Health management system for key equipment of hydropower station based on predictive maintenance
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUODIAN DADU RIVER DAGANGSHAN HYDROPOWER DEV
- Filing Date
- 2026-06-18
- Publication Date
- 2026-07-24
Smart Images
Figure CN122451297A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hydropower station equipment health management technology, specifically a health management system for key hydropower station equipment based on predictive maintenance. Background Technology
[0002] Hydropower station critical equipment generates a large amount of time-series status data during operation. Existing equipment health management technologies typically extract independent features from this data and directly input them into classifiers or threshold comparison models to determine whether the equipment has entered an abnormal state. This approach treats features of different dimensions as isolated criteria for judgment, ignoring the objective information overlap and hierarchical relationships between features. When early signs of equipment degradation appear, the changes in a single feature are often weak and easily masked by noise. Conventional methods struggle to establish an effective information aggregation structure at the feature level, resulting in insufficient early warning capabilities.
[0003] Another shortcoming of existing technical solutions lies in the modeling method of equipment degradation processes. Conventional approaches infer equipment state evolution by fitting statistical trends or setting fixed degradation paths. These linear or quasi-linear assumptions cannot reflect the irregular transitions between failure modes during degradation. Key equipment in hydropower stations is affected by multiple factors, including hydraulic fluctuations and load changes, resulting in degradation paths exhibiting significant non-monotonicity and random jump characteristics. Existing fixed-path models lose probabilistic information during state transitions, leading to discrepancies between early warning level classifications and actual degradation risks.
[0004] To address the issues of the inability to dynamically organize feature relationships and the inability to quantify random migration of degradation paths in existing technologies, a solution is needed that can construct a fusion structure at the feature level that reflects the degree of information overlap and implement state probabilistic walk modeling on this structure, so as to improve the accuracy of early degradation identification and early warning level classification of key equipment in hydropower stations. Summary of the Invention
[0005] This invention provides a health management system for key equipment in hydropower stations based on predictive maintenance. The purpose is to transform the information overlap between multidimensional features into hierarchical node relationships by constructing a feature fusion tree, and to set a degradation mode walking mechanism on the feature fusion tree. The non-monotonic migration process of the equipment degradation path is characterized by the state transition probability sequence of the walking body between each node, thereby realizing the accurate determination of the equipment early warning level and the dynamic configuration of maintenance strategies.
[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention proposes a health management system for key equipment in hydropower stations based on predictive maintenance. By combining time window segmentation of multi-source status data, feature fusion tree construction driven by information overlap, directional walk of degradation patterns, and identification of probability jump points, the present invention achieves quantitative characterization and graded early warning of equipment degradation trends, and adaptively generates a maintenance strategy configuration matrix based on the early warning level.
[0007] In terms of status data acquisition, the system binds an independent data acquisition queue to each monitoring point in the key equipment of the hydropower station. Data is sequentially filled into a sliding window array according to the acquisition timestamp. Once the array is full, it is exported as a multi-source status data block and the array is reset to receive data from the next time window. Preferably, the window length of the sliding window array is adaptively adjusted based on the historical failure cycle data of the key equipment of the hydropower station, so that each data block can cover a time scale that matches the equipment degradation rhythm, thereby ensuring the temporal consistency of subsequent feature extraction and pattern walk analysis.
[0008] In terms of feature fusion, the system extracts time-domain features, frequency-domain features, and energy features from each multi-source state data block, respectively. It calculates the mutual information values between each pair of these features as the first, second, and third mutual information values, and compares each mutual information value with a preset fusion threshold. For two features corresponding to mutual information values greater than the fusion threshold, they are merged into a single fused feature node; for two features corresponding to mutual information values not greater than the fusion threshold, they are retained as independent feature nodes. By building a feature fusion tree from bottom to top based on the merging relationship, redundant representations are effectively eliminated while retaining key feature dimensions sensitive to changes in device state, thus reducing noise interference in degradation pattern mining.
[0009] In terms of degradation pattern tracking, the system maps each feature node in the feature fusion tree to a vertex in a graph structure, sets directed edges between vertices according to the parent-child relationship between nodes, and uses the difference in feature values between two vertices as the migration cost of the walk. The initial position of the walk is set at the vertex corresponding to the root node of the feature fusion tree. In each time window, the feature values of each vertex are updated according to the latest multi-source state data block, the probability of the walk migrating from the current vertex to an adjacent vertex along the directed edge is calculated, and the vertex number where the walk is located at the end of each time window is recorded, thus forming a state transition probability sequence that changes with the window number. As a technical solution of the present invention, the migration cost of the walk is obtained by multiplying the difference in feature values between two feature nodes by the edge weight between the nodes, thereby comprehensively reflecting the structural constraints and numerical fluctuations of the feature space. When calculating the migration probability, preferably, the feature values of each feature node are recalculated for the multi-source state data block corresponding to each time window. The feature value of the vertex where the wandering body is currently located is compared with the feature values of its neighboring vertices. The ratio of the absolute value of the difference between the feature values of the neighboring vertices and the feature value of the current vertex to the migration cost is used as the initial probability. Then, the initial probabilities of all neighboring vertices are normalized to obtain the final probability of the wandering body migrating along each directed edge. This process enables the positional changes of the wandering body to sensitively capture the value fluctuations of different feature nodes in the feature fusion tree, thereby transforming the degradation trajectory of the device in the multi-dimensional feature space into a measurable state transition probability sequence.
[0010] In terms of early warning and maintenance strategy generation, the system marks the window positions where the vertex number jump amplitude between adjacent windows in the state transition probability sequence exceeds the jump threshold as probability jump points, and counts the frequency of probability jump points occurring per unit time. Preferably, the jump threshold is dynamically calculated based on the mean square error of the state transition probability sequence to adapt to the normal fluctuation range under different equipment and different operating conditions. The frequency of probability jump points per unit time is compared with a preset first frequency threshold and a second frequency threshold, where the first frequency threshold is less than the second frequency threshold. When the frequency is less than the first frequency threshold, the early warning level is determined to be no early warning level; when the frequency is between the two, it is determined to be a first-level early warning level; when the frequency is greater than the second frequency threshold, it is determined to be a second-level early warning level. By correlating the discrete jump behavior in the state transition sequence with the intensity of abnormal equipment degradation, adaptive early warning can be achieved without relying on a fixed baseline and manually set thresholds.
[0011] When generating the maintenance strategy configuration matrix, the system sets multiple maintenance operation items for key equipment in the hydropower station, and sets the execution status of each maintenance operation item to either "execute" or "not execute". An initial configuration matrix is constructed using the equipment identifier as the matrix row index and the operation code of the maintenance operation item as the matrix column index. When the warning level is no warning level, all elements in the initial configuration matrix are set to "not execute"; when the warning level is Level 1, the corresponding column index elements are set to "execute" according to the preset light maintenance list; when the warning level is Level 2, the corresponding column index elements are set to "execute" according to the preset deep maintenance list. This generates a maintenance strategy configuration matrix that matches the warning level, allowing maintenance decisions to be directly mapped to specific operation items, improving the accuracy of maintenance resource scheduling.
[0012] As a preferred embodiment of the present invention, the system further includes an implicit association mining module, which extracts the operation codes of maintenance operation items that have been set to execution from the generated maintenance strategy configuration matrix, constructs implicit association edges between maintenance operation items based on the co-occurrence frequency of the operation codes in historical execution records, and feeds back the implicit association edges to the maintenance strategy configuration module. Specifically, each generated maintenance strategy configuration matrix is stored in a historical matrix library. Each matrix is traversed to extract all operation codes of maintenance operation items whose values are executed as a single execution set. The total number of times any two operation codes co-occur in the same single execution set is counted. When the total number of times is greater than a preset co-occurrence threshold, an undirected implicit association edge is established between the maintenance operation items corresponding to the two operation codes, and the weight of the edge is set to the total number of times they co-occur. When generating the maintenance strategy configuration matrix for the next time, the maintenance strategy configuration module completes the initial configuration matrix assignment according to the warning level. It then obtains the operation codes of the maintenance operation items already set to execution as a seed code set, searches for the target operation codes directly connected to the seed codes, and modifies the corresponding elements of the operation items in the initial configuration matrix from "not executed" to "executed." Finally, it outputs the supplemented maintenance strategy configuration matrix. This approach automatically supplements related items based on necessary maintenance items by mining the symbiotic patterns between items in historical maintenance behaviors, avoiding the omission of strongly dependent operations due to improper maintenance list division, thereby improving the completeness and execution security of the maintenance strategy.
[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows:
[0014] The feature fusion tree construction module extracts temporal, frequency, and energy features from each multi-source state data block. It calculates the mutual information value between each pair of these three types of features as a measure of information overlap. Features with mutual information values greater than a fusion threshold are merged into a fused feature node, while features with mutual information values less than the fusion threshold are retained as independent feature nodes. A tree structure is then built from the bottom-level features to the top-level fused features according to the merging relationship. This dynamic node organization method, based on information overlap, ensures that the topology of each feature fusion tree reflects the actual correlation strength between multi-dimensional features within the current time window. This avoids the dulling effect of fixed feature combinations on changes in the intrinsic relationships of the data, and can capture and represent changes in the coupling relationships between features caused by early device degradation through changes in the tree structure.
[0015] The degradation mode walking module maps the nodes of the feature fusion tree to graph vertices and sets directional walking paths. The absolute value of the difference between the feature values of the current vertex and its neighboring vertices is used as a component of the migration cost. In each time window, the probability of the walking entity migrating from the current vertex to its neighboring vertices is calculated based on the latest feature values. The vertex numbers of the walking entity in different time windows are recorded to form a state transition probability sequence. Significant jumps in vertex numbers between adjacent windows in the probability sequence are marked as probability jump points. The equipment warning level is determined by statistically analyzing the frequency of probability jump points per unit time. This walking mechanism models the equipment degradation path as a probabilistic migration between feature nodes, eliminating the need for pre-defined fixed degradation directions or linear paths. This allows the degradation behavior of critical hydropower station equipment, which irregularly switches between multiple failure modes, to be quantified and recorded through vertex jump processes. The frequency of probability jump points directly reflects the instability of the degradation state migration, resulting in a warning level classification that better matches the actual degradation dynamics of the equipment. Meanwhile, after generating the maintenance strategy configuration matrix, an implicit association mining module is introduced to extract the co-occurrence relationship of executed maintenance operation items from the historical maintenance matrix to construct implicit association edges. When generating the maintenance strategy configuration matrix again, the maintenance operation items associated with the implicit association edges are added to the execution state, so that the maintenance strategy can adaptively expand in terms of execution coverage according to historical experience. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0017] Figure 1 This is a schematic diagram of the health management system for key equipment in a hydropower station based on predictive maintenance.
[0018] Figure 2This is a flowchart of the sliding window array filling and multi-source state data block export process;
[0019] Figure 3 This is a flowchart of the feature fusion tree construction method;
[0020] Figure 4 This is a flowchart of the degradation mode roaming module;
[0021] Figure 5 This is a flowchart of the maintenance strategy configuration based on state transition probability jump detection;
[0022] Figure 6 This is a diagram showing the frequency of state transition probability sequences of key equipment in a hydropower station and the classification of early warning levels. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.
[0024] See Figure 1 This invention provides a health management system for key equipment in hydropower stations based on predictive maintenance, comprising a status data acquisition module, a feature fusion tree construction module, a degradation mode walking module, and a maintenance strategy configuration module. The status data acquisition module acquires time-series status data generated by the key equipment in the hydropower station during operation and segments the time-series status data according to the data acquisition time point, forming multi-source status data blocks with time windows as units. The feature fusion tree construction module extracts multi-dimensional features from each multi-source status data block and constructs a feature fusion tree with feature nodes as units based on the information overlap between the extracted multi-dimensional features. The degradation mode walking module sets directional walking paths between the feature nodes of the feature fusion tree and drives the walking body to move along the directional walking paths according to the value fluctuations of the time-series status data at the feature nodes, generating a state transition probability sequence for the walking body. The maintenance strategy configuration module determines the early warning level of the key equipment in the hydropower station based on the probability jump points in the state transition probability sequence and generates a maintenance strategy configuration matrix according to the early warning level.
[0025] Example 1:
[0026] In specific implementation, please refer to Figure 2The status data acquisition module binds an independent data acquisition queue to each monitoring point in the key equipment of the hydropower station. The key equipment of the hydropower station is equipped with multiple monitoring points, each corresponding to a sensor for a specific status parameter, including vibration amplitude, temperature, pressure, rotational speed, flow rate, and displacement. Each monitoring point's data acquisition queue is allocated an independent storage buffer area, and the data acquisition queue adopts a first-in, first-out (FIFO) queue structure. Each element in the queue contains a collection timestamp and the corresponding status parameter value. Each data record collected by the status data acquisition module from each monitoring point through the device interface protocol is pushed to the tail of the corresponding data acquisition queue.
[0027] In practice, the status data acquisition module sequentially fills the values in the data acquisition queue into the sliding window array according to the acquisition timestamp. The sliding window array is a fixed-length numerical array pre-allocated in memory, and its length is adaptively adjusted based on the historical fault cycle data of the hydropower station's key equipment. The process of adaptively adjusting the window length of the sliding window array is as follows: extracting the duration of abnormal pre-fault status symptoms corresponding to each historical fault event from the historical operation and maintenance database of the hydropower station's key equipment, forming a set of symptom durations. ,in For the first The duration of abnormal fluctuations in state parameters recorded in the previous historical fault event, in seconds. Calculate the set of symptom durations. arithmetic mean The calculation method is as follows Window length of the sliding window array Determined by the following formula:
[0028]
[0029] in, The arithmetic mean of the set of durations of symptoms; This is the window length compression factor. The range of values is , The specific value is set according to the type of key equipment in the hydropower station; for equipment such as water turbines... The value is 0.85, for generator-type equipment. The value is 0.75, for equipment of the main transformer type. The value is 0.9; This indicates the floor function. The length of the sliding window array. Configured as an integer value greater than zero, and At least greater than 1. The length of the sliding window array is determined when historical fault data is insufficient to form a valid set of symptom durations. The default values are set according to the equipment type: 3600 for water turbine, 1800 for generator, and 7200 for main transformer.
[0030] In practice, the filling process of the sliding window array is time-driven. The status data acquisition module maintains a sliding window array filling pointer, which initially points to the first position of the sliding window array. The status data acquisition module periodically retrieves data elements from the head of the data acquisition queue of each monitoring point, with the retrieval order consistent with the acquisition timestamp order of the data elements. For each data element retrieved, the status parameter value of that data element is written to the position currently pointed to by the filling pointer in the sliding window array, and then the filling pointer moves one position to the right. When the position pointed to by the filling pointer exceeds the last position of the sliding window array, i.e., the sliding window array is full, all values within the sliding window array are exported as a multi-source status data block. The multi-source status data block contains all status parameter values stored sequentially in the sliding window array. These values come from different monitoring points, and the source of each monitoring point corresponding to each value is implicitly determined by its index position in the sliding window array. The mapping relationship between each index position in the sliding window array and the monitoring point is fixed during the initialization of the sliding window array, and the mapping table is saved by the status data acquisition module for subsequent processing modules to query.
[0031] After the multi-source state data blocks are exported, the state data acquisition module resets the sliding window array to receive data for the next time window. Resetting the sliding window array involves clearing all storage locations to zero or null markers and resetting the fill pointer to the first position of the array. The state data acquisition module then continues to read new data elements sequentially from each data acquisition queue, starting a new filling cycle. Through continuous filling, exporting, and resetting of the sliding window array, the state data acquisition module divides the continuously arriving time-series state data into a series of non-overlapping multi-source state data blocks, each corresponding to a fixed-length time period. There is no data overlap between adjacent time windows, and the sliding step size of the time window is equal to the window length. .
[0032] Example 2:
[0033] In specific implementation, please refer to Figure 3The feature fusion tree construction module receives a multi-source state data block output by the state data acquisition module. This multi-source state data block contains time-series state values generated by all monitoring points on key equipment of the hydropower station within the same time window. The structure of the multi-source state data block is a two-dimensional matrix, with rows corresponding to monitoring points and columns corresponding to sampling times within the same time window. The feature fusion tree construction module extracts time-domain features, frequency-domain features, and energy features from the multi-source state data block.
[0034] In practice, the method for extracting time-domain features is as follows: for each row vector corresponding to a monitoring point in the multi-source state data block, calculate its time-domain statistical index, and use the time-domain statistical index value of each monitoring point as the time-domain feature value of that monitoring point. The time-domain feature values of all monitoring points constitute a time-domain feature sequence. The time-domain statistical indicators use the root mean square value of the signal for monitoring points. corresponding row vector ,in The root mean square value is calculated as the total number of sampling points within the time window. ,Will As monitoring points The time-domain eigenvalues.
[0035] The method for extracting frequency domain features is as follows: Perform a Fast Fourier Transform on the row vector corresponding to each monitoring point in the multi-source state data block to obtain a spectrum sequence. Extract the amplitude corresponding to the main peak of the spectrum sequence as the frequency domain feature value of that monitoring point. The frequency domain feature values of all monitoring points constitute the frequency domain feature sequence. For monitoring points The amplitude spectrum is obtained after fast Fourier transform. ,search The maximum value is used as the frequency domain eigenvalue. ,in Indicates frequency index.
[0036] The method for extracting energy features is as follows: For each row vector corresponding to a monitoring point in the multi-source state data block, calculate its short-time energy value, and use the short-time energy value as the energy feature value of that monitoring point. The energy feature values of all monitoring points constitute an energy feature sequence. For monitoring points row vectors Short-term energy ,Will As monitoring points The energy characteristic value.
[0037] In practice, the feature fusion tree construction module calculates the time-domain feature sequence. With frequency domain feature sequences The first mutual information value between Calculate the first mutual information value. The process is as follows: convert the time-domain feature sequence The values in are used as random variables The implementation sample will be a frequency domain feature sequence. The values in are used as random variables The implementation samples are used to estimate the random variables using the histogram method. and joint probability distribution and marginal probability distribution and Number of bins in the histogram Based on time-domain feature sequences and frequency domain feature sequences The sample size is determined, and the number of bins is determined. ,in The total number of monitoring points. Indicates rounding up. First mutual information value. The calculation formula is:
[0038]
[0039] in, For random variables The set of all value intervals after histogram discretization. For random variables The set of all value intervals after histogram discretization; For random variables Falling into the range And random variables Falling into the range The joint probability estimate; For random variables Falling into the range Marginal probability estimates; For random variables Falling into the range Marginal probability estimates; logarithms The base is 2, and the unit of mutual information value is bits.
[0040] The feature fusion tree construction module uses the first mutual information value to calculate the feature fusion tree construction module. In the same way, calculate the time-domain feature sequence. With energy feature sequence The second mutual information value between and frequency domain feature sequences With energy feature sequence The third mutual information value between In calculating the second mutual information value At that time, and The samples are treated as two random variables, and their joint probability and marginal probability are calculated; the third mutual information value is then calculated. At that time, and The samples were treated as two random variables, and the joint probability and marginal probability were calculated using the same number of histogram bins. .
[0041] In practice, the feature fusion tree construction module will use the first mutual information value Second mutual information value and the third mutual information value Each with a preset fusion threshold Compare. Fusion threshold. The value is 0.6, which is the fusion threshold. The determination method is as follows: acquire multiple multi-source state data blocks generated by key equipment of the hydropower station during its historical normal operation phase, extract time-domain feature sequences, frequency-domain feature sequences, and energy feature sequences from each historical multi-source state data block, calculate the mutual information value between each pair of feature sequences, obtain a mutual information value sample set, and take the 90th percentile of the mutual information value sample set as the fusion threshold. .
[0042] In practical implementation, when the first mutual information value Greater than the fusion threshold At that time, the feature fusion tree construction module merges the time-domain features and frequency-domain features into a single fused feature node, denoted as node . When the first mutual information value Not greater than the fusion threshold At that time, the temporal features are treated as an independent feature node. Frequency domain features are treated as another independent feature node. Similarly, when the second mutual information value Greater than the fusion threshold At that time, the temporal features and energy features are merged into a single fused feature node. When the second mutual information value Not greater than the fusion threshold If the temporal features have not yet been merged, then the temporal features are treated as independent feature nodes; if the energy features have not yet been merged, then the energy features are treated as independent feature nodes. When the third mutual information value... Greater than the fusion threshold At that time, the frequency domain features and energy features are merged into a single fused feature node. When the third mutual information value Not greater than the fusion threshold At this time, frequency domain features or energy features that are not merged are treated as independent feature nodes. If multiple mutual information values are greater than the fusion threshold, ... And cases involving common characteristics, such as and Then the time-domain features, frequency-domain features, and energy features are merged into a single fusion feature node. The feature value of the fused feature node is a weighted concatenation of three feature values.
[0043] In practice, the feature fusion tree construction module builds a feature fusion tree from bottom-level features to top-level fused features according to the merging relationships between feature nodes. Bottom-level feature nodes include time-domain feature nodes, frequency-domain feature nodes, and energy feature nodes. When a fused feature node exists, it serves as the parent node of the merged bottom-level feature nodes. If a fused feature node containing three bottom-level features is formed... The structure of the feature fusion tree is that the root node is directly connected. The root node is a virtual master node; if only some pairwise fusion nodes are formed, the feature fusion tree includes a root node, independent feature nodes, and fused feature nodes, with the corresponding underlying feature nodes attached to the fused feature nodes. The feature fusion tree construction module stores the structure of the completed feature fusion tree and the feature values of each feature node for use by the degradation mode traversal module.
[0044] Example 3:
[0045] In specific implementation, please refer to Figure 4 The degradation mode traversal module obtains the feature fusion tree output by the feature fusion tree construction module. The feature fusion tree contains a root node, at least one fused feature node, and independent feature nodes, as well as parent-child relationships between nodes. The degradation mode traversal module maps each feature node in the feature fusion tree to a vertex in the graph structure and creates a separate vertex for the root node of the feature fusion tree. Each vertex is assigned a unique vertex number, starting from zero and incrementing sequentially, with the vertex number corresponding to the root node being zero.
[0046] In its implementation, the degenerate mode traversal module sets directed edges between vertices according to the parent-child relationships between feature nodes in the feature fusion tree. For a pair of feature nodes with a parent-child relationship in the feature fusion tree, the degenerate mode traversal module sets two directed edges between the vertex corresponding to the parent feature node and the vertex corresponding to the child feature node: a forward edge from the parent vertex to the child vertex, and a reverse edge from the child vertex to the parent vertex. The forward and reverse edges have their own independent edge weights.
[0047] The degenerate mode traversal module calculates the migration cost of each directed edge. The migration cost of the traversal volume is obtained by multiplying the difference in feature values between two feature nodes by the edge weights between the nodes. Before traversal begins, once the feature fusion tree is constructed, the degenerate mode traversal module uses the multi-source state data block corresponding to the first time window to calculate the initial feature values of all feature nodes in the feature fusion tree. The calculation method for the initial feature values of the feature nodes is consistent with the method used by the feature fusion tree construction module to extract feature values. For each directed edge, let the directed edge originate from vertex... Pointing to the vertex ,vertex The initial eigenvalues are ,vertex The initial eigenvalues are Then the magnitude of the difference in eigenvalues Calculated as The weight of a directed edge. The weight of the edge is determined based on the feature correlation between the two feature nodes connected by the directed edge: if the two vertices corresponding to the directed edge originate from feature nodes with a fusion relationship in the feature fusion tree, then the edge weight is determined. The mutual information value between the feature sequences corresponding to these two feature nodes is taken. This mutual information value has been calculated and stored during the feature fusion tree construction stage. If the two vertices corresponding to the directed edge do not have a direct fusion relationship, then the edge weight is... Set to the default value of 0.5. Migration cost. The calculation formula is migration cost It is calculated and fixed before the walk begins, and does not change during the entire degenerate mode walk.
[0048] In practice, the degradation mode walking module sets the initial position of the walk body on the vertex corresponding to the root node of the feature fusion tree, i.e., the vertex with vertex number zero. The position of the walk body is recorded as the current vertex number, which is initially zero.
[0049] The degradation mode traversal module processes each time window sequentially according to the chronological order of the multi-source state data blocks provided by the state data acquisition module. The degradation mode traversal module binds the multi-source state data blocks of each time window to a corresponding window number. Numbering starts from 1 and continues consecutively. For the... The degradation mode walking module operates within the first time window. The feature values of each feature node in the feature fusion tree are recalculated for the multi-source state data block corresponding to the window number. The calculation method is as follows: extract the time-domain feature sequence, frequency-domain feature sequence, and energy feature sequence from the multi-source state data block. According to the node composition rules of the feature fusion tree, for independent feature nodes, directly take the feature value of the corresponding monitoring point in the corresponding type of feature sequence; for fused feature nodes, perform a weighted average of the bottom-level feature values participating in the fusion. The weight of the weighted average is determined according to the proportion of mutual information values corresponding to the bottom-level features. The calculated feature values of each feature node are assigned to the corresponding vertex to obtain the first... Feature values of each vertex within a time window ,in Number the vertices.
[0050] In practice, the degradation mode wandering module is in the first... Within a given time window, calculate the probability that the wandering object will migrate from the current vertex to an adjacent vertex along a directed edge. Obtain the number of the vertex currently occupied by the wandering object. Traverse all vertices in the graph structure Each directed edge originates from a vertex and points to an adjacent vertex. Calculate the walking volume from the current vertex. Migrate to adjacent vertices The initial probability is defined by the following formula:
[0051]
[0052] in, In the first The downstream body moves from the apex within a time window. Migrate to Vertex The initial probability; For the first Vertex under a time window eigenvalues; For the first Vertex under a time window eigenvalues; From the vertex To the top The migration cost of directed edges, migration cost The initialization phase calculates the product of the feature value difference magnitude and the edge weights within the first time window and keeps it unchanged in subsequent processes. When the denominator... If the value is zero, and the molecule is also zero, then set... If the numerator is not zero, then Consider it as an extremely small positive number Perform the calculation.
[0053] In practice, the degradation mode walking module obtains the vertex. After obtaining the initial probabilities of all adjacent vertices, the initial probabilities of all adjacent vertices are normalized to obtain the walking volume starting from the current vertex. The final probability of migrating to each adjacent vertex. The normalization method is to first calculate the sum of the initial probabilities of all adjacent vertices. ,in As vertices The set of all adjacent vertices. If The wandering body will remain at the current vertex. The probability of finding the first vertex is 1, and the probability of migrating to any adjacent vertex is 0. If... Then the final probability The degenerate mode walking module randomly selects a directed edge for migration based on the final probability, updates the position of the walking body to the selected adjacent vertex, and records the position of the walking body at the [number]th [position]. The vertex number at the end of each time window.
[0054] The degradation mode walk module repeats the above process in ascending order of window numbers, processing the multi-source state data blocks corresponding to all time windows sequentially. It records the vertex number at the end of each time window, forming an ordered sequence of vertex numbers changing with the window number. This ordered sequence is the state transition probability sequence of the walk. The state transition probability sequence is output to the maintenance strategy configuration module.
[0055] Example 4:
[0056] In specific implementation, please refer to Figure 5 The maintenance strategy configuration module receives the state transition probability sequence output by the degradation mode traversal module. This sequence is an ordered sequence formed by the change of vertex numbers with window numbers. The maintenance strategy configuration module performs jump detection on the state transition probability sequence and marks the window positions where the jump amplitude of vertex numbers between two adjacent windows exceeds the jump threshold as probability jump points.
[0057] In practice, the transition threshold is dynamically calculated based on the mean square error of the state transition probability sequence. Let the state transition probability sequence contain... The vertex ID value is denoted as ,in Indicates the first The vertex number where the wandering body is located at the end of each time window. The maintenance strategy configuration module calculates the mean squared error of all vertex number values in the state transition probability sequence. The calculation process is as follows: First, calculate the mean value of the vertex number. Then calculate the mean square error. Jump threshold Determined by the following formula:
[0058]
[0059] in, This is the jump sensitivity coefficient; This represents the mean squared deviation of the vertex index values in the state transition probability sequence. The jump sensitivity coefficient. The range of values is Jump sensitivity coefficient The specific value is set according to the operational stability requirements of key equipment in hydropower stations: for key equipment in hydropower stations with high operational stability requirements, the jump sensitivity coefficient is... A value of 1.5 is used for high-stability equipment, including the thrust bearings and guide bearings of hydro-generator units; for key equipment in hydropower stations with medium-stability requirements, the jump sensitivity coefficient is... The value is 2.2. Medium stability level equipment includes main transformers and speed governors; for key equipment in hydropower stations with low stability requirements, the jump sensitivity coefficient is... The value is 3.0. Low stability level equipment includes cooling water pumps and hydraulic devices.
[0060] In practice, the maintenance strategy configuration module checks the vertex numbers of adjacent windows in the state transition probability sequence pair by pair. For the first... The first window and the first Each window, vertex number jump amplitude When the vertex number jumps by an amplitude Exceeding the transition threshold At that time, the first The positions of each window are marked as probability transition points. The maintenance strategy configuration module traverses all window positions in the state transition probability sequence except for the first window, completing the marking of all probability transition points.
[0061] In practice, the maintenance strategy configuration module counts the frequency of probability jump points occurring within a unit of time. The unit of time is set to one hour. The maintenance strategy configuration module converts the number of windows into time units based on the actual duration of the time window. If the duration of a single time window is... If the time interval is seconds, then the total number of windows per unit time is The maintenance strategy configuration module sets up a sliding statistics window, the length of which is equal to... The sliding statistical window slides across the state transition probability sequence in steps of one window position at a time. Within each sliding statistical window, the maintenance strategy configuration module counts the number of window positions marked as probability jump points; this number represents the frequency of probability jump points per unit time within that statistical period, denoted as . .
[0062] In practice, the maintenance strategy configuration module will count the frequency of probability jump points per unit time. It is compared with the first frequency threshold and the second frequency threshold, the first frequency threshold being denoted as... The second frequency threshold is denoted as First frequency threshold Less than the second frequency threshold First frequency threshold The value is 3, the second frequency threshold. The value is 8. First frequency threshold. Second frequency threshold The determination is based on: obtaining state transition probability sequence samples of key equipment in the hydropower station during historical operation, statistically analyzing the frequency distribution of probability jump points per unit time in the historical samples, performing percentile analysis on the historical frequency distribution, and using the thirtieth percentile of the historical frequency distribution as the first frequency threshold. The value of is determined by using the 75th quantile of the historical frequency distribution as the second frequency threshold. The value of .
[0063] In practical implementation, the frequency of probability jump points per unit time Less than the first frequency threshold At that time, the maintenance strategy configuration module determines the warning level to be no warning level. This is based on the frequency of probability jump points per unit time. Between the first frequency threshold Second frequency threshold Between, that is, satisfying At that time, the maintenance strategy configuration module determines the early warning level as Level 1. This is based on the frequency of probability jump points per unit time. Greater than the second frequency threshold At that time, the maintenance strategy configuration module determined the warning level to be a level two warning level.
[0064] In practice, the maintenance strategy configuration module generates a maintenance strategy configuration matrix according to the warning level. This module sets multiple maintenance operation items for key equipment in the hydropower station, including visual inspection, bolt tightening inspection, lubricant replacement, filter replacement, seal replacement, electrical insulation testing, vibration spectrum analysis, bearing clearance adjustment, rotor dynamic balancing correction, cooling system cleaning, winding DC resistance measurement, and partial discharge testing. Each maintenance operation item is assigned a unique operation code, represented by a four-digit numeric character. The execution status of each maintenance operation item is set to either "executed" or "not executed," with "executed" being represented by the number 1 and "not executed" by the number 0.
[0065] In implementation, the maintenance strategy configuration module constructs an initial configuration matrix using the equipment identifiers of key hydropower station equipment as matrix row indices and the operation codes of maintenance operation items as matrix column indices. The equipment identifier is a unique code for the key hydropower station equipment in the asset management system, using a combination of equipment type code and serial number. The number of rows in the initial configuration matrix matches the number of key hydropower station equipment requiring maintenance strategies, and the number of columns matches the total number of maintenance operation items. The initial value of all elements in the initial configuration matrix is 0, corresponding to the non-execution state.
[0066] In practice, when the warning level is no warning level, the maintenance strategy configuration module sets all elements in the initial configuration matrix to "not executed," meaning all elements in the initial configuration matrix are kept to a value of 0. When the warning level is Level 1, the maintenance strategy configuration module sets the corresponding column index elements to "executed" according to the preset light maintenance list. The preset light maintenance list contains a set of operation codes, and the maintenance operation items corresponding to the operation codes in the light maintenance list are: operation codes for visual inspection, operation codes for fastening bolts, operation codes for lubricating oil replacement, operation codes for filter element replacement, and operation codes for electrical insulation testing. The maintenance strategy configuration module obtains the column index corresponding to each operation code in the light maintenance list in the initial configuration matrix and modifies the value of the element corresponding to these column indices in the equipment row from 0 to 1. When the warning level is Level 2, the maintenance strategy configuration module sets the corresponding column index elements to "executed" according to the preset deep maintenance list. The preset deep maintenance list contains operation codes for all maintenance operations. Specifically, the operation codes in the deep maintenance list cover all operations including visual inspection, bolt tightening inspection, lubricant replacement, filter replacement, seal replacement, electrical insulation testing, vibration spectrum analysis, bearing clearance adjustment, rotor dynamic balancing, cooling system cleaning, winding DC resistance measurement, and partial discharge testing. The maintenance strategy configuration module obtains the column index of each operation code in the deep maintenance list in the initial configuration matrix and changes the element values corresponding to these column indices in the equipment row from 0 to 1.
[0067] In practice, the maintenance strategy configuration module outputs the completed matrix as the maintenance strategy configuration matrix. Each row of the maintenance strategy configuration matrix corresponds to a key piece of equipment in a hydropower station, and each column corresponds to a maintenance operation item. A value of 1 indicates that the corresponding maintenance operation item is executed, and a value of 0 indicates that the corresponding maintenance operation item is not executed.
[0068] See Figure 6 In the graph, the horizontal axis represents time in hours, and the vertical axis represents the frequency of probability jumps per unit time. The curve is plotted as a solid line, representing the frequency of probability jump points that change over time. This reflects the changing trend of probability jump points per unit time after the maintenance strategy configuration module performs jump detection on the state transition probability sequence output by the degradation mode walk module. The figure also uses dashed lines and dotted lines to indicate the first frequency threshold, respectively. Second frequency threshold The two thresholds divide the vertical axis region into three warning level intervals.
[0069] The dotted areas at the bottom of the image represent Level 1 warning areas, corresponding to the frequency of probability jump points. Between the first frequency threshold Second frequency threshold Between, that is The diagonal striped area at the top indicates a level-two warning zone, corresponding to the frequency of probability jump points. Greater than the second frequency threshold ,Right now The blank area between the two regions corresponds to no warning level, and the frequency of probability jump points. Less than the first frequency threshold ,Right now .
[0070] The curve trend shows the frequency of probability jump points within the time period from 0 to approximately 400 hours. Basically maintained at the first frequency threshold The curve shows relatively small fluctuations and low frequency, indicating that the equipment status is relatively stable during this stage, and the maintenance strategy configuration module determines it to be at the no-warning level. As time progresses, the curve shows a significant upward trend after 400 hours, with the frequency gradually exceeding the first frequency threshold and entering the first-level warning range, lasting until approximately 700 hours. During this stage, abnormal fluctuations in the equipment status occur, and the maintenance strategy configuration module determines it to be at the first-level warning level, executing a light maintenance operation strategy.
[0071] Frequency of probability jump points after 700 hours It continued to climb rapidly and broke through the second frequency threshold multiple times. Entering the Level 2 warning zone, with a frequency exceeding 10, reaching a maximum of approximately 15, indicates significant abnormal fluctuations in equipment status and a clear degradation trend. The maintenance strategy configuration module determines this as a Level 2 warning, triggering a deep maintenance operation strategy. Overall, the frequency curve of probability jump points smoothly transitions from the no-warning zone to the Level 1 warning range, and then rapidly rises towards the Level 2 warning range, demonstrating the maintenance strategy configuration module's sensitive detection capability for abnormal equipment status and its dynamic adjustment function for warning levels.
[0072] Example 5:
[0073] In practice, the implicit association mining module and the maintenance strategy configuration module work together. Each time the maintenance strategy configuration matrix is generated, the maintenance strategy configuration module outputs the generated matrix to the maintenance execution system, simultaneously transmitting it to the implicit association mining module. The implicit association mining module stores each received maintenance strategy configuration matrix in a historical matrix repository. The historical matrix repository is a persistent storage structure; each record in the repository corresponds to a maintenance strategy configuration matrix generation event, and the record content includes a generation timestamp, the equipment identifier of the key equipment in the hydropower station, and the complete maintenance strategy configuration matrix.
[0074] In practical implementation, the implicit association mining module extracts the operation codes of maintenance operation items that have been set to execution from the maintenance strategy configuration matrix. Based on the co-occurrence frequency of these operation codes in historical execution records, it constructs implicit association edges between maintenance operation items. The implicit association mining module traverses each maintenance strategy configuration matrix in the historical matrix library, extracting the operation codes of all maintenance operation items whose values are set to execution as a single execution set. The extraction method is as follows: for each maintenance strategy configuration matrix stored in the historical matrix library, it traverses row by row. For each row (i.e., each device), it reads the operation code corresponding to the column with a value of 1 in that row, and records the collected operation codes as... ,in Configure the record index of the matrix in the historical matrix library for this maintenance strategy.
[0075] In practice, the implicit association mining module counts the total number of times any two operation codes co-occur in the same single execution set. Let the complete set of all possible operation codes be . ,in To maintain the total number of operation items, the implicit association mining module initializes a... Co-occurrence frequency matrix Co-occurrence frequency matrix Middle elements Operation code With operation code The total number of times they appear together in the same single execution set, initially all The implicit association mining module iterates through all records in the history matrix database, collecting all single execution sets. For each single execution set Take any two different operation codes. and The co-occurrence frequency matrix The corresponding and Simultaneously increment by 1. When The accumulation operation is not performed at this time.
[0076] In practice, the implicit association mining module calculates the total number of times any two operation codes co-occur. Co-occurrence frequency threshold with preset Compare the results. Use the preset co-occurrence threshold. The value is 15. Co-occurrence threshold. The determination is based on the following criteria: All actual maintenance work orders executed within the past two years are obtained from the hydropower station's operation and maintenance history. The set of maintenance operations simultaneously executed in each maintenance work order is extracted as a reference single execution set. The distribution of the number of times any two operation codes co-occur in all reference single execution sets is statistically analyzed, and the median of the co-occurrence distribution is used as the co-occurrence threshold. The value of .
[0077] When the total number of times they appear together Greater than the co-occurrence threshold At that time, the implicit association mining module is in the operation coding Corresponding maintenance operation items and operation codes An undirected implicit association is established between the corresponding maintenance operation items, and the weight of the implicit association is set... Set to the total number of times they appear together Implicitly related edges are stored in a data table, which contains the fields: first operation code, second operation code, and edge weight. All implicitly related edges constitute the implicitly related edge set, and the implicitly related edge mining module feeds this set back to the maintenance strategy configuration module.
[0078] In practical implementation, when generating the maintenance strategy configuration matrix for the next time, the maintenance strategy configuration module supplements and sets the maintenance operation items that have not been set to execution according to the implicit association edges. In the next generation process, the maintenance strategy configuration module first completes the assignment of the initial configuration matrix according to the warning level. The assignment process is consistent with that described in Example 4, thus obtaining the initial configuration matrix. The maintenance strategy configuration module obtains the initial configuration matrix. The operation codes of all maintenance operation items that have been set to be executed are used as the seed code set. The acquisition method is as follows: for the row vectors corresponding to the key equipment of the hydropower station that requires the generation of maintenance strategies, find the initial configuration matrix. Extract the operation codes corresponding to all column indices with an element value of 1 in the current row to form a seed code set. .
[0079] In practice, the maintenance strategy configuration module traverses the seed code set. Each seed code in the code, for each seed code Search for the seed encoding in the set of implicitly related edges constructed by the implicit association mining module. The target operation codes that are directly connected. The search method is: in the implicitly related edge data table, filter out the operation codes whose first operation codes are equal to... Or the second operation code equals The record will contain information that is not equal to... Another operation code as the target operation code The maintenance strategy configuration module encodes all found target operations. Add to candidate supplement set In, and remove the set of seed codes. Repeated operation codes in the middle.
[0080] In practice, the maintenance strategy configuration module traverses the candidate supplement set. Each target operation code in Encode the target operation The corresponding operation items are in the initial configuration matrix. The corresponding element in the code is changed from never being set to execute to being set to execute. Specifically: find the target operation code. In the initial configuration matrix The corresponding column index Index the rows and columns containing the key equipment of the hydropower station. The element value is changed from 0 to 1. After the maintenance strategy configuration module completes the modification of all target operation codes, it outputs the modified initial configuration matrix as the final maintenance strategy configuration matrix. The maintenance strategy configuration module simultaneously sends this final maintenance strategy configuration matrix to the implicit association mining module, which stores the final maintenance strategy configuration matrix in the historical matrix library for use in subsequent implicit association edge construction.
[0081] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A health management system for key equipment in hydropower stations based on predictive maintenance, characterized in that, include: The status data acquisition module is used to acquire the time-series status data generated by the key equipment of the hydropower station during operation, and to segment the time-series status data according to the data acquisition time point to form multi-source status data blocks with time windows as the unit. The feature fusion tree construction module is used to extract multi-dimensional features from each multi-source state data block and construct a feature fusion tree with feature nodes as units based on the information overlap between the extracted multi-dimensional features. The degradation mode walking module is used to set up directional walking paths between feature nodes of the feature fusion tree, and drive the walking body to move on the directional walking path according to the value fluctuation of the time-series state data on the feature nodes, generating the state transition probability sequence of the walking body. The maintenance strategy configuration module is used to determine the early warning level of key equipment in the hydropower station based on the probability jump points in the state transition probability sequence, and generate a maintenance strategy configuration matrix according to the early warning level.
2. The health management system for key equipment in hydropower stations based on predictive maintenance according to claim 1, characterized in that, The status data acquisition module acquires time-series status data generated by key equipment in the hydropower station during operation, and segments the time-series status data according to the data acquisition time point to form multi-source status data blocks based on time windows. The specific implementation method is as follows: The status data acquisition module binds an independent data acquisition queue to each monitoring point in the key equipment of the hydropower station. According to the acquisition timestamp, the values in the data acquisition queue are sequentially filled into the sliding window array. After the sliding window array is full, all the values in the sliding window array are exported as a multi-source status data block. At the same time, the sliding window array is reset to receive data from the next time window.
3. The health management system for key equipment in hydropower stations based on predictive maintenance according to claim 2, characterized in that, The window length of the sliding window array is adaptively adjusted based on the historical failure cycle data of key equipment in the hydropower station.
4. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 1, characterized in that, The feature fusion tree construction module performs multi-dimensional feature extraction on each multi-source state data block, and constructs a feature fusion tree with feature nodes as units based on the information overlap between the extracted multi-dimensional features. The specific implementation method is as follows: The feature fusion tree construction module extracts time-domain features, frequency-domain features, and energy features from the multi-source state data block, calculates the first mutual information value between the time-domain features and the frequency-domain features, the second mutual information value between the time-domain features and the energy features, and the third mutual information value between the frequency-domain features and the energy features. The first, second, and third mutual information values are compared with preset fusion thresholds. The two features corresponding to mutual information values greater than the fusion threshold are merged into one fusion feature node, and the two features corresponding to mutual information values not greater than the fusion threshold are treated as independent feature nodes. The feature fusion tree is built from the bottom-level features to the top-level fusion features according to the merging relationship between the feature nodes.
5. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 1, characterized in that, The degradation mode walking module sets up directional walking paths between feature nodes in the feature fusion tree, and drives the walking volume to move along the directional walking paths based on the value fluctuations of the temporal state data at the feature nodes. The specific implementation of generating the state transition probability sequence of the walking volume is as follows: The degradation mode walking module maps each feature node in the feature fusion tree to a vertex in the graph structure. It sets directed edges between vertices according to the parent-child relationship between feature nodes in the feature fusion tree. The difference in feature values between the two vertices connected by each directed edge is used as the migration cost of the walking body. The initial position of the walking body is set on the vertex corresponding to the root node of the feature fusion tree. The feature values of each vertex are updated according to the multi-source state data block corresponding to each time window. The probability of the walking body migrating from the current vertex to the adjacent vertex along the directed edge is calculated. The vertex number where the walking body is located at the end of each time window is recorded. The sequence formed by the change of vertex number with the window number is used as the state transition probability sequence of the walking body.
6. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 5, characterized in that, The migration cost of the walk is obtained by multiplying the difference in eigenvalues between two feature nodes by the edge weights between the nodes.
7. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 5, characterized in that, The specific implementation of the degradation mode walking module, which updates the feature values of each vertex according to the multi-source state data block corresponding to each time window, and calculates the probability of the walking volume migrating from the current vertex to an adjacent vertex along the directed edge, is as follows: The degradation mode walking module binds the multi-source state data blocks of each time window to the corresponding window number. For the multi-source state data blocks corresponding to each window number, it recalculates the feature values of each feature node in the feature fusion tree. It compares the feature value of the current vertex of the walk with the feature values of the adjacent vertices, calculates the absolute value of the difference between the feature values of the adjacent vertices and the feature value of the current vertex, and uses the ratio of the absolute value of the difference to the migration cost of the walk from the current vertex to the adjacent vertex as the initial probability of the walk from the current vertex to the adjacent vertex along the directed edge. Then, it normalizes the initial probabilities of all adjacent vertices to obtain the final probability of the walk from the current vertex to each adjacent vertex.
8. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 1, characterized in that, The specific implementation method of the maintenance strategy configuration module in determining the early warning level of key equipment in the hydropower station based on the probability jump points in the state transition probability sequence is as follows: The maintenance strategy configuration module marks the window positions where the vertex number jump amplitude between two adjacent windows in the state transition probability sequence exceeds the jump threshold as probability jump points. It counts the frequency of probability jump points occurring within a unit of time and compares the frequency of probability jump points within a unit of time with a first frequency threshold and a second frequency threshold, where the first frequency threshold is less than the second frequency threshold. When the frequency of probability jump points within a unit of time is less than the first frequency threshold, the warning level is determined to be no warning level. When the frequency of probability jump points within a unit of time is between the first and second frequency thresholds, the warning level is determined to be a level one warning level. When the frequency of probability jump points within a unit of time is greater than the second frequency threshold, the warning level is determined to be a level two warning level.
9. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 8, characterized in that, The transition threshold is dynamically calculated based on the mean squared error of the state transition probability sequence.
10. The health management system for key equipment in a hydropower station based on predictive maintenance according to claim 8, characterized in that, The specific implementation method of the maintenance strategy configuration module generating the maintenance strategy configuration matrix according to the early warning level is as follows: The maintenance strategy configuration module sets multiple maintenance operation items for key equipment in the hydropower station. The execution status of each maintenance operation item is set to either "execute" or "not execute". An initial configuration matrix is constructed using the equipment identifier of the key equipment in the hydropower station as the matrix row index and the operation code of the maintenance operation item as the matrix column index. When the warning level is no warning level, all elements in the initial configuration matrix are set to "not execute". When the warning level is Level 1, the corresponding column index element is set to "execute" according to the preset light maintenance list. When the warning level is Level 2, the corresponding column index element is set to "execute" according to the preset deep maintenance list. The matrix after the assignment is completed is output as the maintenance strategy configuration matrix.