Intelligent fault diagnosis and analysis system and method for power secondary equipment

Abstract

The invention relates to the technical field of power system automation, and particularly discloses an intelligent fault diagnosis and analysis system and method for power secondary equipment, and the method comprises the following steps: constructing a sensor network, collecting the operation data of a power system, and extracting the characteristics related to the fault of the power secondary equipment in the operation data; constructing a fault diagnosis model of the power secondary equipment, and obtaining faults of the power secondary equipment in the monitoring time period; constructing an association period according to the fault occurrence time point, generating coordinate points according to the fault occurrence sequence in the association period, and clustering the coordinate points to obtain a cluster; and according to the density of the clustering clusters, strong correlation faults of the faults are obtained, and when the faults occur, prompt information is sent to maintain the strong correlation faults corresponding to the occurring faults in advance. According to the invention, the real-time accurate diagnosis, strong correlation early warning and interpretable visual root cause analysis of the power secondary equipment fault are realized, and the safe operation of the system is effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of power system automation, and in particular to an intelligent fault diagnosis and analysis system and method for power secondary equipment. Background Art

[0002] With the continuous development and increasing intelligence of power systems, substation secondary equipment plays an increasingly important role in grid operation. However, traditional secondary equipment fault diagnosis methods often rely on manual inspections and empirical judgment, resulting in low diagnostic efficiency, low accuracy, and slow response. In the complex and changing power grid operating environment, secondary equipment failures can trigger a chain reaction, posing a serious threat to the safe and stable operation of the grid.

[0003] In recent years, the development of new technologies such as intelligent sensing, big data, and artificial intelligence has provided new means for diagnosing faults in power secondary equipment. By monitoring the operating status of secondary equipment in real time and leveraging big data analysis and artificial intelligence, rapid and accurate diagnosis of secondary equipment faults can be achieved. However, existing technologies mostly focus on real-time or short-term data analysis and lack the ability to deeply mine and utilize long-term time series data. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent fault diagnosis and analysis system and method for power secondary equipment to solve the above technical problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solutions: The intelligent fault diagnosis and analysis method for power secondary equipment includes the following steps: Build a sensor network to collect power system operation data, and use big data technology to extract features related to power secondary equipment failures in the operation data; Based on the extracted features, a fault diagnosis model for the power secondary equipment is constructed, and the faults occurring in the power secondary equipment within a preset monitoring period are obtained through the fault diagnosis model; A correlation cycle is constructed based on the time point of the fault occurrence, and coordinate points are generated according to the order of fault occurrence within the correlation cycle. The coordinate points are clustered to obtain clusters; The strongly correlated faults of the faults are obtained according to the density of the clusters. When a fault is detected, prompt information is triggered to perform early maintenance on the strongly correlated faults corresponding to the fault that has occurred.

[0006] A further solution: before extracting features related to power secondary equipment faults in the operating data, it also includes: cleaning, denoising and normalizing the collected operating data.

[0007] Further solution: The steps of generating coordinate points include: Select a fault as the target fault, record the time point t1 when it occurs, and set the associated period [t1, t1 + Δt], where Δt represents the length of the preset associated period; Identify other faults that occur within the associated period, and generate coordinate points (A1, A2,..., An) in the order of occurrence, where An represents the order in which the nth type of fault occurs within the associated period.

[0008] Further solution: When there is a target fault within the associated period, discard the current associated period.

[0009] Further solution: The steps to obtain the clustering clusters include: Set the clustering radius R, and cluster the coordinate points through the DBSCAN algorithm to obtain the initial clusters; Obtain the minimum circumscribed polyhedron of the initial cluster, divide the minimum circumscribed polyhedron into m target bodies with the same volume and shape, where m is the preset quantity; Statistically analyze the density of the coordinate points within a single target body. If the density of the coordinate points is less than the preset value, delete the coordinate points within the target body; Re-cluster the coordinate points within the minimum circumscribed polyhedron through the DBSCAN algorithm, and repeat the above steps until no more coordinate points are deleted to obtain the clustering clusters.

[0010] Further solution: Determining the strongly associated faults includes: Calculate the theoretical centroid of the clustering cluster with the maximum density, and obtain the values of each dimension in the coordinate points corresponding to the theoretical centroid. Take the type of the fault corresponding to the smallest value as the strongly associated fault of the current fault.

[0011] Further solution: After a fault occurs, it also includes: Use deep learning technology to analyze the cause of the fault; Visually display the occurred faults and the causes of the faults.

[0012] The intelligent fault diagnosis and analysis system for secondary power equipment includes: Acquisition module: Construct a sensor network to collect the operation data of the power system, and use big data technology to extract the features related to the faults of secondary power equipment from the operation data; Diagnosis module: Build a fault diagnosis model for secondary power equipment based on the extracted features, and obtain the faults that occur in secondary power equipment within the preset monitoring period through the fault diagnosis model; Analysis module: Construct an associated period according to the time point when the fault occurs, generate coordinate points based on the order of fault occurrence within the associated period, and cluster the coordinate points to obtain the clustering clusters Collaboration Module: Obtain strongly associated faults of a fault based on the density of the clustering cluster. When a fault is detected, trigger a prompt message to perform preventive maintenance on the strongly associated faults corresponding to the occurred fault.

[0013] Advantages of the present invention: Compared with the prior art: 1) The present invention constructs a fault diagnosis model based on a sensor network, maps the target fault to multi-dimensional coordinates within an associated period, and combines density adaptive DBSCAN clustering to dynamically mine long-term time series data. This process not only retains the sequence of faults but also automatically eliminates low-value samples in historical data with high noise, continuously updates the clustering cluster structure, and realizes the full-life cycle tracking and trend recognition of the operating state of secondary equipment; 2) The present invention extracts the theoretical centroid of clustering theory and determines strongly associated faults by the minimum value method of coordinate dimensions, and can automatically screen trigger links from complex fault relationships. After detecting the initial fault, the system can predict the subsequent faults that are most likely to be caused by it and immediately send maintenance prompts, realizing the transformation from "post-event emergency" to "pre-event prevention", blocking the fault propagation path, and reducing the risk of large-area outages; 3) The present invention combines a deep learning model to perform multi-dimensional feature explanations on the causes of faults, and presents key indicators, causal links, and time evolution processes with a visual dashboard. The visualization layer converts the output of abstract algorithms into intuitive charts, heat maps, and flowcharts, enabling operation and maintenance personnel to quickly grasp the essence of faults and locate weak links without a special data science background. Description of the Drawings

[0014] The present invention will be further described below with reference to the drawings.

[0015] Figure 1 It is a schematic flow chart of the intelligent fault diagnosis and analysis method for power secondary equipment of the present invention. Detailed Embodiments

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0017] Please refer to Figure 1 As shown, the present invention is an intelligent fault diagnosis and analysis system and method for power secondary equipment, including: Deploy multiple types of intelligent sensors such as temperature, current, voltage, partial discharge, and operating displacement at key parts of secondary equipment such as protection devices, measurement and control devices, relays, and switchgear to collect corresponding operation data, and quasi-real-time access the original measurement values and events to the edge gateway through IEC61850-9-2LE sampled values or synchronous Ethernet; to avoid interference on the primary high-voltage side, select wireless temperature patches and fiber optic current sensors for the secondary circuit of current transformers as a redundant acquisition solution. The edge gateway is built-in with a GPS timing module and a PTP precision clock to timestamp all data to ensure millisecond-level alignment; then, on the gateway side, use the Kafka streaming platform to distinguish different device tags by topic dimension and push the data to the upper-layer HDFS distributed file system, and use the Parquet columnar storage format for compression and storage to reduce disk I / O; SparkStreaming integrates DeltaLake to perform CDC queue comparison on the arriving data, and automatically eliminates records with abnormal formats, CRC check failures, and missing fields; in the feature engineering stage, for continuous quantities such as current harmonic distortion rate, bus voltage amplitude drift, and breaker trip coil resistance change, extract the mean, peak, and skewness within the sliding window, and calculate the event frequency and average duration for discrete quantities such as protection start flags and telemetry jitter counts; combine ST-LSTM for multi-scale embedding of time series features, then complete dimension compression through an autoencoder and extract local texture evolution features using a gray-level co-occurrence matrix to capture potential early weak signals of faults; the entire process is orchestrated in a DAG manner on the Yarn cluster. When receiving the PMU wide-area measurement wake-up signal, the system automatically triggers the FlinkCEP rule to combine the voltage sag mode sequence and write it into the ClickHouse OLAP engine in real time for the upper-layer diagnostic model to call, realizing a seamless closed-loop from bottom-layer perception to high-layer feature extraction; In a preferred embodiment of the present invention, on the edge gateway side, the timestamps of different sensor streams are precisely aligned by PTP and duplicate frames with disordered time sequences are removed. Then, for missing fields, multi-channel correlation regression is used for interpolation. For example, when a certain phase voltage has an instantaneous packet loss, a regression model is constructed using the in-phase historical window and the adjacent phase voltage to fill it; for waveforms with sudden changes in the sampling period, they are first resampled at the least common multiple sampling rate, and then the continuous waveform is restored using cubic Hermite interpolation; for slow variables such as temperature and vibration, five-point Savitzky–Golay filtering is used to smooth the trend, and for switch quantity jitter, the Hough transform is used to identify and merge pseudo-jumps; for current signals with mixed power frequency and high-frequency harmonics, wavelet soft threshold adaptive denoising is used, and the threshold is selected based on Stein's non-linear risk estimation. Electromagnetic interference spikes are determined to be abnormal by the three-sigma principle in cooperation with the Hampel filter and then replaced with the sliding median; to remove the differences in sensor dimensions and range drifts, continuous variables are uniformly standardized by Z-score and the mean and variance parameters are recorded for subsequent online inference reuse. Discrete features of event counting types are logarithmically scaled according to the device category. All processing results are appended and written into the cleaning layer in the Parquet column storage format for feature engineering to call; It can be understood that the purpose of cleaning, denoising, and normalization is to minimize the interference of sensor noise, communication jitter, and dimensional inconsistency on subsequent feature extraction and model training, so that the same physical quantity remains comparable and additive in different sampling channels and different time periods. This ensures that statistical features, frequency domain features, and time series models are more focused on the health status of the device itself rather than data defects when capturing weak abnormal patterns, thereby providing a stable and reliable input basis for fault clustering analysis, strong correlation fault determination, and deep learning cause tracing, and promoting the accuracy and robustness of the overall diagnosis chain; After completing feature engineering, first clean and organize the historical labeled secondary equipment fault logs, pair the feature vectors corresponding to each record with the fault type labels one by one, and use stratified random sampling to divide the data into a training set, a validation set, and a hold-out test set; then use the maximum correlation minimum redundancy and mutual information double screening to select the feature subset with the highest information gain. For the problem of extremely unbalanced fault samples, use the SMOTE-Tomek combination method to oversample rare labels such as disconnection and braking failure, and at the same time delete the noise synthetic samples to keep the distribution reasonable; for long sequence features such as temperature, current harmonics, and partial discharge amplitude, construct a bidirectional LSTM-Attention subnetwork to capture the temporal dependencies before and after, and for discrete event sequences such as relay action positions and telemetry signal jitters, build a one-dimensional convolutional subnetwork to extract local patterns. Then, in the fusion layer, use multi-head self-attention to align and splice the two-way features and output them to the fully connected layer. Finally, obtain the probability distribution of each fault category through Softmax; in the training stage, use the AdamW optimizer combined with the cosine annealing learning rate, calculate the FocalLoss and macro-average F1 on the validation set after each round and save the optimal weights according to the early stopping strategy. After training, convert the network to ONNX and deploy it on the edge inference engine. During online operation, grab the latest feature tensors with a five-second sliding window to perform forward inference. If the same label is output in three consecutive windows, it is determined that there is a corresponding fault in the window and written to the event bus. After the monitoring period ends, the background task automatically aggregates all fault events for subsequent correlation analysis; It should be noted that the construction and deployment of the diagnostic model is to quickly map high-dimensional and heterogeneous real-time features into clear fault categories, so that fault detection can be completed on the edge side closest to the data source, avoiding bandwidth occupation and latency caused by the transmission of original massive data; reduce the interference of redundant information and class imbalance on the discrimination boundary through feature screening and sample balancing, and improve the generalization ability of the model in new scenarios; adopt a multi-subnetwork fusion strategy to simultaneously depict the slow change trend of continuous quantities and the mutation mode of discrete quantities, so that the diagnostic logic can identify complex coupling relationships rather than relying solely on single-threshold decision-making, thus being more sensitive to early mild anomalies; after encapsulating the model as a lightweight inference service, the result can be given in milliseconds, providing an accurate timestamp and clear class label for subsequent clustering and strong correlation fault mining, ensuring that the entire intelligent fault diagnosis chain is based on reliable and real-time input, and realizing the closed-loop from data to decision-making; Construct an association period according to the time point when the fault occurs, generate coordinate points according to the order of fault occurrences within the association period, and cluster the coordinate points to obtain clustering clusters; In another preferred embodiment of the present invention, generating coordinate points includes: First, load a fault type mapping table in the event stream database. Arrange all concerned secondary equipment faults in a fixed order to form a vector dimension index list. For example, the first dimension corresponds to the operation of bus differential protection, the second dimension corresponds to the failure of standby power supply input, the third dimension corresponds to the refusal of sectional switch operation, the fourth dimension corresponds to feeder overload, and it continues to all fault types. Subsequently, select a certain fault as the target fault and record its occurrence time t1 of any time. Then, read the associated cycle length Δt according to the configuration file to generate a time window [t1, t1 + Δt]. The Flink job sorts and scans all fault trigger records within this time window according to event time, calculates the sequence number of each record within the window. When a certain fault event is scanned, immediately find its corresponding fixed dimension position k in the state backend and write the value j (i.e., the sequence) to the k-th element of the vector. If the same fault appears again within the window, only keep the earliest sequence number without overwriting. For the fault dimensions that do not appear within the window, uniformly fill zero placeholder values. Finally, obtain a sparse coordinate vector with a fixed dimension order but the numerical values of each element changing according to the sequence, such as (1, 3, 2,...), and output it to Kafka for subsequent clustering steps to call; It can be understood that keeping the coordinate dimensions fixed and only allowing the numerical values on the dimensions to change with the fault occurrence sequence can ensure that the vectors generated in different time windows have exactly the same geometric coordinate system and are thus directly comparable. This enables clustering algorithms based on distance or density to accurately measure the similarity between fault patterns without being affected by errors caused by dimension exchange. In addition, the sparse vector maintains structural stability in the high-dimensional space, facilitating fast indexing and incremental calculation, which helps DBSCAN efficiently aggregate fault clusters with similar evolution trajectories from large-scale historical data, and further provides a reliable and unified data basis for extracting strongly correlated faults and formulating preventive maintenance strategies; It should be noted that when the target fault exists within the associated cycle, discard the current associated cycle; Another preferred embodiment of the present invention, obtaining the clustering clusters includes: On the distributed Spark nodes, the cluster radius R is selected using the inflection point of the k-distance curve of the historical coordinate points, and the scikit-learn DBSCAN algorithm is called to scan all coordinate points in the broadcast variable in sequence. Points with a distance less than R and a number of neighborhood samples of not less than minPts are classified into the same initial cluster. After the initial clustering is completed, a three-dimensional (or multi-dimensional in high-dimensional scenarios) convex hull approximation of the coordinate set of each cluster is calculated, and the minimum circumscribed polyhedron is further obtained by the fast rotating caliper algorithm. If the dimension exceeds three, it degenerates into a hyperrectangular bounding box. Then, according to the geometric segmentation rule of m=64 (which can be other values), the circumscribed polyhedron is divided into the same number of segments on each coordinate axis, and the circumscribed polyhedron is subdivided into m targets with identical volumes and shapes are identified, and a point-to-grid mapping is established using a grid index hash table. The system traverses all grids, counting the number of internal coordinate points and dividing the result by the grid volume to obtain the density. This density is then compared to the sparsity threshold configured by the operations engineer in the core node. Any grid with a density below the threshold is considered a noise area, and its internal coordinate points are removed and the hash table is updated. After this removal, only the remaining coordinate points are retained, and DBSCAN is restarted. The cycle of calculating the minimum circumscribed polyhedron, dividing the grid, and removing low-density grids is repeated on the new point set until no coordinate points are removed in two consecutive iterations. The output cluster labels are then written to Hive for subsequent strong correlation fault algorithm reading. It should be noted that the addition of minimum circumscribed polyhedron gridding and sparse iterations after DBSCAN is intended to further remove data points that are captured by the initial radius but have sparse distribution edges and are likely to disrupt subsequent centroid calculations, while maintaining its natural adaptability to clusters of arbitrary shapes. Fine-grained density checks can shrink the cluster boundaries to the core area that is most representative of the fault mode, improving the stability and accuracy of the centroid coordinates when inferring strongly correlated faults. Multiple iterations ensure that the removal of sparse grids does not destroy the overall connectivity of the cluster, allowing the cluster shape to maintain physical interpretability during the gradual "convergence" process. The resulting clusters can more realistically reflect the evolution trajectory of the target fault in the time-series coordinate space, providing a more reliable data basis for proactive maintenance strategies. Obtain strongly correlated faults based on cluster density. When a fault occurs, send prompts to perform proactive maintenance on the strongly correlated faults corresponding to the fault. In another preferred embodiment of the present invention, obtaining strongly correlated faults of a fault includes: Traverse all clustering clusters, divide the number of coordinate points within the cluster by the volume of the minimum bounding polyhedron of the same cluster to obtain the density index, and select the cluster with the largest density value as the representative cluster; then calculate the arithmetic mean of the values of all coordinate points in the same dimension in this cluster to obtain the theoretical centroid, and then call the KD-Tree to search for the actual coordinate point P closest to C within the cluster, ensuring that the coordinates used satisfy the discrete sequential coding rule; read the integer values p1, p2,... of P in each dimension according to the predefined fault dimension mapping table, traverse these values to find the smallest element p_min greater than zero, and the corresponding dimension index k_min is mapped to the specific fault type. For example, if p2 is the smallest and equal to 1, it means that the fault in the second dimension appears first in the representative cluster. Then, mark this fault type as the strongly associated fault of the current target fault, and write the mapping result into Redis for the alarm subsystem to subscribe to. The reason for selecting the cluster with the largest density and using its theoretical centroid as the analysis base point is that this cluster can best represent the most frequent and typical associated scenarios of the target fault in historical data, and the centroid vector reflects the order of occurrence of various faults on average; in this sequential coding system, the smaller the value, the earlier the trigger. Therefore, taking the smallest and non-zero dimension can intuitively lock the fault type that is most often used as the cause or precursor. Through this method, the key fault with a strong predictive effect on the target fault can be quickly refined without a complex causal model, providing a one-hop maintenance pointer for the alarm system, helping the operation and maintenance personnel prepare maintenance resources in advance and cut off the fault chain, thereby ensuring the stability of the secondary system. A preferred embodiment of the present invention further includes, after a fault occurs: After a certain fault event is written into the event bus, the analysis module immediately traces back the multi-source original waveforms, protection start flags, telecontrol signal states, and topology switching information in a few minutes before and after the fault occurs, converts these data into a tensor sequence through the feature pipeline and combines it with the in-station topology Figure 1It is sent into the pre-trained time-series-topology joint deep model offline; at the front end of the model, bidirectional LSTM is used to encode the transient and steady-state features of each sensing channel. In the middle layer, the graph convolutional network is used to map the device connection relationship into an adjacency matrix to capture the mutual influence. The output layer gives the probability distribution of possible causes and the attention weights of key channels simultaneously under the multi-task framework. Subsequently, two interpretability algorithms, SHAP and integrated-gradients, are called to decompose the feature contribution of the model output, and several indicators and device nodes with the highest scores are obtained as suspected root causes, and the results are written into the time-series database together with the event ID. The visualization service reads these data and generates an interactive dashboard through ECharts. On the left side of the interface, a waveform superposition diagram with the timeline as the horizontal axis and fault trigger point markers are drawn. In the middle, a single-line diagram of the substation colored according to the attention weights is displayed. On the right side, a causal chain Sankey diagram and a SHAP bar chart are shown. Users can click on any node to view the corresponding numerical curve or historical screenshot, and the whole page supports adaptive switching between mobile phones and large screens; It should be noted that introducing deep learning root cause analysis can automatically induce complex correlation patterns by integrating time domain, frequency domain, and topology information without relying on artificial rules, enabling the system to quickly point to the most likely source even in the face of multiple concurrent faults or hidden causes. Then, presenting the interpretability results of the model in an intuitive graph can help operation and maintenance personnel understand the algorithm decision logic, avoid black-box guessing, and improve the trust in the model output. At the same time, integrating real-time waveforms, topology coloring, and causal chains in the same view through an interactive dashboard facilitates the duty personnel to complete positioning, comparison, and dispatching communication within dozens of seconds, providing a credible and traceable basis for formulating subsequent maintenance plans, arranging outage windows, and evaluating risks, thus enabling the entire intelligent fault diagnosis system of secondary equipment to form a closed loop of self-detection, self-interpretation, and self-display.

[0018] Intelligent Fault Diagnosis and Analysis System for Power Secondary Equipment, including: Acquisition module: Construct a sensor network to collect the operation data of the power system, and use big data technology to extract the features related to the faults of power secondary equipment from the operation data; Diagnosis module: Build a fault diagnosis model for power secondary equipment based on the extracted features, and obtain the faults that occurred in the power secondary equipment during the preset monitoring period through the fault diagnosis model; Analysis module: Construct an association period according to the time point when the fault occurs, generate coordinate points based on the order of fault occurrence within the association period, and perform clustering on the coordinate points to obtain clustering clusters; Collaboration module: Obtain the strongly associated faults of the fault according to the density of the clustering clusters, and when a fault is detected, trigger a prompt message to perform preventive maintenance on the strongly associated faults corresponding to the occurred fault.

[0019] The above has described in detail an embodiment of the present invention, but the above content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made in accordance with the scope of the application of the present invention shall still fall within the scope covered by the patent of the present invention.

Claims

1. An intelligent fault diagnosis and analysis method for secondary power equipment, characterized in that, It includes the following steps: Construct a sensor network to collect the operation data of the power system, and use big data technology to extract the features related to the faults of secondary power equipment from the operation data; Based on the extracted features, construct a fault diagnosis model for secondary power equipment, and obtain the faults that occurred in the secondary power equipment during the preset monitoring period through the fault diagnosis model; Construct an association period according to the time point when the fault occurs, generate coordinate points according to the order of fault occurrence within the association period, and cluster the coordinate points to obtain clustering clusters; Obtain the strongly associated faults of the fault according to the density of the clustering cluster. When a fault is detected, trigger a prompt message to perform preventive maintenance on the strongly associated faults corresponding to the occurred fault.

2. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 1, characterized in that Before extracting the features related to the faults of secondary power equipment from the operation data, it also includes: cleaning, denoising and normalizing the collected operation data.

3. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 1, wherein Generating coordinate points includes: Select a type of fault as the target fault, record the time point t1 when it occurs, and set the association period [t1, t1+Δt], where Δt represents the length of the preset association period; Identify other faults that occurred within the association period, and generate coordinate points (A1, A2,..., An) according to the occurrence order, where An represents the order of occurrence of the nth type of fault within the association period.

4. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 3, characterized in that When there is a target fault within the association period, discard the current association period.

5. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 1, characterized in that Obtaining the clustering cluster includes: Set the clustering radius R, and cluster the coordinate points through the DBSCAN algorithm to obtain the initial cluster; Obtain the minimum circumscribed polyhedron of the initial cluster, divide the minimum circumscribed polyhedron into m target bodies with the same volume and shape, where m is the preset quantity; Statistically analyze the density of the coordinate points within a single target body. If the density of the coordinate points is less than the preset value, delete the coordinate points within the target body; Re-cluster the coordinate points within the minimum circumscribed polyhedron through the DBSCAN algorithm, and repeat the above steps until no more coordinate points are deleted to obtain the clustering cluster.

6. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 1, characterized in that Determining the strongly associated faults includes: Calculate the theoretical centroid of the clustering cluster with the maximum density, and obtain the values of each dimension in the coordinate points corresponding to the theoretical centroid. Take the type of the fault corresponding to the smallest value as the strongly associated fault of the current fault.

7. The intelligent fault diagnosis and analysis method for secondary power equipment according to claim 1, characterized in that, After a fault occurs, it also includes: Use deep learning technology to analyze the cause of the fault; Visually display the occurred fault and the cause of the fault.

8. Intelligent Fault Diagnosis and Analysis System for Secondary Electrical Equipment, characterized in that, It includes: Acquisition module: Construct a sensor network to collect the operation data of the power system, and use big data technology to extract the features related to the faults of secondary power equipment from the operation data; Diagnosis module: Based on the extracted features, construct a fault diagnosis model for secondary power equipment, and obtain the faults that occurred in the secondary power equipment during the preset monitoring period through the fault diagnosis model; Analysis module: Construct an association period according to the time point when the fault occurs, generate coordinate points according to the order of fault occurrence within the association period, and cluster the coordinate points to obtain clustering clusters; Collaboration module: Obtain the strongly associated faults of the fault according to the density of the clustering cluster. When a fault is detected, trigger a prompt message to perform preventive maintenance on the strongly associated faults corresponding to the occurred fault.

Images