Early fault early warning method, system, equipment and medium for group hydropower station

Abstract

The invention discloses a group hydropower station early fault early warning method, system and device and a medium. The method comprises the steps of obtaining original monitoring data and meta-information of a hydropower station; extracting a multi-dimensional key index from the preprocessed data, and generating a standardized feature vector; based on the standardized feature vector, obtaining an encryption gradient parameter through local model training and parameter encryption; the encryption gradient parameters are uploaded to a cloud end, and an optimized global model is obtained through federal learning and an optimization algorithm; on the basis of the optimized global model, when the local model detects abnormal features, a multi-level verification mechanism is triggered, the risk level is evaluated, and a corresponding early warning signal is generated, and on the premise that the data sovereignty and privacy security are guaranteed, the fault early warning capability is qualitatively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault early warning, and in particular to a method, system, equipment and medium for early warning of faults in a group hydropower station. Background Art

[0002] The hydropower industry is currently at a critical stage of digital transformation. However, due to industry characteristics and safety regulations, as well as the particular complexity and sensitivity of fault warnings for hydropower equipment, traditional centralized warning solutions are unable to meet actual needs. Three key issues urgently need to be addressed:

[0003] First, regarding the scarcity of fault warning data, hydropower equipment fault data is not only limited in quantity but also extremely unevenly distributed, exhibiting a typical small sample size characteristic. For example, for Francis turbine generator sets, the frequency of bearing wear failures is closely related to operating conditions. Under high-load continuous operation (>90% of rated load), the incidence of bearing wear failures far exceeds that under normal operating conditions. However, for safety reasons, power plants typically avoid operating units under long-term high-load conditions, further reducing the accumulation of fault samples. Furthermore, the sample size disparity between different fault modes is significant: the sample size ratio for common faults (such as bearing wear) to rare faults (such as runner blade cracks) can reach as high as 50:1, resulting in a model's accuracy rate for identifying rare faults below 20%. Furthermore, the average annual accumulation of valid fault samples for a single power plant is less than three, while traditional methods require at least 200 samples to build a reliable model. Federated learning is needed to address the scarcity of fault samples, balance learning outcomes for common and rare faults, enable cross-plant knowledge sharing, and design early warning model training methods for small sample sizes.

[0004] Secondly, in terms of communication security and efficiency, the special network environment of hydropower stations brings unique challenges. Most power stations are located in remote areas and rely on satellite communication links, which have limited bandwidth (2-4Mbps), high latency (≥500ms) and high costs. For traditional federated learning solutions, the amount of single model parameter updates far exceeds the carrying capacity of satellite links, and at the same time will squeeze communication resources for key businesses such as water situation monitoring. This patent needs to solve the problems of communication complexity and inefficiency, while ensuring the integrity and confidentiality of model updates, and keeping communication overhead within an acceptable range.

[0005] Finally, regarding data privacy and compliance, the issue extends beyond the prohibition on cross-site data transmission and encompasses deeper contradictions. The power industry is subject to strict regulations under the Cybersecurity Multi-Level Protection Regulations and the Power Monitoring System Security Protection Regulations, requiring production data to be stored within local security zones. An audit report from a provincial power grid revealed that 85% of hydropower stations refused to transmit core monitoring data, such as vibration and temperature, due to compliance requirements. However, traditional centralized early warning solutions often struggle to meet these requirements. Centralized data storage increases the risk of attack; Furthermore, third-party technical service providers involved in model development may be able to reverse engineer sensitive information from model parameters. The risk of information leakage during model parameter transmission must be addressed, and localized data storage at each power station must be implemented to ensure compliance with relevant security regulations. Summary of the Invention

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

[0007] Therefore, the present invention provides a group hydropower station early fault warning method, system, equipment and medium to solve the problems of the existing method such as scarce and uneven distribution of fault samples, low communication efficiency, and high risk of data privacy leakage.

[0008] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0009] In a first aspect, the present invention provides a method for early warning of faults in a group hydropower station, comprising:

[0010] Obtaining original monitoring data and meta-information of hydropower stations;

[0011] Extract multi-dimensional key indicators from pre-processed data and generate standardized feature vectors;

[0012] Based on the standardized feature vector, the encrypted gradient parameters are obtained through local model training and parameter encryption;

[0013] Upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms;

[0014] Based on the optimized global model, when the local model detects abnormal features, a multi-level verification mechanism is triggered to assess the risk level and generate corresponding early warning signals.

[0015] As a preferred solution of the group hydropower station early fault warning method described in the present invention, the local model training and parameter encryption include:

[0016] Based on the standardized feature vectors, the diagnostic model is trained locally at each hydropower station;

[0017] Generate gradient parameter information of the local model during training;

[0018] Differential coding technology is used to compress model parameter differences;

[0019] Use homomorphic encryption technology to encrypt and obtain encrypted gradient parameters.

[0020] The beneficial effect of this preferred technical solution is that by locally training the diagnostic model based on standardized feature vectors at each hydropower station and generating, compressing and homomorphically encrypting gradient parameters during the training process, efficient transmission and aggregation model updates are achieved while protecting data privacy, thereby improving the accuracy of fault prediction and the overall security of the system.

[0021] As a preferred solution of the group hydropower station early fault warning method of the present invention, the uploading of the encrypted gradient parameters to the cloud includes:

[0022] Selective parameter transmission technology is used to filter the most important parameters for transmission, and intelligent channel scheduling is combined to optimize transmission efficiency;

[0023] The encrypted gradient parameters are transmitted to the federated learning cloud via satellite link.

[0024] As a preferred solution of the group hydropower station early fault warning method described in the present invention, the cloud-based training and optimization algorithm through federated learning includes:

[0025] The cloud platform receives the encrypted gradient parameters transmitted by each power station, performs weighted aggregation and decryption, and forms a new global model;

[0026] Dynamic importance-aware federated aggregation, comparative knowledge distillation, and fusion physical mechanisms are applied in the cloud to obtain an optimized global model.

[0027] The beneficial effect of this preferred technical solution is that the encrypted gradient parameters transmitted by each power station are received and weightedly aggregated through the cloud platform. After decryption, optimization algorithms such as dynamic importance-aware federated aggregation, comparative knowledge distillation, and fusion physical mechanisms are applied to form and continuously optimize the global model, thereby significantly improving the model's generalization ability and fault prediction accuracy while ensuring data privacy.

[0028] As a preferred solution of the group hydropower station early fault warning method described in the present invention, the triggering of the multi-level verification mechanism includes:

[0029] When the local model detects an abnormal feature, it conducts a horizontal comparison in the local historical database and eliminates instantaneous interference by analyzing historical data under similar working conditions;

[0030] If it is confirmed to be a true anomaly, the remaining useful life of the equipment is predicted.

[0031] As a preferred solution of the group hydropower station early fault warning method described in the present invention, the encryption using homomorphic encryption technology includes:

[0032] Generate public and private keys in the cloud;

[0033] Distribute the generated public key to the edge nodes of each hydropower station;

[0034] The scaling factor is used to convert the gradient parameter in floating point form into integer form for scaling processing;

[0035] The edge node uses the received public key to perform homomorphic encryption on the scaled fully connected layer gradient parameters to obtain encrypted gradient parameters.

[0036] As a preferred solution of the group hydropower station early fault warning method of the present invention, the risk level assessment includes:

[0037] If the abnormal feature intensity is lower than the preset warning threshold and does not continue to deteriorate, and the similarity with the known failure mode is not in the same score range, it is judged as low risk;

[0038] If the abnormal feature intensity reaches the preset warning threshold, the similarity with the typical fault feature library is within a score range, and the key operating parameters are outside the normal range, it is judged as medium risk;

[0039] If either the key indicator exceeds the safety limit or the system's predicted remaining safe operating time is less than a specific time, it is judged to be high risk.

[0040] In a second aspect, the present invention provides an early fault warning system for a group hydropower station, comprising:

[0041] Acquisition module, used to obtain the original monitoring data and meta-information of the hydropower station;

[0042] Feature extraction module, used to extract multi-dimensional key indicators from preprocessed data and generate standardized feature vectors;

[0043] The model training module is used to obtain encrypted gradient parameters through local model training and parameter encryption based on the standardized feature vector;

[0044] The optimization module is used to upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms;

[0045] The anomaly detection module is used to trigger a multi-level verification mechanism based on the optimized global model, evaluate the risk level and generate corresponding early warning signals when the local model detects abnormal features.

[0046] In a third aspect, the present invention provides an electronic device, comprising:

[0047] Memory, used to store programs;

[0048] The processor is used to execute the computer executable instructions, which, when executed by the processor, implement the steps of the group hydropower station early fault warning method.

[0049] In a fourth aspect, the present invention provides a computer-readable storage medium, comprising: when the program is executed by a processor, the steps of implementing the group hydropower station early fault warning method.

[0050] The beneficial effects of the present invention are as follows: the present invention uses feature extraction and standardization technology to extract high-quality feature vectors from data, thereby improving the accuracy of fault diagnosis; uses homomorphic encryption and differential coding technology to achieve efficient and secure data transmission while ensuring data privacy; and uses cloud aggregation and optimization algorithm technology to achieve continuous optimization of the global model, significantly improving the accuracy and reliability of fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them:

[0052] Figure 1 A schematic diagram of the basic flow of a method for early fault warning of a hydropower station group provided by one embodiment of the present invention;

[0053] Figure 2 An operation flow chart of an early fault warning method for a group hydropower station provided by one embodiment of the present invention;

[0054] Figure 3 A schematic diagram of federated learning for a hydropower station providing an early warning method for failures in a group of hydropower stations according to an embodiment of the present invention;

[0055] Figure 4 A diagram of a federated learning training optimization algorithm for an early fault warning method for a group hydropower station provided by one embodiment of the present invention;

[0056] Figure 5 A diagram of a communication efficiency optimization system for a group hydropower station early fault warning method provided by one embodiment of the present invention;

[0057] Figure 6A multi-dimensional privacy protection system diagram of a group hydropower station early fault warning method provided by one embodiment of the present invention;

[0058] Figure 7 A system hardware and software layer model diagram of an early fault warning method for a group hydropower station provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0059] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0060] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a method for early fault warning of a group hydropower station, comprising:

[0061] S100: obtaining original monitoring data and meta-information of the hydropower station;

[0062] S200: Extract multi-dimensional key indicators from preprocessed data and generate standardized feature vectors;

[0063] S300: Based on the standardized feature vector, encrypted gradient parameters are obtained through local model training and parameter encryption;

[0064] S400: Upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms;

[0065] S500: Based on the optimized global model, when the local model detects abnormal features, a multi-level verification mechanism is triggered to evaluate the risk level and generate corresponding early warning signals.

[0066] It should be noted that hydropower stations face a range of operational challenges, including aging equipment, complex operating environments (such as high humidity and severe vibration), frequent load fluctuations, and potential operational errors. These factors can lead to equipment failures and even system crashes, seriously impacting power generation efficiency and safety. Early fault warning systems monitor key parameters (such as vibration, temperature, and operating condition data) in real time. They leverage intelligent sensor terminals and advanced data analysis technologies (such as feature extraction, federated learning, and homomorphic encryption) to conduct distributed collaborative monitoring and fault prediction while ensuring data privacy. These systems can proactively identify abnormal conditions, trigger multi-level verification mechanisms, and issue timely warning signals. This not only helps hydropower stations implement preventive maintenance measures and reduce unplanned downtime, but also optimizes maintenance costs, extends equipment life, and ensures safe and stable operation, thereby improving overall operational efficiency and economic benefits. Therefore, establishing an efficient and reliable early fault warning system is crucial to the sustainable development of hydropower stations.

[0067] Therefore, in order to address the problems of scarce and uneven distribution of fault samples, low communication efficiency, and high risk of data privacy leakage in the above-mentioned existing methods, through steps S100-S500, the intelligent sensor terminal is used to collect and preprocess the original data, and multi-dimensional key indicators are extracted from it to generate standardized feature vectors; based on these feature vectors, the diagnosis model is trained locally in each hydropower station, and the gradient parameters are encrypted and protected by homomorphic encryption technology; differential coding technology is used to compress the model parameter differences and reduce the amount of transmitted data; the cloud platform receives the encrypted gradient parameters, performs weighted aggregation and decryption, and applies optimization algorithms such as dynamic importance-aware federated aggregation, comparative knowledge distillation and fusion physical mechanism to form and continuously optimize the global model, ultimately achieving high-precision fault prediction and timely warning signal output, ensuring the safe and stable operation of hydropower station equipment.

[0068] Example 2, reference Figure 2-Figure 7 , which is an embodiment of the present invention, provides a group hydropower station early fault warning method based on the previous embodiment, including:

[0069] In the embodiment of this application, a distributed architecture is used to achieve coordinated monitoring and early warning of the equipment status of each hydropower station within the group, and an intelligent fault early warning system is built under the premise of ensuring data privacy and security. The entire process covers the entire process from data collection to early warning decision-making, and each module works together to form a closed-loop management, such as Figure 2 shown.

[0070] In an embodiment of the present application, obtaining the original monitoring data of the hydropower station in step S100 includes deploying an intelligent sensor terminal on site at the hydropower station to collect high-frequency vibration signals, temperature time series data and operating parameters in real time; the metadata includes equipment attributes and environmental parameters.

[0071] In an embodiment of the present application, in step S200, the raw data is pre-processed on the edge side, adaptive filtering is used to eliminate power grid frequency interference, and then wavelet decomposition technology is used to separate the noise components in the signal. The feature extraction module extracts key indicators from multiple dimensions in the time domain, frequency domain, and time-frequency domain, such as calculating the energy distribution of a specific frequency band, the pulse characteristics of the statistical waveform, etc., to generate a standardized feature vector. This stage ensures that sensitive data does not leave the station, and all raw signals are desensitized locally.

[0072] In an embodiment of the present application, a diagnostic model is deployed in each hydropower station in step S300 and trained based on the pre-processed feature data. The model adopts a federated learning framework, and in the initial stage, a pre-trained meta-model is sent down from the cloud. The model learns common fault features through the group's historical data. During local training, the model is dynamically decoupled into a common layer and an individual layer: the common layer participates in global knowledge sharing, and the individual layer retains equipment specificity. The gradient parameters generated by the training are screened for importance, and only the parts that contribute significantly to the model update are retained, and then encrypted through the algorithm. The encrypted parameter fragments are bound to the device fingerprint, which is generated by the inherent vibration characteristics of the unit to ensure that each update can be traced back to a specific device. The processed data is transmitted via a satellite link, and the volume is compressed using differential coding technology, with priority given to ensuring the real-time transmission of key parameters.

[0073] In the embodiment of the present application, in step S400, the cloud platform receives the model update parameters encrypted and transmitted by each power station, verifies the legitimacy of the device fingerprint, and filters out unauthorized nodes. The aggregation server securely decrypts the multi-source parameters, combines the data quality score with dynamic weighting, and improves the contribution weight of small sample power stations and high signal-to-noise ratio data. The updated model is encrypted and sent to each power station to achieve knowledge sharing and model iteration, such as Figure 3 shown.

[0074] In the embodiment of the present application, uploading the encrypted gradient parameters to the cloud in step S400 includes:

[0075] Selective parameter transmission technology is used to filter the most important parameters for transmission, and intelligent channel scheduling is combined to optimize transmission efficiency;

[0076] The encrypted gradient parameters are transmitted to the federated learning cloud via satellite link.

[0077] In an optional implementation, the selective parameter transmission technology in step S400 may be Top-K sparsification technology; may also be differential privacy; may also be gradient quantization;

[0078] In an optional embodiment, the Top-K sparsification technique includes retaining only the most important top 10% of the parameters of the locally calculated gradient tensor, setting the remaining parameters to zero, and uploading them to the cloud for aggregation processing.

[0079] In an optional embodiment, differential privacy includes adding noise that conforms to a specific distribution to the locally calculated gradient tensor, and uploading these noisy gradients to the cloud for aggregation processing.

[0080] In an optional embodiment, gradient quantization includes converting the locally calculated gradient tensor into a low-precision format, reducing the number of bits required to represent the gradient through quantization, and then uploading it to the cloud for aggregation processing.

[0081] It should be noted that the Top-K sparsification technique employed in this paper offers significant advantages over differential privacy and gradient quantization, primarily in its efficient data compression and transmission. By retaining only the most important 10% of gradient parameters and setting the rest to zero, Top-K sparsification not only significantly reduces the amount of data required for transmission, lowering communication costs, but also maintains model training accuracy and convergence speed, avoiding the potential loss of accuracy caused by adding noise (differential privacy) or the degradation of numerical precision caused by the quantization process. Furthermore, Top-K sparsification directly targets the most important parts of the gradient, enabling prioritized and accurate transmission of critical information, enhancing the effectiveness and real-time nature of model updates in federated learning. In contrast, while differential privacy offers stronger privacy protection, it may increase computational complexity and impact model performance. While gradient quantization can effectively reduce the size of transmitted data, it may not be the optimal choice for scenarios requiring high precision. Therefore, Top-K sparsification offers a more balanced solution in terms of ensuring communication efficiency and model training effectiveness.

[0082] In this embodiment of the present application, when the local model detects an abnormal feature in step S600, a multi-level verification mechanism is triggered. A horizontal comparison is performed in the local historical database to eliminate transient interference and predict the remaining useful life of the device. The warning information is determined to a risk level based on the following criteria:

[0083] (1) Observation level (low risk): The abnormal characteristics are weak and do not continue to deteriorate. The degree of match with known failure modes is low, and the overall operating parameters of the equipment remain stable.

[0084] (2) Group level (medium-high risk): The abnormal characteristics reach the preset warning threshold, have a high degree of match with the typical fault feature library, and are accompanied by abnormal changes in key operating parameters;

[0085] (3) Emergency level: The characteristic anomaly deteriorates rapidly, and key indicators that may trigger a chain reaction exceed the limit, or the system predicts that the remaining safe operation time is insufficient.

[0086] In the embodiment of the present application, step S400 uses the federated learning cloud training and optimization algorithm, including:

[0087] The cloud platform receives the encrypted gradient parameters transmitted by each power station, performs weighted aggregation and decryption, and forms a new global model;

[0088] Dynamic importance-aware federated aggregation, comparative knowledge distillation, and fusion physical mechanisms are applied in the cloud to obtain an optimized global model.

[0089] In the embodiment of this application, the federated learning training optimization algorithm includes core technologies such as physical-guided prototype space construction, cross-station comparative knowledge distillation, and dynamic importance-aware federated aggregation, which can achieve efficient knowledge transfer and improve rare fault detection capabilities without leaving the local data. Figure 4 As shown, this solves the data scarcity problem unique to the hydropower industry.

[0090] (1) Fusion physical mechanism

[0091] Based on the operating mechanism of hydropower equipment, a dual-channel feature mapping network is established:

[0092] ①Mechanism characteristic channel:

[0093] By building a priori knowledge base through the equipment physical equations (such as the rotor dynamics model), the vibration signal is decomposed into:

[0094] x phy =Φ(vib|w,H)

[0095] Where Φ is a dynamic feature extraction function based on the rotational speed ω and the water head H, which captures the energy distribution of the frequency band that is strongly related to the physical characteristics of the equipment.

[0096] ②Data-driven channel:

[0097] Use deep convolutional networks to automatically learn time-frequency features:

[0098] x data =CNN(vib)

[0099] The prototype vector of each fault category is synthesized by weighting the two-channel features:

[0100]

[0101] The weight λ is dynamically adjusted according to the fault type (e.g., λ = 0.7 for bearing wear and λ = 0.3 for electrical fault), so that the physical prior and data features complement each other organically.

[0102] (2) Cross-station comparative knowledge distillation

[0103] To address the issue of data distribution differences among power plants, a two-way contrastive distillation strategy was designed. Under the federated learning framework, the central server maintains a global prototype pool (containing common feature representations for various types of faults). Each local power plant downloads the global prototype pool and adjusts the local model through horizontal contrastive learning. The contrastive loss function is used to force the local fault prototype vector to align with the global prototype and away from prototypes of other categories, thereby reducing feature differences between power plants. For rare faults with very few samples (such as fewer than 5 samples), a vertical distillation strategy is adopted. The central server aggregates the common features of multiple power plants and generates a synthetic feature vector as a "teacher model"; based on these synthetic features, each local power plant trains a "student model" to match the probability distribution of the teacher's output (through KL divergence loss), thereby migrating rare fault knowledge to the local area.

[0104] ①Horizontal comparison:

[0105] Building a global prototype pool on the federated server After downloading, each power station will compare and learn with the local prototype:

[0106]

[0107] Among them, the molecular part Calculate the cosine similarity between the local prototype vector and the global prototype vector, and construct the optimization target through negative logarithmic loss: the higher the similarity, the smaller the loss, driving the local prototype to align with the global prototype of the corresponding category. This loss function includes similarity calculations for all categories. The optimization process automatically suppresses the similarity between local prototypes and global prototypes of other categories, forming a separation in the feature space. When the temperature coefficient τ approaches 0, the function pays more attention to difficult samples (negative samples with close similarity). As the temperature coefficient τ increases, the function smoothes the similarity distribution between categories. This loss function forces local prototypes to move closer to the global consensus and away from prototypes of other categories.

[0108] ②Vertical distillation:

[0109] For rare fault classes (sample size < 5), a teacher-student model architecture is used:

[0110] Teacher model: Aggregate common features of multiple power stations to generate synthetic feature vectors

[0111] Student model: learns the teacher’s knowledge through KL divergence loss:

[0112]

[0113] Under the premise of protecting data privacy, rare fault characteristics can be migrated across power plants.

[0114] (3) Dynamic Importance-Aware Federated Aggregation

[0115] Traditional federated averaging aggregates model parameters across power plants, assigning weights based solely on data volume and ignoring differences in data quality. To address this issue, a multi-dimensional quality assessment metric is introduced to dynamically adjust aggregation weights. Each power plant calculates its own quality assessment vector, which includes three dimensions: ① Data completeness (percentage of valid samples), which measures data availability; ② Feature discrimination (the ratio of intra-class spacing to inter-class spacing), which assesses the feature's ability to distinguish fault categories; and ③ Physical consistency (the degree of match between the generated features and the physical theoretical model), which verifies whether the features conform to the device mechanism.

[0116] The central server maintains a global query vector (used to guide the direction of weight distribution). Each power station concatenates its own quality vector with the global query vector and inputs it into the gating network (a neural network with a weight matrix). The sigmoid function generates a dynamic aggregation weight:

[0117] w k =σ(W g [q||h k ])

[0118] Among them, q is the global query vector, h k is the characteristic quality vector of power station k, and σ is the sigmoid function.

[0119] This weighting takes into account both data quality and global requirements, rather than relying solely on data volume. For example, if a small power plant demonstrates high discrimination or strong physical consistency in a rare fault category, its weight will be significantly increased, ensuring that its local characteristics are fully represented in the global model. Ultimately, the server performs a weighted aggregation based on the dynamic weights of each power plant, updates the global model, and distributes it to each power plant for the next round of training.

[0120] In the embodiments of the present application, the system communication efficiency optimization system focuses on the optimization method of communication efficiency in the federated learning process, especially the efficient data transmission strategy in a bandwidth-constrained environment; the system communication efficiency optimization system includes gradient value-driven intelligent compression (such as Top-K screening), differential coding and entropy compression collaboration, channel-aware intelligent scheduling and other technical means, aiming to reduce the amount of data transmission and improve transmission efficiency.

[0121] In the embodiment of the present application, the system communication efficiency optimization system, such as Figure 5 As shown, the following steps are included:

[0122] (1) Intelligent compression driven by gradient value

[0123] Traditional federated learning directly transmits complete gradients, and large-scale data transmission can easily lead to congestion in satellite communication links. This paper constructs a parameter importance assessment model that calculates a weighted score of the absolute value of the gradient and its historical update frequency to identify the parameters that contribute most to model updates. Gradient value-driven intelligent compression introduces a dynamic TOP-K screening mechanism, where the K value is the ratio threshold for retaining parameters. The dynamic TOP-K screening process is as follows:

[0124] During the initialization phase, at the beginning of federated training, the model is not yet stable. A loose screening strategy is adopted, retaining the top 15% of high-value gradients with a K value. A protective screening is implemented on the physical association layer, forcing the retention of the top 30% of its parameters. The model convergence speed is evaluated after each round of training. If the loss decrease rate is <5%, the K value is automatically relaxed to 20%.

[0125] During the stable training phase, when the model enters the stable convergence period, adaptive adjustment is started, the K value is dynamically adjusted according to the channel quality, and a momentum compensation mechanism is introduced. For the gradients that are discarded for three consecutive rounds, the update amount is compensated according to the historical average to prevent long-term neglect of important parameters.

[0126] If a severe environment emergency mode occurs, such as heavy rain or snow weather, the communication is monitored to be degraded (such as bit error rate > 10 -3 ): The K value is compressed to 5%, and only the classification layer and physical core layer gradients are transmitted. The local cache queue is enabled to temporarily store the filtered middle-layer features. After the channel is restored, incremental retransmission is performed. The aggregation period is dynamically extended from the standard 1 hour to 3 hours to reduce resource competition caused by frequent updates.

[0127] (2) Differential Coding and Entropy Compression Collaboration

[0128] To solve the redundant transmission problem of continuous updates, a spatiotemporal differential compression mechanism is designed:

[0129] ①Incremental encoding

[0130] Only the parameter difference between the current model and the baseline is transmitted each time, and 32-bit floating point data is converted to 8-bit integers through signed quantization. Combined with run-length encoding technology, continuous zero-value areas are efficiently compressed.

[0131] ②Dynamic dictionary optimization

[0132] A Huffman coding dictionary is constructed based on historical transmission data, assigning short codewords to frequently occurring gradient patterns. Small parameter adjustments (±1 to 3) use a 2-bit code, while large adjustments (>±10) use a 4-bit code.

[0133] (3) Channel-aware intelligent scheduling

[0134] Establish a dynamic communication quality perception system to achieve multi-service collaborative transmission;

[0135] ①Bandwidth prediction model

[0136] By monitoring the signal-to-noise ratio (SNR) and bit error rate (BER), the available bandwidth can be estimated in real time. -3 ), it automatically switches to simplified transmission mode: only the classification layer gradient is sent, and the data volume is reduced to 20% of the normal mode.

[0137] ②Multi-level priority control

[0138] Build a three-level transmission queue management system:

[0139] The emergency queue is used to transmit control instructions, ensuring that the end-to-end delay is controlled within 100ms to meet real-time response requirements;

[0140] The priority queue is dedicated to model update data and is allocated 30% of fixed bandwidth resources to ensure the efficiency and stability of parameter synchronization during the federated learning process.

[0141] The regular queue carries monitoring data and uses the remaining bandwidth in the network for transmission, completing the upload of device status information without affecting high-priority tasks.

[0142] In the embodiment of the present application, the system's multi-dimensional privacy protection system covers the application of homomorphic encryption technology, device fingerprint generation, digital signature, layered encryption processing and time-space binding mechanisms to ensure the security and privacy of data throughout its entire life cycle.

[0143] In an optional embodiment, the homomorphic encryption technology in step S300 may be a homomorphic encryption scheme based on ring fault-tolerant learning (CKKS), or a BFV homomorphic encryption scheme; or a TFHE homomorphic encryption scheme;

[0144] In an optional embodiment, a homomorphic encryption scheme based on ring-based fault-tolerant learning (CKKS) includes applying a scaling factor to the fully connected layer gradients to convert them into integer form to reduce the impact of encryption noise, encrypting these gradients using CKKS and performing weighted average aggregation calculations directly on the cloud in the ciphertext state, and decrypting them with a private key to obtain the global gradients and restore the original numerical accuracy.

[0145] In an optional embodiment, the BFV homomorphic encryption scheme includes applying appropriate scaling processing to the fully connected layer gradients to adapt to the integer operation requirements of BFV, encrypting these processed gradient data using BFV, and performing weighted average aggregation calculation in the encrypted state on the cloud, obtaining the global gradient through private key decryption and restoring the original value.

[0146] In an optional embodiment, the TFHE homomorphic encryption scheme includes appropriately quantizing and encoding the fully connected layer gradients to adapt to the Boolean operation or finite field operation requirements of TFHE, encrypting these processed gradient data using TFHE, and performing logical operations or specific forms of aggregation calculations in the encrypted state on the cloud, and decrypting with a private key to restore the global gradient and recover the original value.

[0147] It should be noted that the present invention adopts a homomorphic encryption scheme based on ring fault-tolerant learning (CKKS), which has significant advantages over BFV and TFHE homomorphic encryption schemes, especially in processing floating-point operations and high-precision requirements; CKKS supports efficient addition and multiplication operations directly on ciphertext and can maintain high numerical accuracy (error is controlled within 1e-5), which is very suitable for computationally intensive tasks such as gradient aggregation in deep learning models. Compared with BFV, although BFV performs well in integer operations, it is not as effective as CKKS in processing scenarios requiring high-precision floating-point numbers; and although TFHE provides extremely high flexibility and security in fine-grained logical operations, its complexity and computational overhead are high, making it unsuitable for large-scale data processing or applications with strong real-time requirements. Therefore, CKKS can not only meet the needs of efficient and accurate model updates in federated learning environments, but also ensure a good balance between communication efficiency and privacy protection.

[0148] In the embodiments of this application, Figure 6 As shown in the figure, to ensure the security of device status information and model parameters during federated learning, the system has built a multi-dimensional privacy protection system, covering data encryption, gradient protection, device identity authentication and secure transmission mechanism during the model training phase.

[0149] In an embodiment of the present application, during the model training phase, a homomorphic encryption scheme based on ring-based fault-tolerant learning (CKKS) is used to provide special protection for sensitive device state information contained in the gradient of the fully connected layer. This scheme supports floating-point operations, and the numerical precision error is controlled within 1e-5, meeting the high requirements of deep learning for computational accuracy while having good scalability and efficient computing capabilities. Through this encryption method, the cloud can directly perform weighted average aggregation operations on the ciphertext gradient without decryption, effectively preventing reverse inference of device status caused by parameter leakage.

[0150] Gradient sensitivity analysis shows that fully connected layers are highly correlated with device characteristics and fault classification results, posing a high risk of reverse engineering and therefore being identified as key protection targets. The system employs a selective encryption strategy: homomorphic encryption is applied only to fully connected layer gradients, while lightweight obfuscation is used for convolutional layers, ensuring overall security while balancing computational overhead.

[0151] The encryption implementation process is as follows: The cloud first completes parameter initialization, generates a public key (pk) and a private key (sk), and distributes the public key to each edge node. The private key is kept by the secure hardware module. The edge side locally converts the gradient of the fully connected layer The data is multiplied by the scaling factor (1e6) and converted to integer form to reduce the impact of encryption noise before encryption. The cloud receives the encrypted gradient data and performs aggregation calculations. Finally, the private key is used to decrypt the global gradient and restore the original scale.

[0152] In addition, the system has designed a device fingerprint generation mechanism as a unique digital identifier for physical devices. Each hydro-generator unit will produce individual vibration characteristics during steady-state operation (such as rated speed). The system collects signals through high-precision sensors and extracts the following three key physical characteristics after adaptive filtering to eliminate environmental noise:

[0153] Frequency stability: Analyzes the standard deviation of the fundamental frequency (1X frequency) amplitude fluctuation to reflect the assembly characteristics of the mechanical structure.

[0154] Sideband fingerprint: Calculates the energy proportion of the 50Hz bandwidth on both sides of the rotation frequency to capture subtle differences in shaft misalignment

[0155] High-frequency resonance markers: Statistical analysis of energy distribution in the 2-5kHz frequency band to identify early signs of blade cavitation and other damage.

[0156] These physical characteristics are combined with the device serial number and iterated through the nationally encrypted SM3 hash algorithm to generate an irreversible, unique device fingerprint. Even for the same model, fingerprint similarity can be less than one in a thousand due to differences in manufacturing tolerances, installation stress, and other factors.

[0157] After the edge device completes local training and generates gradient parameters, it starts the multi-level security packaging process:

[0158] Digital Signature: Use the device-specific SM2 asymmetric key to jointly sign the gradient data hash value, device fingerprint, and nanosecond-accurate timestamp. This signature acts like an electronic seal; any tampering will result in verification failure.

[0159] Layered encryption: For the fully connected layer gradients that reflect device status details, CKKS homomorphic encryption that supports ciphertext operations is used to ensure that the aggregation process can be calculated without decryption; the gradients of other layers are encrypted through SM4 blocks, balancing efficiency and security.

[0160] Space-time binding: Encrypted data, signature fingerprints, timestamps, and other elements are encapsulated into standardized data packets and transmitted to the cloud via satellite links. Each data packet acts like an encrypted letter with an anti-counterfeiting mark, concealing its content and identifying its source.

[0161] In the embodiment of the present application, the system realizes all-round security protection of device status information, model parameters and identity identification in the federated learning process by constructing a multi-dimensional privacy protection system of "encryption-signature-encapsulation", effectively supporting the trusted collaborative training and intelligent decision-making applications of the group-level knowledge sharing network.

[0162] In the embodiments of this application, Figure 7 As shown, the system utilizes an edge-cloud collaborative architecture, enabling localized data storage and processing at the physical level and the construction of a group-level knowledge sharing network at the logical level. The hardware-level functional architecture includes a data acquisition unit, an edge computing unit, and a cloud-based training unit. The data acquisition unit comprises a vibration signal acquisition module, a temperature monitoring module, and an operating parameter acquisition module. The vibration signal acquisition module includes an integrated accelerometer, signal conditioning circuitry, and an anti-aliasing filter, collecting real-time vibration signals from key rotating components (such as bearings and shafting). The temperature monitoring module uses a multi-channel temperature sensor and cold-junction compensation circuit to monitor the temperature of key components such as thrust bearings and stator windings. The operating parameter acquisition module simultaneously records operating parameters such as head height, power generation load, and speed, and involves devices such as current / voltage transformers, magnetoelectric speed sensors, dynamic pressure sensors, and flow switches. The edge computing unit, deployed on the local host computer at the hydropower station, receives data signals from various sensors, performs real-time signal processing and feature extraction, and stores both raw and feature data. It also manages multi-mode communication links, supporting satellite communication interfaces and Beidou short message backup channels. This unit also deploys a lightweight federated learning early warning model, features local model update capabilities, and runs multi-level fault warning decision logic. The cloud-based training unit, comprised of a central federated learning server equipped with high-performance CPU and GPU resources, receives encrypted model parameters or gradient information uploaded from distributed power plants, performs model aggregation and global updates, and distributes the optimized parameters to each edge node.

[0163] In the embodiment of the present application, the functional architecture of the software layer includes a real-time data stream processing engine, a local model inference framework, a communication encryption and privacy protection module, a cloud-based federated learning service cluster and an intelligent early warning decision-making system; the real-time data stream processing engine includes a signal preprocessing layer, which uses a digital filter group and an outlier detection algorithm to reduce noise and clean the original data to ensure the accuracy of subsequent analysis; the local model inference framework is equipped with a neural network inference engine, runs a lightweight fault diagnosis model, and combines an adaptive learning module to implement local incremental training, including gradient calculation and parameter optimization mechanism, to continuously improve model accuracy; the communication encryption and privacy protection module is responsible for the management of encrypted communication processes, adopts a national secret algorithm system to ensure transmission security, and designs a data compression module, applies gradient compression and differential coding technology, optimizes communication efficiency and reduces bandwidth occupancy; the cloud-based federated learning service cluster coordinates the distributed training process, has functions such as gradient decryption and weighted average calculation, evaluates model performance and reliability, and generates a global knowledge base to guide the update of each edge node model; the intelligent early warning decision-making system executes multi-level fault identification and early warning judgment logic based on local model output and cloud knowledge base, providing timely and accurate status feedback for operation and maintenance.

[0164] In the embodiment of the present application, in terms of input design, the system integrates the original monitoring data collected locally with the model parameter information required for federated training; the local input data includes original monitoring data and meta-information input; the original monitoring data includes high-frequency vibration signals, temperature time series data and operating parameters; the high-frequency vibration signal includes a sampling rate of not less than 10kHz, which is used to capture the transient characteristics of rotating parts such as bearings and gears; the temperature time series data includes recording the temperature change curves of key parts such as thrust bearings and stator windings; the operating parameters include speed (rpm), load (MW), head height (m) and guide vane opening (%), reflecting the real-time operating status of the unit. Meta-information input includes equipment attributes and environmental parameters; equipment attributes include unit model, operating years and maintenance records (processed using One-hot encoding); environmental parameters include water temperature, air pressure, and lubricant type (converted into a model-recognizable form through categorical variable encoding); the federated training input mainly includes the local model update information uploaded by each edge node, and the local model parameters include gradient tensors and model differences; the gradient tensor is expressed as: After Top-K sparsification (retaining the top 10% of important parameters), the communication load is reduced; the model difference is expressed as: Differential coding compression technology is used to optimize transmission and further improve communication efficiency.

[0165] In terms of output design, the system achieves the dual goals of knowledge sharing and intelligent decision-making, including federated aggregate output and early warning decision output; the federated aggregate output includes global model parameters, which are expressed as: Among them, ak is the weight enhancement factor of small sample power stations.

[0166] The early warning decision outputs multi-level early warning signals as shown in Table 1:

[0167] Table 1: Grading warning mechanism

[0168]

[0169] It should be noted that this invention overcomes the data leakage dilemma of traditional centralized learning. By deeply integrating vibration signature fingerprints with national secret algorithms, it creates an innovative paradigm of "data remains constant, knowledge is shared." Device fingerprints are generated based on the inherent physical characteristics of the unit, acting like irreplaceable digital genes. This ensures that every model update can be accurately traced back to a specific device, completely eliminating malicious activities such as identity forgery and model contamination. Combining homomorphic encryption with blockchain technology, a full-link protection system of "edge desensitization - transmission encryption - cloud-based signature verification - behavior verification" is constructed to meet the highest security and compliance requirements of the power industry.

[0170] It should be noted that the present invention is aimed at the pain point of limited bandwidth of satellite links in remote hydropower stations, and is the first to create a mechanism-guided gradient value assessment mechanism. By analyzing the mapping relationship between fault characteristic frequency bands and neural network parameters, it intelligently screens parameter updates that are critical to the early warning task and eliminates redundant data transmission. The integration of differential coding and adaptive compression strategies doubles the efficiency of model interaction, ensuring real-time access to key early warning information even in harsh channel environments such as heavy rain. The system has environmental perception capabilities, can dynamically adjust transmission strategies, and seamlessly switch to local cache mode when communication is interrupted to ensure that the early warning function remains online.

[0171] It should be noted that this invention deeply embeds the physical mechanisms of equipment into a machine learning framework, addressing the common problem of small-sample learning in industrial scenarios. Virtual fault data generated by digital twins significantly expands the cognitive boundaries of rare defects; the cross-station collaborative verification mechanism effectively suppresses false alarm interference, making early warning decisions both sensitive and reliable. The system automatically adapts to the differences between different models and operating conditions, achieving group-level knowledge transfer without complex parameter adjustments, providing universal intelligent protection for various types of hydropower units, such as mixed flow and axial flow.

[0172] It should be noted that this invention utilizes a lightweight edge computing architecture and modular design, making it compatible with the smooth upgrade of existing monitoring systems and significantly reducing deployment costs. From large power stations in the Yangtze River Basin to small water plants in mountainous areas, a "local autonomy-cloud collaboration" early warning network can be rapidly established, driving the intelligent transformation of the hydropower industry across the board. This technological system, based on "safety as the foundation, efficiency as the accompaniment, and precision as the foundation," not only reshapes the equipment health management model but also sets an industry benchmark for predictive maintenance of clean energy facilities, contributing to the implementation of the "dual carbon" strategy.

[0173] Example 3 is an embodiment of the present invention. This embodiment is different from the first embodiment in that it provides an early fault warning system for a group hydropower station.

[0174] It should be noted that the technical solution of the early fault warning system of the group hydropower station and the technical solution of the early fault warning method of the group hydropower station mentioned above belong to the same concept. For the details not described in detail in the technical solution of the early fault warning system of the group hydropower station in this embodiment, please refer to the description of the technical solution of the early fault warning method of the group hydropower station mentioned above.

[0175] In this embodiment, a group hydropower station early warning system includes:

[0176] Acquisition module, used to obtain the original monitoring data and meta-information of the hydropower station;

[0177] Feature extraction module, used to extract multi-dimensional key indicators from preprocessed data and generate standardized feature vectors;

[0178] The model training module is used to obtain encrypted gradient parameters through local model training and parameter encryption based on the standardized feature vector;

[0179] The optimization module is used to upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms;

[0180] The anomaly detection module is used to trigger a multi-level verification mechanism based on the optimized global model, evaluate the risk level and generate corresponding early warning signals when the local model detects abnormal features.

[0181] This embodiment further provides an electronic device applicable to a method for early fault warning of a group hydropower station, including:

[0182] Memory and processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement a group hydropower station early fault warning method proposed in the above embodiment.

[0183] This embodiment further provides a storage medium on which a computer program is stored. When the program is executed by a processor, an early fault warning method for a group hydropower station as proposed in the above embodiment is implemented.

[0184] The storage medium proposed in this embodiment and the method for implementing an early fault warning method for a group hydropower station proposed in the above embodiment belong to the same inventive concept. The technical details not described in detail in this embodiment can be referred to the above embodiment, and this embodiment has the same beneficial effects as the above embodiment.

[0185] Through the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented with the help of software and necessary general-purpose hardware, and of course it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a computer's floppy disk, read-only memory (ROM), random access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., including a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0186] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for early warning of failures of a group hydropower station, characterized in that: include: Obtaining original monitoring data and meta-information of hydropower stations; Extract multi-dimensional key indicators from pre-processed data and generate standardized feature vectors; Based on the standardized feature vector, the encrypted gradient parameters are obtained through local model training and parameter encryption; Upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms; Based on the optimized global model, when the local model detects abnormal features, a multi-level verification mechanism is triggered to assess the risk level and generate corresponding early warning signals.

2. The early warning method for failure of a group hydropower station according to claim 1, characterized in that: The local model training and parameter encryption include: Based on the standardized feature vectors, the diagnostic model is trained locally at each hydropower station; Generate gradient parameter information of the local model during training; Differential coding technology is used to compress model parameter differences; Use homomorphic encryption technology to encrypt and obtain encrypted gradient parameters.

3. The early warning method for failure of a group hydropower station according to claim 1 or 2, characterized in that: The step of uploading the encrypted gradient parameters to the cloud includes: Adopt selective parameter transmission technology to filter the most important parameters for transmission, and combine it with channel intelligent scheduling to optimize transmission efficiency; The encrypted gradient parameters are transmitted to the federated learning cloud via satellite link.

4. The early warning method for failure of a group hydropower station according to claim 3, characterized in that: The federated learning and optimization algorithms include: The cloud platform receives the encrypted gradient parameters transmitted by each power station, performs weighted aggregation and decryption, and forms a new global model; Dynamic importance-aware federated aggregation, comparative knowledge distillation, and fusion physical mechanisms are applied in the cloud to obtain an optimized global model.

5. The early warning method for failure of a group hydropower station according to claim 4, characterized in that: The triggering multi-level verification mechanism includes: When the local model detects an abnormal feature, it conducts a horizontal comparison in the local historical database and eliminates instantaneous interference by analyzing historical data under similar working conditions; If it is confirmed to be a true anomaly, the remaining useful life of the equipment is predicted.

6. The early warning method for failure of a group hydropower station according to claim 5, characterized in that: The encryption using homomorphic encryption technology includes: Generate public and private keys in the cloud; Distribute the generated public key to the edge nodes of each hydropower station; The scaling factor is used to convert the gradient parameter in floating point form into integer form for scaling processing; The edge node uses the received public key to perform homomorphic encryption on the scaled fully connected layer gradient parameters to obtain encrypted gradient parameters.

7. The early warning method for failure of a group hydropower station according to claim 6, characterized in that: The assessed risk level includes: If the abnormal feature intensity is lower than the preset warning threshold and does not continue to deteriorate, and the similarity with the known failure mode is not in the same score range, it is judged as low risk; If the abnormal feature intensity reaches the preset warning threshold, the similarity with the typical fault feature library is within a score range, and the key operating parameters are outside the normal range, it is judged as medium risk; If either the key indicator exceeds the safety limit or the system's predicted remaining safe operating time is less than a specific time, it is judged to be high risk.

8. An early fault warning system for a group hydropower station, applying the method according to any one of claims 1 to 7, characterized in that: include: Acquisition module, used to obtain the original monitoring data and meta-information of the hydropower station; Feature extraction module, used to extract multi-dimensional key indicators from preprocessed data and generate standardized feature vectors; The model training module is used to obtain encrypted gradient parameters through local model training and parameter encryption based on the standardized feature vector; The optimization module is used to upload the encrypted gradient parameters to the cloud and obtain the optimized global model through federated learning and optimization algorithms; The anomaly detection module is used to trigger a multi-level verification mechanism based on the optimized global model, evaluate the risk level and generate corresponding early warning signals when the local model detects abnormal features.

9. An electronic device, characterized in that: include: Memory, used to store programs; A processor, configured to load the program to execute the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images