Power equipment fault intelligent diagnosis method and system based on deep learning

Abstract

The invention discloses a power equipment fault intelligent diagnosis method and system based on deep learning, and relates to the technical field of fault prediction and health management. An initial data set is constructed through multi-dimensional feature extraction, time sequence data smoothing processing and time-space neighborhood data analysis, and a feature change rate is combined to obtain a fault diagnosis result; and accurately identifying potential state transition candidate points of the equipment. Furthermore, a conversion type label is generated through noise filtering and historical path mode matching, the position of a key transfer node is determined by adopting a node position calibration technology, high-risk transfer nodes are marked in combination with fluctuation value gradual increase detection and risk accumulation judgment rules, and finally an early warning signal sequence of future state change is generated. According to the invention, through fusion of multi-level data processing and dynamic analysis logic, real-time monitoring and risk early warning of equipment state change are realized, a dynamic conversion identification mechanism is established, key nodes of state transition are automatically identified from continuous monitoring data, and the fault pre-judgment capability is improved.

Description

Technical Field

[0001] This invention relates to the field of fault prediction and health management technology, and in particular to a method and system for intelligent fault diagnosis of power equipment based on deep learning. Background Technology

[0002] As the core pillar of power grid operation, the reliable operation of power equipment is directly related to energy security, supply security, and stable socio-economic development. With the rapid expansion of power grid scale and the increasing complexity of the operating environment, equipment failure has become the primary hidden danger restricting system stability. Failures are often not sudden events, but rather a process that gradually evolves from minor anomalies into serious damage. This progressive degradation characteristic requires monitoring methods to capture the continuous trajectory of state changes, thereby enabling early intervention.

[0003] However, current diagnostic technologies are mostly limited to static identification of single fault types, failing to address the dynamic fluctuations in equipment health, leading to the accumulation of potential risks and even large-scale power outages. While existing diagnostic methods can handle common problems, they reveal significant shortcomings when dealing with differences between equipment. These methods are typically based on general thresholds or empirical rules, making it difficult to accurately separate and judge early partial discharge signals in transformer oil-paper insulation systems or microscopic wear signs on circuit breaker contact surfaces. Because different equipment have fundamentally different structures and workloads—for example, the interaction between the oil gap and solid insulation inside a transformer differs from the friction mechanism of the mechanical components in a circuit breaker—the fixed characteristic indicators used in current technologies often ignore these uniquenesses, resulting in missed or misjudged early signs, further allowing problems to quietly worsen.

[0004] This equipment specificity is further amplified in capturing the evolution patterns of faults, creating an even more challenging situation. Faults in different devices do not remain isolated but rather follow specific paths, gradually transforming. For example, partial discharge in a transformer originates from concentrated electric fields within the insulation cavity, initially manifesting as intermittent, weak pulses. However, as the cavity expands and gas accumulates in the oil, it gradually erodes the surrounding paper insulation, eventually leading to a through-breakdown, causing a surge in tank pressure and winding burnout. Similarly, contact wear in circuit breakers begins with the shedding of particles from the metal surface, initially only slightly increasing resistance. However, continuous friction evolves into unstable contact, generating localized high temperatures, triggering arc jumping and overheating melting, and even affecting the arc-extinguishing chamber, causing operational blockage. The signal characteristics of this transformation process are highly dependent on the equipment material and operating conditions. For instance, fluctuations in transformer oil temperature can mask discharge pulses, while the circuit breaker's opening and closing frequency accelerates the wear rate. Current methods, lacking fine-grained tracking of these specific patterns, cannot extract the critical transition points from massive amounts of time-series data, resulting in diagnostics remaining at a superficial level.

[0005] Therefore, how to establish a dynamic conversion identification mechanism to automatically identify key nodes of state transition from continuous monitoring data, targeting the unique evolution paths of transformer partial discharge to main insulation breakdown, circuit breaker contact wear to poor contact and then to overheating, has become a key issue in improving fault prediction capabilities. Summary of the Invention

[0006] This invention provides a method and system for intelligent fault diagnosis of power equipment based on deep learning. It establishes a dynamic transformation identification mechanism to identify key nodes of state transition from continuous monitoring data, thereby improving fault prediction capabilities, in response to the unique evolution paths of transformer partial discharge to main insulation breakdown, circuit breaker contact wear to poor contact and then to overheating.

[0007] This invention provides a deep learning-based intelligent fault diagnosis method for power equipment, executed by a computer, comprising: Based on the equipment operation data of power equipment, multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment are extracted. Based on the multi-dimensional feature vectors, an initial data set is generated by using a time-series point information smoothing method combined with spatiotemporal neighborhood data extraction technology. Based on the initial data set, the characteristic change rate calculation method is used to analyze the change trend within the local data range, determine the characteristic change rate, and determine potential state transition candidate points based on the characteristic change rate. Based on the potential state transition candidate points, data corresponding to the surrounding spatiotemporal neighborhood are selected to form a local data subset. Based on the local data subset, the device historical path pattern matching method is used to generate the corresponding device conversion type label. Based on the device conversion type label, key transfer features of the device are extracted. Based on the key transfer features of the device, node position calibration technology is used to perform continuous scanning to determine the device node position and the corresponding fluctuation value interval division result. Based on the location of the device node, the corresponding device signal fluctuation sequence is obtained. If the device signal fluctuation sequence is in a continuous upward state and meets the device risk accumulation judgment rule, it is marked as a high-risk transfer node of the device. Based on the high-risk transfer nodes of the equipment and combined with the results of the fluctuation value interval division, the time-series information of the high-risk transfer nodes of the equipment is smoothed to generate a warning signal sequence for future changes in the state of the equipment.

[0008] This invention provides a deep learning-based intelligent fault diagnosis method and system for power equipment. Addressing the business scenario problem of identifying potential risks and predicting future state changes during the gradual evolution of equipment states, it constructs an initial dataset through multi-dimensional feature extraction, time-series data smoothing, and spatiotemporal neighborhood data analysis. Combined with feature change rates, it accurately identifies potential state transition candidate points. Furthermore, it generates transition type labels through noise filtering and historical path pattern matching, determines the locations of key transition nodes using node position calibration technology, and marks high-risk transition nodes by combining fluctuation value increment detection and risk accumulation judgment rules. Finally, it generates a sequence of early warning signals for future state changes. This invention, through the integration of multi-level data processing and dynamic analysis logic, achieves real-time monitoring and risk warning of equipment state changes, significantly improving equipment operational safety and prediction accuracy. It provides intelligent decision support for equipment management, establishing a dynamic transition identification mechanism for unique evolution paths such as transformer partial discharge leading to main insulation breakdown, and circuit breaker contact wear leading to poor contact and overheating. This mechanism automatically identifies key nodes of state transition from continuous monitoring data, enhancing fault prediction capabilities. Attached Figure Description

[0009] Figure 1 This is one of the flowcharts of the intelligent fault diagnosis method for power equipment based on deep learning provided in the embodiments of the present invention; Figure 2 This is the second flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention; Figure 3 This is the third flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention; Figure 4 This is the fourth flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention; Figure 5 This is the fifth flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention; Figure 6 This is the sixth flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention; Figure 7 This is the seventh flowchart of the intelligent fault diagnosis method for power equipment based on deep learning provided in this embodiment of the invention. Detailed Implementation

[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] Reference Figure 1 This invention provides a deep learning-based intelligent fault diagnosis method for power equipment, comprising the following steps: Step 100: Based on the equipment operation data of the power equipment, extract the multidimensional feature vector related to the progressive evolution attributes of the power equipment; based on the multidimensional feature vector, generate an initial data set by using a time-series point information smoothing processing method combined with spatiotemporal neighborhood data extraction technology. First, multi-dimensional, continuous time-series equipment operation data generated during the operation of power equipment is collected. Feature information reflecting the gradual evolution of equipment status over time is extracted from this data, resulting in a multi-dimensional feature vector. This feature information is a multi-dimensional feature vector constructed by fusing different operational dimensions, characterizing the degradation process of equipment health status. Next, the extracted multi-dimensional feature vector is processed using a time-series information smoothing method to suppress fluctuations in the original data caused by noise, transient interference, or unstable acquisition, while preserving the main trends and key changes in equipment status evolution. Time-series smoothing makes the feature curves more stable and continuous, reducing the interference of random fluctuations on subsequent analysis, thus more clearly presenting the evolution path of equipment status.

[0012] After time-series smoothing, a spatiotemporal neighborhood data extraction technique is introduced. This technique considers not only the feature information at a single point in time but also the data from adjacent time periods (temporal neighborhood) and data from other potentially related monitoring points or feature dimensions within the same time period (spatial neighborhood), thus enhancing the context of the current feature vector. By fusing relevant information within the spatiotemporal neighborhood, the operating state of the device at a specific moment can be more comprehensively characterized, reducing deviations caused by random factors and generating an initial dataset. This initial dataset contains a smoothed feature sequence reflecting the device's own state evolution, as well as incorporating spatiotemporal context information.

[0013] Step 200: Based on the initial data set, the characteristic change rate calculation method is used to analyze the change trend within the local data range, determine the characteristic change rate, and determine potential state transition candidate points based on the characteristic change rate. Based on the obtained initial dataset, feature change rate calculation is introduced to analyze the changing trends of each feature dimension within a local data range. The feature change rate calculation method transforms static multidimensional feature points into rate vectors containing dynamic change information by calculating the instantaneous or average rate of change of feature values ​​within a local time window, thereby obtaining the feature change rate and directly characterizing the speed and acceleration of the evolution of equipment health status.

[0014] The calculation principle of the characteristic change rate calculation method is as follows: the equipment operation data contained in the initial data set is regarded as a continuous and dynamic evolution process. The characteristic change rate is obtained by calculating the change of the equipment operation data characteristics (such as vibration amplitude, leakage current, partial discharge intensity, etc.) in the initial data set per unit time. This change is the first derivative or discrete difference.

[0015] After obtaining a continuous sequence of characteristic change rates, potential candidate points for device state transitions are determined based on these rates. These candidate points are not simply extreme points, but rather key locations where the characteristic change rate undergoes a significant and sustained shift within a local area. For example, when the evolution rate of a key characteristic suddenly changes from a long-term low-level stable fluctuation to a sustained rapid upward or downward trend, its inflection point is considered a potential candidate point for state transition. Through synergistic analysis of multidimensional characteristic change rate curves, spatiotemporal regions where multiple characteristic rates simultaneously undergo trend changes can be identified.

[0016] Step 300: Based on the potential state transition candidate points, select the data corresponding to the surrounding spatiotemporal neighborhood to form a local data subset, and based on the local data subset, use the device historical path pattern matching method to generate the corresponding device conversion type label. Based on the identified potential state transition candidate points, for each candidate point, relevant data is selected from its surrounding spatiotemporal neighborhood to construct a local data subset that fully reflects the contextual evolution information of that point. This local data subset includes the candidate point's own characteristics and rate information, and expands the description of the candidate point from a temporal perspective to an event fragment describing the local state transition process. Therefore, in this way, the resulting local data subset can be a complete data entity representing a possible state transition.

[0017] Subsequently, for these local data subsets encapsulating event fragments, a device historical path pattern matching method is employed for analysis. This method compares and matches currently observed state transition event fragments with a library of typical state transition patterns pre-learned and abstracted from a large amount of historical operational data. This library contains spatiotemporal feature data of various standard paths, including known fault development, operational mode switching, and health state degradation. During the matching process, considering evolution trends, morphological similarities, and the collaborative change patterns of multiple features, the method identifies which type of historical transition pattern best matches the current fragment in the overall evolutionary trajectory. Through this pattern matching, a specific device transition type label can be generated for each potential state transition candidate event. This label indicates the transition category, such as "gradual degradation of insulation performance," "intermittent jamming initiation of mechanical components," or "smooth switching of operating load states."

[0018] It should be noted that the algorithm principle of the equipment historical path pattern matching method is as follows: First, the evolution process feature trajectory of the entire evolution process before, during, and after the fault is extracted from the massive equipment status data of power equipment (including normal degradation and various fault cases). For each evolution process feature trajectory, a historical path pattern is generated. This historical path pattern contains dynamic information such as the temporal sequence of state evolution, the rate of change, and the collaborative relationships of multiple features. Based on a local data subset, for each potential state transition candidate point, data within a time window before and after it is extracted from the local data subset, forming the current event path segment, i.e., the state transition event segment. Then, the state transition event segment is matched with the historical path pattern. For example, dynamic time warping, sequence similarity measurement, or methods based on temporal deep learning models can be used to calculate the similarity between the current state transition event segment and each historical path pattern in terms of overall shape, change trend, and key phase. The historical path pattern with the highest similarity is used as the corresponding equipment transformation type label.

[0019] Step 400: Extract key transfer features of the equipment based on the equipment conversion type label, and perform continuous scanning using node position calibration technology based on the key transfer features of the equipment to determine the equipment node position and the corresponding fluctuation value interval division result. Based on the generated equipment transition type labels, each classified state transition event is analyzed in depth to extract key equipment transition features that reflect the specific transition mode. These key equipment transition features are screened and refined; for example, they may be unique combinations of changes in the amplitude and duration of certain sensor readings during the transition event, or specific proportional relationships between different physical quantities before and after the transition. The purpose of extracting key transition features is to construct a highly condensed and discriminative feature signature for each identified transition mode, thereby condensing complex event fragments into a set of core indicators that can be used for comparison and quantitative evaluation.

[0020] Subsequently, leveraging key transition features, node position calibration technology is employed to continuously scan and precisely locate the entire time series, thereby correcting and verifying the previously identified, relatively coarse, potential state transition candidate points in time and space. Based on typical waveforms or variation templates of the key transition features, the node position calibration technology searches for the most accurate position with the highest matching degree within the neighborhood of the candidate point, thus determining the location of the device node where a fundamental state transition has occurred. This positioning process effectively eliminates potential timing deviations or noise interference from earlier steps, ensuring the accuracy of the state transition event on the timeline.

[0021] After obtaining the device node locations, further statistical analysis is performed on the fluctuation range of key transition characteristics before and after each node. By aggregating the feature values ​​of a large number of similar transition events, the typical numerical distribution of the key transition characteristic at the time of the transition is determined, thereby dividing the corresponding fluctuation value intervals and obtaining the fluctuation value interval division results. For example, for a transition such as "slight insulation degradation," its key transition characteristic "leakage current increment" may be defined within a specific numerical range. This fluctuation value interval division result assigns a quantified characteristic fluctuation representation to each state transition node, enabling not only the determination of what kind of transition occurred but also the assessment of the intensity or typicality of the transition.

[0022] Step 500: Obtain the corresponding equipment signal fluctuation sequence based on the equipment node location. If the equipment signal fluctuation sequence is in a continuously rising state and meets the equipment risk accumulation judgment rule, then mark it as a high-risk equipment transfer node. Based on the location of each device node, the associated device signal fluctuation sequence is extracted. This sequence directly or indirectly reflects key signal indicators in the device fault development process, such as the insulation loss value, the intensity of partial discharge, or the energy of a specific vibration spectrum. The extracted device signal fluctuation sequence is normalized and context-aligned, clearly characterizing the trend trajectory of state changes caused by node events.

[0023] Subsequently, the evolution pattern of the device's signal fluctuation sequence was analyzed. When it was identified that the device's signal fluctuation sequence showed a continuous upward trend, that is, the signal value maintained a monotonic or overall upward trend in subsequent observation windows, it indicated that the state transition triggered by this node might not be an isolated event, but rather the activation of some irreversible degradation mechanism within the device.

[0024] To conduct risk assessment, a pre-defined equipment risk accumulation judgment rule is invoked. This rule is a complex set of logic that includes checking whether the signal is continuously rising, comprehensively evaluating whether the rate of increase is accelerating, whether the upward trends of multiple signals are coordinated, and whether the current level is approaching a known historical fault threshold. This equipment risk accumulation judgment rule effectively distinguishes between benign, recoverable state fluctuations and malignant trends pointing towards fault evolution.

[0025] If the signal fluctuation sequence of a device corresponding to a node simultaneously meets the continuously rising pattern and passes the risk accumulation judgment rule, then that node will be marked as a high-risk transfer node. This means that this state transfer is highly likely to be the starting point or a critical deterioration point of a continuous failure chain. High-risk transfer nodes represent a higher level of real-time risk, requiring priority attention and potentially triggering higher-level warnings or intervention checks.

[0026] Step 600: Based on the high-risk transfer nodes of the equipment and combined with the fluctuation value interval division results, smooth the time-series information of the high-risk transfer nodes of the equipment to generate a warning signal sequence for future changes in the state of the equipment.

[0027] Based on the marked high-risk equipment transfer nodes and their corresponding fluctuation value intervals, the final early warning information generation stage begins. First, according to the transfer type of each high-risk transfer node and its corresponding fluctuation value interval, the risk level of that node is preliminarily quantified. The interval division results provide a benchmark for judging the intensity level of current signal fluctuations; for example, by positioning the node's characteristic fluctuation values ​​within the mild, moderate, or severe degradation range, a static risk intensity label is assigned to it.

[0028] Subsequently, time-series information was applied again to smooth the high-risk transfer nodes of these devices and their continuously monitored, associated signal fluctuation data. This process suppressed high-frequency noise and random interference, extracted the true and stable evolution trend in the risk signals, and merged discrete high-risk nodes with subsequent continuous trend changes into a coherent and smooth risk evolution trajectory. Through this processing, the risk indication, which might have appeared abrupt or discontinuous, was transformed into a curve that smoothly changes over time. This curve more clearly reveals the continuity and directionality of the fault accumulation process.

[0029] Based on the smoothed trajectory, a sequence of early warning signals for future changes in equipment status is generated by combining the initial risk intensity of nodes defined by the fluctuation range with the evolution trend of the trajectory. This early warning signal sequence is a signal stream that evolves over time and exhibits varying intensity. For example, the sequence may begin with a low-level warning corresponding to a moderate risk range and a stable trend; subsequently, as the smoothed curve shows the trend accelerating upwards and breaking through a higher fluctuation range threshold, the early warning signal automatically escalates to a higher level.

[0030] The intelligent fault diagnosis method for power equipment based on deep learning provided by this invention achieves real-time monitoring and risk warning of equipment status changes through the integration of multi-level data processing and dynamic analysis logic, thereby improving the safety of equipment operation and the accuracy of prediction. Its established dynamic transformation identification mechanism automatically identifies key nodes of state transition from continuous monitoring data, thereby improving the fault prediction capability.

[0031] In one embodiment, please refer to Figure 2 The method involves extracting multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment from the equipment operation data. Based on these multi-dimensional feature vectors, an initial data set is generated using a time-series information smoothing method combined with spatiotemporal neighborhood data extraction technology. This initial set includes: Step 101: Based on the equipment operation data of the power equipment, determine the original time series dataset containing state timing information; Step 102: Based on the original time-series dataset, extract multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment; Step 103: Based on the multidimensional feature vector, a time-series point information smoothing method is used to perform denoising operation to obtain smoothed time-series data; Step 104: If there are abnormal fluctuations in the smoothed time series data, then based on the smoothed time series data and combined with spatiotemporal neighborhood information, spatiotemporal neighborhood data extraction technology is used to perform local data correction and determine the corrected time series sequence. Step 105: Based on the corrected time series, the support vector machine algorithm is applied for classification processing to classify the device operating status and obtain the initial data set.

[0032] First, multi-source, continuous monitoring signals generated by power equipment during operation are collected and integrated to construct a complete original time-series dataset that preserves timestamps and state dependencies. Then, multi-dimensional feature vectors closely related to the progressive evolution of equipment health status are extracted from the original time-series dataset. These multi-dimensional feature vectors are used to characterize the potential trajectory of equipment performance degradation from different physical dimensions, thereby constructing a feature space that can represent the long-term evolution trend of the equipment.

[0033] To improve data quality, a time-series information smoothing method is introduced to denoise the multi-dimensional feature vectors. This effectively filters out random fluctuations caused by measurement noise and transient interference, while retaining trend information reflecting the actual changes in the equipment's state, resulting in a smoother and more stable time-series data trajectory. If abnormal fluctuations are still detected in the smoothed data, a correction mechanism is further activated. This mechanism combines the spatiotemporal neighborhood information of the current data point, considering data from temporally adjacent points and related monitoring points, and uses spatiotemporal neighborhood data extraction technology to locally correct abnormal fluctuations. Suspected anomalies are placed within their context for rationality evaluation and correction, ultimately generating a corrected time-series sequence.

[0034] The algorithm principle of the spatiotemporal neighborhood data extraction technology is as follows: For smoothed time-series data, any target data point is placed in a three-dimensional neighborhood composed of both a time axis and a spatial axis for analysis. This three-dimensional neighborhood includes data from the target data point over a past period and at a predetermined future time (time neighborhood), as well as data from other monitoring points associated with the target data point within the same time period (spatial neighborhood). From the three-dimensional data blocks of this neighborhood, spatiotemporal event features that reflect how the target data point is generated, propagates, and is affected are extracted. Finally, the spatiotemporal neighborhood information and these spatiotemporal event features are used to correct outliers in the smoothed time-series data, thereby obtaining the corrected time-series sequence.

[0035] Finally, the corrected time series is input into the Support Vector Machine (SVM) algorithm for classification. Leveraging its strength in handling high-dimensional, non-linear data, the SVM, based on historical learning experience, accurately categorizes the operating status of equipment, distinguishing between different state modes such as normal, sub-healthy, and early abnormal, thereby generating an initial dataset with clear state labels and excellent data quality.

[0036] For example, when obtaining a time-series sequence of equipment status from continuously monitored equipment operation data, sensors can collect equipment operation parameters every minute, such as vibration frequency, temperature, and current value. Suppose a device collects 1440 data points over 24 hours, with an average vibration frequency of 5.2 Hz, an average temperature of 45.3 degrees Celsius, and an average current of 10.5 amperes. Based on this data, a time-series dataset containing timestamps and multiple parameters is formed. Next, multi-dimensional feature vectors are extracted for the progressive evolution attributes of the equipment. Principal component analysis (PCA) can be used to reduce the dimensionality of the vibration frequency, temperature, and current data. The variance contribution rate of the first principal component is calculated to be 85%. The first two principal components are extracted as feature vectors, resulting in two-dimensional feature values ​​for each time point. For example, the feature vector for a certain time point might be (3.7, 1.2). Subsequently, a time-series information smoothing method was used, employing a sliding window averaging approach to smooth the feature vector. The window size was set to 5 minutes, meaning the average of five consecutive data points was taken. The smoothed feature vector at a given time point was calculated as (3.5, 1.1), thus reducing random fluctuations in the data. Next, combining spatiotemporal neighborhood data extraction technology, a spatiotemporal neighborhood matrix was constructed, considering data from three spatially adjacent devices and data from 10 minutes before and after them. Assuming the mean deviation of vibration frequencies within the neighborhood is 0.3 Hz, a weighted averaging algorithm (with weights equal to the reciprocal of distance) was used to fuse the data, generating an initial dataset after noise filtering. For example, the filtered value at a given time point was (3.4, 1.0).

[0037] In this embodiment, intelligent data cleaning that integrates spatiotemporal context understanding improves the reliability and physical consistency of the input data. Furthermore, by placing support vector machine classification on high-quality data, the boundaries of state division become clearer and more accurate, effectively overcoming the interference of noisy data on the classifier's performance.

[0038] In one embodiment, please refer to Figure 3 The step of analyzing the change trend within the local data range using a feature change rate calculation method based on the initial data set, determining the feature change rate, and determining potential state transition candidate points based on the feature change rate includes: Step 201: Based on the initial data set, the characteristic change rate calculation method is used to analyze the change trend within the local data range to obtain the characteristic change rate of the power equipment in different time periods. Step 202: Based on the characteristic change rate, compare the data of each time period with a preset rate threshold and a comparison standard to locate abnormal points that exceed the preset rate threshold. Step 203: If the abnormal point exists, mark the abnormal point as a candidate point for potential state transition, and obtain the transition time point and time context features of the candidate point based on the feature change rate. Step 204: Based on the transfer time point and the time context features, a support vector machine algorithm is used for classification to obtain the classification result; Step 205: Based on the classification results, select the points that are confirmed as state transitions to obtain the potential state transition candidate points.

[0039] This step is used to identify key nodes where the operating state of power equipment undergoes fundamental changes from dynamic evolution trends. First, based on the initial dataset, a feature change rate calculation method is used to analyze the evolution trend of each feature dimension within a local data window to calculate the instantaneous or average change rate of equipment features over a short time scale. This transforms the static state feature description into a dynamic quantitative indicator of change trend, obtaining the feature change rate of power equipment in different time periods.

[0040] After obtaining a continuous sequence of characteristic change rates, a preset rate threshold and multi-dimensional comparison criteria are introduced. The change rate calculated for each time period is compared with these thresholds to quickly scan and locate anomalies where the rate exceeds the normal fluctuation range. These anomalies indicate that the device state may be undergoing drastic or accelerated changes. Subsequently, these initially screened anomalies are marked as candidate points for potential state transitions, and the time of their occurrence and their corresponding temporal context features are recorded, preserving complete contextual information for subsequent in-depth verification.

[0041] To effectively distinguish between genuine, meaningful state transitions and transient disturbances or noise, a Support Vector Machine (SVM) algorithm is employed for secondary classification of candidate points. This SVM algorithm, based on patterns learned from extensive historical data, comprehensively analyzes the transition time and data characteristics of each candidate point to determine whether it constitutes a genuine and credible state transition event. The SVM algorithm leverages the powerful non-linear classification capabilities of machine learning models to identify complex pattern boundaries, thereby significantly improving the accuracy of the classification. Finally, based on the classification results of the SVM, points confirmed as genuine state transitions are selected, forming a set of potential state transition candidate points.

[0042] For example, based on an initial dataset that has undergone noise filtering, the system first analyzes the changing trend of the equipment status using a feature change rate calculation method. Specifically, the change rate is quantified by calculating the Euclidean distance difference between feature vectors at adjacent time points. For instance, if the feature vector at a certain time point is (3.4, 1.0) and the next time point is (3.6, 1.1), the calculated change rate is 0.22. Subsequently, the change rate over a continuous 30-minute period is cumulatively averaged to obtain a local average change rate of 0.18. Simultaneously, the system defines the local data range as a 15-minute data window. Within this window, the system calculates the fluctuations in the feature change rate, assuming a standard deviation of 0.05, indicating a relatively stable trend. Next, the calculated average rate of change is compared with a preset threshold. Assuming the threshold standard is 0.2, if the average rate of change exceeds the threshold, the time point is automatically marked as a potential state transition candidate point. For example, if the average rate of change within a certain window is 0.21, which exceeds the threshold of 0.2, the point is determined to be a candidate point, and its timestamp and feature value (3.6, 1.1) are recorded for subsequent analysis.

[0043] In this embodiment, by using the support vector machine model to identify outliers in the initial screening, the advantages of machine learning in complex pattern recognition are not only brought into play, but also the problems of high computational cost and sensitivity to noise when directly processing the original sequence are avoided, thus achieving a balance between efficiency and accuracy.

[0044] In one embodiment, please refer to Figure 4 The step involves selecting data corresponding to the surrounding spatiotemporal neighborhood based on the potential state transition candidate points to form a local data subset, and generating corresponding device transition type labels based on the local data subset using a device historical path pattern matching method, including: Step 301: Based on the potential state transition candidate points, select data corresponding to the surrounding spatiotemporal neighborhood from the spatiotemporal neighborhood range to form an initial device data set; Step 302: Based on the initial device data set, group the data according to the time and space dimensions to obtain the local data subset; Step 303: Based on the local data subset, noise filtering technology is used to clean the data, remove outliers and irrelevant interference items, and generate an optimized device data subset; Step 304: Based on the optimized subset of device data and combined with the corresponding historical path information, extract trajectory segments related to the potential state transition candidate points to obtain a path feature set; Step 305: Based on the path feature set, a pre-established device historical path pattern matching method is used for comparison and analysis to obtain the path pattern matching result; Step 306: Based on the path pattern matching results, analyze the state transition patterns of candidate points and generate preliminary type classification identifiers; Step 307: Based on the preliminary type classification identifier and combined with the corresponding context information, the preliminary type classification identifier is further refined to obtain the device conversion type label.

[0045] First, based on the identified potential state transition candidate points, surrounding data is extracted from their associated spatiotemporal neighborhoods to form an initial device dataset. This initial dataset is then structured and grouped according to time and space dimensions, forming a series of local data subsets. Next, noise filtering techniques are introduced to clean these local data subsets, effectively removing outliers and irrelevant noise introduced by measurement errors or transient interference, generating an optimized device data subset that significantly improves data quality and signal-to-noise ratio. Based on this optimized subset, and combined with historical path information of the devices, state evolution trajectory fragments directly related to the current candidate point are extracted, thus obtaining a set of path features characterizing the dynamic process of this potential transition event.

[0046] After acquiring the path feature set, a pre-established device historical path pattern matching method is invoked to intelligently compare and analyze the current event's path feature set with a rich library of known faults or state transition patterns, resulting in a quantified path pattern matching result. Based on this result, the state transition patterns of candidate points are analyzed, and preliminary type classification labels are generated, such as initial classification as "overheating tendency" or "insulation degradation." To achieve more refined diagnosis, the specific contextual information of the event, such as load conditions and environmental parameters, is further integrated to refine and calibrate the preliminary classification labels. For example, preliminary insulation degradation is refined into surface creepage initiation caused by moisture or internal defect development due to aging. Finally, a high-precision device transition type label is output.

[0047] Furthermore, corresponding contextual information is introduced to refine and contextualize the initial identification. This contextual information is a multi-dimensional dataset that can include the equipment's real-time operating conditions, external environmental conditions, related subsystem states, and the spatiotemporal evolution characteristics of the abnormal event itself. By fusing and analyzing the initial classification identification with the contextual information, it is possible to trace the root cause of the fault mode and determine its context, ultimately generating a device conversion type label. The equipment's real-time operating conditions can include the load rate, voltage level, and operating mode (start-up, steady state, shutdown process) of the equipment at the time of the anomaly; the external environmental conditions can include ambient temperature, humidity, and pollution level; the related subsystem states are data from other subsystems physically or functionally associated with the anomaly, such as the initial identification of bearing vibration anomalies; the spatiotemporal evolution attributes of the abnormal event itself refer to the development characteristics of the anomaly in time and equipment spatial structure, such as whether it is a rapid abrupt change or a slow gradual change, whether it is limited to a single point or has spread spatially.

[0048] For example, in processing potential state transition candidate points, firstly, spatiotemporal neighborhood data extraction technology is used. Assuming the candidate point's location is latitude and longitude (116.39, 39.91) and the timestamp is October 1, 2023, at 10:00:00, the system automatically extracts surrounding device movement data with a radius of 500 meters and a time window of 30 minutes before and after this point. This forms a local data subset containing 100 trajectory records. These records include speed (average 5.2 m / s), direction (30 degrees north), and time interval (average 10 seconds). Subsequently, data subset noise filtering technology is applied. The Kalman filter algorithm is used to smooth the speed and direction data, removing outliers (records greater than 15 m / s, accounting for approximately 5%), resulting in 80 filtered valid records, ensuring data quality. Next, combining the device's historical path pattern matching method, the filtered subset was compared with 1000 path patterns stored in the historical trajectory database. A dynamic time warping algorithm was used to calculate the similarity, with a threshold of 0.8. The five path patterns with the highest similarity were selected, revealing that three of these patterns indicated a device transitioning from a moving state to a stationary state, with an average dwell time of 20 minutes. Finally, based on the matching results and dwell time analysis, the device transition type was automatically labeled as "temporary dwelling," and a confidence level of 0.85 was recorded.

[0049] In this embodiment, by placing noise filtering and trajectory segment extraction before pattern matching, the data quality of the input matching model is ensured, thereby greatly improving the reliability and accuracy of historical pattern comparison results. Furthermore, by generating labels in two stages—preliminary classification and contextual refinement—the clarity of the classification framework is guaranteed, and the diagnostic conclusions are refined and specified.

[0050] In one embodiment, please refer to Figure 5 The step of continuously scanning using node position calibration technology based on the key transfer characteristics of the equipment to determine the device node positions and the corresponding fluctuation value interval division results includes: Step 401: Based on the key transfer features of the equipment, a pre-established classification model is used to determine the preliminary distribution status of the equipment features; Step 402: Based on the preliminary distribution state, perform continuous scanning using node position calibration technology to locate the initial coordinate points of the device nodes; Step 403: Based on the initial coordinate point, obtain the signal fluctuation data of the corresponding device node, and based on the signal fluctuation data, use a quantization method to calculate the range of fluctuation value changes to obtain the signal fluctuation interval; Step 404: Based on the signal fluctuation range, extract key change points of fluctuation value changes. If the fluctuation value of the key change point exceeds a preset fluctuation threshold, calibrate the initial coordinate point to determine the location of the device node. Step 405: Based on the device node location and the signal fluctuation range, generate the fluctuation value range distribution of the device node, and determine the fluctuation value range division result based on the fluctuation value range distribution.

[0051] First, based on the extracted key transfer features of the equipment, a pre-trained classification model is used to analyze the aggregation patterns of these features, thereby determining the initial distribution of the equipment features in the state space. Then, using node position calibration technology, continuous scanning and probing are performed across the entire time-series data according to the initial distribution, precisely locating the initial coordinates of the equipment nodes representing significant state changes. After obtaining the initial coordinates, the raw signal fluctuation data of the corresponding equipment nodes is extracted based on these coordinates. A quantitative calculation method is then used to statistically analyze the amplitude range of the signal fluctuations within the node's neighborhood, thereby obtaining a quantitative signal fluctuation interval to initially characterize the intensity profile of the event at that node.

[0052] The algorithm principle of the node position calibration technology is as follows: For the initial distribution state, within the spatiotemporal range guided by the initial distribution state, sliding analysis is performed with a high temporal resolution (higher than the original data sampling interval). A small data segment (window) is continuously scanned and analyzed, and a preset calibration rule is applied for evaluation. It should be noted that this scanning process is dynamic and adaptive; its window size and scan step size can be dynamically adjusted according to the data characteristics of the initial distribution state (such as signal-to-noise ratio and rate of change). The calibration rule can be a multi-factor fusion decision function, such as whether the curve shape of the current data window (such as the steepness of the rising edge and a specific oscillation pattern) matches the preset typical waveform of the state transition initiation. Through the above scanning and evaluation, the precise initial coordinates of the device node are finally output.

[0053] To ensure the accuracy of node positioning, the dynamic details within the fluctuation range are further analyzed. Key change points in the fluctuation value process are extracted, and it is determined whether the fluctuation values ​​of these key change points exceed preset fluctuation thresholds set based on historical experience or physical models. If the fluctuation values ​​of key change points exceed the preset thresholds, indicating significant abnormal fluctuations, the initial coordinate points are fine-tuned and calibrated to ultimately determine more accurate device node positions that better reflect the physical moment of signal abrupt changes. Finally, all calibrated device node positions and their corresponding signal fluctuation ranges are integrated to generate a holistic distribution of fluctuation value ranges for the device nodes during the state evolution process. This fluctuation value range distribution not only reflects the intensity of individual events but also characterizes the statistical regularity and hierarchical relationship of fluctuation amplitudes for different types of nodes. Therefore, the final fluctuation value range division result can be determined based on this fluctuation value range distribution.

[0054] In this embodiment, the calibration of node locations and the division of their fluctuation intensity ranges are integrated, enabling location information and quantitative intensity information to corroborate each other and be generated collaboratively. Finally, the generated fluctuation value range distribution results provide a standardized quantitative benchmark for equipment condition assessment, allowing subsequent advanced analyses such as risk grading and trend prediction to be built on a unified metric foundation, thus improving the accuracy of intelligent diagnosis.

[0055] In one embodiment, please refer to Figure 6 If the device signal fluctuation sequence is in a continuously rising state and meets the device risk accumulation judgment rules, it is marked as a high-risk transfer node, including: Step 501: Based on the device signal fluctuation sequence, perform weighted average fusion processing on the signal fluctuation values ​​of adjacent time points through a sliding window to obtain a smoothed signal fluctuation sequence; Step 502: Based on the smoothed signal fluctuation sequence, calculate the difference between adjacent time points to obtain the signal fluctuation difference sequence; Step 503: Based on the signal fluctuation difference sequence, a linear regression algorithm is used to fit the slope of the signal fluctuation difference sequence; Step 504: If the slope is positive and the period of time during which the positive value is maintained exceeds a preset length, then it is determined that the device signal fluctuation sequence is in a continuously rising state. Step 505: If the device signal fluctuation sequence shows a continuous upward trend, then based on the signal fluctuation difference sequence, determine the upward trend segment and calculate the ratio of the cumulative increase of the signal fluctuation difference within the upward trend segment to its duration to obtain the risk accumulation rate. Step 506: If the risk accumulation rate exceeds the preset risk accumulation judgment threshold, then mark the corresponding device node as a high-risk device transfer node, wherein the device risk accumulation judgment rule is that the corresponding risk accumulation rate exceeds the risk accumulation judgment threshold.

[0056] First, by applying a sliding window weighted average fusion process to the device signal fluctuation sequence, random noise and short-term interference are effectively suppressed, generating a smooth and coherent signal fluctuation sequence that highlights the true risk evolution trend. Then, the difference between adjacent time points in this smooth sequence is calculated to obtain the signal fluctuation difference sequence. This signal fluctuation difference sequence directly reflects the amount of change in fluctuation per unit time and is the basis for judging the trend direction and intensity. After obtaining the signal fluctuation difference sequence, a linear regression algorithm is used to perform a holistic quantitative analysis of its changing trend. The slope obtained by fitting the data objectively characterizes the average change tendency of the fluctuation sequence over a period of time. Only when the slope is positive, and this positive state lasts for more than a preset duration, is the device signal fluctuation sequence confirmed to be in a continuous upward trend. This dual condition effectively filters out short-term technical rebounds or fluctuations, ensuring the continuity and stability of the identified upward trend.

[0057] To further assess the urgency of risk accumulation, based on the confirmed continuous upward trend, the upward trend segment was precisely located, and the risk accumulation rate was obtained by calculating the ratio of the cumulative increase in signal fluctuation difference within this segment to its duration. This risk accumulation rate not only reflects whether things are deteriorating but also precisely quantifies how fast the deterioration is occurring. Finally, a dynamic risk accumulation judgment rule was introduced: the calculated risk accumulation rate was compared with a preset risk accumulation judgment threshold; only when the risk accumulation rate exceeds this threshold was the corresponding equipment node marked as a high-risk transfer node.

[0058] This embodiment provides a precise and quantifiable basis for early warning classification and predictive maintenance decisions by generating the risk accumulation rate indicator, thereby improving the intelligence level and economic benefits of equipment health management.

[0059] In one embodiment, please refer to Figure 7 The step of smoothing the time-series information of the high-risk transfer nodes of the equipment, based on the fluctuation value interval division results, and generating a warning signal sequence for future changes in the equipment state, includes: Step 601: Based on the high-risk transfer nodes of the equipment, obtain the node tag data related to the high-risk transfer nodes and the key identification content in the node tag data, and perform preliminary screening based on the key identification content using preset classification rules to obtain the priority ranking result of the high-risk transfer nodes of the equipment. Step 602: Based on the priority ranking result and the fluctuation value interval division result, perform data standardization processing on the fluctuation range in the division result to determine the stability evaluation value of the fluctuation value interval. Step 603: Based on the stability evaluation value, obtain the time point information of the conversion start identifier; based on the time point information, use the time window division method to determine the effective conversion identifier point. Step 604: Based on the effective transition markers, obtain the time sequence information of the effective transition markers, and based on the time sequence information, perform noise reduction processing using a smoothing method to obtain the processed time sequence data sequence. Step 605: Based on the processed time series data sequence and combined with the prediction requirements of future state changes, determine the trend characteristics of state change prediction. Step 606: Based on the trend characteristics, a support vector machine algorithm is used for classification processing to determine the potential risk level of the state change, and the early warning signal sequence data is generated based on the potential risk level.

[0060] First, all marked high-risk equipment transfer nodes are aggregated, and their node tagging data and key identifiers are extracted. Based on the key identifiers, a preliminary screening and priority ranking of the high-risk equipment transfer nodes is performed using preset classification rules, such as risk accumulation rate, fluctuation range, and node type, resulting in a priority ranking of the high-risk equipment transfer nodes.

[0061] After prioritizing, the data within each fluctuation range is standardized based on the previously defined ranges, and a stability assessment value is calculated for each range. This stability assessment value quantifies the concentration and reliability of the data distribution within the range, thus providing a basis for judging whether the volatility characteristics of the risk node are stable and reliable. Subsequently, based on this stability assessment value, the starting time of the transition event is located, and a time window segmentation method is used to define effective transition markers on the timeline that can effectively represent this state transition process.

[0062] Subsequently, for effective transformation markers, detailed time-series information is extracted, and smoothing and noise reduction processing is applied again to eliminate high-frequency noise, resulting in a processed time-series data sequence. Based on the processed time-series data sequence and considering the need to predict future state changes, key trend features that can indicate future evolution directions are extracted, such as trend persistence, acceleration, or morphological inflection points. Finally, a support vector machine algorithm is used to classify these trend features, not only determining whether a risk exists but also accurately identifying its potential future evolution patterns, and determining the dynamic potential risk level accordingly, such as high risk accelerating deterioration and medium risk stable evolution. Ultimately, based on this potential risk level, corresponding early warning signal sequence data with different urgency levels are automatically generated.

[0063] In this embodiment, trend feature extraction is combined with risk evolution pattern classification based on machine learning, so that the early warning signal is no longer based solely on whether the current state exceeds the limit, but on intelligent prediction of how the risk may evolve in the future, thus realizing trend early warning.

[0064] The following describes the intelligent fault diagnosis system for power equipment based on deep learning provided by the present invention. The intelligent fault diagnosis system for power equipment based on deep learning described below can be referred to in correspondence with the intelligent fault diagnosis method for power equipment based on deep learning described above.

[0065] This invention also provides a deep learning-based intelligent fault diagnosis system for power equipment, comprising: The smoothing module is used to extract multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment based on the equipment operation data of the power equipment, and generate an initial data set based on the multi-dimensional feature vectors by using a time-series point information smoothing processing method combined with spatiotemporal neighborhood data extraction technology. The rate of change detection module is used to analyze the change trend within the local data range based on the initial data set using a characteristic rate of change calculation method, determine the characteristic rate of change, and determine potential state transition candidate points based on the characteristic rate of change. The conversion type label generation module is used to select data corresponding to the surrounding spatiotemporal neighborhood based on the potential state transition candidate points, form a local data subset, and generate corresponding device conversion type labels based on the local data subset using the device historical path pattern matching method. The node location determination module is used to extract key transfer features of the device based on the device conversion type label, and to perform continuous scanning using node location calibration technology based on the key transfer features of the device to determine the device node location and the corresponding fluctuation value interval division result. The high-risk node identification module is used to obtain the corresponding equipment signal fluctuation sequence based on the location of the equipment node. If the equipment signal fluctuation sequence is in a continuously rising state and meets the equipment risk accumulation judgment rule, it is marked as a high-risk transfer node of the equipment. The early warning signal generation module is used to smooth the time-series information of the high-risk transfer nodes of the equipment based on the high-risk transfer nodes of the equipment and the results of the fluctuation value interval division, and generate an early warning signal sequence for future changes in the state of the equipment.

Claims

1. A method for intelligent fault diagnosis of power equipment based on deep learning, characterized in that, Executed by a computer, including: Based on the equipment operation data of power equipment, multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment are extracted. Based on the multi-dimensional feature vectors, an initial data set is generated by using a time-series point information smoothing method combined with spatiotemporal neighborhood data extraction technology. Based on the initial data set, the characteristic change rate calculation method is used to analyze the change trend within the local data range, determine the characteristic change rate, and determine potential state transition candidate points based on the characteristic change rate. Based on the potential state transition candidate points, data corresponding to the surrounding spatiotemporal neighborhood are selected to form a local data subset. Based on the local data subset, the device historical path pattern matching method is used to generate the corresponding device conversion type label. Based on the device conversion type label, key transfer features of the device are extracted. Based on the key transfer features of the device, node position calibration technology is used to perform continuous scanning to determine the device node position and the corresponding fluctuation value interval division result. Based on the location of the device node, the corresponding device signal fluctuation sequence is obtained. If the device signal fluctuation sequence is in a continuous upward state and meets the device risk accumulation judgment rule, it is marked as a high-risk transfer node of the device. Based on the high-risk transfer nodes of the equipment and combined with the results of the fluctuation value interval division, the time-series information of the high-risk transfer nodes of the equipment is smoothed to generate a warning signal sequence for future changes in the state of the equipment.

2. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, The equipment operation data based on the power equipment is used to extract multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment. Based on these multi-dimensional feature vectors, an initial data set is generated using a time-series information smoothing method combined with spatiotemporal neighborhood data extraction technology. This set includes: Based on the equipment operation data of power equipment, determine the original time series dataset containing state timing information; Based on the original time-series dataset, extract multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment; Based on the multidimensional feature vector, a time-series point information smoothing method is used to perform denoising operations to obtain smoothed time-series data. If there are abnormal fluctuations in the smoothed time series data, then based on the smoothed time series data and combined with spatiotemporal neighborhood information, spatiotemporal neighborhood data extraction technology is used to perform local data correction and determine the corrected time series sequence. Based on the corrected time series, the support vector machine algorithm is applied for classification processing to classify the device operating status and obtain the initial data set.

3. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, The step of analyzing the change trend within a local data range using a characteristic change rate calculation method based on the initial data set, determining the characteristic change rate, and determining potential state transition candidate points based on the characteristic change rate includes: Based on the initial data set, the characteristic change rate calculation method is used to analyze the change trend within the local data range to obtain the characteristic change rate of the power equipment in different time periods. Based on the characteristic change rate, the data of each time period is compared with a preset rate threshold and a comparison standard to locate abnormal points that exceed the preset rate threshold. If the abnormal point exists, the abnormal point is marked as a candidate point for potential state transition, and the transition time point and time context features of the candidate point are obtained based on the feature change rate. Based on the transfer time point and the time context features, a support vector machine algorithm is used for classification to obtain the classification result. Based on the classification results, points that are confirmed as state transitions are selected, and potential state transition candidate points are obtained.

4. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, Based on the potential state transition candidate points, data corresponding to the surrounding spatiotemporal neighborhood is selected to form a local data subset. Based on this local data subset, a device historical path pattern matching method is used to generate corresponding device transition type labels, including: Based on the potential state transition candidate points, data corresponding to the surrounding spatiotemporal neighborhoods are selected from the spatiotemporal neighborhood range to form an initial device data set; Based on the initial device data set, the local data subset is obtained by grouping the data according to the time and space dimensions; Based on the aforementioned local data subset, noise filtering technology is used to clean the data, remove outliers and irrelevant interference items, and generate an optimized device data subset. Based on the optimized subset of device data and combined with the corresponding historical path information, trajectory segments related to the potential state transition candidate points are extracted to obtain a path feature set. Based on the aforementioned path feature set, a pre-established device historical path pattern matching method is used for comparison and analysis to obtain path pattern matching results. Based on the path pattern matching results, the state transition patterns of candidate points are analyzed, and preliminary type classification identifiers are generated. Based on the preliminary type classification identifier, and combined with the corresponding context information, the preliminary type classification identifier is further refined to obtain the device conversion type label.

5. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, The step of continuously scanning using node position calibration technology based on the key transfer characteristics of the equipment to determine the device node positions and the corresponding fluctuation value interval division results includes: Based on the key transfer characteristics of the equipment, a pre-established classification model is used to determine the preliminary distribution status of the equipment characteristics; Based on the initial distribution state, continuous scanning is performed using node position calibration technology to locate the initial coordinate points of the device nodes; Based on the initial coordinate point, obtain the signal fluctuation data of the corresponding device node, and based on the signal fluctuation data, use a quantization method to calculate the range of fluctuation value changes to obtain the signal fluctuation interval; Based on the signal fluctuation range, key change points of fluctuation value change are extracted. If the fluctuation value of the key change point exceeds the preset fluctuation threshold, the initial coordinate point is calibrated to determine the position of the device node. Based on the device node location and the signal fluctuation range, a fluctuation value range distribution for the device node is generated, and the fluctuation value range division result is determined based on the fluctuation value range distribution.

6. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, If the device signal fluctuation sequence is in a continuously rising state and meets the device risk accumulation judgment rules, it is marked as a high-risk transfer node, including: Based on the device signal fluctuation sequence, a weighted average fusion process is performed on the signal fluctuation values ​​of adjacent time points through a sliding window to obtain a smoothed signal fluctuation sequence. Based on the smoothed signal fluctuation sequence, the difference between adjacent time points is calculated to obtain the signal fluctuation difference sequence. Based on the signal fluctuation difference sequence, a linear regression algorithm is used to fit the slope of the signal fluctuation difference sequence. If the slope is positive and remains positive for a period of time exceeding a preset length, then the device signal fluctuation sequence is determined to be in a continuously rising state. If the device signal fluctuation sequence shows a continuous upward trend, then based on the signal fluctuation difference sequence, the upward trend segment is determined, and the ratio of the cumulative increase of the signal fluctuation difference within the upward trend segment to its duration is calculated to obtain the risk accumulation rate. If the risk accumulation rate exceeds a preset risk accumulation judgment threshold, the corresponding device node is marked as a high-risk transfer node, wherein the device risk accumulation judgment rule is that the corresponding risk accumulation rate exceeds the risk accumulation judgment threshold.

7. The intelligent fault diagnosis method for power equipment based on deep learning according to claim 1, characterized in that, Based on the high-risk transfer nodes of the equipment and combined with the fluctuation value interval division results, the time-series information of the high-risk transfer nodes of the equipment is smoothed to generate a warning signal sequence for future changes in the equipment state, including: Based on the high-risk transfer nodes of the equipment, obtain the node tag data related to the high-risk transfer nodes and the key identification content in the node tag data, and perform preliminary screening based on the key identification content using preset classification rules to obtain the priority ranking result of the high-risk transfer nodes of the equipment. Based on the priority ranking result and the fluctuation value interval division result, the fluctuation range in the division result is subjected to data standardization processing to determine the stability evaluation value of the fluctuation value interval. Based on the stability evaluation value, the time point information of the conversion start identifier is obtained. Based on the time point information, the effective conversion identifier point is determined by using a time window division method. Based on the effective transition markers, the time-series point information of the effective transition markers is obtained, and based on the time-series point information, a smoothing method is used for noise reduction to obtain the processed time-series data sequence. Based on the processed time series data sequence, and combined with the prediction requirements for future state changes, the trend characteristics of state change prediction are determined. Based on the trend characteristics, a support vector machine algorithm is used for classification to determine the potential risk level of state changes, and the early warning signal sequence data is generated based on the potential risk level.

8. A deep learning-based intelligent fault diagnosis system for power equipment, characterized in that, include: The smoothing module is used to extract multi-dimensional feature vectors related to the progressive evolution attributes of the power equipment based on the equipment operation data of the power equipment, and generate an initial data set based on the multi-dimensional feature vectors by using a time-series point information smoothing processing method combined with spatiotemporal neighborhood data extraction technology. The rate of change detection module is used to analyze the change trend within the local data range based on the initial data set using a characteristic rate of change calculation method, determine the characteristic rate of change, and determine potential state transition candidate points based on the characteristic rate of change. The conversion type label generation module is used to select data corresponding to the surrounding spatiotemporal neighborhood based on the potential state transition candidate points, form a local data subset, and generate corresponding device conversion type labels based on the local data subset using the device historical path pattern matching method. The node location determination module is used to extract key transfer features of the device based on the device conversion type label, and to perform continuous scanning using node location calibration technology based on the key transfer features of the device to determine the device node location and the corresponding fluctuation value interval division result. The high-risk node identification module is used to obtain the corresponding equipment signal fluctuation sequence based on the location of the equipment node. If the equipment signal fluctuation sequence is in a continuously rising state and meets the equipment risk accumulation judgment rule, it is marked as a high-risk transfer node of the equipment. The early warning signal generation module is used to smooth the time-series information of the high-risk transfer nodes of the equipment based on the high-risk transfer nodes of the equipment and the results of the fluctuation value interval division, and generate an early warning signal sequence for future changes in the state of the equipment.

Images