Fault diagnosis method for cooling system of high-voltage direct-current converter valve

Abstract

The invention discloses a fault diagnosis method for a cooling system of a high-voltage direct-current converter valve, which relates to the technical field of detection of the cooling system of the high-voltage direct-current converter valve, and breaks through the judgment limitation of a single parameter threshold by synchronously acquiring parameters such as temperature, flow and pressure by adopting a sensor group. Accidental noise or local disturbance is effectively filtered through correlation analysis and cross validation between parameters, and false alarms are reduced; a secondary fusion model constructed based on multi-source data can deeply learn a parameter dynamic balance relationship under a normal working condition, and an alarm is triggered only during multi-parameter collaborative deviation, so that the fault judgment accuracy is greatly improved; the model can capture complex nonlinear coupling characteristics among parameters, automatically adapt to dynamic working conditions such as power fluctuation and environment temperature change, dynamically track system state evolution through real-time data input, sensitively recognize transient impact caused by load abrupt change and gradual deterioration caused by scaling, and improve the stability of the system. The problem that a traditional method is insufficient in non-linear change adaptive capacity is solved, and the complex adaptability is improved.

Description

Technical Field

[0001] This invention relates to the field of high-voltage DC converter valve cooling system testing technology, and in particular to a fault diagnosis method for high-voltage DC converter valve cooling systems. Background Technology

[0002] High-voltage direct current converter valves are core equipment in high-voltage direct current transmission systems. Their operational stability directly determines the power conversion efficiency and grid security. As a key auxiliary system of the converter valve, the cooling system removes the heat generated during valve operation through circulating cooling medium. Abnormal operating conditions may lead to overheating and damage of the converter valve or even system shutdown. Therefore, real-time and accurate fault diagnosis of the high-voltage direct current converter valve cooling system has significant engineering value.

[0003] In the prior art, the fault diagnosis methods of the converter valve cooling system mainly include the following types: (1) Threshold comparison method: alarms are triggered by setting a fixed threshold for a single parameter (such as temperature or pressure); this method is simple to implement, but lacks analysis of the correlation characteristics of multiple parameters, is easily affected by instantaneous interference, and has a high false alarm rate; (2) Statistical model method: a parameter statistical distribution model is established based on historical data, and the fault is judged by deviation analysis; this type of method has high requirements for data stability and is difficult to adapt to the nonlinear changes under the dynamic working conditions of the cooling system. Summary of the Invention

[0004] In view of this, the present invention proposes a fault diagnosis method for a high-voltage DC converter valve cooling system, which can effectively solve the problems of high false alarm rate and difficulty in adapting to nonlinear changes in the dynamic operating conditions of the cooling system in the existing technology.

[0005] The technical solution of this invention is implemented as follows:

[0006] A fault diagnosis method for a high-voltage DC converter valve cooling system includes:

[0007] Construct a two-level fusion model;

[0008] Real-time parameter data of the high-voltage DC converter valve cooling system are acquired based on a sensor array.

[0009] The two-level fusion model calculates the comprehensive state assessment value of the high-voltage DC converter valve cooling system based on real-time parameter data;

[0010] The fault level is determined based on the comprehensive condition assessment value, and fault handling suggestions are output based on the fault level determination.

[0011] As a further optional solution to the fault diagnosis method for the high-voltage DC converter valve cooling system, the acquisition of real-time parameter data of the high-voltage DC converter valve cooling system based on the sensor group specifically includes:

[0012] The sensor array collects pressure, flow, liquid level, and temperature data of the high-voltage DC converter valve cooling system, and simultaneously collects the operating status signals of each sensor.

[0013] The 3σ criterion is used to identify outliers in pressure, flow, level, and temperature data. Based on the sensor's operating status signal, the outlier data is corrected to obtain the corrected parameter data.

[0014] Based on the corrected parameter data, data filtering and standardization are performed to obtain standardized parameter data;

[0015] The final parameter data is obtained by extracting time-domain and statistical features based on standardized parameter data.

[0016] As a further optional solution to the fault diagnosis method for the high-voltage DC converter valve cooling system, the step of correcting abnormal data based on the sensor's operating status signal to obtain corrected parameter data specifically includes:

[0017] When the sensor's operating status signal is normal, linear interpolation is used to correct abnormal data;

[0018] When the sensor's operating status signal is abnormal, the abnormal data is marked as missing data, and the K-nearest neighbor interpolation method based on historical similar operating conditions is used to supplement it.

[0019] As a further optional solution to the fault diagnosis method for the high-voltage DC converter valve cooling system, the step of performing data filtering and standardization processing based on the corrected parameter data to obtain standardized parameter data specifically includes:

[0020] The corrected parameter data is filtered using a filter combination algorithm, which includes a Kalman filter algorithm and a wavelet threshold filter algorithm. The Kalman filter algorithm is used to suppress low-frequency drift error, and the wavelet threshold filter algorithm is used to eliminate high-frequency vibration noise.

[0021] The filtered parameter data is standardized using the min-max standardization formula to obtain standardized parameter data.

[0022] As a further optional scheme of the fault diagnosis method for the high voltage DC converter valve cooling system, the first level of the two-level fusion model is a feature extraction layer, which uses a parallel structure of convolutional neural network and gated recurrent unit to extract the spatial and temporal features of the data. The second level of the two-level fusion model is a fusion decision layer, which concatenates the feature vectors output by the convolutional neural network and the gated recurrent unit and inputs them into an attention mechanism. Key features are highlighted through attention weight allocation. Finally, a comprehensive state evaluation value is output through a fully connected layer.

[0023] As a further optional solution to the fault diagnosis method for the high-voltage DC converter valve cooling system, the step of determining the fault level based on the comprehensive state assessment value and outputting fault handling suggestions based on the fault level determination specifically includes:

[0024] Based on the comprehensive status evaluation value output by the two-level fusion model, the status of the high-voltage DC converter valve cooling system is classified into levels, including excellent, good, normal, warning, and fault.

[0025] When the level is classified as a warning or a fault level, the fault diagnosis unit combines the parameter deviation coefficient and attention weight to locate the main fault source and automatically generate corresponding fault handling suggestions.

[0026] The fault location results and fault handling suggestions are transmitted to the main control unit of the cooling high-voltage DC converter valve cooling system, and a local audible and visual warning is triggered.

[0027] A fault diagnosis system for a high-voltage DC converter valve cooling system includes:

[0028] The two-level fusion model construction module is used to build the two-level fusion model;

[0029] The real-time parameter acquisition module is used to acquire real-time parameter data of the high-voltage DC converter valve cooling system based on the sensor group;

[0030] The comprehensive state assessment value calculation module is used to calculate the comprehensive state assessment value of the high voltage DC converter valve cooling system based on real-time parameter data in the secondary fusion model.

[0031] The fault diagnosis and decision-making module is used to determine the fault level based on the comprehensive status assessment value and output fault handling suggestions based on the fault level determination.

[0032] As a further optional solution for the fault diagnosis system of the high voltage DC converter valve cooling system, the first level of the two-level fusion model is a feature extraction layer, which uses a parallel structure of convolutional neural network and gated recurrent unit to extract the spatial and temporal features of the data. The second level of the two-level fusion model is a fusion decision layer, which concatenates the feature vectors output by the convolutional neural network and the gated recurrent unit and inputs them into an attention mechanism. Key features are highlighted through attention weight allocation. Finally, a comprehensive state evaluation value is output through a fully connected layer.

[0033] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the fault diagnosis method for the high-voltage DC converter valve cooling system described above.

[0034] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the fault diagnosis method for the high-voltage DC converter valve cooling system described above.

[0035] The beneficial effects of this invention are as follows: By collecting multi-dimensional parameter data such as temperature, flow rate, and pressure through a sensor array, it overcomes the limitations of single-parameter threshold judgment. Correlation analysis and cross-validation between different parameters can effectively filter false anomalies caused by occasional noise or local disturbances, reducing false triggers. The two-level fusion model built based on multi-source data can learn the dynamic balance relationship between parameters under normal operating conditions, triggering an alarm only when multiple parameters deviate from the baseline mode, significantly improving the reliability of fault judgment. The two-level fusion model can capture the complex nonlinear coupling relationship between cooling system parameters, automatically learn the system behavior characteristics under dynamic operating conditions such as power fluctuations and ambient temperature changes. By continuously inputting real-time parameter data, the model can dynamically track the evolution trajectory of the system state, and has sensitive identification capabilities for transient impacts (such as sudden load changes in the converter valve) and gradual degradation (such as cooling efficiency reduction caused by scaling), effectively solving the problem that existing technologies are difficult to adapt to nonlinear changes under dynamic operating conditions of cooling systems, and improving complexity adaptability. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.

[0037] Figure 1 This is a flowchart illustrating a fault diagnosis method for a high-voltage DC converter valve cooling system, as shown in the embodiment.

[0038] Figure 2 This is a schematic diagram illustrating the composition of a fault diagnosis system for a high-voltage DC converter valve cooling system, as shown in the embodiment.

[0039] Figure 3 This is a schematic diagram of the composition of a computing device according to an embodiment. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0041] refer to Figures 1 to 3A fault diagnosis method for a high-voltage DC converter valve cooling system, comprising:

[0042] A two-level fusion model is constructed, wherein the two-level fusion model is used to calculate the comprehensive state evaluation value of the high-voltage DC converter valve cooling system; in some embodiments, the first level of the two-level fusion model is a feature extraction layer, which uses a parallel structure of convolutional neural network and gated recurrent unit to extract the spatial and temporal features of the data; the second level of the two-level fusion model is a fusion decision layer, which concatenates the feature vectors output by the convolutional neural network and the gated recurrent unit and inputs them into an attention mechanism, highlights key features through attention weight allocation, and finally outputs the comprehensive state evaluation value through a fully connected layer.

[0043] Specifically, the convolutional neural network contains two convolutional layers (3×3 kernels, 16 and 32 respectively) and one pooling layer (2×2 pooling size) to extract spatial features of the data (such as coupling and correlation features between parameters); the gated recurrent unit contains two hidden layers (64 and 32 neurons respectively) to extract temporal features of the data (such as temperature trends over time); the second level is the fusion decision layer, which concatenates the feature vectors output by the convolutional neural network and the gated recurrent unit and inputs them into the Attention mechanism. Key features are highlighted through attention weight allocation (such as pressure mutation features under fault conditions). Finally, a 1-dimensional comprehensive state evaluation value S of the cooling system (0-1, the closer to 1, the better the state) is output through a fully connected layer.

[0044] Training sample preparation: A cooling system simulation experimental platform was built to simulate 20 typical operating conditions, including normal operation, excessive pressure (≥2.0MPa), excessive flow (≤5m³ / h), insufficient liquid level (≤100mm), excessive temperature (≥60℃), and multiple fault coupling. 1000 sets of data were collected for each condition, and processed to form 20,000 samples, which were divided into a training set (14,000 sets), a validation set (4,000 sets), and a test set (2,000 sets) in a 7:2:1 ratio. Simultaneously, a corresponding comprehensive state evaluation value S was assigned to each sample (calculated by combining expert scoring with system efficiency; S=0.9-1.0 for normal operation, S=0.5-0.8 for a single fault, and S=0.1-0.4 for multiple fault coupling).

[0045] Model training and optimization: The Adam optimizer was used to train the model, with an initial learning rate of 0.001. A learning rate decay strategy was adopted (decaying to 0.9 of the previous round every 100 rounds). The mean squared error (MSE) loss function was used: Loss = (1 / N)×Σ(S - S)², where N is the number of samples, S is the model prediction value, and S is the labeled value. An early stopping strategy was adopted during training (training was stopped if the validation set loss did not decrease for 20 consecutive rounds) to avoid overfitting. After training, the model parameters were stored in the Flash memory of the neural network computing unit.

[0046] It should be noted that convolutional neural networks effectively extract spatial distribution features (such as correlation patterns of multi-sensor parameters and regional anomaly propagation paths) from parameter data through local perception and weight sharing mechanisms, revealing potential fault correlations in the physical topology of the cooling system; gated recurrent units use gating mechanisms to capture the temporal dependencies of parameters (such as the hysteresis effect of temperature changes and the periodicity of flow fluctuations), accurately modeling the dynamic evolution of the system; and parallel structures simultaneously retain spatial neighborhood features and temporal series memory, solving the problem of insufficient spatiotemporal correlation modeling capabilities of single models.

[0047] After concatenating the feature vectors output by CNN and GRU, the importance weights of each feature dimension are calculated through an attention mechanism. The system dynamically focuses on the features that have the greatest impact on the current state assessment (such as the temporal features of a sudden temperature rise in the early stage of a fault or the spatial features of local pressure anomalies). This avoids the limitations of fixed weight allocation and can still highlight key features under complex operating conditions (such as load changes and multiple fault coupling), thereby improving the accuracy of the assessment value in representing the real system state.

[0048] The first level of parallel extraction involves independent computation of CNN and GRU, fully utilizing the parallelism of computing resources to shorten the feature extraction time. The second level of lightweight fusion ensures that the computational complexity of the attention mechanism and fully connected layers is controllable, compressing the time of a single evaluation to the millisecond level (e.g., ≤10ms), meeting the real-time requirements of fault diagnosis in high-voltage DC systems. While ensuring the ability to represent high-dimensional features, redundant computations are reduced through hierarchical computation and attention filtering, achieving a synergistic optimization of "high precision and low latency".

[0049] Real-time parameter data of the high-voltage DC converter valve cooling system are acquired based on a sensor array, specifically including:

[0050] A sensor array compatible with multiple types of sensors, including pressure (0-2.5MPa), flow rate (0-50m³ / h), liquid level (0-500mm), and temperature (0-100℃), is used to synchronously acquire multiple core parameter data of the high-pressure DC converter valve cooling system at a frequency of 10Hz, including pressure P, flow rate Q, liquid level H, and temperature T. At the same time, the operating status signals of each sensor (such as power supply voltage and self-test signal) are also acquired.

[0051] The 3σ criterion is used to identify outliers in pressure, flow, level, and temperature data. The mean μ and standard deviation σ of each parameter are calculated. If the data value exceeds the range of [μ-3σ, μ+3σ], it is determined to be an outlier. Based on the sensor's working status signal, the outlier data is corrected to obtain the corrected parameter data.

[0052] Based on the corrected parameter data, data filtering and standardization are performed to obtain standardized parameter data;

[0053] The final parameter data is obtained by extracting time-domain and statistical features based on standardized parameter data.

[0054] Specifically, it synchronously collects core parameters such as pressure, flow rate, liquid level, and temperature, and acquires sensor operating status signals (such as power supply voltage and self-test status) in real time. Through multi-parameter joint monitoring, it comprehensively captures the operating status of the cooling system, avoids misjudgment of a single parameter, and the synchronous acquisition of sensor status signals can identify sensor faults (such as abnormal voltage and self-test failure) in real time, avoiding data distortion caused by sensor failure and improving the reliability of the data source.

[0055] The 3σ statistical criterion is used to identify abnormal values ​​of parameters, and the abnormal data is corrected by combining the sensor's working status signal (such as checking whether the sensor is faulty). The 3σ criterion effectively eliminates abnormal values ​​caused by random interference or instantaneous fluctuations of the sensor, reducing false diagnosis. Combined with the correction mechanism of the sensor status signal, it can distinguish between real system abnormalities and sensor faults, avoid false alarms caused by sensor problems, and significantly reduce the false alarm rate.

[0056] By extracting time-domain features (such as mean, variance, and peak value) and statistical features (such as differential entropy and kurtosis) from standardized data, a high-dimensional feature vector is constructed. The time-domain features capture the dynamic changes of parameters (such as the slope feature of a sudden temperature rise), while the statistical features reflect the distribution characteristics of the data (such as the probability density of flow fluctuations). The combination of the two can comprehensively characterize the system's operating status, and the high-dimensional feature vector provides richer fault information, supporting the accurate identification of fault modes under complex operating conditions by two-level fusion models (such as CNN-GRU-Attention).

[0057] In some embodiments, the step of correcting abnormal data based on the sensor's operating status signal to obtain corrected parameter data specifically includes:

[0058] When the sensor's operating status signal is normal, linear interpolation is used to correct abnormal data;

[0059] When the sensor's operating status signal is abnormal, the abnormal data is marked as missing data, and the K-nearest neighbor interpolation method based on historical similar operating conditions is used to supplement it.

[0060] Specifically, by monitoring the sensor's operating status signals in real time (such as power supply voltage and self-test status), it is possible to clearly distinguish whether the data anomaly is due to a fault in the sensor itself or a change in the actual state of the cooling system. If the sensor is in an abnormal state, the data source is determined to be unreliable, thus avoiding misjudging sensor faults as system faults and reducing false alarms. If the sensor is in a normal state, the data anomaly reflects the actual fluctuations in the system and needs to be corrected to retain valid information. This differentiation mechanism significantly improves the targeting of data correction and avoids information loss or misprocessing caused by a "one-size-fits-all" approach.

[0061] When a sensor is functioning normally but its data momentarily becomes abnormal, the outlier is filled in by linear interpolation using data from adjacent time points. This method is simple to calculate, highly real-time, and suitable for brief disturbances (such as instantaneous numerical jumps caused by electromagnetic pulses). It eliminates noise while preserving the data trend, ensuring the continuity of the input data. When a sensor malfunctions, the data is marked as missing, and a replacement value is generated by weighted averaging the K nearest neighbor samples of similar operating conditions in the historical database. The missing data is then filled in using historical data from similar operating conditions, ensuring the consistency between the supplemented data and the actual operating mode of the system. This avoids the bias introduced by simple filling (such as mean filling) and improves the accuracy of the data.

[0062] In some embodiments, the step of performing data filtering and standardization based on the corrected parameter data to obtain standardized parameter data specifically includes:

[0063] The corrected parameter data is filtered using a filter combination algorithm, which includes a Kalman filter algorithm and a wavelet threshold filter algorithm. The Kalman filter algorithm is used to suppress low-frequency drift error, and the wavelet threshold filter uses a db4 wavelet basis and a 3-level decomposition to eliminate high-frequency vibration noise.

[0064] The filtered parameter data is mapped to the [0,1] interval using the min-max normalization formula to obtain standardized parameter data.

[0065] Specifically, Kalman filtering suppresses low-frequency drift: through dynamic system modeling and recursive estimation, it corrects low-frequency errors caused by environmental temperature drift, aging, etc. in real time, effectively eliminating the slow drift of parameters such as pressure and temperature (such as the deviation of temperature sensors caused by the cumulative thermal effect of the environment), ensuring long-term data stability, and avoiding misdiagnosis caused by low-frequency noise.

[0066] Wavelet threshold filtering eliminates high-frequency interference: By utilizing the multi-scale analysis capability of wavelet transform, high-frequency noise (such as electromagnetic interference and instantaneous spikes caused by mechanical vibration) is thresholded and reduced, while retaining the effective features of parameter abrupt changes (such as sudden pressure rise caused by faults) and eliminating random high-frequency noise, thereby improving the reliability of transient data changes.

[0067] The two-level fusion model calculates the comprehensive state assessment value of the high-voltage DC converter valve cooling system based on real-time parameter data.

[0068] Specifically, model loading and data input: After the neural network computing unit is powered on, it automatically loads the trained two-level fusion model; the data preprocessing unit inputs the real-time processed 15-dimensional feature vector into the model at a frequency of 10Hz.

[0069] Real-time fusion computing: The FPGA chip adopts a pipelined parallel computing architecture to perform parallel convolution and loop calculations on the input feature vector. The Attention mechanism dynamically assigns weights (weight range 0-1) according to the importance of features. The fully connected layer outputs a real-time comprehensive state evaluation value S.

[0070] Based on the comprehensive condition assessment value, the fault level is determined, and fault handling suggestions are output based on the fault level determination, specifically including:

[0071] Based on the comprehensive status evaluation value output by the two-level fusion model, the status of the high-voltage DC converter valve cooling system is classified into levels, including excellent, good, normal, warning, and fault.

[0072] When the level is classified as a warning or a fault level, the fault diagnosis unit combines the parameter deviation coefficient and attention weight to locate the main fault source and automatically generate corresponding fault handling suggestions.

[0073] The fault location results and fault handling suggestions are transmitted to the main control unit of the cooling high-voltage DC converter valve cooling system, and a local audible and visual warning is triggered.

[0074] Specifically, the real-time comprehensive status assessment value S is divided into 5 levels: excellent (S≥0.9), good (0.7≤S<0.9), normal (0.5≤S<0.7), warning (0.3≤S<0.5), and fault (S<0.3), and a status assessment report is generated, which includes the original values, standardized values ​​and key feature weights of each parameter;

[0075] Fault localization: When the status level is warning or fault, the fault diagnosis unit calculates the deviation coefficient C of each parameter, and calculates the fault contribution F=C×w by combining the feature weight w output by the Attention mechanism; the parameter with the largest F is the main fault source. If multiple parameters F>0.3, it is determined to be multiple fault coupling.

[0076] Anomaly warning output: The fault diagnosis unit transmits the fault location results and status assessment report to the cooling system main control unit via Ethernet, and simultaneously triggers local audible and visual warnings; if the fault level is determined, it automatically outputs fault handling suggestions (such as checking the pressure relief valve if the pressure is too high);

[0077] Model adaptive update: 1000 sets of normal working condition data are automatically extracted each month and processed as incremental training samples; the incremental learning algorithm is used to fine-tune the model, only updating the weight parameters of the fully connected layer and the Attention mechanism to avoid fusion interruption caused by full retraining; if the loss of the model test set decreases by ≥10% after incremental training, the new parameters are fixed.

[0078] It should be noted that, based on the comprehensive status assessment value (S), the system status is divided into five levels: excellent, good, normal, warning, and fault, establishing a continuous status spectrum from "healthy" to "failure". Different levels correspond to different handling procedures (such as triggering inspections at the warning level and initiating emergency shutdowns at the fault level), avoiding the economic losses caused by "one-size-fits-all" shutdowns. Maintenance personnel can quickly identify the degree of system degradation through level labels, shortening the fault assessment time (e.g., the response time for fault level can be shortened to the minute level).

[0079] Quantify the deviation of the actual values ​​of each parameter from the normal range, identify abnormal parameters, reuse the feature importance score in the secondary fusion model, locate key fault-related parameters, the parameter deviation coefficient identifies anomalies at the data level, and the attention weight reveals the fault correlation at the model level. The combination of the two can cross-validate the fault source (e.g., if the pressure parameter deviation coefficient is high in a certain area and the attention weight is concentrated, then it is determined that the area is blocked), avoiding misjudgment caused by a single method (e.g., relying only on threshold comparison).

[0080] Based on the fault level and location results, the system automatically matches handling suggestions (such as adjusting operating parameters and isolating faulty equipment) from the pre-set rule base. The handling suggestions are directly output to the main control unit, reducing reliance on expert experience. This is especially suitable for remote areas or unattended sites, ensuring that similar faults are handled in a unified manner and avoiding secondary risks caused by differences in human operation.

[0081] The fault location results and handling suggestions are transmitted to the main control unit, and the local audible and visual alarm device is triggered simultaneously. The audible and visual warning is triggered immediately after the diagnosis is completed, reminding on-site personnel to intervene quickly (e.g., the fault-level warning response time is ≤3 seconds). Timely warnings prevent the fault from spreading (e.g., if pipeline leaks are not dealt with in time, it may cause the converter valve to overheat and shut down).

[0082] A fault diagnosis system for a high-voltage DC converter valve cooling system includes:

[0083] The two-level fusion model construction module is used to build the two-level fusion model;

[0084] The real-time parameter acquisition module is used to acquire real-time parameter data of the high-voltage DC converter valve cooling system based on the sensor group;

[0085] The comprehensive state assessment value calculation module is used to calculate the comprehensive state assessment value of the high voltage DC converter valve cooling system based on real-time parameter data in the secondary fusion model.

[0086] The fault diagnosis and decision-making module is used to determine the fault level based on the comprehensive status assessment value and output fault handling suggestions based on the fault level determination.

[0087] In some embodiments, the first level of the two-level fusion model is a feature extraction layer, which uses a parallel structure of convolutional neural network and gated recurrent unit to extract spatial and temporal features of the data. The second level of the two-level fusion model is a fusion decision layer, which concatenates the feature vectors output by the convolutional neural network and the gated recurrent unit and inputs them into an attention mechanism. Key features are highlighted through attention weight allocation, and finally, a comprehensive state evaluation value is output through a fully connected layer.

[0088] A computing device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the fault diagnosis method for the high-voltage DC converter valve cooling system described above.

[0089] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the fault diagnosis method for the high-voltage DC converter valve cooling system described above.

[0090] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of diagnosing a fault in a high voltage direct current converter valve cooling system, characterized by, The application relates to a high-voltage direct-current converter valve cooling system fault diagnosis method and device. The application comprises the following steps: a secondary fusion model is constructed; real-time parameter data of a high-voltage direct-current converter valve cooling system is acquired based on a sensor group; a comprehensive state evaluation value of the high-voltage direct-current converter valve cooling system is calculated based on the real-time parameter data by the secondary fusion model; 2. The method of claim 1, wherein a fault level is determined according to the comprehensive state evaluation value, and a fault treatment suggestion is output based on the fault level determination. The real-time parameter data of the high-voltage direct-current converter valve cooling system is acquired based on the sensor group, and the method comprises the following steps: pressure data, flow data, liquid level data and temperature data of the high-voltage direct-current converter valve cooling system are collected based on the sensor group, and working state signals of the sensors are synchronously collected; abnormal values of the pressure data, the flow data, the liquid level data and the temperature data are identified by adopting a 3sigma criterion, and the abnormal data is corrected based on the working state signals of the sensors to obtain corrected parameter data; data filtering and standardization processing are performed according to the corrected parameter data to obtain standardized parameter data; 3. The method according to claim 2, characterized in that time domain features and statistical features are extracted based on the standardized parameter data to obtain final parameter data. The abnormal data is corrected based on the working state signals of the sensors to obtain the corrected parameter data, and the method comprises the following steps: when the working state signals of the sensors are normal, the abnormal data is corrected by adopting a linear interpolation method; 4. The method of claim 2, wherein when the working state signals of the sensors are abnormal, the abnormal data is marked as data missing, and is supplemented by adopting a K-nearest neighbor interpolation method based on historical similar working conditions. The data filtering and standardization processing are performed according to the corrected parameter data to obtain the standardized parameter data, and the method comprises the following steps: the corrected parameter data is filtered by adopting a filter combination algorithm, wherein the filter combination algorithm comprises a Kalman filter algorithm and a wavelet threshold filter algorithm, the Kalman filter algorithm is used for inhibiting low-frequency drift error, and the wavelet threshold filter algorithm is used for eliminating high-frequency vibration noise; 5. The method of claim 1, wherein, the filtered parameter data is standardized by a min-max standardization formula to obtain the standardized parameter data.

6. The method of claim 1, wherein, The first level of the secondary fusion model is a feature extraction layer, spatial features and time sequence features of data are extracted by adopting a convolutional neural network and a gated recurrent unit parallel structure, the second level of the secondary fusion model is a fusion decision layer, feature vectors output by the convolutional neural network and the gated recurrent unit are spliced and then input into an attention mechanism, key features are highlighted by attention weight distribution, and finally a comprehensive state evaluation value is output by a full connection layer. The fault level is determined according to the comprehensive state evaluation value, and the fault treatment suggestion is output based on the fault level determination, and the method comprises the following steps: the state of the high-voltage direct-current converter valve cooling system is divided into levels according to the comprehensive state evaluation value output by the secondary fusion model, and the levels comprise excellent, good, normal, early warning and fault; when the level division is the early warning or fault level, a fault diagnosis unit locates a main fault source in combination with a parameter deviation coefficient and an attention weight, and automatically generates a corresponding fault treatment suggestion; 7. A high voltage direct current converter valve cooling system fault diagnostic system characterized by, the fault positioning result and the fault treatment suggestion are transmitted to a main control unit of the high-voltage direct-current converter valve cooling system, and a local sound-light early warning is triggered. The application comprises the following steps: a secondary fusion model is constructed; The real-time parameter acquisition module is configured to acquire real-time parameter data of the HVDC converter valve cooling system based on the sensor group. The comprehensive state evaluation value calculation module is configured to calculate a comprehensive state evaluation value of the HVDC converter valve cooling system based on the real-time parameter data. The fault diagnosis decision module is configured to determine a fault level based on the comprehensive state evaluation value and output a fault processing suggestion based on the fault level determination.

8. The high voltage DC power valve cooling system fault diagnostic system of claim 7, wherein, The first level of the secondary fusion model is a feature extraction layer, which adopts a parallel structure of a convolutional neural network and a gated recurrent unit to extract spatial features and time sequence features of data.

9. A computing device, comprising: The computer program stored in the memory and executable on the processor implements the steps of the HVDC converter valve cooling system fault diagnosis method of any one of claims 1-6 when the processor executes the computer program.

10. A computer-readable storage medium, characterized in that, The computer program stored in the memory and executable on the processor implements the steps of the HVDC converter valve cooling system fault diagnosis method of any one of claims 1-6 when the processor executes the computer program.

Images