Hydropower station equipment fault alarm method and system

Abstract

The invention relates to the technical field of hydropower station equipment monitoring, and provides a hydropower station equipment fault alarm method and system, and the method comprises the steps: obtaining a plurality of pieces of multi-source heterogeneous monitoring data of hydropower station equipment under a plurality of health historical working conditions; determining a reference feature template according to the plurality of health historical working conditions and the plurality of corresponding multi-source heterogeneous monitoring data; acquiring real-time heterogeneous monitoring data of the hydropower station equipment under the current working condition; according to the current working condition, the corresponding real-time heterogeneous monitoring data, the reference feature template and the fault feature knowledge base, identifying the early fault type of the hydropower station equipment; and multi-stage fault alarm is carried out according to the early fault type identification result. The problems that a traditional hydropower station equipment alarm method depends on a fixed threshold value and is difficult to adapt to working condition changes, multi-source heterogeneous monitoring data lacks effective fusion, early fault features are not sensitive, and false alarm and missing alarm are likely to happen are solved.

Description

Methods and systems for alarming equipment faults in hydropower stations Technical Field

[0001] This application relates to the field of hydropower station equipment monitoring technology, specifically to a method and system for alarming faults in hydropower station equipment. Background Technology

[0002] Hydropower stations are a crucial component of my country's power supply system, and the reliability and stability of their equipment operation directly impact grid security and power quality. To ensure the normal operation of critical equipment such as turbine generators, transformers, and speed control systems, various sensors are typically used to monitor multiple parameters in real time, including vibration, temperature, pressure, current, and voltage. Traditional equipment condition monitoring systems often rely on preset fixed thresholds to trigger alarms. Once a parameter exceeds its limit, the system issues an alarm, prompting maintenance personnel to intervene and inspect. This method can, to some extent, reflect obvious equipment anomalies and has become a common monitoring tool in hydropower stations.

[0003] However, existing alarm methods perform poorly when facing early or potential faults. Due to the complex structure and variable operating conditions of hydropower station equipment, data from a single sensor often fails to fully reflect the true state of the equipment, resulting in the system being insensitive to subtle fault characteristics. The lack of an effective fusion and analysis mechanism between heterogeneous data from multiple sources makes it impossible to identify cross-modal anomaly patterns holistically, leading to insufficient fault early warning capabilities. Furthermore, fixed thresholds are difficult to adapt to dynamic changes in equipment under different loads and operating stages, easily causing false alarms or missed alarms, reducing the reliability and practicality of the alarm system. Therefore, there is an urgent need for a new technical method that can deeply integrate multi-source monitoring data, accurately identify early anomalies, and make dynamic alarm decisions. Summary of the Invention

[0004] This application provides a method and system for alarming faults in hydropower station equipment, which solves the problems of traditional hydropower station equipment alarm methods relying on fixed thresholds, difficulty in adapting to changes in operating conditions, lack of effective fusion of multi-source heterogeneous data, insensitivity to early fault characteristics, and easy false alarms and missed alarms.

[0005] This application adopts the following technical solution: Firstly, this application provides a method for alarming faults in hydropower station equipment. The method includes: acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions; normalizing the multiple historical health conditions to form multiple normalized historical health conditions; fusing the multiple heterogeneous monitoring data to form multiple unified standard feature vectors; establishing a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors; acquiring real-time heterogeneous monitoring data of the hydropower station equipment under the current operating condition; normalizing the current operating condition to form a current normalized condition; fusing the real-time heterogeneous monitoring data to form a real-time standard feature vector; identifying early fault types of the hydropower station equipment based on the current normalized condition, the real-time standard feature vector, the benchmark feature template, and a fault feature knowledge base; and performing multi-level fault alarms based on the early fault type identification results.

[0006] Secondly, this application provides a hydropower station equipment fault alarm system, which includes: a data acquisition module for acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple health historical operating conditions; a data preprocessing and fusion module for normalizing multiple health historical operating conditions to form multiple health historical normalized operating conditions, and fusing multiple multi-source heterogeneous monitoring data to form multiple unified standard feature vectors; a benchmark feature template management module for establishing benchmark feature templates based on multiple health historical normalized operating conditions and corresponding multiple unified standard feature vectors; the data acquisition module is also used to acquire real-time heterogeneous monitoring data of hydropower station equipment under the current operating condition; the data preprocessing and fusion module is also used to normalize the current operating condition to form the current normalized operating condition, and fusing the real-time heterogeneous monitoring data to form a real-time standard feature vector; and an anomaly identification and alarm decision module for identifying early fault types of hydropower station equipment based on the current normalized operating condition, real-time standard feature vectors, benchmark feature templates, and fault feature knowledge base, and performing multi-level fault alarms based on the early fault type identification results.

[0007] The above-mentioned technical solutions adopted in this application can achieve the following beneficial effects: This invention obtains a unified standard feature vector by performing modal-specific preprocessing and hierarchical attention fusion on multi-source data such as vibration, temperature, pressure, and electrical data. It constructs and incrementally updates the adaptive benchmark feature template based on operating conditions by combining operating parameters such as head and load, and abandons fixed thresholds to adapt to changes and aging of equipment operating conditions. It amplifies weak abnormal components by using residual networks and achieves accurate early abnormality identification by combining fault feature knowledge base. It eliminates instantaneous interference and single-parameter misjudgment through three-level dynamic alarm decision-making, and then uses operation and maintenance verification feedback reinforcement learning for closed-loop optimization, which effectively improves alarm accuracy, solves the problems of insufficient early warning capability and low reliability of existing systems, and ensures the stable and reliable operation of key equipment in hydropower stations. Attached Figure Description

[0008] The accompanying drawings, which are included to provide a further understanding of this application and constitute a part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 shows a flowchart of a hydropower station equipment fault alarm method according to an embodiment of this application; Figure 2 shows a structural schematic diagram of a hydropower station equipment fault alarm system according to an embodiment of this application; Figure 3 shows a structural schematic diagram of a computer device according to an embodiment of this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] Figure 1 shows a flowchart of a hydropower station equipment fault alarm method according to an embodiment of this application. Referring to Figure 1, the method includes steps S110 to S150.

[0011] Step S110: Obtain multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions.

[0012] This embodiment first acquires the health history data of the hydropower station equipment to establish a baseline feature template. The health history data includes at least two parts: one part is the health history operating conditions, and the other part is multi-source heterogeneous monitoring data, with each health history operating condition corresponding to one set of multi-source heterogeneous monitoring data.

[0013] In some optional implementations, step S110 involves acquiring multiple multi-source heterogeneous monitoring data of the hydropower station equipment under multiple historical health conditions, including: acquiring multiple historical health conditions of the hydropower station equipment; wherein each historical health condition includes at least: head parameters and load rate parameters; acquiring multiple multi-source heterogeneous monitoring data of the hydropower station equipment; wherein each multi-source heterogeneous monitoring data includes at least: vibration signals, temperature signals, pressure signals, and electrical signals; wherein each signal includes multiple signal values; and establishing a correspondence between each historical health condition and each multi-source heterogeneous monitoring data.

[0014] This embodiment acquires multiple historical health conditions of the hydropower station equipment. These historical health conditions include: head parameters, load rate parameters, and may also include operating duration parameters, etc.

[0015] For example, obtain the health history of hydropower station equipment throughout its entire lifecycle, from initial commissioning to stable operation.

[0016] For example, multiple health history operating conditions of hydropower station equipment were collected from the start of commissioning to five years of stable operation. These health history operating conditions can cover different heads (e.g., 50~150m), different load rates (e.g., 10~100%), and different operating durations (e.g., 0~18000h). A total of 10,000 health history operating conditions were collected, each including head parameters, load rate parameters, and operating duration parameters.

[0017] This embodiment acquires multiple heterogeneous monitoring data from hydropower station equipment. The heterogeneous monitoring data includes vibration signals, temperature signals, pressure signals, and electrical signals.

[0018] For example, multi-source heterogeneous monitoring data is collected through various sensors deployed on key equipment in hydropower stations (such as turbine generators, transformers, and guide vanes). The types of multi-source heterogeneous monitoring data include vibration signals, temperature signals, pressure signals, and electrical signals (which can include current and voltage electrical signals). Each type of signal includes multiple signal values.

[0019] For example, multiple vibration sensors are installed at the stator and rotor bearing housings of the hydro-generator unit to collect vibration signals; temperature sensors are installed at the transformer windings and core to collect temperature signals; pressure sensors are installed at the pressure oil pipes and water guiding mechanisms to collect pressure signals; and electrical sensors are installed at the generator output terminals and the high and low voltage sides of the transformer to collect electrical signals.

[0020] During the acquisition of historical operating conditions and multi-source heterogeneous monitoring data, the correspondence between each historical operating condition and each multi-source heterogeneous monitoring data is recorded.

[0021] Step S120: Normalize multiple historical health conditions to form multiple normalized historical health conditions; fuse multiple heterogeneous monitoring data from multiple sources to form multiple unified standard feature vectors; and establish a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors.

[0022] This embodiment then uses a normalization algorithm to normalize each historical operating condition into a single historical normalized operating condition. Through an intelligent fusion algorithm, each multi-source heterogeneous monitoring data point is transformed into a unified standard feature vector. Based on the correspondence between historical operating conditions and multi-source heterogeneous monitoring data, a correspondence between historical normalized operating conditions and the unified standard feature vector is established. Finally, a baseline feature template is generated based on this correspondence.

[0023] In some optional implementations, step S120 involves normalizing multiple historical health conditions to form multiple normalized historical health conditions, fusing multiple heterogeneous monitoring data from multiple sources to form multiple unified standard feature vectors, and establishing a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors. This includes: for a single historical health condition; normalizing multiple parameters of the historical health condition to obtain the normalized historical health condition; for a single heterogeneous monitoring data set; preprocessing and standardizing each signal value of each type of signal to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals; wherein each heterogeneous standardized feature vector set includes multiple standardized feature vectors corresponding to the number of signal values ​​for the corresponding signal type; fusing the multiple heterogeneous standardized feature vector sets to obtain a unified standard feature vector; and using a domain-adaptive transfer learning algorithm to train and generate a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors.

[0024] The same processing is performed for each health history condition, so the normalization process for one health history condition is used as an example for illustration.

[0025] Taking a healthy historical operating condition as an example, each parameter included in the healthy historical operating condition is normalized to obtain a healthy historical normalized operating condition.

[0026] For example, taking the head parameter as an example, the following min-max normalization formula (1) is used during normalization: Formula (1); where, This represents the value of a head parameter under a healthy historical operating condition. This represents the historical normalized operating condition value of the head parameter under the health historical operating condition. This indicates the historical maximum value of the head parameter under the specified health historical operating conditions. This represents the historical minimum value of the head parameter under the health historical operating conditions.

[0027] The same normalization process is applied to parameters such as load rate and runtime, which will not be elaborated here.

[0028] For example, given a health history operating condition, normalization processing yields a health history normalized operating condition. = ,in, This represents the historical normalized operating condition value of the head parameter under the health historical operating condition. This represents the historical normalized operating condition value of the health history load rate parameter. This represents the historical normalized operating condition value of the health history operating condition runtime parameter.

[0029] The same processing is performed for each multi-source heterogeneous monitoring data point, so the processing flow of one multi-source heterogeneous monitoring data point is used as an example for illustration.

[0030] For a multi-source heterogeneous monitoring dataset, including vibration, temperature, pressure, and electrical signals, with each signal comprising multiple signal values, preprocessing and standardization are performed on each signal value for each signal to obtain multiple heterogeneous standardized feature vector sets corresponding to the various signals.

[0031] For example, for multi-source heterogeneous monitoring data, the vibration signal includes multiple signal values, which are then preprocessed and standardized to obtain a heterogeneous standardized feature vector set. This set contains multiple standardized feature vectors corresponding to each signal value of the vibration signal. The same preprocessing and standardization processes are performed on temperature, pressure, and electrical signals, and will not be elaborated further here.

[0032] After preprocessing and standardization, the resulting sets of multiple heterogeneous standardized feature vectors are fused to obtain a unified standard feature vector.

[0033] In other words, for a set of multi-source heterogeneous monitoring data, a corresponding unified standard feature vector is obtained after processing. Each set of multi-source heterogeneous monitoring data corresponds to a unified standard feature vector.

[0034] In some optional implementations, each signal value of each type of signal is preprocessed and standardized to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals. This includes: performing wavelet transform on multiple vibration signal values ​​of the vibration signal to extract multiple vibration time-frequency features; performing spatiotemporal correlation correction on multiple temperature signal values ​​of the temperature signal to obtain multiple temperature correction features; performing spatiotemporal correlation correction on multiple pressure signal values ​​of the pressure signal to obtain multiple pressure correction features; performing harmonic component decomposition on multiple electrical signal values ​​of the electrical signal to obtain multiple electrical frequency domain features; and performing standardization on multiple vibration time-frequency features, multiple temperature correction features, multiple pressure correction features, and multiple electrical frequency domain features to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals.

[0035] For a single vibration signal value, wavelet transform is used to extract a vibration time-frequency feature that includes both time-domain and frequency-domain characteristics. This process is repeated for each vibration signal value.

[0036] For example, for a vibration signal value, a 5-level wavelet transform is performed using the db4 wavelet basis function to decompose it into approximation coefficients and detail coefficients, thereby extracting vibration time-domain features (such as peak value, root mean square, kurtosis, etc.) and vibration frequency-domain features (such as center frequency, frequency band energy, etc.).

[0037] Taking the root mean square vibration time-domain characteristics as an example, the root mean square vibration time-domain characteristics are expressed by the following formula (2): Formula (2); where, This represents the time-domain characteristics of the root mean square vibration. Indicates the number of vibration signal sampling points. Indicates the first Vibration amplitude at each sampling point.

[0038] Taking the vibration frequency domain characteristics at the center frequency as an example, the vibration frequency domain characteristics at the center frequency are represented by the following formula (3): Formula (3); where, This represents the frequency domain characteristics of vibration at the center frequency. Indicates the number of frequency bands. Indicates the first The center frequency of each frequency band Indicates the first Energy in each frequency band.

[0039] For example, for a vibration signal value, a 6-level wavelet transform is performed using the db6 wavelet basis function to decompose it into one level of approximation coefficients reflecting the low-frequency trend of the signal and six levels of detail coefficients reflecting the high-frequency fluctuations of the signal. Vibration time-domain features (such as peak value, peak-to-peak value, impulse index, margin index, etc.) and vibration frequency-domain features (such as frequency band energy ratio, peak frequency, etc.) are extracted.

[0040] For example, a vibration signal value of a rotor bearing housing of a hydro-generator unit can be decomposed and the "energy ratio of the 500~1000Hz frequency band" can be extracted as a key vibration frequency domain feature. This feature can effectively reflect the high-frequency fluctuations of early bearing wear.

[0041] For a single temperature signal value, spatiotemporal correlation correction is performed to obtain a temperature correction feature. The same processing is applied to each temperature signal value.

[0042] For example, for a temperature signal value, spatiotemporal correlation correction is performed using Kriging interpolation to fill in missing values ​​in the temperature sensor data and eliminate spatial distribution differences. The temperature correction characteristics are obtained through the following formula (4): Formula (4); where, express Location at Temperature correction characteristics at any time, Indicates the number of effective temperature sensors in the vicinity. This represents the result obtained by calculating the first semivariogram. Interpolation weights for each temperature sensor, Indicates the first A temperature sensor in Location Temperature signal value at any given time.

[0043] For example, for a temperature signal value of a transformer core, eight temperature sensors at different locations on the core are selected, and the temperature value of the central region of the core is calculated using Kriging interpolation. Time series smoothing (using 3rd exponential smoothing) is introduced during the correction process to eliminate instantaneous temperature fluctuations. The temperature correction formula is as follows: Formula (5): Formula (5); where, Indicates the first The third exponential smoothing value (i.e., temperature correction value) at time 3. Indicates the first The second exponential smoothing value at time 1, Indicates the first The first exponentially smoothed value at time 1, Indicates the first The temperature signal value at any given time. Temperature correction ensures that the temperature correction feature accurately reflects the temperature status of the core area of ​​the equipment.

[0044] For a single pressure signal value, spatiotemporal correlation correction is performed to obtain a pressure correction feature. The same processing is applied to each pressure signal value.

[0045] The correction method for pressure signals is the same as that for temperature signals, and will not be repeated here.

[0046] For a given electrical signal value, harmonic component decomposition yields an electrical frequency domain characteristic. This process is repeated for each electrical signal value.

[0047] For example, for a current electrical signal value, the Fast Fourier Transform is used to decompose the harmonic components, extracting the amplitude and phase characteristics of the fundamental, third, and fifth harmonics. The harmonic component decomposition formula is as follows (6): Formula (6); where, Indicates at angular frequency Current amplitude in the frequency domain at that point This indicates the number of sampling points in the Fast Fourier Transform. Indicates the first The electrical signal value of the current at each sampling point.

[0048] For example, for a current electrical signal value of the generator stator, decompose it to obtain the current amplitude of the fundamental frequency (50Hz), the third harmonic (150Hz), and the fifth harmonic (250Hz), and calculate the harmonic distortion rate as the current frequency domain characteristic. The harmonic distortion rate is calculated using the following formula (7): Formula (7); where, Indicates harmonic distortion rate. This indicates the amplitude of the third harmonic current. This indicates the amplitude of the 5th harmonic current. express The amplitude of the second harmonic current. The harmonic distortion rate can effectively reflect whether there is a short circuit fault in the generator winding.

[0049] The decomposition method for voltage electrical signals is the same as that for current electrical signals, and will not be repeated here.

[0050] Each preprocessed feature is standardized to transform it into a set of heterogeneous standardized feature vectors for each signal.

[0051] For example: standardization is performed using the z-score standardization formula (8): Formula (8); where, For a given signal value, a heterogeneous standardized feature vector is... This represents the characteristic value of the signal after preprocessing. This represents the historical characteristic mean of the signal type to which the signal value belongs. This represents the historical standard deviation of the signal type to which the signal value belongs.

[0052] Standardization ensures that features of different dimensions can be fused and calculated.

[0053] For example, standardization is performed on the original range of 0.1~5mm for the "peak-to-peak value" feature in vibration signals and the original range of 0.1~5% for the "harmonic distortion rate" feature in electrical signals, so that both feature values ​​are mapped to an interval with a mean of 0 and a standard deviation of 1.

[0054] After preprocessing and standardization, a set of multi-source heterogeneous monitoring data is transformed into a heterogeneous standardized feature vector set for vibration signals (the heterogeneous standardized feature vector set for vibration signals includes multiple standardized feature vectors corresponding to the number of vibration signal values), a heterogeneous standardized feature vector set for temperature signals (the heterogeneous standardized feature vector set for temperature signals includes multiple standardized feature vectors corresponding to the number of temperature signal values), a heterogeneous standardized feature vector set for pressure signals (the heterogeneous standardized feature vector set for pressure signals includes multiple standardized feature vectors corresponding to the number of pressure signal values), and a heterogeneous standardized feature vector set for electrical signals (the heterogeneous standardized feature vector set for electrical signals includes multiple standardized feature vectors corresponding to the number of electrical signal values).

[0055] In some optional implementations, multiple heterogeneous standardized feature vector sets are fused to obtain a unified standard feature vector, including: data-level fusion and feature-level fusion of multiple heterogeneous standardized feature vector sets using a hierarchical attention fusion mechanism; data-level fusion includes: constructing multiple adaptive allocation weights corresponding to multiple heterogeneous standardized feature vector sets; fusing multiple standardized feature vectors of each heterogeneous standardized feature vector set according to the corresponding adaptive allocation weights to obtain each modality feature vector; feature-level fusion includes: constructing cross-modal attention allocation weights; fusing multiple modality feature vectors according to the cross-modal attention allocation weights to obtain a unified standard feature vector.

[0056] When fusing multiple heterogeneous normalized feature vector sets, a two-level fusion approach is adopted. The first-level fusion involves fusing multiple normalized feature vectors from a single heterogeneous normalized feature vector set to obtain the modal feature vector corresponding to each signal. The second-level fusion involves fusing multiple modal feature vectors to obtain a unified standard feature vector.

[0057] When fusing multiple heterogeneous standardized feature vector sets through a hierarchical attention fusion mechanism, data-level fusion is performed first.

[0058] Data-level fusion targets sensors of the same type, i.e., a heterogeneous set of normalized feature vectors. The aim of data-level fusion is to use an adaptive weight allocation algorithm to suppress redundancy and enhance complementarity, while eliminating interference, for sensors of the same type (such as vibration sensors at different locations).

[0059] For example, six vibration sensors are installed on the stator core of a hydroelectric generator set, located at the top, bottom, left, right, front, and rear of the stator. Six vibration signal values ​​are collected, and after preprocessing and standardization, six standardized feature vectors of the vibration signals are obtained. The heterogeneous normalized feature vector set that makes up the vibration signal.

[0060] An adaptive weighting method is used to fuse the standardized feature vectors of six vibration signals. The weighting is determined based on the signal-to-noise ratio (SNR) of the vibration sensors. Vibration sensors with higher SNRs receive greater weights to achieve redundancy suppression and complementary enhancement.

[0061] The signal-to-noise ratio of each vibration sensor is calculated using the following formula (9): Formula (9); where, Indicates the first The signal-to-noise ratio of each vibration sensor Indicates the first Useful power of each vibration sensor Indicates the first Noise power of a vibration sensor.

[0062] Based on the signal-to-noise ratio, the weight of each vibration sensor is calculated using the following formula (10): Formula (10); where, Indicates the first The weight of each vibration sensor, To ensure weight normalization.

[0063] The six vibration signal standardized feature vectors are weighted and fused using the following formula (11) to obtain the fused vibration mode feature vector: Formula (11); where, Represents the eigenvectors of vibration modes. Indicates the first A standardized feature vector of a vibration signal.

[0064] Furthermore, during the data-level fusion process, if the signal-to-noise ratio (SNR) of a vibration sensor is lower than a set threshold (e.g., 10dB), the data from that vibration sensor is considered interference data, and its weight is set to 0 to remove the corresponding vibration signal standardized feature vector. For example, if the SNR of the vibration sensor at the 4th position drops to 8dB due to loose installation, that vibration sensor is removed, and only the standardized feature vectors of the vibration signals from the remaining 5 vibration sensors are used for fusion.

[0065] For heterogeneous normalized feature vector sets corresponding to other signal types, data-level fusion is performed using the same method as the fusion of normalized feature vectors for vibration signals described above; this will not be elaborated further here. For heterogeneous normalized feature vector sets corresponding to other signal types, adaptive weight allocation can be determined based on factors such as the criticality of the sensor's location.

[0066] When fusing multiple heterogeneous standardized feature vector sets through a hierarchical attention fusion mechanism, feature-level fusion is then performed.

[0067] Feature-level fusion targets feature vectors from different modalities. It aims to fuse feature vectors from different modalities using cross-modal attention weight allocation to ultimately obtain a unified standard feature vector. ,in, This represents the dimension of the modal feature vector.

[0068] Specifically, the modal feature vector dimensions include vibration modes, temperature modes, pressure modes, and electrical modes. Dynamic attention weights are assigned to each modal feature vector to enhance effective modal features under key operating conditions. The allocation of cross-modal attention weights can be based on the following criteria: increasing the weight of electrical mode feature vectors under high load conditions and increasing the weight of vibration mode feature vectors under low load conditions.

[0069] For example, multi-source heterogeneous monitoring data from a hydro-generator unit can be fused to obtain a unified standard feature vector containing data such as "vibration frequency domain energy, temperature correction value, current harmonic distortion rate, and pressure stability value." This unified standard feature vector can comprehensively reflect the multi-dimensional status of the equipment. Through data-level and feature-level fusion, the problem of isolated analysis of multi-source heterogeneous data is solved, complementary information from various data types is integrated, and the completeness and accuracy of the unified standard feature vector are improved, providing comprehensive feature support for subsequent anomaly identification.

[0070] For example, construct a cross-modal attention network, whose input is a feature vector for each modality, including vibration mode feature vectors. (e.g., vibration frequency domain energy, vibration time domain peak-to-peak value, etc.), temperature modal eigenvectors (e.g., core correction temperature, winding correction temperature, etc.), pressure mode eigenvector (e.g., pressure stability value) and electrical modal eigenvectors (For example, current harmonic distortion rate, voltage fluctuation value, etc.).

[0071] The cross-modal attention allocation weights are determined based on the health history operating parameters (such as load rate) of the device. In other words, for a multi-modal feature vector obtained by processing multi-source heterogeneous monitoring data, the corresponding cross-modal attention allocation weights are assigned according to the health history operating parameters of the multi-source heterogeneous monitoring data, thereby merging the multi-modal feature vector data into a unified standard feature vector.

[0072] For example, the load rate parameters of health historical operating conditions corresponding to multiple modal feature vectors are normalized using min-max normalization to obtain the normalized load rate. .

[0073] The cross-modal attention allocation weights are then calculated using the following formula (12): , , , Formula (12); where, The attention weights represent the eigenvectors of vibration modes. The attention weights represent the eigenvectors of the temperature modes. The attention weights represent the feature vectors of the stress mode. The attention weights represent the electrical modal feature vectors, and =1. For example, the adjustment coefficient. Take 0.8, Take 1.0, Take 1.2, etc.

[0074] Under high load conditions (such as load factor) ), corresponding to normalized load factor According to formula (12), the attention weights of the electrical mode feature vectors This will significantly improve (e.g., increase to over 0.5) the attention weights of the vibration mode feature vectors. It will be significantly reduced (for example, reduced to below 0.2). At this point, the electrical modal feature vector is more sensitive to equipment faults, and the ability to identify electrical anomalies can be enhanced by increasing its weight.

[0075] Under low load conditions (such as load factor) ), corresponding to normalized load factor According to formula (12), the attention weights of the vibration mode eigenvectors This will significantly improve the attention weights of the electrical modal feature vectors. The value will be significantly reduced, and at this time the vibration mode feature vector can better reflect the mechanical fault of the equipment. By increasing its weight, the ability to identify mechanical abnormalities can be enhanced.

[0076] The feature vectors of each modality are fused according to the cross-modal attention allocation weights, and a unified standard feature vector is obtained by the following formula (13): Formula (13); where, This represents a unified standard feature vector. Additionally, the "+" sign here indicates fusion.

[0077] A domain-adaptive transfer learning algorithm is used to map the unified standard feature vectors under each health history working condition to the feature space, generating a working condition-adaptive baseline feature template. The mapping process is achieved by constructing the "working condition-feature" mapping function of the following formula (14): Formula (14); where, Represents the reference feature template. This represents the transfer learning mapping matrix. This indicates the bias term.

[0078] By training on each standardized feature vector corresponding to 10,000 historical health conditions, the parameters of the mapping function are determined, and a baseline feature template is generated.

[0079] Step S130: Obtain real-time heterogeneous monitoring data of the hydropower station equipment under the current operating conditions, normalize the current operating conditions to form the current normalized operating conditions, and fuse the real-time heterogeneous monitoring data to form a real-time standard feature vector.

[0080] This embodiment continues to acquire real-time data from the hydropower station equipment. The real-time data includes at least two parts: one part is the current operating condition, and the other part is real-time heterogeneous monitoring data.

[0081] Obtain the current operating status of the hydropower station equipment. The current operating status includes at least the current head and current load rate, and may also include the current operating time.

[0082] For the current operating condition, the multiple parameters of the current operating condition are normalized to obtain the current normalized operating condition.

[0083] For example: For the current operating condition, after normalization, we obtain the current normalized operating condition. = ,in, This represents the current normalized working condition value of the head parameters under the current operating conditions. This represents the current normalized operating condition value of the load rate parameter under the current operating condition. This represents the current normalized operating condition value of the runtime parameter for the current operating condition.

[0084] Real-time heterogeneous monitoring data of hydropower station equipment is acquired. This data includes real-time vibration signals, real-time temperature signals, real-time pressure signals, and real-time electrical signals. Each real-time signal comprises multiple real-time signal values.

[0085] For real-time heterogeneous monitoring data, each real-time signal value for each type of real-time signal is preprocessed and standardized to obtain multiple real-time heterogeneous standardized feature vector sets corresponding to various real-time signals. Each real-time heterogeneous standardized feature vector set includes multiple real-time standardized feature vectors corresponding to the number of real-time signal values ​​for the corresponding real-time signal type. These multiple real-time heterogeneous standardized feature vector sets are then fused to obtain the real-time standard feature vector.

[0086] The method for preprocessing and standardizing each real-time signal value for each type of real-time signal is the same as the aforementioned process for preprocessing and standardizing each signal value for each type of signal, and will not be repeated here. The method for fusing multiple real-time heterogeneous standardized feature vector sets is the same as the aforementioned process for fusing multiple heterogeneous standardized feature vector sets, and will not be repeated here.

[0087] For example: the real-time standard feature vector is represented as ,in Represents the real-time standard feature vector. This represents the dimension of the modal feature vector.

[0088] Step S140: Identify early fault types of hydropower station equipment based on the current normalized operating conditions, real-time standard feature vectors, benchmark feature templates, and fault feature knowledge base.

[0089] This embodiment then determines the target reference feature vector under the current normalized operating condition based on the current normalized operating condition and the reference feature template; compares the real-time standard feature vector and the target reference feature vector to determine the amplification deviation of the real-time standard feature vector and the target reference feature vector under the current operating condition; when the amplification deviation exceeds the deviation threshold, it compares the real-time deviation feature between the real-time standard feature vector and the target reference feature vector with the multimodal feature combination of the fault feature knowledge base to identify the target early fault type of the hydropower station equipment.

[0090] In some optional implementations, step S140, identifying early fault types of hydropower station equipment based on the current normalized operating condition, real-time standard feature vector, benchmark feature template, and fault feature knowledge base, includes: determining the target benchmark feature vector under the current normalized operating condition through linear interpolation based on the current normalized operating condition and benchmark feature template; determining the deviation degree under the current normalized operating condition based on the real-time standard feature vector and the target benchmark feature vector, and amplifying the deviation degree through multi-scale amplification to obtain the amplified deviation degree; when the amplified deviation degree exceeds the deviation degree threshold, calculating the real-time deviation feature based on the real-time standard feature vector and the target benchmark feature vector; constructing a fault feature knowledge base; wherein the fault feature knowledge base stores the correspondence between each early fault type and each multimodal feature combination; calculating the similarity between the real-time deviation feature and each multimodal feature combination, and matching the corresponding target early fault type in the fault feature knowledge base based on the similarity.

[0091] The current normalized chemical condition is expressed as The target baseline eigenvector under the current normalized working condition is determined by linear interpolation, and is expressed as follows: .

[0092] For example: current water head (Within the healthy historical head parameters of 80~120m), current load rate (Between 50% and 70% of the historical health load rate), the target baseline eigenvector under the current normalized working condition is obtained by linear interpolation, expressed by formula (15): Formula (15); where, This represents the baseline feature template under operating conditions of 80m head and 50% load rate. This represents the baseline feature template under operating conditions of 120m head and 50% load rate. This represents the baseline feature template under operating conditions of 80m head and 70% load rate. This represents the baseline feature template under the operating conditions of 120m head and 70% load rate.

[0093] Calculate the deviation between the real-time standard eigenvector and the target baseline eigenvector under the current normalized working conditions.

[0094] Real-time standard feature vector representation is as follows The target baseline feature vector is represented as The deviation is then calculated using the following formula (16): Formula (16); where, Indicates the degree of deviation. Indicates the first The weights of the modal eigenvectors and , The first eigenvector represents the real-time standard feature vector. Modal feature vectors, The first eigenvector representing the target baseline eigenvector There are several modal eigenvectors. In some cases, the weight of the vibration modal eigenvector is 0.3, the weight of the temperature modal eigenvector is 0.2, the weight of the pressure modal eigenvector is 0.25, and the weight of the electrical modal eigenvector is 0.25.

[0095] When amplifying the bias at multiple scales, a residual network is constructed, consisting of an input layer, three residual blocks, and an output layer. The input to the residual network is the bias. The output is the magnification deviation after amplification. Each residual block contains a convolutional layer (3×3 kernel size), a batch normalization layer, and a ReLU activation layer. Residual connections directly append the input to the residual block output. This amplifies the bias. This can be expressed by the following formula (17): Formula (17); where, They represent the first The weight matrix and batch normalization parameters of the layer residual blocks, This indicates the bias term.

[0096] If the amplification deviation does not exceed the deviation threshold, it indicates that the real-time heterogeneous monitoring data under the current normalized chemical condition is healthy, and no further early fault diagnosis is required. If the amplification deviation exceeds the deviation threshold, it indicates that the real-time heterogeneous monitoring data under the current normalized chemical condition is unhealthy, and early fault identification is necessary. The deviation threshold can be determined based on the actual situation.

[0097] For example, the deviation threshold can be 0.3. When the deviation... Traditional methods struggle to identify biases, but the bias is amplified by a residual network. This significantly highlights subtle anomalies and can solve the problem of insensitivity to early fault characteristics.

[0098] The fault feature knowledge base can be pre-built based on actual conditions. When building the fault feature knowledge base, historical fault types of hydropower station equipment are collected, such as bearing wear, generator winding short circuits, transformer core overheating, governor jamming, and pressure oil pipe leakage. For each historical fault type, historical monitoring data is collected and fused to extract the corresponding multimodal feature combinations, ultimately forming a correspondence between early fault types and multimodal feature combinations. The fusion processing of historical monitoring data can refer to the aforementioned process for fusion processing of multi-source heterogeneous monitoring data, and will not be repeated here.

[0099] For example: the multi-mode characteristic combination corresponding to bearing wear is "increased amplitude of high-frequency harmonic vibration + increased bearing temperature + current fluctuation"; the multi-mode characteristic combination corresponding to speed governor jamming is "pressure fluctuation of water guide mechanism exceeds baseline by 20% + speed deviation exceeds baseline by 15% + amplitude of low-frequency vibration component increases by 10%"; the multi-mode characteristic combination corresponding to transformer core overheating is "core temperature exceeds baseline by 10℃ + no-load loss increases by 5% + voltage waveform distortion rate increases by 3%".

[0100] When the amplified deviation exceeds the deviation threshold, the real-time deviation feature is calculated based on the actual standard feature vector and the target reference feature vector, and is expressed by the following formula (18): Formula (18); where, This indicates the real-time deviation characteristic. Additionally, the "-" here is used to indicate the calculation deviation.

[0101] The real-time deviation features are compared with the multimodal features in the fault feature knowledge base using the cosine similarity matching algorithm. The similarity is expressed by the following formula (19): Formula (19); where, Indicates real-time deviation characteristics and the first Similarity of multimodal feature combinations Indicates the first A combination of multimodal features.

[0102] By ranking the similarity between real-time deviation features and each multimodal feature combination, and selecting the highest similarity, a correspondence is formed between real-time deviation features, the most similar multimodal feature combination, and the early fault type. This allows for the determination of the target early fault type corresponding to the real-time heterogeneous monitoring data under the current operating condition.

[0103] For example, if the similarity between the real-time deviation feature and the multimodal feature combination corresponding to bearing wear is 0.82, and the similarity between the multimodal feature combination corresponding to other early fault types is less than 0.7, then the target early fault type is determined to be bearing wear.

[0104] By refining the early fault type identification process, subtle fault characteristics can be captured more accurately and anomaly judgments can be made, providing a basis for subsequent alarm decisions.

[0105] Step S150: Perform multi-level fault alarms based on the early fault type identification results.

[0106] This embodiment finally performs multi-level dynamic alarm decision-making based on the early fault type identification results. The decision-making process of multi-level dynamic alarm can be divided into three levels: Level 1 decision, which determines whether the deviation of any single modal feature in the real-time deviation features exceeds the corresponding modal health threshold range and the duration reaches the duration threshold, in order to avoid instantaneous interference; Level 2 decision, which verifies the abnormal correlation of cross-modal feature deviations to eliminate misjudgments caused by single fluctuations; Level 3 decision, which combines the historical fault frequency and fault pattern of the target's early fault type, outputs the alarm level through a fuzzy inference engine, and generates auxiliary judgments of fault location and possible causes.

[0107] In some optional implementations, step S150, which involves performing multi-level fault alarms based on the early fault type identification results, includes: dynamically generating multiple modal health threshold intervals under the current normalized operating conditions based on the target baseline feature vector; determining whether any target single-modal feature deviation in the real-time deviation features exceeds the corresponding modal health threshold interval and the duration reaches a duration threshold; if so, determining whether the related single-modal feature deviations associated with the target single-modal feature deviation in the real-time deviation features exceed the corresponding modal health threshold interval; if so, determining the alarm level using a fuzzy inference engine based on the historical fault frequency and fault pattern of the target early fault type.

[0108] Based on the determined target benchmark feature vector It can dynamically generate multiple modal health threshold ranges under the current normalized working condition, expressed by formula (20): , Formula (20); where, This indicates the first normalized condition. The lower limit of the modal health threshold range This indicates the first normalized condition. The upper limit of a modal health threshold range, Indicates the confidence interval. This represents the first eigenvector of the target baseline under the current normalized chemical condition. The standard deviation of each modality. A value of 2.33 can be used to represent a 99% confidence interval.

[0109] For example, when the equipment load rate increases from 50% to 80%, the modal health threshold range will be dynamically adjusted accordingly. This raises the upper limit of the modal health threshold range for current-related characteristics to adapt to normal current fluctuations under high load, avoiding false alarms caused by changes in operating conditions. Through a refined process of constructing dynamic modal health threshold ranges, changes in equipment operating status can be more accurately adapted, further reducing the limitations of fixed thresholds.

[0110] Level 1 decision: Determine whether the deviation of any target single-modal feature in the real-time deviation features exceeds the corresponding modal health threshold range and the duration reaches the duration threshold.

[0111] For example: Electrical modal characteristic deviation exceeds If this continues for 5 minutes, a Level 1 decision will be triggered.

[0112] Secondary decision: Determine whether the relevant single-modal feature deviations associated with the target single-modal feature deviation in the real-time deviation features exceed the corresponding modal health threshold range.

[0113] For example: when the vibration mode characteristic deviation exceeds At that time, detect whether the temperature modal characteristic deviation exceeds Or, when the electrical modal characteristic deviation exceeds At that time, check whether the deviation of the pressure modal characteristic exceeds If an anomaly in the correlation is found, a secondary decision will be triggered and passed.

[0114] Level 3 decision-making: Based on the historical failure frequency and failure pattern of the target's early failure types, the alarm level is determined using a fuzzy inference engine.

[0115] Based on the historical failure frequency and failure pattern of the early failure type of the target, the alarm level is output through a fuzzy inference engine. The fuzzy inference engine uses a triangular membership function to calculate the membership degree of the alarm level, as shown in the following formula (21): , , Formula (21); where, This indicates a warning. This indicates a Level 1 alarm. This indicates an emergency alarm. Indicates other situations, , They represent the weighting coefficients and , Indicates the historical failure frequency. Indicates the fault pattern, These represent the threshold values ​​of the membership function, for example... , , .

[0116] The alarm level is determined based on the principle of maximum membership, and auxiliary judgments are generated to determine the fault location (such as rotor bearing housing) and possible causes (such as insufficient lubricant) that lead to the early fault type (bearing wear) of the target.

[0117] In some optional implementations, the above method further includes: periodically acquiring multiple heterogeneous monitoring data of hydropower station equipment under multiple health update conditions; normalizing the multiple health update conditions to form multiple normalized health update conditions, and fusing the multiple heterogeneous monitoring data to form multiple update standard feature vectors; and updating the baseline feature template by incremental learning based on the multiple normalized health update conditions and the corresponding multiple update standard feature vectors.

[0118] Health update data for hydropower station equipment should be incorporated regularly (e.g., monthly). The health update data should include at least multiple health update conditions and multiple update heterogeneous monitoring data.

[0119] For example: Health update data is collected at the end of each month. Multiple health update conditions are normalized to form multiple health update normalized conditions. The normalization method is the same as the process described above for normalizing multiple historical health conditions to form multiple historical health normalized conditions, and will not be repeated here. Multiple updated heterogeneous monitoring data are fused to form multiple updated standard feature vectors. The fusion method is the same as the process described above for fusion of multiple multi-source heterogeneous monitoring data to form multiple unified standard feature vectors, and will not be repeated here. The corresponding associated health update normalized conditions and updated standard feature vectors are used as the update dataset.

[0120] Incremental learning algorithm is used to optimize the baseline feature template to solve the baseline offset problem caused by equipment aging. Incremental learning adopts incremental support vector machine algorithm, and the parameter update of the baseline feature template is as follows (22): , Formula (22); where, This represents the updated transfer learning mapping matrix. This represents the transfer learning mapping matrix before the update. This represents the updated bias term. This indicates the bias term before the update. Indicates the learning rate. This indicates the amount of data in the updated dataset. Indicates the first Each corresponding health update normalized condition and update standard feature vector Mark as health data.

[0121] For example, when the equipment is in its 36th month of operation, the baseline feature template is optimized by incorporating 280 sets of health update data from that month. After the update, the baseline value of "bearing vibration peak-to-peak value" in the baseline feature template is adjusted from 0.5mm to 0.7mm, and the corresponding modal health threshold range is adjusted from [0.3, 0.7] to [0.5, 0.9], which adapts to the normal increase in vibration amplitude caused by equipment aging and avoids false alarms caused by baseline offset.

[0122] Incremental updates ensure that the baseline feature template always remains consistent with the current health status of the device, guaranteeing the accuracy of the baseline feature template and effectively solving the baseline offset problem throughout the device's lifecycle.

[0123] In some optional implementations, the above method further includes: adjusting the cross-modal attention allocation weights and / or modal health threshold ranges and / or multi-level fault alarm rules using reinforcement learning based on the on-site verification results.

[0124] After conducting on-site inspections, maintenance personnel will provide feedback on the verification results. Verification results can fall into three categories: false alarms (no fault detected during on-site inspection but an alarm is triggered); missed alarms (no alarm was triggered but a fault was found on-site); and confirmed alarms (the alarm information matches the on-site fault).

[0125] For example: if the fault alarm is "early warning - bearing wear", and the maintenance personnel confirm that the bearing is not worn after on-site inspection, then it is reported as a false alarm; if the system does not alarm but the maintenance personnel find that the speed controller is stuck, then it is reported as a missed alarm; if the fault alarm is not "Level 1 alarm - winding short circuit", and the maintenance personnel confirm that the winding is short circuit after on-site inspection, then it is reported as a confirmation.

[0126] Transform the verification results into reward signals for reinforcement learning. For example, when the reward value rule is set to "confirm". "False alarm" "Underreporting" Because the risk of underreporting is higher than that of false reporting, the penalties are more severe.

[0127] The Q-learning reinforcement learning algorithm is used to adjust cross-modal attention assignment weights and / or modal health threshold ranges and / or multi-level fault alarm rules. The state of the Q-learning reinforcement learning algorithm... Defined as "Current Normalized Operating Condition - Real-time Deviation Characteristics", Action The alarm levels are defined as "no alarm, early warning, level one alarm, and emergency alarm". The value represents the state. Next action The expected cumulative reward. The value is updated using the following formula (23): Formula (23); where, Indicates the discount factor. Indicates the execution of an action The next state after that, This represents the optimal action for the next state.

[0128] For example, if multiple false alarms occur ("vibration characteristic deviation combined with related single-mode characteristic deviation triggers an alarm but no abnormality is found on site"), reinforcement learning is used to reduce the attention weight of the vibration mode feature vector, for example, from 0.3 to 0.2, while adjusting the duration of the vibration characteristic deviation in the first-level decision, for example, extending it from 5 minutes to 8 minutes. If multiple false alarms occur ("current characteristic deviation combined with related single-mode characteristic deviation does not trigger an alarm but a short circuit fault is found on site"), reinforcement learning is used to increase the attention weight of the electrical mode feature vector, for example, from 0.25 to 0.35, while reducing the upper limit of the electrical mode health threshold range, for example, from 2% to 1.8%.

[0129] By regularly optimizing the closed-loop system based on on-site verification results, the accuracy of alarms can be further improved.

[0130] Figure 2 shows a schematic diagram of the structure of a hydropower station equipment fault alarm system proposed in one embodiment of this application. Please refer to Figure 2. The system includes: a data acquisition module 210, used to acquire multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple health historical operating conditions; a data preprocessing and fusion module 220, used to normalize multiple health historical operating conditions to form multiple health historical normalized operating conditions, and to fuse multiple multi-source heterogeneous monitoring data to form multiple unified standard feature vectors; a benchmark feature template management module 230, used to establish benchmark feature templates based on multiple health historical normalized operating conditions and corresponding multiple unified standard feature vectors; the data acquisition module 210 is also used to acquire real-time heterogeneous monitoring data of hydropower station equipment under the current operating condition; the data preprocessing and fusion module 220 is also used to normalize the current operating condition to form the current normalized operating condition, and to fuse the real-time heterogeneous monitoring data to form a real-time standard feature vector; and an anomaly identification and alarm decision module 240, used to identify early fault types of hydropower station equipment based on the current normalized operating condition, real-time standard feature vectors, benchmark feature templates, and fault feature knowledge base, and to perform multi-level fault alarms based on the early fault type identification results.

[0131] In some optional implementations, the data acquisition module 210 is specifically used to: acquire multiple historical health conditions of the hydropower station equipment; wherein each historical health condition includes at least: head parameters and load rate parameters; acquire multiple multi-source heterogeneous monitoring data of the hydropower station equipment; wherein each multi-source heterogeneous monitoring data includes at least: vibration signal, temperature signal, pressure signal, and electrical signal; wherein each signal includes multiple signal values; and establish a correspondence between each historical health condition and each multi-source heterogeneous monitoring data.

[0132] In some optional implementations, the data preprocessing and fusion module 220 is specifically used for: for a health history operating condition; normalizing multiple parameters of the health history operating condition to obtain a normalized health history operating condition; for a multi-source heterogeneous monitoring data; preprocessing and standardizing each signal value of each type of signal to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals; wherein, a heterogeneous standardized feature vector set includes multiple standardized feature vectors corresponding to the number of signal values ​​of the corresponding signal type; fusing multiple heterogeneous standardized feature vector sets to obtain a unified standard feature vector.

[0133] In some optional implementations, the data preprocessing and fusion module 220 is specifically used to: perform wavelet transform on multiple vibration signal values ​​of the vibration signal to extract multiple vibration time-frequency features; perform spatiotemporal correlation correction on multiple temperature signal values ​​of the temperature signal to obtain multiple temperature correction features; perform spatiotemporal correlation correction on multiple pressure signal values ​​of the pressure signal to obtain multiple pressure correction features; perform harmonic component decomposition on multiple electrical signal values ​​of the electrical signal to obtain multiple electrical frequency domain features; and perform standardization processing on multiple vibration time-frequency features, multiple temperature correction features, multiple pressure correction features, and multiple electrical frequency domain features to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals.

[0134] In some optional implementations, the data preprocessing and fusion module 220 is specifically used to: perform data-level fusion and feature-level fusion on multiple heterogeneous standardized feature vector sets using a hierarchical attention fusion mechanism; data-level fusion includes: constructing multiple adaptive allocation weights corresponding to multiple heterogeneous standardized feature vector sets; fusing multiple standardized feature vectors of each heterogeneous standardized feature vector set according to the corresponding adaptive allocation weights to obtain each modality feature vector; feature-level fusion includes: constructing cross-modal attention allocation weights; fusing multiple modality feature vectors according to the cross-modal attention allocation weights to obtain a unified standard feature vector.

[0135] In some optional implementations, the baseline feature template management module 230 is specifically used to: generate a baseline feature template by using a domain-adaptive transfer learning algorithm based on multiple health history normalized conditions and corresponding multiple unified standard feature vectors.

[0136] In some optional implementations, the anomaly identification and alarm decision module 240 is specifically used for: determining the target baseline feature vector under the current normalized working condition by linear interpolation based on the current normalized working condition and the baseline feature template; determining the deviation degree under the current normalized working condition based on the real-time standard feature vector and the target baseline feature vector, and amplifying the deviation degree by multi-scale magnification to obtain the amplified deviation degree; when the amplified deviation degree exceeds the deviation degree threshold, calculating the real-time deviation feature based on the real-time standard feature vector and the target baseline feature vector; constructing a fault feature knowledge base; wherein the fault feature knowledge base stores the correspondence between each early fault type and each multimodal feature combination; calculating the similarity between the real-time deviation feature and each multimodal feature combination, and matching the corresponding target early fault type in the fault feature knowledge base based on the similarity.

[0137] In some optional implementations, the anomaly identification and alarm decision module 240 is specifically used to: dynamically generate multiple modal health threshold intervals under the current normalized operating conditions based on the target baseline feature vector; determine whether any target single-modal feature deviation in the real-time deviation features exceeds the corresponding modal health threshold interval and the duration reaches the duration threshold; if so, determine whether the related single-modal feature deviations associated with the target single-modal feature deviation in the real-time deviation features exceed the corresponding modal health threshold interval; if so, determine the alarm level using a fuzzy inference engine based on the historical fault frequency and fault pattern of the target early fault type.

[0138] In some alternative implementations, the data acquisition module 210 is specifically used to periodically acquire multiple update heterogeneous monitoring data of hydropower station equipment under multiple health update conditions.

[0139] In some optional real-time modes, the data preprocessing and fusion module 220 is specifically used to: normalize multiple health update conditions to form multiple health update normalized conditions, and fuse multiple update heterogeneous monitoring data to form multiple update standard feature vectors.

[0140] In some optional implementations, the baseline feature template management module 230 is specifically used to: update the baseline feature template using incremental learning based on multiple health update normalized conditions and corresponding multiple update standard feature vectors.

[0141] In some optional implementations, the system also includes a feedback optimization module for adjusting cross-modal attention allocation weights and / or modal health threshold ranges and / or multi-level fault alarm rules based on on-site verification results using reinforcement learning.

[0142] The hydropower station equipment fault alarm system achieves real-time processing and accurate alarm of multi-source data through the collaboration of a data acquisition module, a data preprocessing and fusion module, a benchmark feature template management module, an anomaly identification and alarm decision module, and a feedback optimization module.

[0143] The data acquisition module deploys multiple types of sensors: piezoelectric accelerometers for vibration, platinum resistance thermometers for temperature, strain gauge sensors for pressure, and Hall effect transformers for current and voltage. All sensors are installed in critical parts of equipment such as hydro-generators and transformers. It is paired with an industrial-grade data acquisition terminal to collect data in real time. The terminal's sampling frequency is set to 1024Hz, and the AD conversion accuracy is 16 bits. Simultaneously, a self-testing module monitors the sensor status, determining sensor functionality based on signal range, noise, and changes, effectively preventing faulty sensors from affecting data quality.

[0144] The data preprocessing and fusion module is deployed on the server side and implemented using the Python programming language combined with the TensorFlow deep learning framework. First, modality-specific preprocessing is performed on each monitoring data, including wavelet denoising of vibration signals, spatiotemporal correction of temperature and pressure signals, and harmonic decomposition of electrical signals. Then, z-score standardization and [0,1] interval normalization are used to eliminate the dimensional differences between different types of data. Finally, data fusion is achieved through hierarchical attention fusion, which includes data-level fusion based on adaptive weight allocation and feature-level fusion based on cross-modal attention weight allocation. The final output is a 64-dimensional unified feature vector, and the fusion time for a single set of data is no more than 1 second.

[0145] The baseline feature template management unit is developed based on a C++ real-time computing engine, adopts the GBT algorithm and is implemented based on the XGBoost library. It combines the equipment's full life cycle health data to build a "condition-feature" mapping model. Through the condition transfer learning algorithm, feature vectors under different health conditions are mapped to a unified feature space. New health operation data are added to the updated model every month through an incremental learning algorithm to avoid baseline offset problems caused by equipment aging.

[0146] The anomaly identification and alarm decision-making module is implemented using a hybrid programming approach of Python and MATLAB. First, it calculates the deviation between the real-time fused features and the target baseline feature vector, assigning higher weights to key features during the calculation. Then, it uses a ResNet-18 residual network to amplify the deviation at multiple scales, highlighting weak anomaly components. Next, it compares the deviation with a fault feature knowledge base containing 15 types of faults to preliminarily determine the anomaly type, employing a cosine similarity matching algorithm. Finally, it executes a three-level dynamic alarm decision-making process: single-modal continuous deviation judgment, cross-modal feature correlation verification, and alarm level determination based on fuzzy reasoning. The entire decision-making process can be completed within one second, outputting alarm information including the fault location and possible causes, which is simultaneously pushed to the maintenance personnel's terminal.

[0147] The feedback optimization module is built on the OpenAI Gym reinforcement learning framework. It receives the on-site verification results from maintenance personnel via a web interface, including false alarms, missed alarms, and fault confirmation. The verification results are associated with the corresponding alarm records and stored to form a feedback dataset. The feedback dataset is used for closed-loop optimization. The optimized parameters are synchronized to other modules every week during the system's idle period, such as 2-4 am.

[0148] In summary, this invention obtains a unified standard feature vector by performing modal-specific preprocessing and hierarchical attention fusion on multi-source data such as vibration, temperature, pressure, and electrical data. It then constructs and incrementally updates an adaptive baseline feature template based on operating conditions, combining this with parameters such as head and load, thus abandoning fixed thresholds to adapt to changes in equipment operating conditions and aging. Furthermore, it amplifies weak anomaly components using residual networks and achieves accurate early anomaly identification by combining a fault feature knowledge base. Finally, it eliminates instantaneous interference and single-parameter misjudgments through a three-level dynamic alarm decision-making process, and utilizes operation and maintenance verification feedback reinforcement learning for closed-loop optimization, effectively improving alarm accuracy and solving the problems of insufficient early warning capabilities and low reliability in existing systems, thereby ensuring the stable and reliable operation of key equipment in hydropower stations.

[0149] It should be noted that the above-mentioned hydropower station equipment fault alarm system can implement the fault alarm methods of hydropower station equipment one by one, which will not be elaborated further.

[0150] In one embodiment, a computer device is provided, the internal structure of which can be shown in Figure 3. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external devices via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a hydropower station equipment fault alarm method.

[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program performs the following steps: acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions; normalizing the multiple historical health conditions to form multiple normalized historical health conditions, fusing the multiple heterogeneous monitoring data to form multiple unified standard feature vectors, and establishing a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors; acquiring real-time heterogeneous monitoring data of hydropower station equipment under the current operating condition, normalizing the current operating condition to form the current normalized condition, and fusing the real-time heterogeneous monitoring data to form a real-time standard feature vector; identifying early fault types of hydropower station equipment based on the current normalized condition, the real-time standard feature vector, the benchmark feature template, and a fault feature knowledge base; and performing multi-level fault alarms based on the early fault type identification results.

[0152] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0153] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0154] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0155] The above-described 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for alarming faults in hydropower station equipment, characterized in that, The method includes: acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions; normalizing the multiple historical health conditions to form multiple normalized historical health conditions; fusing the multiple multi-source heterogeneous monitoring data to form multiple unified standard feature vectors; establishing a benchmark feature template based on the multiple normalized historical health conditions and the corresponding multiple unified standard feature vectors; acquiring real-time heterogeneous monitoring data of hydropower station equipment under the current operating condition; normalizing the current operating condition to form the current normalized condition; fusing the real-time heterogeneous monitoring data to form a real-time standard feature vector; identifying early fault types of hydropower station equipment based on the current normalized condition, real-time standard feature vector, benchmark feature template, and fault feature knowledge base; and performing multi-level fault alarms based on the early fault type identification results.

2. The method for alarming faults in hydropower station equipment according to claim 1, characterized in that, The acquisition of multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions includes: acquiring multiple historical health conditions of hydropower station equipment; wherein each historical health condition includes at least: head parameters and load rate parameters; acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment; wherein each multi-source heterogeneous monitoring data includes at least: vibration signal, temperature signal, pressure signal, and electrical signal; wherein each signal includes multiple signal values; and establishing a correspondence between each historical health condition and each multi-source heterogeneous monitoring data.

3. The method for alarming faults in hydropower station equipment according to claim 2, characterized in that, The process involves normalizing multiple historical health conditions to form multiple normalized health conditions, fusing multiple heterogeneous monitoring data from multiple sources to form multiple unified standard feature vectors, and establishing a benchmark feature template based on the normalized health conditions and corresponding unified standard feature vectors. This includes: for a single historical health condition; normalizing multiple parameters of the historical health condition to obtain the normalized health condition; for a single heterogeneous monitoring data set; preprocessing and standardizing each signal value for each type of signal to obtain multiple sets of heterogeneous standardized feature vectors corresponding to various signals; wherein each set of heterogeneous standardized feature vectors includes multiple standardized feature vectors corresponding to the number of signal values ​​for the corresponding signal type; fusing the multiple sets of heterogeneous standardized feature vectors to obtain a unified standard feature vector; and using a domain-adaptive transfer learning algorithm to train and generate a benchmark feature template based on the normalized health conditions and corresponding unified standard feature vectors.

4. The hydropower station equipment fault alarm method according to claim 3, characterized in that, The process of preprocessing and standardizing each signal value of each type of signal to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals includes: performing wavelet transform on multiple vibration signal values ​​of vibration signals to extract multiple vibration time-frequency features; performing spatiotemporal correlation correction on multiple temperature signal values ​​of temperature signals to obtain multiple temperature correction features; performing spatiotemporal correlation correction on multiple pressure signal values ​​of pressure signals to obtain multiple pressure correction features; performing harmonic component decomposition on multiple electrical signal values ​​of electrical signals to obtain multiple electrical frequency domain features; and standardizing the multiple vibration time-frequency features, multiple temperature correction features, multiple pressure correction features, and multiple electrical frequency domain features to obtain multiple heterogeneous standardized feature vector sets corresponding to multiple signals.

5. The hydropower station equipment fault alarm method according to claim 3, characterized in that, The process of fusing multiple heterogeneous standardized feature vector sets to obtain a unified standard feature vector includes: performing data-level fusion and feature-level fusion on multiple heterogeneous standardized feature vector sets using a hierarchical attention fusion mechanism; data-level fusion includes: constructing multiple adaptive allocation weights corresponding to multiple heterogeneous standardized feature vector sets; fusing multiple standardized feature vectors of each heterogeneous standardized feature vector set according to the corresponding adaptive allocation weights to obtain a feature vector for each modality; feature-level fusion includes: constructing cross-modal attention allocation weights; fusing multiple modal feature vectors according to the cross-modal attention allocation weights to obtain a unified standard feature vector.

6. The hydropower station equipment fault alarm method according to claim 5, characterized in that, The process of identifying early fault types of hydropower station equipment based on the current normalized operating conditions, real-time standard feature vectors, benchmark feature templates, and a fault feature knowledge base includes: determining the target benchmark feature vector under the current normalized operating conditions through linear interpolation based on the current normalized operating conditions and benchmark feature templates; determining the deviation degree under the current normalized operating conditions based on the real-time standard feature vectors and the target benchmark feature vectors, and amplifying the deviation degree through multi-scale amplification; when the amplified deviation degree exceeds the deviation degree threshold, calculating the real-time deviation feature based on the real-time standard feature vectors and the target benchmark feature vectors; constructing a fault feature knowledge base; wherein the fault feature knowledge base stores the correspondence between each early fault type and each multimodal feature combination; calculating the similarity between the real-time deviation feature and each multimodal feature combination, and matching the corresponding target early fault type in the fault feature knowledge base based on the similarity.

7. The hydropower station equipment fault alarm method according to claim 6, characterized in that, The multi-level fault alarm based on the early fault type identification results includes: dynamically generating multiple modal health threshold intervals under the current normalized operating conditions based on the target baseline feature vector; determining whether the deviation of any target single-modal feature in the real-time deviation features exceeds the corresponding modal health threshold interval and the duration reaches the duration threshold; if so, determining whether the related single-modal feature deviations associated with the target single-modal feature deviation in the real-time deviation features exceed the corresponding modal health threshold interval; if so, determining the alarm level using a fuzzy inference engine based on the historical fault frequency and fault pattern of the target early fault type.

8. The method for alarming faults in hydropower station equipment according to claim 1, characterized in that, The method further includes: periodically acquiring multiple heterogeneous monitoring data of hydropower station equipment under multiple health update conditions; normalizing the multiple health update conditions to form multiple normalized health update conditions; fusing the multiple heterogeneous monitoring data to form multiple update standard feature vectors; and using incremental learning to update the baseline feature template based on the multiple normalized health update conditions and the corresponding multiple update standard feature vectors.

9. The method for alarming faults in hydropower station equipment according to claim 7, characterized in that, The method further includes: adjusting the cross-modal attention allocation weights and / or modal health threshold ranges and / or multi-level fault alarm rules using reinforcement learning based on the on-site verification results.

10. A fault alarm system for hydropower station equipment, characterized in that, The system includes: a data acquisition module for acquiring multiple multi-source heterogeneous monitoring data of hydropower station equipment under multiple historical health conditions; a data preprocessing and fusion module for normalizing multiple historical health conditions to form multiple normalized historical health conditions and fusing multiple heterogeneous monitoring data to form multiple unified standard feature vectors; a benchmark feature template management module for establishing benchmark feature templates based on multiple normalized historical health conditions and corresponding unified standard feature vectors; the data acquisition module is also used to acquire real-time heterogeneous monitoring data of hydropower station equipment under current operating conditions; the data preprocessing and fusion module is also used to normalize the current operating conditions to form the current normalized operating conditions and fuse the real-time heterogeneous monitoring data to form real-time standard feature vectors; and an anomaly identification and alarm decision module for identifying early fault types of hydropower station equipment based on the current normalized operating conditions, real-time standard feature vectors, benchmark feature templates, and a fault feature knowledge base, and for performing multi-level fault alarms based on the early fault type identification results.

Images