Sensor networks include voltage sensors, current sensors, temperature sensors, and vibration sensors; A synchronous acquisition unit is provided, which uses GPS timestamps to synchronize multi-source sensor data (error < 1ms) to ensure the time consistency of multi-source sensor data. The data preprocessing unit uses a sliding window statistical method to identify and remove outliers in the original data, and uses linear interpolation to fill in missing data values, generating clean data that is then transmitted to the data fusion module.

[0028] It also includes a data storage unit: multi-source sensor data is temporarily stored in the local cache of the edge device and marked in the format of "device ID - time - parameter type" for easy traceability later.

[0029] In a more specific embodiment of the present invention, the data fusion module employs a Kalman filter to fuse the output data from multiple sensors, generating a more accurate system state assessment value. The data fusion module introduces an adaptive optimization mechanism, dynamically adjusting the noise covariance matrix of the Kalman filter based on long-term historical data and real-time operating status, thereby improving noise suppression capabilities under different environments and sensitivity to abnormal data.

[0030] The data fusion module includes: The state vector construction unit is used to construct a system state vector with voltage, current, temperature and vibration amplitude as components based on the purification data. Receive cleanup data and use the cleanup data to construct a system state vector. (U is voltage, I is current, T is temperature, and A is vibration amplitude), the subscript K represents the k-th discrete time point. For example, take k=1000 (corresponding to t=10s, the 1000th sampling point) as the initial time.

[0031] The Kalman filter unit is used to filter and fuse the system state vector and output the system state evaluation value. State equations (process models):

[0032] in: It is the state transition matrix, which describes how the state evolves from time K-1 to time K; It is process noise, with a mean of zero and a covariance matrix of... The normal distribution, i.e. .

[0033] Observation equations (measurement model):

[0034] in, It is the observation vector; It is the observation matrix; It is observation noise, which follows a pattern with zero mean and a covariance matrix of... The normal distribution, i.e. This represents the measurement error of the sensor.

[0035] Initial state estimation: The initial state estimate is passed through the previous state estimation. Calculation of the moving window mean of the raw data at time (time) To initialize the window length, e.g., 10 seconds.

[0036]

[0037] in, The original observation values ​​are N, which is the number of sampling points within the window (e.g., N=1000).

[0038] Initial error covariance: The initial error covariance matrix reflects the uncertainty of the initial state estimation and is calculated by the variance of the data within the window.

[0039]

[0040] in, For voltage window variance, This represents the variance of the current window, and so on.

[0041] Process noise covariance: reflects the uncertainty of system state evolution. The initial value is set through historical data statistics (such as state fluctuations during load changes).

[0042]

[0043] in, For voltage process noise variance, This represents the variance of the current process noise, and so on.

[0044] Observation noise covariance: reflects the sensor measurement error, and the initial value is obtained from the sensor manual or calibration experiment.

[0045]

[0046] in, For the voltage sensor noise variance, This represents the noise variance of the current sensor, and so on.

[0047] Output: Initialization parameter set [ , , , , , ].

[0048] An adaptive optimization unit is used to dynamically adjust the process noise covariance matrix and the observation noise covariance matrix during the Kalman filtering process.

[0049] 1. Forecasting Phase (Time Update) Based on the state equation and the observation equation, predict the state and error covariance at the current moment.

[0050] Input: Initialization parameter set [ , , , , , ]; State prediction: ; Error covariance prediction: ; Output: Predicted state and prediction error covariance ; 2. Update Phase (Measurement Update) Input: Predicted state and prediction error covariance and the observations from the previous time step

[0051] Combined with the observations at the current moment That is, the sensor measurement vector (measurements of voltage, current, temperature, and vibration) at time k corrects the predicted state and suppresses noise interference.

[0052]

[0053] Calculate the filter residual: ; Calculate residual covariance: ; Calculate the Kalman gain: ; Kalman gain Determine the weights of the predicted and observed values: The larger the value, the greater the impact of the observation on the state estimate.

[0054] Updated state estimate: ; Update error covariance: ; Output: Evaluated state value after fusion and updated error covariance ; 3. Adaptive adjustment of noise covariance; Traditional Kalman filter assumptions and While known and fixed, sensor noise and environmental interference can vary with operating conditions in practice. Adaptive optimization dynamically adjusts the filter residual by calculating its statistical characteristics in real time. and This improves the robustness of filtering.

[0055] Input: The residual at time k=1001 , and the residual sequence of the first 60 sampling points.

[0056] residual The root mean square (RMS) reflects the degree of deviation between the observed and predicted values, and is defined as:

[0057] Where M is the length of the sliding window (e.g., M=60, corresponding to 1 second of data, sampling rate 100Hz).

[0058] Process noise covariance adjustment:

[0059] in The initial process noise covariance, For adjustment coefficients (such as) =0.1).

[0060] Observation noise covariance adjustment:

[0061] in For the initial observation noise covariance, For adjustment coefficients (such as) =0.05).

[0062] Output: Adaptively adjusted noise covariance , This is used for filtering calculations at the next time step, k=1002.

[0063] The final output is the fused state evaluation value. This refers to the optimal estimates of voltage, current, temperature, and vibration amplitude, which generate system state assessment values ​​for subsequent fault prediction modules.

[0064] In a more specific embodiment provided by this invention, based on the fused system state evaluation values, the PC algorithm (Peter-Clark algorithm) or the NOTEARS algorithm (a causal discovery algorithm based on continuous optimization) is introduced to learn the causal relationship graph between variables from historical data. Based on the pre-trained causal graph, only variables that have a direct causal relationship with the target fault (such as "current increase to over-temperature fault") are selected as feature inputs of the LSTM model, irrelevant features are eliminated, and the root cause path of the fault is output to assist maintenance personnel in quickly locating the fault source.

[0065] The fault prediction module includes: The causal graph construction unit constructs a causal graph reflecting the causal relationships between variables based on the system state evaluation value; In the initial stage of system deployment, an initial causal graph G = (V, E) is constructed using historical normal operation and fault data (at least one year, including normal operating conditions and typical fault conditions, such as battery overcurrent, PCS overtemperature, and other labeled fault data). Where: node V represents all sensor variables and derived feature variables; directed edge E represents the causal relationship between variables ("current increase to overtemperature fault").

[0066] Common types of variables in energy storage systems are shown in Table 1: Table 1

[0067] Output cause-effect graph Stored as "nodes (variables) + directed edges (causal relationships)," for example: .

[0068] The feature filtering unit is used to filter out features directly related to the target fault from the real-time input system state evaluation values ​​based on the cause-effect graph. Classified by fault type, the direct causal parent node of each fault (i.e., the input features of subsequent LSTM) is recorded. Common fault types are shown in Table 2: Table 2

[0069] Based on the features that are directly related to the target fault, a basic LSTM model is trained, which has a preliminary fault prediction capability.

[0070] The LSTM prediction unit is used to predict the fault risk value sequence for the next hour based on the selected features and output the fault risk value.

[0071] (1) Data acquisition and preprocessing Input: Data output from the data fusion module

[0072] Data segmentation: Combining a sliding window (e.g., window width = 60s, step size = 1s) to evaluate state values. Divide into segments.

[0073] Feature extraction: Extract features within the window (such as temperature slope) Current fluctuation Voltage fluctuations Vibration mean wait).

[0074] Output: Feature matrix [ , , , ,...] (2) Online incremental causal discovery (dynamic adaptation of causal reasoning) During system operation, changes in the causal structure are continuously monitored. When a new, stable causal edge is detected (e.g., an increase in vibration leading to an anomaly in temperature, where this edge is statistically significant), it is added to the causal graph, indicating that the system may have developed a new failure mode or a performance degradation path.

[0075] (3) Causal feature selection (LSTM input optimization) This approach filters out core features directly related to the currently predicted fault from real-time features, reducing the computational cost of LSTM. Example of causal feature filtering (using "battery overcurrent" as an example): Input: Feature matrix [T_slope, I_range, U_range, A_mean, ...] and fault label "battery overcurrent"; Cause-effect graph reasoning: Find the direct parent node V = {peak charging / discharging current, temperature slope (dT / dt), current slope (dI / dt)} related to "battery overcurrent" in the cause-effect graph. Output: Feature matrix of the filtered direct causal parent nodes [peak charging / discharging current, temperature slope (dT / dt), current slope (dI / dt)] (4) Real-time fault risk prediction using LSTM model Long Short-Term Memory (LSTM) networks are used to perform trend analysis on long-term series data to establish a model of system health status degradation.

[0076] LSTM model training: Input: The filtered direct causal parent node feature matrix Output: A sequence of fault risk values ​​for the next hour (3600 points, format: timestamp, risk value), i.e., fault risk values.

[0077] Adaptive optimization: Every day at midnight, local historical data (last 7 days) is used to simultaneously optimize the causal graph and LSTM model to avoid performance drift.

[0078] In another embodiment of the present invention, the present invention also provides a fault early warning method based on multi-data fusion and causal reasoning mechanism. The system described above is characterized by comprising the following steps: S1. Through the data acquisition module, multi-source sensor data of the high-voltage cascaded energy storage converter system are collected in real time and synchronously, and the multi-source sensor data is preprocessed to obtain purified data. S2. The purification data is fused using an adaptive Kalman filter algorithm through the data fusion module to obtain the system status evaluation value. S3. Through the fault prediction module, the pre-constructed causal graph is used to filter features directly related to the target fault from the system state evaluation value, and the filtered features are input into the pre-trained LSTM model for prediction to obtain the fault risk value. S4. The fault risk value is received through the dynamic threshold adjustment module, and the fault judgment threshold is dynamically calculated in combination with the power grid load and environmental parameters. S5. The early warning module compares the fault risk value with the fault judgment threshold. When the fault risk value exceeds the fault judgment threshold, an early warning message is generated and displayed.

[0079] Step S1 ensures the temporal consistency and numerical reliability of the original data through GPS synchronization and sliding window preprocessing. Step S2 dynamically adjusts the noise covariance using adaptive Kalman filtering, ensuring that the system state assessment values ​​maintain high accuracy and robustness under different operating conditions. Step S3 relies on causal graphs to filter direct parent node features, significantly reducing the dimensionality of the LSTM input and improving model interpretability, while avoiding the introduction of noise by irrelevant variables. Step S4 introduces grid load and environmental parameters to construct a dynamic baseline, enabling the fault judgment threshold to have adaptive capabilities under operating conditions. Step S5 transforms abstract risk values ​​into actionable operation and maintenance instructions through multimodal visualization and interactive troubleshooting guidance. Each step is interconnected: the cleaned data in S1 forms the basis for the fusion in S2; the state assessment values ​​in S2 are the input for causal filtering in S3; the risk values ​​in S3 are the basis for threshold calculation in S4; and the dynamic thresholds in S4 are the criteria for early warning decisions in S5.

[0080] In a more specific embodiment provided by the present invention, the specific process of step S2 is as follows: S21. Based on the purification data, construct a system state vector with voltage, current, temperature and vibration amplitude as components; S22. Perform Kalman filtering iteration on the system state vector and output the fused state evaluation value; S23. During the Kalman filter iteration process, the process noise covariance matrix and the observation noise covariance matrix are dynamically adjusted based on the statistical characteristics of the filter residuals.

[0081] The specific process of S23 is as follows: S231. Calculate the root mean square value of the filtered residual within the set sliding window; S232. Based on the root mean square value, the process noise covariance matrix and the observation noise covariance matrix are adjusted proportionally or exponentially.

[0082] In a more specific embodiment provided by the present invention, the specific process of S3 is as follows: S31. Based on the real-time acquired system state evaluation value, extract the feature variables within the time window; S32. Based on the initial causal graph, select direct causal parent node variables that are directly related to the target fault from the feature variables.

[0083] This application achieves the following: In the fault prediction task of high-voltage cascaded energy storage converter system, to address the shortcomings of existing technologies such as noise sensitivity due to full feature input, submersion of early weak fault signals, and insufficient model generalization ability, a mechanism for extracting physically interpretable features within a limited time window is adopted, combined with strict causal path screening using a pre-constructed causal graph. This ensures that the input space of the LSTM model is precisely constrained to the set of direct causal variables of the target fault. Since there is a clear causal dependency between the input variables and the occurrence of the fault, the model's modeling of abnormal evolution trends is more focused, significantly improving the early identification sensitivity of slow degradation faults such as IGBT thermal stress accumulation and increased capacitor dielectric loss.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fault early warning system based on multi-data fusion and causal reasoning mechanisms for fault prediction in high-voltage cascaded energy storage converter systems, characterized in that, The system includes: The data acquisition module is used to simultaneously acquire data from multiple sensor sources, perform data preprocessing, and output purification data. The data fusion module is used to fuse the purification data through adaptive Kalman filtering to generate system status assessment values; The fault prediction module is used to filter out features directly related to the target fault from the system state evaluation value based on the pre-constructed causal graph, and input the filtered features into the pre-trained LSTM model to output the fault risk value. The dynamic threshold adjustment module is used to receive the fault risk value and, in combination with the power grid load and environmental parameters, dynamically calculate and output the fault judgment threshold. The early warning module is used to compare the fault risk value with the fault determination threshold, and generate and display early warning information when the fault risk value exceeds the fault determination threshold; The data acquisition module and the data fusion module are deployed at the edge computing end of the high-voltage cascaded energy storage converter system, while the fault prediction module, the dynamic threshold adjustment module, and the early warning module are deployed in the cloud.

2. The fault early warning system based on multi-data fusion and causal reasoning mechanism according to claim 1, characterized in that, The data acquisition module includes: Sensor networks include voltage sensors, current sensors, temperature sensors, and vibration sensors; A synchronous acquisition unit, wherein the synchronous acquisition unit uses GPS timestamps to synchronize multi-source sensor data; The data preprocessing unit uses a sliding window statistical method to identify and remove outliers in the original data, and uses linear interpolation to fill in missing data values ​​to generate cleaned data.

3. The fault early warning system based on multi-data fusion and causal reasoning mechanism according to claim 2, characterized in that, The data fusion module includes: The state vector construction unit is used to construct a system state vector with voltage, current, temperature and vibration amplitude as components based on the purification data. The Kalman filter unit is used to filter and fuse the system state vector and output the system state evaluation value. An adaptive optimization unit is used to dynamically adjust the process noise covariance matrix and the observation noise covariance matrix during the Kalman filtering process.

4. A fault early warning system based on multi-data fusion and causal reasoning mechanism according to claim 3, characterized in that, The adaptive optimization unit calculates the root mean square value of the filtered residual through a sliding window, and adjusts the process noise covariance matrix and the observation noise covariance matrix according to the root mean square value.

5. A fault early warning system based on multi-data fusion and causal reasoning mechanism according to claim 3, characterized in that, The fault prediction module includes: The causal graph construction unit constructs a causal graph reflecting the causal relationships between variables based on the system state evaluation value; The feature filtering unit is used to filter out features directly related to the target fault from the real-time input system state evaluation values ​​based on the cause-effect graph. The LSTM prediction unit is used to predict the fault risk value sequence for the next hour based on the selected features and output the fault risk value.

6. A fault early warning system based on multi-data fusion and causal reasoning mechanism according to claim 5, characterized in that, The causal graph construction unit uses the PC algorithm or the NOTEARS algorithm to construct the initial causal graph.

7. A fault early warning method based on multi-data fusion and causal reasoning mechanism, using the system described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Through the data acquisition module, multi-source sensor data of the high-voltage cascaded energy storage converter system are collected in real time and synchronously, and the multi-source sensor data is preprocessed to obtain purified data. S2. The purification data is fused using an adaptive Kalman filter algorithm through the data fusion module to obtain the system status evaluation value. S3. Through the fault prediction module, the pre-constructed causal graph is used to filter features directly related to the target fault from the system state evaluation value, and the filtered features are input into the pre-trained LSTM model for prediction to obtain the fault risk value. S4. The fault risk value is received through the dynamic threshold adjustment module, and the fault judgment threshold is dynamically calculated in combination with the power grid load and environmental parameters. S5. The early warning module compares the fault risk value with the fault judgment threshold. When the fault risk value exceeds the fault judgment threshold, an early warning message is generated and displayed.

8. The fault early warning method based on multi-data fusion and causal reasoning mechanism according to claim 7, characterized in that, The specific process of step S2 is as follows: S21. Based on the purification data, construct a system state vector with voltage, current, temperature and vibration amplitude as components; S22. Perform Kalman filtering iteration on the system state vector and output the fused state evaluation value; S23. During the Kalman filter iteration process, the process noise covariance matrix and the observation noise covariance matrix are dynamically adjusted based on the statistical characteristics of the filter residuals.

9. A fault early warning method based on multi-data fusion and causal reasoning mechanism according to claim 8, characterized in that, The specific process of S23 is as follows: S231. Calculate the root mean square value of the filtered residual within the set sliding window; S232. Based on the root mean square value, the process noise covariance matrix and the observation noise covariance matrix are adjusted proportionally or exponentially.

10. A fault early warning method based on multi-data fusion and causal reasoning mechanism according to claim 1, characterized in that, The specific process of S3 is as follows: S31. Based on the real-time acquired system state evaluation value, extract the feature variables within the time window; S32. Based on the initial causal graph, select direct causal parent node variables that are directly related to the target fault from the feature variables.

Images