A method and system for remote diagnosis of hydroelectric faults based on multi-modal feature fusion

Abstract

The application discloses a kind of water and electricity fault remote diagnosis method and system based on multi-modal feature fusion, it is related to the technical field of equipment condition monitoring and fault diagnosis, the present application is by vibration, temperature, pressure signal is quantified as symbol sequence, and high-density operation partition is carried out in combination with head-load probability density distribution, so that benchmark causal diagram can adapt to the dynamic characteristics under different working conditions, significantly reduce the false alarm and omission caused by unified model in all working conditions;Simplified granger causality test using rolling prediction, without building complex vector autoregressive model, significantly reduce the computational complexity, suitable for edge node real-time deployment;And by comparing real-time causal diagram with partition benchmark causal diagram, the disappearance, appearance and direction reversal of directed edge can be detected simultaneously, which can effectively identify early faults that are difficult to detect by traditional threshold method, and improve the sensitivity and interpretability of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the technical field of equipment condition monitoring and fault diagnosis, and in particular to a remote diagnosis method and system for hydropower faults based on multimodal feature fusion. Background Technology

[0002] As complex large-scale rotating machinery, hydro-generator units are affected by the coupling of multiple operating parameters such as head and load, and the physical quantities such as vibration, temperature, and pressure exhibit nonlinear and time-varying causal transmission relationships. In recent years, data-driven fault diagnosis methods have been extensively studied, and traditional techniques such as threshold alarms, pattern matching, and feature trend analysis have been applied to some extent in practical engineering. However, most existing methods use static models or empirical rules uniform across all operating conditions, which are difficult to adapt to the significant differences in the dynamic characteristics of the unit under different head-load combinations. At the same time, traditional causal inference methods usually rely on vector autoregressive models of continuous numerical sequences, which have high computational complexity and are sensitive to noise, making it difficult to achieve real-time online inference at edge nodes.

[0003] For example, CN104331631A discloses a remote diagnostic decision-making method for the operating status of hydropower units. By establishing a basic database containing normal values ​​and alarm values, as well as a fault database containing fault causes, the method employs three types of mode analysis methods: unit start-up mode, operating mode, and shutdown mode. Combined with specific frequency band analysis and trend analysis, the method compares the collected unit operating data with the values ​​in the basic database. If the values ​​exceed the range, the method matches them in the fault database to diagnose the fault. However, this method has the following shortcomings: First, fault diagnosis relies on pre-set fixed thresholds and manually constructed fault rule bases, lacking the ability to adapt to the unit's operating conditions (head, load). When the unit operates in an operating condition area not fully covered by the historical database for a long time, the false alarm rate and missed alarm rate increase significantly. Second, this method only performs threshold comparison and trend analysis on a single physical quantity (such as vibration amplitude), failing to reveal the causal coupling relationship between multi-source signals such as vibration, temperature, and pressure, and making it difficult to identify structural faults caused by the reversal of causal direction or the appearance / disappearance of edges. Third, the specific frequency band analysis only focuses on the 10-200Hz frequency band of the top cover vibration measurement point, without considering the differences in causal transmission modes under different operating zones, resulting in the neglect of the operating condition dependence of the diagnostic conclusions.

[0004] CN109407634A discloses a business architecture for a remote diagnostic platform for hydropower units in a power generation group. It constructs a three-tiered platform at the plant, regional, and group levels, realizing business functions such as condition monitoring, energy efficiency analysis, fault diagnosis, operation optimization, and maintenance decision-making. While this architecture has advantages in system integration and data flow management, the fault diagnosis module still employs traditional threshold judgment and rule matching methods, lacking the ability to model the dynamic causal relationships of the units. Furthermore, this architecture primarily focuses on platform hierarchy and business process design, lacking specific implementation schemes for core algorithms such as underlying signal processing, adaptive operating condition partitioning, and online causal graph construction, making it difficult to support refined fault location and early warning. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0006] In view of the aforementioned existing problems, the present invention is proposed.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a remote diagnosis method for hydropower faults based on multimodal feature fusion, characterized by: collecting unit vibration signals, temperature signals, and pressure signals; calculating the increment of each signal relative to the previous moment, and quantizing the increment into symbols to obtain vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences; at edge nodes, constructing a probability density distribution based on head and load values ​​from historical normal operation data, dividing the head-load plane into multiple high-density operation zones, and pre-storing a reference directed causal graph under normal conditions corresponding to each high-density operation zone; inputting the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence into the edge nodes, determining the target zone based on the current head and load values, performing a simplified Granger causality test on each symbol sequence within the target zone, and constructing a real-time directed causal graph; comparing the real-time directed causal graph with the reference directed causal graph corresponding to the target zone, detecting direction reversal of directed edges or the appearance and disappearance of edges, and outputting the fault type and abnormal directed edges.

[0008] As a preferred embodiment of the present invention, the construction of the probability density distribution includes: collecting water head values ​​during historical normal operation. and load value Multiple sample points Establish the head axis in the memory of the edge node. and load shaft A two-dimensional grid is used, with the water head axis divided into cells at fixed intervals; the number of sample points in each cell is counted. ,reserve The cells, where, The minimum number of samples to retain for a cell; the division into multiple high-density running partitions includes: merging adjacent retained cells into connected regions using a four-connectivity rule, with each connected region marked as a high-density running partition, resulting in a total of... High-density operating zones .

[0009] As a preferred embodiment of the present invention, the generation of the reference directed causal graph includes: for each partition The complete vibration, temperature, and pressure symbol sequences corresponding to all historical normal operating times are extracted, and each continuous segment is processed separately in chronological order. Rolling prediction is performed within each continuous segment, without matching across segments. A simplified Granger causality test method based on rolling prediction is used to calculate the directed causal relationship between each symbol pair: Let the symbol sequence to be tested... and The lengths are all ;in, The lag order is given by the time interval; for any given time interval... : In sequence All indexes satisfy and The location, after counting these locations The symbol frequency is used to determine the highest frequency symbol. If no conditions are met. If the result is 0, then the prediction is 0; in the sequence All indexes satisfy and The location, after counting these locations The highest frequency of a symbol is taken as the symbol frequency. If no match is found, the prediction is 0; iterate through the data. arrive Statistical analysis based on Number of correct predictions and based on Number of times self-prediction is correct ;like Then it is determined that a directed edge exists, where, To tolerate offsets, repeat the above process for each distinct pair of symbolic variables, and aggregate all determined-to-existing directed edges into a baseline directed causal graph. .

[0010] In a preferred embodiment of the present invention, determining the target partition includes: reading the current head value. and load value traversal A high-density running partition is checked for the existence of partitions. satisfy: in, This is the lower boundary of the zone's water head. The upper limit of the waterway, This is the lower limit of the load. The upper limit of the load is defined; if it exists, the partition that meets the condition is selected as the target partition; if it does not exist, the head boundary distance of each partition is calculated. : Distance from load boundary : choose Minimum partition, and requirements Less than or equal to the first distance threshold in the direction of water head. If the distance is less than or equal to the first distance threshold in the load direction, it is selected as the target partition; otherwise, the current operating condition is determined to be in a low-density area, an unmatched partition is output, and the current detection is skipped.

[0011] As a preferred embodiment of the present invention, the construction of the real-time directed causal graph includes: directly reading a sliding window sequence of length n from the real-time vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence; in the target partition Within this process, the simplified Granger causality test for rolling prediction is performed again for each pair of symbol sequences, but the pattern lookup source data used for prediction is limited to each prediction time in the current sliding window sequence. The previous part; in which, the tolerance offset Keep them the same; aggregate all determined-existing directed edges into a real-time directed causal graph. and will and target partition The identifiers are associated and stored in the memory of the edge nodes.

[0012] In a preferred embodiment of the present invention, the comparison includes: reading the real-time directed causal graph from the edge node memory. and according to the target partition The identifier is read from the non-volatile memory to obtain the corresponding reference directed causal graph. Extract separately and The set of all directed edges in and Each directed edge is denoted as ,in ;in, These are identifiers for vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences, respectively.

[0013] In a preferred embodiment of the present invention, the comparison further includes traversing all directed edges; the directed edges include: Assign ternary state values ​​in priority order; concatenate the directed edges of the abnormality with state values ​​of 1, 2 or 3 in order to form a fault code, and output the fault code and the corresponding directed edge description: state value 1 corresponds to disappearance, state value 2 corresponds to appearance, and state value 3 corresponds to reversal.

[0014] As a preferred embodiment of the present invention, the step of assigning a ternary state value includes: if If it is, then assign the value 3; otherwise, if If it is, then assign the value 1; otherwise, if If the value is 2, then assign it the value 2; otherwise, assign it the value 0.

[0015] On the other hand, the present invention also provides a remote diagnostic system for hydropower faults based on multimodal feature fusion, including a symbol incremental encoding module, which collects vibration signals, temperature signals, and pressure signals of the unit, calculates the increment of each signal relative to the previous moment, and quantizes the increment into symbols to obtain vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences; a partitioned causal graph library module, which constructs a probability density distribution based on the head and load values ​​in historical normal operation data at the edge nodes, divides the head-load plane into multiple high-density operation partitions, and pre-stores the reference directed causal graph under normal conditions corresponding to each high-density operation partition; a real-time causal construction module, which inputs the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence into the edge nodes, determines the target partition according to the current head value and current load value, performs a simplified Granger causality test on each symbol sequence within the target partition, and constructs a real-time directed causal graph; and a causal graph comparison and diagnosis module, which compares the real-time directed causal graph with the reference directed causal graph corresponding to the target partition, detects the direction reversal of directed edges or the appearance and disappearance of edges, and outputs the fault type and abnormal directed edges.

[0016] The beneficial effects of this invention are as follows: By quantizing vibration, temperature, and pressure signals into symbol sequences and combining them with the head-load probability density distribution for high-density operation partitioning, the baseline causal graph can adapt to the dynamic characteristics under different operating conditions, significantly reducing false alarms and missed alarms caused by the unified model for all operating conditions. The simplified Granger causality test using rolling prediction eliminates the need to construct a complex vector autoregression model, greatly reducing computational complexity and making it suitable for real-time deployment of edge nodes. Furthermore, by comparing the real-time causal graph with the partitioned baseline causal graph, it can simultaneously detect three structural anomalies: the disappearance, appearance, and direction reversal of directed edges. This can effectively identify early faults that are difficult to detect using traditional threshold methods (such as causal direction reversal and the generation of new coupling relationships), improving the sensitivity and interpretability of fault diagnosis. It also achieves refined, low-latency remote diagnosis of the operating status of hydro-generator units. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating a remote diagnosis method for hydropower faults based on multimodal feature fusion, as shown in this invention.

[0018] Figure 2 This is a schematic diagram of the real-time signal acquisition and symbol quantization process shown in this invention.

[0019] Figure 3 The present invention presents a structural diagram of a remote hydropower fault diagnosis system based on multimodal feature fusion. Detailed Implementation

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0022] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0023] According to an embodiment of the present invention, in combination Figure 1 and Figure 2 The flowchart shown illustrates a remote fault diagnosis method for hydropower based on multimodal feature fusion, comprising: S1: Collect vibration, temperature and pressure signals from the unit, calculate the increment of each signal from the previous moment to the current moment, and quantize the increment into symbols to obtain vibration symbol sequences, temperature symbol sequences and pressure symbol sequences.

[0024] Specifically, the vibration signal output by the unit's vibration sensor, the temperature signal output by the temperature sensor, and the pressure signal output by the pressure sensor are collected synchronously at a sampling frequency of 200Hz.

[0025] The vibration signals are dimensionless after normalization, with temperature signals measured in degrees Celsius (°C) and pressure signals in megapascals (MPa). The vibration, temperature, and pressure signal values ​​acquired each time are pushed into the tails of preset length 100 vibration, temperature, and pressure buffer queues, respectively. When the length of any queue exceeds 100, the oldest signal value at the head of that queue is automatically removed to maintain a constant queue length. The buffer queues store the raw signal values ​​from the most recent 100 sampling times.

[0026] Furthermore, starting from the second sampling point (i.e., the current sampling index) The current signal value is read from the vibration buffer queue, temperature buffer queue, and pressure buffer queue, respectively. and the signal value at the previous moment Calculate the increments of each physical quantity using the following formula: For the first sampling point ( Since there is no signal value from the previous moment, incremental calculation and symbol generation operations are not performed, and the system enters a waiting state until the next sampling moment arrives. This initial skipping process ensures the boundary integrity of the incremental calculation and avoids data anomalies caused by referencing null values.

[0027] A preferred approach is to quantize the increment into a sign, which includes the following steps: Vibration increment quantization includes: a preset vibration positive dead zone threshold. (Normalized value), vibration negative dead zone threshold .like Output the ascending sign +; if < If the output is decreasing, it will output a - sign; otherwise, it will output a steady sign 0. Furthermore, the preset temperature dead zone threshold is +0.5 to -0.3, and the preset pressure dead zone threshold is +0.1 to -0.05. Temperature increment quantization and pressure increment quantization use the same comparison logic as vibration increment quantization to output temperature and pressure signs respectively.

[0028] The aforementioned dead zone threshold can be pre-calibrated based on historical data under normal operating conditions using statistical methods (such as taking twice the standard deviation of the noise amplitude) and kept fixed after system deployment. It should be noted that the asymmetric dead zone can effectively filter out small fluctuations in the signal near the equilibrium position, while retaining positive and negative changes with trend significance.

[0029] Finally, the generated vibration, temperature, and pressure symbols for the current moment are appended to the end of the vibration, temperature, and pressure symbol sequences, respectively. Then, the current length of each symbol sequence is checked: if the length of any symbol sequence exceeds a preset window length (e.g., 50), the earliest symbol at the beginning of that sequence is removed, ensuring the sequence length is strictly equal to 50. After completing these operations, the process returns to continue processing the data for the next sampling moment, forming a real-time sliding window update mechanism for the symbol sequences.

[0030] This mechanism ensures that at any given time, all three symbol sequences contain the symbol values ​​of the most recent 50 sample points.

[0031] It should be noted that the length of the buffer queue (100) is twice the length of the sliding window (50). That is, when step S3 triggers the causal graph construction once every 50 new sampling points (corresponding to 250ms), the incremental calculation in S1 needs to access the signal values ​​of the current time step and the previous time step. The buffer queue of length 100 ensures that even under high-frequency triggering, the signal value of the previous time step is still retained in the queue and will not be discarded prematurely due to the update of the sliding window. At the same time, the buffer queue of length 100 provides sufficient historical data redundancy for incremental calculation, avoiding data truncation errors caused by an excessively short queue.

[0032] Specifically, the buffer queue adopts a first-in-first-out (FIFO) structure. Each time a new sampling point is pushed in, the first element of the queue is automatically popped out, ensuring that the queue always stores the most recent 100 original signal values, decoupling it from the sliding window update of the symbol sequence.

[0033] S2: At the edge node, a probability density distribution is constructed based on the head and load values ​​in the historical normal operation data. The head-load plane is divided into multiple high-density operation partitions, and the baseline directed causal graph under normal conditions corresponding to each high-density operation partition is pre-stored.

[0034] It should be noted that in existing technologies, the operational status monitoring of hydro-generator units typically employs a unified fault discrimination model covering all operating conditions, neglecting the impact of head and load variations on the unit's dynamic characteristics. Because the causal transmission relationships between vibration, temperature, and pressure differ significantly under different head and load combinations, a single model cannot cover the entire operating condition range, resulting in high false alarm and false negative rates for fault detection. Furthermore, traditional operating condition zoning methods often use uniform rectangular grid division, failing to consider the probability density distribution of actual operating data, leading to overly detailed low-density areas and insufficient high-density areas. To address these issues, the specific operation of this invention is as follows: S2.1: Collect head values ​​during historical normal operation. (Unit: m) and load value (Unit: MW), constituting multiple sample points .

[0035] Establish the head axis in the memory of the edge node. and load shaft A two-dimensional grid is used, with the water head axis divided into cells at fixed intervals. Among them, , These are the minimum and maximum values ​​of historical water head data, respectively; Similarly, the head axis is divided into equal intervals at fixed intervals, for example, with a grid spacing of 1m, and the load axis is divided into equal intervals at a grid spacing of 1MW, forming several rectangular cells.

[0036] Count each cell Number of sample points within This statistical result constitutes the empirical probability density distribution (unnormalized) on the head-load plane.

[0037] reserve The cells designated as "high-density" are excluded; the remaining cells are considered low-density areas and will not participate in subsequent partitioning and merging. The minimum number of samples to retain for a cell is set to 50 by default.

[0038] Through the above density screening steps, operating conditions that rarely occur during long-term unit operation are filtered out, thus avoiding interference from sparse areas on the partition boundaries.

[0039] S2.2: Using the four-connectivity rule (adjacent vertically, horizontally, and vertically), adjacent retained cells are merged into connected regions. Each connected region is marked as a high-density running partition, resulting in a total of... High-density operating zones .

[0040] Specifically, starting from any unvisited reserved cell, recursively or iteratively search all reserved cells that are 4-connected to it, marking these cells as the same connected region. After traversing all reserved cells, each connected region is defined as a high-density running partition, resulting in a total of... One partition.

[0041] For each partition Calculate the lower limit of the water head Water head boundary Lower limit of load Upper limit of load Take the minimum and maximum values ​​of the corresponding axes for all cells within the partition.

[0042] It should be noted that in the actual implementation, attention needs to be paid to handling boundary cells. That is, for cells located at the edge of the head axis or load axis, if the density meets the standard but lacks four-connected neighbors, they should form a separate partition, which may contain only a single cell. In this way, the final set of partitions can preserve the local causal characteristics of high-density areas to the greatest extent, while reducing storage and computational overhead.

[0043] S2.3: For each partition Extract complete vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences corresponding to all historical normal operation times (not limited to 50 characters, but covering the entire historical time range within the partition).

[0044] Since historical data may contain multiple discontinuous time segments (such as data gaps caused by unit start-up and shutdown), each continuous segment needs to be processed separately in chronological order: Within each consecutive segment, the vibration, temperature, and pressure symbol sequences are time-aligned and then spliced ​​independently, but pattern matching across segments is strictly prohibited; that is, for any prediction time... Historical Index Only taken from the same continuous segment The previous moment.

[0045] A better approach is to use a simplified Granger causality test with rolling prediction to calculate the directed causal relationship between each pair of symbols. The specific process is as follows: Suppose the symbol sequence to be tested and The lengths are all ;in, The lag order is used in the embodiments of the present invention. .

[0046] For any time : The lag pattern is as follows: (1) Based on Self-prediction (use (Lag pattern) In sequence All indexes satisfy and The location, after counting these locations symbol frequency (i.e.) At any moment The value of the highest-frequency sign is taken as the value of the highest-frequency sign. If no conditions are met. If the result is 0, then the prediction is 0.

[0047] (2) Based on Self-prediction : In sequence All indexes satisfy and The location, after counting these locations The highest frequency of a symbol is taken as the symbol frequency. If there is no match, the prediction is 0.

[0048] Traversal arrive Statistical analysis based on Number of correct predictions (Right now Equal to actual (number of times) and based on Number of times self-prediction is correct (Right now Equal to actual (number of times); if Then it is determined that a directed edge exists, where, To tolerate offset, in this embodiment of the invention, .

[0049] For six possible combinations of directed edges Perform the above verification process separately, and compile all the directed edges that are determined to exist into a baseline directed causal graph. .

[0050] The upper and lower bounds of the head and load of each partition are associated with the corresponding baseline directed cause-effect graph and stored in the non-volatile memory of the edge node.

[0051] It should be noted that traditional methods typically calculate cross-correlation or vector autoregression models for the entire sequence, which easily introduces future information (i.e., using...). Data prediction after time point (Time). This invention limits the predicted time. The historical index is used, and only historical data within the same continuous segment is used, strictly adhering to the temporal directionality of causality testing. In practice, a prefix pattern lookup table needs to be maintained for each continuous segment: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] from Increment to The pattern matching index is dynamically updated to ensure that only previously processed historical positions are accessed during each prediction. This implementation has lower time complexity (symbol set size of 3, lag order of 3, and a maximum of 27 patterns), thus it is computationally efficient and suitable for offline execution on edge nodes.

[0052] S3: Input the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence into the edge node, determine the target zone according to the current head value and the current load value, perform a simplified Granger causality test on each symbol sequence within the target zone, and construct a real-time directed causal graph.

[0053] First, after each complete sliding window update (i.e., after collecting 50 new sampling points, corresponding to 250ms), S3 is triggered once. The trigger time is recorded as the current sampling time. Simultaneously read the water head value at that moment. and load value The head value and load value should be time-aligned with the last symbol of the sliding window.

[0054] S3.1: Read the current head value and load value traversal A high-density running partition is checked for the existence of partitions. satisfy: in, This is the lower boundary of the zone's water head. The upper limit of the waterway, This is the lower limit of the load. This represents the upper limit of the load.

[0055] If a partition exists, select the partition that meets the criteria (if multiple partitions exist, choose the one with the smallest product of head and load interval length, i.e., the more compact partition) as the target partition; if no partition exists, calculate the head boundary distance for each partition. : Distance from load boundary : In the above distance definition, if the current head value falls within the partition head interval, then Otherwise, take the distance to the nearest boundary.

[0056] choose Minimum partition, and requirements The distance is less than or equal to the first distance threshold in the direction of the water head (preset to 2m). The distance must be less than or equal to a first distance threshold in the load direction (preset to 2MW). Further selection is possible when multiple candidate zones have the same minimum combined distance. Smaller partitions (preferably head matching) or random selection are used, but this invention defaults to using the first partition scanned to ensure determinism.

[0057] If the conditions are met, the target partition is selected; otherwise, the current operating condition is determined to be in a low-density area (i.e., far from all historical high-density operating areas), and the unmatched partition is output and the current detection is skipped, or the global default baseline map is used, which can be selected according to application requirements.

[0058] S3.2: Directly read the current sliding window sequence of length n from the real-time vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence (in this example, n=50, corresponding to the trigger time). The previous 50 symbols are denoted as follows: The symbols in the window are arranged in chronological order from earliest to latest.

[0059] In the identified target partition Within, for each pair of symbol sequences (i.e. and , and , and The simplified Granger causality test for the rolling forecast in S2.3 is performed again. The difference from offline baseline plotting is that the pattern lookup source data used for forecasting is limited to each forecast time step in the current sliding window sequence. The previous part, that is, for the index within the window... At that moment, historical index Only retrieved from the same window and The subsequence. Since the window length is 50 and the lag order is 3, the number of effective predictions is 50-3=47. Tolerance offset =5 is kept the same value as in the offline phase to ensure consistency between real-time judgment and benchmark construction.

[0060] Specifically, for each pair traverse the window arrive (Corresponding to symbols 4 through 50 within the window), calculate based on the rules in S2.3. The number of correct predictions and based on The number of times a prediction is correct is used to determine whether a directed edge exists.

[0061] After performing the above judgment on each of the six possible combinations of directed edges, all the directed edges that exist are gathered into a directed graph, which is the real-time directed causal graph under the current working condition. .

[0062] Finally, With target partition The identifier is associated and stored in the memory of the edge node. If S3.1 determines that a partition does not match and the global default baseline map is not used, then no partition is generated. The current detection cycle is skipped, and the system waits for the next trigger moment.

[0063] It should be noted that, yes, this invention uses the exact same rolling prediction rules and the same tolerance offset as S2.3 during real-time construction, ensuring... and Comparability in terms of statistical scope.

[0064] S4: Compare the real-time directed cause-effect graph with the reference directed cause-effect graph corresponding to the target partition, detect the direction reversal of directed edges or the appearance and disappearance of edges, and output the fault type and abnormal directed edges.

[0065] S4.1: Read the real-time directed causal graph from the edge node memory. and according to the target partition The identifier is read from the non-volatile memory to obtain the corresponding reference directed causal graph. .

[0066] Extract separately and The set of all directed edges in and Each directed edge is denoted as ,in ;in, These are identifiers for vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences, respectively.

[0067] S4.2: Traverse all directed edges, where directed edges include: Ternary status values ​​are assigned in the following priority order: like If the value is 3 (direction reversed), then the value is assigned to 3; this state indicates that there is a state in the reference diagram from arrive The edge, while the real-time graph has the opposite direction (from arrive The edge of ) reflects the reversal of the causal direction.

[0068] Otherwise, if If the value is 1 (edge ​​disappears), this state indicates that the directed edges that exist in the baseline graph no longer appear in the real-time graph.

[0069] Otherwise, if If the value is 2 (edge ​​appears), this state indicates that a directed edge that did not exist in the baseline graph has appeared in the real-time graph.

[0070] Otherwise, assign a value of 0 (normal), indicating that the existence of the directed edge is consistent in both the baseline graph and the real-time graph and there is no direction reversal.

[0071] The above priority order ensures that the direction reversal state can be identified first, because direction reversal may simultaneously satisfy the conditions of edge disappearance and edge appearance (e.g., the baseline graph has...). But no Real-time images are available. But no If we first determine whether something has disappeared or appeared, we will lose the reversal information. By prioritizing the reversal judgment, we can accurately capture changes in the causal direction.

[0072] S4.3: Arrange the directed edges of anomalies with state values ​​of 1, 2, or 3 in order (e.g., lexicographical order). The fault code is constructed by concatenating the states (or other predefined sequences) into an edge identifier: a sequence of state values. For example, in this embodiment of the invention, six directed edges are traversed in a fixed order, with the state value of each edge represented by a number (0, 1, 2, 3), which can be compressed into a 6-bit ternary number or directly output as a comma-separated string. For example, in sequence... , If the status values ​​are [0,1,0,2,3,0], then fault codes 0,1,0,2,3,0 or V→P:1; T→P:2; P→V:3 can be output. This invention recommends outputting natural language descriptions to reduce the difficulty of interpretation for maintenance personnel.

[0073] Simultaneously, output the corresponding natural language description: state value 1 corresponds to disappearance, state value 2 corresponds to appearance, and state value 3 corresponds to reversal. If there are no abnormal directed edges (all state values ​​are 0), output a normal or empty fault code.

[0074] like Figure 3As shown, the present invention also provides a remote diagnostic system for hydropower faults based on multimodal feature fusion, including a symbol increment encoding module, which collects unit vibration signals, temperature signals and pressure signals, calculates the increment of each signal relative to the previous time, and quantizes the increment into symbols to obtain vibration symbol sequences, temperature symbol sequences and pressure symbol sequences; The partitioned causal graph library module constructs a probability density distribution based on the head and load values ​​in historical normal operation data at the edge nodes, divides the head-load plane into multiple high-density operation partitions, and pre-stores the baseline directed causal graph under normal conditions corresponding to each high-density operation partition. The real-time causal construction module inputs the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence into the edge node, determines the target zone based on the current head value and current load value, performs a simplified Granger causality test on each symbol sequence within the target zone, and constructs a real-time directed causal graph. The cause-effect graph comparison and diagnosis module compares the real-time directed cause-effect graph with the reference directed cause-effect graph corresponding to the target partition, detects the reversal of the direction of directed edges or the appearance and disappearance of edges, and outputs the fault type and abnormal directed edges.

[0075] The system also includes one or more processors and memory.

[0076] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of a remote hydropower fault diagnosis method based on multimodal feature fusion as described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0077] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of a remote hydropower fault diagnosis method based on multimodal feature fusion according to the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0078] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0079] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0080] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0081] In any case, the language can be either compiled or interpreted.

[0082] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0083] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0084] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0085] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0086] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0087] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0088] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A remote fault diagnosis method for hydropower based on multimodal feature fusion, characterized in that: include: The vibration, temperature, and pressure signals of the unit are collected. The increment of each signal relative to the previous moment is calculated, and the increment is quantized into a symbol to obtain the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence. At the edge nodes, a probability density distribution is constructed based on the head and load values ​​in the historical normal operation data. The head-load plane is divided into multiple high-density operation partitions, and a baseline directed causal graph under normal conditions corresponding to each high-density operation partition is pre-stored. The vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence are input into the edge node. The target zone to which the node belongs is determined based on the current head value and the current load value. A simplified Granger causality test is performed on each symbol sequence within the target zone to construct a real-time directed causal graph. The real-time directed causal graph is compared with the reference directed causal graph corresponding to the target partition. The direction reversal of directed edges or the appearance and disappearance of edges are detected, and the fault type and abnormal directed edges are output.

2. The remote fault diagnosis method for hydropower based on multimodal feature fusion as described in claim 1, characterized in that: The construction probability density distribution includes: Collect historical head values ​​during normal operation and load value Multiple sample points Establish the head axis in the memory of the edge node. and load shaft A two-dimensional grid, with the water head axis divided into cells at fixed intervals; Count the number of sample points in each cell ,reserve The cells, where, The minimum number of samples to retain for a cell; The division into multiple high-density operating partitions includes: The four-connectivity rule is used to merge adjacent retained cells into connected regions. Each connected region is marked as a high-density running partition, resulting in a total of High-density operating zones .

3. The remote diagnosis method for hydropower faults based on multimodal feature fusion as described in claim 2, characterized in that: The generation of the baseline directed causal graph includes: For each partition The system extracts complete vibration, temperature, and pressure symbol sequences corresponding to all historical normal operating times, and processes each consecutive segment separately in chronological order. Rolling prediction is performed within each segment, without matching across segments. A simplified Granger causality test method based on rolling prediction is used to calculate the directed causal relationship between each symbol pair. Suppose the symbol sequence to be tested and The lengths are all ;in, The lag order; For any time : In sequence All indexes satisfy and The location, after counting these locations The symbol frequency is used to determine the highest frequency symbol. If no conditions are met. If , then the prediction is 0; In sequence All indexes satisfy and The location, after counting these locations The highest frequency of a symbol is taken as the symbol frequency. If there is no match, the prediction is 0. Traversal arrive Statistical analysis based on Number of correct predictions and based on Number of times self-prediction is correct ;like Then it is determined that a directed edge exists, where, Tolerance offset; Repeat the testing process for each distinct pair of symbolic variables, and compile all the directed edges that are determined to exist into a baseline directed causal graph. .

4. The remote diagnosis method for hydropower faults based on multimodal feature fusion as described in claim 3, characterized in that: The target partition to which the determination belongs includes: Read the current head value and load value traversal A high-density running partition is checked for the existence of partitions. satisfy: in, This is the lower boundary of the zone's water head. The upper limit of the waterway, This is the lower limit of the load. This is the upper limit of the load. If it exists, select the partition that meets the criteria as the target partition; if it does not exist, calculate the head boundary distance for each partition. : Distance from load boundary : choose Minimum partition, and requirements Less than or equal to the first distance threshold in the direction of water head. If the distance is less than or equal to the first distance threshold in the load direction, it is selected as the target partition; otherwise, the current operating condition is determined to be in a low-density area, an unmatched partition is output, and the current detection is skipped.

5. The remote diagnosis method for hydropower faults based on multimodal feature fusion as described in claim 4, characterized in that: The construction of the real-time directed causal graph includes: Read the current sliding window sequence of length n directly from the real-time vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence; In the target partition Within this process, the simplified Granger causality test for rolling prediction is performed again for each pair of symbol sequences, but the pattern lookup source data used for prediction is limited to each prediction time in the current sliding window sequence. The previous part; Among them, tolerance offset Keep it the same; All determined directed edges are aggregated into a real-time directed causal graph. and will and target partition The identifiers are associated and stored in the memory of the edge nodes.

6. The remote fault diagnosis method for hydropower based on multimodal feature fusion as described in claim 5, characterized in that: The comparison includes: The real-time directed causal graph is read from the memory of the edge node. and according to the target partition The identifier is read from the non-volatile memory to obtain the corresponding reference directed causal graph. ; Extract separately and The set of all directed edges in and Each directed edge is denoted as ,in ;in, These are identifiers for vibration symbol sequences, temperature symbol sequences, and pressure symbol sequences, respectively.

7. The remote diagnosis method for hydropower faults based on multimodal feature fusion as described in claim 6, characterized in that: The comparison also includes: Traverse all directed edges; The directed edges include: , assign ternary status values ​​in order of priority; Concatenate the directed edges of the anomalies with state values ​​of 1, 2 or 3 in sequence to form a fault code, and output the fault code and the corresponding directed edge description: state value 1 corresponds to disappearance, state value 2 corresponds to appearance, and state value 3 corresponds to reversal.

8. The remote fault diagnosis method for hydropower based on multimodal feature fusion as described in claim 7, characterized in that: The assignment of the ternary state value includes: like If so, then the value is assigned to 3; Otherwise, if If , then the value is assigned to 1; Otherwise, if If , then the value is assigned to 2; Otherwise, assign a value of 0.

9. A remote hydropower fault diagnosis system based on multimodal feature fusion, based on the remote hydropower fault diagnosis method based on multimodal feature fusion as described in any one of claims 1 to 8, characterized in that: Also includes: The symbol increment encoding module collects vibration, temperature and pressure signals from the unit, calculates the increment of each signal relative to the previous moment, and quantizes the increment into symbols to obtain vibration symbol sequences, temperature symbol sequences and pressure symbol sequences. The partitioned causal graph library module constructs a probability density distribution based on the head and load values ​​in historical normal operation data at the edge nodes, divides the head-load plane into multiple high-density operation partitions, and pre-stores the baseline directed causal graph under normal conditions corresponding to each high-density operation partition. The real-time causal construction module inputs the vibration symbol sequence, temperature symbol sequence, and pressure symbol sequence into the edge node, determines the target zone based on the current head value and current load value, performs a simplified Granger causality test on each symbol sequence within the target zone, and constructs a real-time directed causal graph. The cause-effect graph comparison and diagnosis module compares the real-time directed cause-effect graph with the reference directed cause-effect graph corresponding to the target partition, detects the reversal of the direction of directed edges or the appearance and disappearance of edges, and outputs the fault type and abnormal directed edges.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the remote diagnosis method for hydropower faults based on multimodal feature fusion as described in any one of claims 1 to 8.

Images