Horizontal double-runner impulse hydro-generator state monitoring method and system

Abstract

The application discloses a horizontal double-runner impulse water turbine generator state monitoring method and system, relates to the technical field of intelligent state monitoring, and comprises the following steps: collecting double-runner multi-modal data of a water turbine generator for preprocessing, generating two groups of standardized data, calculating a difference vector, and constructing a multi-dimensional feature matrix; performing linear transformation based on the multi-dimensional feature matrix, generating a query, key and value matrix, performing near and far field division, obtaining a near field and a far field, calculating near field attention on the basis of the near field, constructing a binary tree on the basis of the far field, calculating far field attention, fusing the near and far field attention, obtaining a final output feature vector, performing state prediction by using a full connection layer, outputting a state label for parameter iterative training and feedback, and obtaining a prediction result; and the application improves the accuracy of state monitoring and identification through effective feature fusion.

Description

Technical Field

[0001] This invention relates to the field of intelligent condition monitoring technology, and in particular to a condition monitoring method and system for a horizontal double-runner impulse turbine generator. Background Technology

[0002] With the continuous development of intelligent technologies, condition monitoring has become an important component of management for many industrial equipment, especially in high-risk industries such as power and machinery. Traditional condition monitoring methods mostly rely on single physical signals (such as vibration, temperature, and current) for fault diagnosis. These methods generally identify the operating status of equipment by collecting operating parameters and performing simple threshold judgments. However, with the increasing complexity of equipment, especially equipment like horizontal double-rotor impulse turbine generators which contain multiple moving parts and involve multi-physics interactions, data from a single sensor can no longer accurately reflect its operating status. In recent years, the emergence of multimodal sensor technology has greatly enhanced the accuracy of equipment condition monitoring systems. Multimodal sensor fusion technology can simultaneously collect different types of signals, such as vibration, temperature, sound waves, and electrical signals, providing more comprehensive data support and thus improving the reliability of fault detection.

[0003] Existing technologies are often time-consuming and memory-intensive when processing long-term series data and high-resolution data. They also rely heavily on neural network-based models and lack effective mechanisms to handle long-term dependencies and complex interactions between multimodal data. This results in certain limitations in the identification of device operating status. Furthermore, the attention allocation for key features is not precise enough. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a method and system for monitoring the condition of a horizontal double-runner impulse turbine generator, which solves the problems of the lack of an effective mechanism in the prior art to handle the complex interaction between long-term dependencies and multimodal data, and the insufficient precision in the attention allocation of key features.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a condition monitoring method for a horizontal double-runner impulse turbine generator, comprising, Multimodal data of the dual runners of the hydro-generator were collected, preprocessed, and two sets of standardized data were generated. The difference vector was calculated, and a multidimensional feature matrix was constructed. Linear transformation is performed based on the multidimensional feature matrix to generate query, key, and value matrices. Near and far field partitioning is performed to obtain near and far fields. Near field attention is calculated based on the near field, and far field attention is calculated after constructing a binary tree based on the far field. The near and far field attentions are fused to obtain the final output feature vector. A fully connected layer is used for state prediction, and the output state label is used for parameter iterative training and feedback to obtain the prediction result. Implement corresponding early warning strategies based on the forecast results.

[0007] As a preferred embodiment of the horizontal double-rotor impulse turbine generator condition monitoring method of the present invention, the method involves: performing a linear transformation based on a multi-dimensional feature matrix to generate a query, key, and value matrix, performing near-field and far-field partitioning to obtain the near-field and far-field regions; calculating the near-field attention based on the near-field region; and constructing a binary tree based on the far-field region to calculate the far-field attention. The near-field and far-field attention are then fused to obtain the final output feature vector, including: A uniformly distributed random initialization method is used to generate a query, key, and value mapping matrix; By using a query, key, and value mapping matrix, a linear transformation is performed on the multidimensional feature matrix to generate a query, key, and value matrix; Based on the query matrix, the query vector is extracted, and the distance between the query vector and each multidimensional feature vector in the multidimensional feature matrix is ​​calculated using Euclidean distance. A distance threshold is set using a rule of thumb. If the distance is less than or equal to the distance threshold, the current multidimensional feature vector is considered to be the near field of the query vector; otherwise, the current multidimensional feature vector is considered to be the far field of the query vector. Based on the near field, near field attention is calculated at full resolution; Based on the far field, all multidimensional feature vectors in the far field are used as the initial set, defined as the root node, and a binary tree is constructed on the root node using a recursive binary splitting method, with a maximum depth set to [value missing]. ; K-means clustering is used to divide the key vector corresponding to the root node into two subsets to form left and right child nodes. When the maximum depth is reached, the splitting stops and leaf nodes are formed. For each leaf node, the aggregation formula is used to obtain the aggregated representations of the key vector and the value vector, resulting in the aggregated key vector and the aggregated value vector. Calculate far-field attention based on aggregated bond vectors; The final output feature vector of the query vector is generated by fusing near-field and far-field attention.

[0008] As a preferred embodiment of the horizontal double-rotor impulse turbine generator condition monitoring method of the present invention, wherein: the condition prediction using a fully connected layer and the output of condition labels include: Based on the final output feature vector, the state of the hydro-generator is predicted through a fully connected layer to obtain the state distribution probability, and the highest probability is selected as the prediction label, including normal, minor fault, and severe fault.

[0009] As a preferred embodiment of the horizontal double-rotor impulse turbine generator condition monitoring method of the present invention, wherein: the parameter iterative training and feedback to obtain the prediction result includes: Define loss functions, including classification cross-entropy loss and contrastive loss function; The loss functions are weighted and fused to generate the final loss function, and the value of the final loss function is minimized. The Adam optimizer is used to iteratively optimize the parameters. During the iteration process, backpropagation is performed. When the loss value of the final loss function no longer decreases significantly, the iteration stops and the iteratively optimized parameters are fed back. The parameters include the query, key-value mapping matrix, weight matrix of the fully connected layer, and bias terms; The new multidimensional feature matrix is ​​linearly transformed until the predicted label is output.

[0010] As a preferred embodiment of the horizontal double-runner impulse turbine generator condition monitoring method X of the present invention, wherein: the step of executing a corresponding early warning strategy based on the prediction results includes: Predictive labels will be used as the basis for monitoring the condition of hydro-generators; When the predicted label is normal, the current monitoring results will be sent to the monitoring personnel. When the predicted label is a minor fault, an alert message should be sent to the monitoring personnel immediately, including an abnormality in the hydro generator, requesting the monitoring personnel to check it immediately. When the predicted label indicates a serious fault, an immediate warning is issued via an audible and visual alarm.

[0011] As a preferred embodiment of the horizontal double-runner impulse turbine generator condition monitoring method of the present invention, wherein: the collected dual-runner multimodal data of the turbine generator is preprocessed to generate two sets of standardized data, including: Using two sets of sensors, multimodal data of the two runners in the impulse turbine generator were collected at each time point, resulting in two sets of multimodal data, including... Multimodal data of the rotor Multimodal data of the rotor; The sensors include vibration, temperature, current, voltage, and water pressure sensors, and the multimodal data includes vibration, temperature, current, voltage, and water pressure data; Denoising and standardization were performed on the two sets of multimodal data to obtain two sets of standardized data, including Standardized data of the rotor Standardized data for the rotor.

[0012] As a preferred embodiment of the horizontal double-runner impulse turbine generator condition monitoring method of the present invention, wherein: the calculation of the difference vector and the construction of a multi-dimensional feature matrix include: Based on two sets of standardized data, their respective standardized vectors are constructed to obtain the standardized vectors for each time point. Rotary normalized vector sum The wheel-standardized vector is calculated by subtraction to determine the difference between two standardized vectors, forming a difference vector at each time point. All data at each time point are concatenated to form a multidimensional feature vector for each time point. These multidimensional feature vectors are then integrated to form a multidimensional feature matrix.

[0013] Secondly, the present invention provides a condition monitoring system for a horizontal double-runner impulse turbine generator, comprising, The data acquisition and construction module is used to collect multimodal data of the dual runners of the hydro-generator, preprocess the data, generate two sets of standardized data, calculate the difference vector, and construct a multidimensional feature matrix. The partitioning and fusion module performs linear transformations based on the multi-dimensional feature matrix, generating query, key, and value matrices. It then performs near-field and far-field partitioning, obtaining near-field and far-field features. Based on the near-field feature, it calculates near-field attention, and based on the far-field feature, it constructs a binary tree and calculates far-field attention. Finally, it fuses the near-field and far-field attention to obtain the final output feature vector. The training and prediction module is used to predict the state using a fully connected layer, output state labels for iterative parameter training and feedback, and obtain the prediction results. The strategy execution module is used to execute corresponding early warning strategies based on the prediction results.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the horizontal double-rotor impulse turbine generator condition monitoring method as described in the first aspect of the present invention.

[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the horizontal double-rotor impulse turbine generator condition monitoring method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By generating a query, key, and value mapping matrix and combining Euclidean distance and distance threshold to effectively distinguish near-field and far-field features of data, this invention can efficiently process long sequences and large-scale data through a hierarchical attention mechanism while preserving global contextual information between multimodal data. Furthermore, by fusing near-field and far-field attention, the attention to key features can be improved. Therefore, this invention can not only extract important information from multimodal data more accurately, but also improve the accuracy of state monitoring and recognition through effective feature fusion. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the condition monitoring method for a horizontal double-runner impulse turbine generator in Example 1.

[0019] Figure 2 This is a structural diagram of the horizontal double-runner impulse turbine generator condition monitoring system in Example 1.

[0020] Figure 3 This is a flowchart of the corresponding early warning strategy executed in Example 1. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

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

[0023] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0024] Example 1, referring to Figures 1-3This is the first embodiment of the present invention, which provides a method for monitoring the condition of a horizontal double-runner impulse turbine generator, including the following steps: S1. Collect multimodal data of the dual runners of the hydro-generator, preprocess it to generate two sets of standardized data, calculate the difference vector, and construct a multidimensional feature matrix; Specifically, multimodal data from the dual runners of the hydro-generator were collected and preprocessed to generate two sets of standardized data, including: Using two sets of sensors, multimodal data of the two runners in the impulse turbine generator were collected at each time point, resulting in two sets of multimodal data, including... Multimodal data of the rotor Multimodal data of the rotor; The sensors include vibration, temperature, current, voltage, and water pressure sensors, and the multimodal data includes vibration, temperature, current, voltage, and water pressure data; Denoising and standardization were performed on the two sets of multimodal data to obtain two sets of standardized data, including Standardized data of the rotor Standardized data for the rotor.

[0025] By simultaneously acquiring multimodal data such as vibration, temperature, current, voltage, and water pressure using two sets of sensors, the operating status of the equipment can be comprehensively reflected. Furthermore, through standardization and noise reduction processing, scale differences and external interference between data points are eliminated, improving data quality and accuracy and providing a reliable foundation for subsequent analysis.

[0026] Furthermore, the difference vector is calculated, and a multidimensional feature matrix is ​​constructed, including: Based on two sets of standardized data, their respective standardized vectors are constructed to obtain the standardized vectors for each time point. Rotary normalized vector sum The wheel-standardized vector is calculated by subtraction to determine the difference between two standardized vectors, forming a difference vector at each time point. All data at each time point are concatenated to form a multidimensional feature vector for each time point. These multidimensional feature vectors are then integrated to form a multidimensional feature matrix.

[0027] By calculating the difference vector between two rotors, the coordinated changes between them can be accurately captured, especially when a fault occurs, as the differences between the rotors are significant, thus improving the sensitivity of fault identification. Furthermore, concatenating and integrating multidimensional feature vectors into a multidimensional feature matrix provides rich input data for subsequent machine learning models, enhancing the model's ability to diagnose equipment faults.

[0028] S2. Perform linear transformation based on the multidimensional feature matrix to generate query, key, and value matrices. Perform near-field and far-field partitioning to obtain near-field and far-field. Calculate near-field attention based on the near-field and far-field attention based on the far-field. Fuse the near-field and far-field attention to obtain the final output feature vector. Use a fully connected layer for state prediction, output state labels for parameter iterative training and feedback, and obtain the prediction result. Specifically, a linear transformation is performed based on the multidimensional feature matrix to generate a query, key, and value matrix. Near-field and far-field partitioning is then performed to obtain the near-field and far-field features. Near-field attention is calculated based on the near-field feature matrix, and far-field attention is calculated after constructing a binary tree based on the far-field feature matrix. The near-field and far-field attention are then fused to obtain the final output feature vector, including: A uniformly distributed random initialization method is used to generate a query, key, and value mapping matrix; By using a uniformly distributed random initialization method to generate the query, key, and value mapping matrix, the overfitting problem in traditional methods is avoided. Especially in large-scale datasets and complex problems, the uniform distribution can distribute weights more evenly, improving the model's generalization ability. This method allows the model to maintain good exploratory behavior in the initial stage, providing a more diverse parameter space for subsequent optimization.

[0029] By using a query, key, and value mapping matrix, a linear transformation is performed on the multidimensional feature matrix to generate a query, key, and value matrix; Based on the query matrix, the query vector is extracted, and the distance between the query vector and each multidimensional feature vector in the multidimensional feature matrix is ​​calculated using Euclidean distance. Calculating the distance between the query vector and each feature vector in the multidimensional feature matrix using Euclidean distance helps to accurately identify which data points are most relevant to the current query vector. This step refines the similarity of each data point, thereby improving the accuracy of fault identification.

[0030] A distance threshold is set using a rule of thumb. If the distance is less than or equal to the distance threshold, the current multidimensional feature vector is considered to be the near field of the query vector; otherwise, the current multidimensional feature vector is considered to be the far field of the query vector. Based on the near field, near-field attention is calculated at full resolution using the following formula:

[0031] In the formula, Indicates the first The query vector and the first Attention between the key vectors corresponding to the multidimensional feature vectors Represents an exponential function. Indicates the first A query vector, This indicates the transpose operation. Indicates the first A key vector of multidimensional feature vectors This represents the dimension of the query or key vector. Indicates the index of a multidimensional feature vector. Indicates belonging to, Indicates the first Near field of each query vector Indicates the first A key vector of multidimensional feature vectors; The method distinguishes between the near and far fields by using a distance threshold set by an empirical rule. This approach improves model processing speed by simplifying the computation process. The near-field portion utilizes full-resolution attention calculations to accurately capture relevant information, while the subsequent far-field portion employs aggregated computations to reduce computational complexity, thus saving significant amounts of memory and computing resources.

[0032] Based on the far field, all multidimensional feature vectors in the far field are used as the initial set, defined as the root node. A binary tree is then constructed based on the root node using a recursive binary splitting method. The maximum depth is set according to relevant book knowledge. ( ); Using K-means clustering ( The key vector corresponding to the root node is divided into two subsets to form left and right child nodes. When the maximum depth is reached, the splitting stops and leaf nodes are formed. For each leaf node, the aggregation formula is used to obtain the aggregated representations of the key vector and the value vector, resulting in the aggregated key vector and the aggregated value vector. The far-field attention is calculated based on the aggregated bond vector, using the following formula:

[0033] In the formula, Indicates the first The first in the layer Query vectors and nodes attention between, Indicates the hierarchical level (corresponding to the depth of the binary tree). Indicates the first Index of nodes in the layer Indicates the first The query vector at the _th ... The set of nodes in the layer (which can be statistically obtained). Represents a node aggregated bond vector, Represents a node The aggregated bond vector; K-means clustering is used to group far-field nodes, and a binary tree is recursively constructed based on the root node, which effectively partitions and manages far-field information. Through this hierarchical aggregation process, the model reduces computational load and maintains high processing efficiency in large-scale data processing. Simultaneously, this method dynamically adjusts the data partitioning recursively, improving the accuracy of data aggregation and reducing accuracy loss due to limited computing resources.

[0034] The final output feature vector of the query vector is generated by fusing near-field and far-field attention, as shown in the formula:

[0035] In the formula, Indicates the first The final output feature vector of each query vector. Indicates the first The value vector of a multidimensional feature vector This indicates the total number of levels (i.e., the maximum depth). Represents a node The aggregated value vector.

[0036] By fusing near-field and far-field attention, this invention effectively balances computational accuracy and efficiency. Near-field attention provides detailed feature information, while far-field attention reduces computational load through aggregation. This strategy ensures comprehensive computation while avoiding performance bottlenecks caused by excessive computation in traditional methods when processing large-scale, high-dimensional data.

[0037] Furthermore, a fully connected layer is used for state prediction, outputting state labels, including: Based on the final output feature vector, the state of the hydro-generator is predicted through a fully connected layer to obtain the state distribution probability, and the highest probability is selected as the prediction label, including normal, minor fault, and severe fault. The formula for predicting the state of the hydro-generator through the fully connected layer is as follows:

[0038] In the formula, Represents the probability distribution of the state. This represents the Softmax activation function. This represents the weight matrix of the fully connected layer. This represents the bias term of the fully connected layer.

[0039] By using a fully connected layer to predict the state of the final output feature vector of the query vector and then using a Softmax activation function to output the state distribution probability, this process effectively transforms the latent information in multimodal data into specific classification labels. This method enables efficient and accurate equipment fault early warning, allowing for timely identification of equipment anomalies based on real-time monitoring data.

[0040] Furthermore, through iterative parameter training and feedback, prediction results are obtained, including: Define loss functions, including classification cross-entropy loss and contrastive loss function; The classification cross-entropy loss is calculated as follows:

[0041] In the formula, This represents the value of the classification cross-entropy loss function. Indicates the total number of time points. This indicates the total number of status categories (1: normal, 2: minor fault, 3: serious fault). Indicates the first Labels for standardized data (belonging to categories) (The value is 1 if it is 1, otherwise it is 0). Represents the natural logarithm function. Indicates the first Each standardized data belongs to the category The predicted probability; The classification cross-entropy loss accurately measures the difference between the model's predicted class probabilities and the true labels. By minimizing the cross-entropy loss, the model can obtain a reasonable probability distribution among the classes, thereby enhancing the accuracy of equipment status prediction. This loss function is particularly suitable for multi-class status classification tasks, such as determining whether a hydroelectric generator is in a normal, minor, or severe fault state. Using this method, the system can efficiently perform real-time monitoring and react promptly.

[0042] The contrast loss function is formulated as follows:

[0043] In the formula, This represents the comparison loss function value. Indicates the first Comparison labels for standardized data (when) The standardized data vector of the rotor and When the normalized data vectors of the rotating wheel correspond to the same real state category, , indicates a positive sample pair, otherwise (representing negative sample pairs) Indicates the first Distance to standardized data, This indicates the operation of retrieving the maximum value. This indicates the maximum tolerance distance (the value of this parameter can be set from 0 to infinity, and its default value can be set to 2). No. The formula for obtaining the distance from standardized data is:

[0044] In the formula, Represents Euclidean distance. Indicates the first In standardized data The standardized data vector of the rotor, Indicates the first In standardized data The standardized data vector of the rotor; Contrastive loss enables the model to effectively distinguish between similar and different samples during training. Through contrastive loss, the system can not only identify the classification differences between normal and faulty systems, but also accurately distinguish between different types of faults. For example, contrastive loss can improve the detection capability of minor faults in hydro-generators during operation, increasing the sensitivity of fault warnings. Secondly, optimizing contrastive loss allows the model to focus more on effective features, thereby improving the accuracy of fault diagnosis. Furthermore, Euclidean distance, as a standard method for measuring the similarity between samples, can directly calculate the "physical" distance between samples. It allows the model to determine whether samples belong to the same category based on the distance between them, with the help of contrastive loss. When the turbine runner of a hydro-generator deviates, Euclidean distance can effectively reflect the subtle differences between the runners, thus helping to determine the operating status of the equipment.

[0045] The loss functions are weighted and fused to generate the final loss function. The final loss function value is minimized by the following formula:

[0046] In the formula, This represents the final loss function value. This represents the value of the classification cross-entropy loss function. This represents the weight of the contrastive loss function value (the value of this weight can be set from 0.1 to 10; to make the contrastive loss contribute more to the final loss and thus enhance the model's ability to distinguish between positive and negative sample pairs, a default value of 0.5 can be used). The Adam optimizer is used to iteratively optimize the parameters. During the iteration process, backpropagation is performed. When the loss value of the final loss function no longer decreases significantly, the iteration stops and the iteratively optimized parameters are fed back. The parameters include the query, key-value mapping matrix, weight matrix of the fully connected layer, and bias terms; The new multidimensional feature matrix is ​​linearly transformed until the predicted label is output.

[0047] By introducing a weighted fusion mechanism, the impact of classification cross-entropy loss and contrastive loss on the final training objective can be flexibly adjusted. Specifically, appropriate weights are set. This allows the model to focus more on classification accuracy or sample discrimination when processing multimodal data. For example, in fault diagnosis of hydro-generators, when the model needs to focus more on fault classification, the weight of the classification cross-entropy loss can be increased; while when the fault type is complex or difficult to distinguish, the weight of the contrastive loss can be increased, thereby enhancing the model's ability to identify minor faults.

[0048] S3. Implement corresponding early warning strategies based on the forecast results; Specifically, based on the forecast results, corresponding early warning strategies will be implemented, including: Predictive labels will be used as the basis for monitoring the condition of hydro-generators; When the predicted label is normal, the current monitoring results will be sent to the monitoring personnel. When the predicted label is a minor fault, an alert message should be sent to the monitoring personnel immediately, including an abnormality in the hydro generator, requesting the monitoring personnel to check it immediately. When the predicted label indicates a serious fault, an immediate warning is issued via an audible and visual alarm.

[0049] By implementing corresponding early warning strategies, this invention forms a hierarchical response mechanism, which not only effectively improves the accuracy and response speed of fault detection, but also reduces manual intervention and improves equipment safety and operation and maintenance efficiency.

[0050] This embodiment also provides a horizontal double-runner impulse turbine generator condition monitoring system, including: The data acquisition and construction module is used to collect multimodal data of the dual runners of the hydro-generator, preprocess the data, generate two sets of standardized data, calculate the difference vector, and construct a multidimensional feature matrix. The partitioning and fusion module performs linear transformations based on the multi-dimensional feature matrix, generating query, key, and value matrices. It then performs near-field and far-field partitioning, obtaining near-field and far-field features. Based on the near-field feature, it calculates near-field attention, and based on the far-field feature, it constructs a binary tree and calculates far-field attention. Finally, it fuses the near-field and far-field attention to obtain the final output feature vector. The training and prediction module is used to predict the state using a fully connected layer, output state labels for iterative parameter training and feedback, and obtain the prediction results. The strategy execution module is used to execute corresponding early warning strategies based on the prediction results.

[0051] This embodiment also provides a computer device applicable to the condition monitoring method of a horizontal double-rotor impulse turbine generator, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the condition monitoring method of the horizontal double-rotor impulse turbine generator as proposed in the above embodiment. The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0052] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the method for monitoring the condition of a horizontal double-rotor impulse turbine generator as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0053] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for condition monitoring of a horizontal double-runner impulse turbine generator, characterized in that: include, Multimodal data of the dual runners of the hydro-generator were collected, preprocessed, and two sets of standardized data were generated. The difference vector was calculated, and a multidimensional feature matrix was constructed. Linear transformation is performed based on the multidimensional feature matrix to generate query, key, and value matrices. Near and far field partitioning is performed to obtain near and far fields. Near field attention is calculated based on the near field, and far field attention is calculated after constructing a binary tree based on the far field. The near and far field attentions are fused to obtain the final output feature vector. A fully connected layer is used for state prediction, and the output state label is used for parameter iterative training and feedback to obtain the prediction result. Implement corresponding early warning strategies based on the forecast results.

2. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 1, characterized in that: The process involves performing a linear transformation based on a multidimensional feature matrix to generate a query, key, and value matrix, followed by near-field and far-field partitioning. Near-field attention is calculated based on the near-field data, and far-field attention is calculated after constructing a binary tree based on the far-field data. The near-field and far-field attention are then fused to obtain the final output feature vector, which includes: A uniformly distributed random initialization method is used to generate a query, key, and value mapping matrix; By using a query, key, and value mapping matrix, a linear transformation is performed on the multidimensional feature matrix to generate a query, key, and value matrix; Based on the query matrix, the query vector is extracted, and the distance between the query vector and each multidimensional feature vector in the multidimensional feature matrix is ​​calculated using Euclidean distance. A distance threshold is set using a rule of thumb. If the distance is less than or equal to the distance threshold, the current multidimensional feature vector is considered to be the near field of the query vector; otherwise, the current multidimensional feature vector is considered to be the far field of the query vector. Based on the near field, near field attention is calculated at full resolution; Based on the far field, all multidimensional feature vectors in the far field are used as the initial set, defined as the root node, and a binary tree is constructed on the root node using a recursive binary splitting method, with a maximum depth set to [value missing]. ; K-means clustering is used to divide the key vector corresponding to the root node into two subsets to form left and right child nodes. When the maximum depth is reached, the splitting stops and leaf nodes are formed. For each leaf node, the aggregation formula is used to obtain the aggregated representations of the key vector and the value vector, resulting in the aggregated key vector and the aggregated value vector. Calculate far-field attention based on aggregated bond vectors; The final output feature vector of the query vector is generated by fusing near-field and far-field attention.

3. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 2, characterized in that: The process of using fully connected layers for state prediction and outputting state labels includes: Based on the final output feature vector, the state of the hydro-generator is predicted through a fully connected layer to obtain the state distribution probability, and the highest probability is selected as the prediction label, including normal, minor fault, and severe fault.

4. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 3, characterized in that: The parameters are iteratively trained and fed back to obtain prediction results, including: Define loss functions, including classification cross-entropy loss and contrastive loss function; The loss functions are weighted and fused to generate the final loss function, and the value of the final loss function is minimized. The Adam optimizer is used to iteratively optimize the parameters. During the iteration process, backpropagation is performed. When the loss value of the final loss function no longer decreases significantly, the iteration stops and the iteratively optimized parameters are fed back. The parameters include the query, key-value mapping matrix, weight matrix of the fully connected layer, and bias terms; The new multidimensional feature matrix is ​​linearly transformed until the predicted label is output.

5. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 4, characterized in that: The step of implementing corresponding early warning strategies based on prediction results includes: Predictive labels will be used as the basis for monitoring the condition of hydro-generators; When the predicted label is normal, the current monitoring results will be sent to the monitoring personnel. When the predicted label is a minor fault, an alert message should be sent to the monitoring personnel immediately, including an abnormality in the hydro generator, requesting the monitoring personnel to check it immediately. When the predicted label indicates a serious fault, an immediate warning is issued via an audible and visual alarm.

6. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 5, characterized in that: The collected dual-runner multimodal data of the hydro-generator are preprocessed to generate two sets of standardized data, including: Using two sets of sensors, multimodal data of the two runners in the impulse turbine generator were collected at each time point, resulting in two sets of multimodal data, including... Multimodal data of the rotor Multimodal data of the rotor; The sensors include vibration, temperature, current, voltage, and water pressure sensors, and the multimodal data includes vibration, temperature, current, voltage, and water pressure data; Denoising and standardizing were performed on the two sets of multimodal data to obtain two sets of standardized data, including Standardized data of the rotor Standardized data for the rotor.

7. The condition monitoring method for a horizontal double-runner impulse turbine generator as described in claim 6, characterized in that: The calculation of the difference vector and the construction of the multidimensional feature matrix include: Based on two sets of standardized data, their respective standardized vectors are constructed to obtain the standardized vectors for each time point. Rotary normalized vector sum The wheel-standardized vector is calculated by subtraction to determine the difference between two standardized vectors, forming a difference vector at each time point. All data at each time point are concatenated to form a multidimensional feature vector for each time point. These multidimensional feature vectors are then integrated to form a multidimensional feature matrix.

8. A condition monitoring system for a horizontal double-runner impulse turbine generator, based on the condition monitoring method for a horizontal double-runner impulse turbine generator according to any one of claims 1 to 7, characterized in that: include, The data acquisition and construction module is used to collect multimodal data of the dual runners of the hydro-generator, preprocess the data, generate two sets of standardized data, calculate the difference vector, and construct a multidimensional feature matrix. The partitioning and fusion module performs linear transformations based on the multi-dimensional feature matrix, generating query, key, and value matrices. It then performs near-field and far-field partitioning, obtaining near-field and far-field features. Based on the near-field feature, it calculates near-field attention, and based on the far-field feature, it constructs a binary tree and calculates far-field attention. Finally, it fuses the near-field and far-field attention to obtain the final output feature vector. The training and prediction module is used to predict the state using a fully connected layer, output state labels for iterative parameter training and feedback, and obtain the prediction results. The strategy execution module is used to execute corresponding early warning strategies based on the prediction results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the horizontal double-rotor impulse turbine generator condition monitoring method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the horizontal double-rotor impulse turbine generator condition monitoring method according to any one of claims 1 to 7.

Images