Open source osmon-based pumped storage power station capital construction period equipment parameter multi-source data acquisition and monitoring system

Abstract

The application discloses a kind of based on open source Hong Meng's pumped storage power station capital construction period equipment parameter multi-source data acquisition and monitoring system, belong to electric power engineering digitization and intelligent construction technical field. Including distributed intelligent acquisition layer and multi-source data fusion and intelligent analysis layer, the distributed intelligent acquisition layer is based on the data interface and control interface of pre-set connection multi-source data fusion and intelligent analysis layer, the distributed intelligent acquisition layer is used to collect multi-source equipment parameters, edge pre-processing and preliminary diagnosis, the multi-source data fusion and intelligent analysis layer are used to the fusion and depth intelligent analysis of multi-source heterogeneous data from distributed intelligent acquisition layer. The application realizes unified, efficient, intelligent data acquisition network, provides advanced, precise equipment state prediction capability, enhances the depth correlation and explainability of fault diagnosis, optimizes overall system resources and efficiency.

Description

Technical Field

[0001] This invention relates to the field of digital and intelligent construction technology for power engineering, specifically to a multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of a pumped storage power station based on the open-source HarmonyOS. Background Technology

[0002] Pumped storage power station infrastructure projects are massive in scale, involving high-intensity collaborative operations of various key equipment such as tunnel boring machines, tower cranes, and large pump sets in complex environments. Real-time and accurate monitoring of the operational status of these devices (such as vibration, temperature, and current parameters) is crucial for ensuring construction safety, progress, and quality. Currently, monitoring methods in this field mainly face the following challenges:

[0003] First, there is a data silo problem at the data acquisition level. Various devices are often equipped with proprietary acquisition modules with different protocols, resulting in data that cannot be uniformly accessed and exchanged, forming data silos and making it difficult to conduct global collaborative analysis.

[0004] Secondly, the existing monitoring systems have limited intelligence. Most systems only achieve remote telemetry of data and simple threshold alarms, which is a reactive mode. They cannot provide early warning of equipment degradation trends, let alone predictive fault diagnosis. Manual fault diagnosis relies on experience, which is inefficient and difficult to guarantee accuracy.

[0005] Secondly, there is a lack of in-depth analysis capabilities regarding equipment status. Equipment failures are often the result of the combined effects of multiple source parameters (time-series signals) and inter-device relationships (spatial topology). Existing methods either focus on signal processing for single devices or lack the ability to model the interrelationships in complex device networks, making it difficult to uncover deep-seated failure mechanisms and propagation paths from massive, heterogeneous data.

[0006] The open-source HarmonyOS operating system, with its distributed architecture, kernel security, and seamless cross-device collaboration, offers new possibilities for building a unified industrial data acquisition framework. Meanwhile, deep learning technologies, represented by convolutional neural networks, graph neural networks, and Transformers, have demonstrated powerful advantages in processing time-series and graph-structured data. Therefore, how to deeply integrate the distributed hardware collaboration capabilities of open-source HarmonyOS with advanced deep learning algorithms to build an infrastructure-phase equipment monitoring system capable of intelligent sensing, integrated diagnostics, and predictive maintenance has become an urgent technical problem to be solved.

[0007] To address the aforementioned issues, there is an urgent need for a multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of pumped storage power stations based on the open-source HarmonyOS, which can solve the problems existing in traditional methods. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of a pumped storage power station based on the open-source HarmonyOS. This system realizes a unified, efficient, and intelligent data acquisition network, provides advanced and accurate equipment status prediction capabilities, enhances the deep correlation and interpretability of fault diagnosis, and optimizes the overall system resources and efficiency.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of a pumped storage power station based on the open-source HarmonyOS includes: a distributed intelligent acquisition layer and a multi-source data fusion and intelligent analysis layer. The distributed intelligent acquisition layer is connected to the multi-source data fusion and intelligent analysis layer based on a preset data interface and control interface.

[0011] The distributed intelligent acquisition layer is used to acquire parameters from multiple sources, perform edge preprocessing, and conduct preliminary diagnosis.

[0012] The multi-source data fusion and intelligent analysis layer is used to fuse and perform deep intelligent analysis on multi-source heterogeneous data from the distributed intelligent acquisition layer. The deep intelligent analysis includes performing deep analysis on the fused multi-source heterogeneous data based on a spatiotemporal fusion deep learning model, and outputting device health status, parameter prediction and anomaly warning.

[0013] The spatiotemporal fusion deep learning model includes a temporal feature extraction branch, a spatial correlation mining branch, and a global correlation mining branch. The temporal feature extraction branch is constructed based on an improved one-dimensional convolutional neural network structure, the spatial correlation mining branch is constructed based on an improved graph convolutional network structure, and the global correlation mining branch is constructed based on a Transformer encoder structure.

[0014] Furthermore, the distributed intelligent acquisition layer includes multiple HarmonyOS intelligent acquisition terminals and one or more edge intelligent clusters. The HarmonyOS intelligent acquisition terminals are connected to physical sensors to perform signal acquisition, local feature extraction and lightweight diagnosis, and output terminal perception data packets.

[0015] Each edge intelligent cluster is connected to a group of HarmonyOS intelligent acquisition terminals serving the same master device or the same construction area, for fusing multiple terminal perception data packets received and outputting cluster-level fused feature data packets.

[0016] Furthermore, the terminal-aware data packet includes at least a unique terminal identifier, a timestamp based on precise clock synchronization, a device micro-feature vector, and local diagnostic results, wherein the local diagnostic results include local anomaly confidence and preliminary anomaly type identifier.

[0017] Furthermore, the cluster-level fused feature data package includes at least: a unique identifier for the device cluster, a cluster-level timestamp, a device cluster health index, a multi-dimensional fused feature vector, and a summary of the cluster-level anomaly diagnosis report.

[0018] Furthermore, the multi-source data fusion and intelligent analysis layer includes a multi-source data fusion center, a spatiotemporal fusion deep learning model, and a dynamic knowledge graph and decision module. The edge intelligent cluster is connected to the multi-source data fusion center, the multi-source data fusion center is connected to the spatiotemporal fusion deep learning model, and the spatiotemporal fusion deep learning model is connected to the dynamic knowledge graph and decision module.

[0019] The multi-source data fusion center is used to receive and process cluster-level fusion feature data packets from all edge intelligent clusters, perform global spatiotemporal alignment and topology association, and output standardized spatiotemporal feature tensors and global device topology adjacency matrices.

[0020] The spatiotemporal fusion deep learning model is used to perform deep analysis on the spatiotemporal feature tensor and the global device topology adjacency matrix, and output device health status, parameter prediction and anomaly warning.

[0021] The dynamic knowledge graph and decision-making module is used to perform causal reasoning and decision generation on the model output, and generate control instructions that are fed back to the distributed intelligent acquisition layer.

[0022] Furthermore, the improved one-dimensional convolutional neural network structure includes an input layer, a hidden layer, and an output layer. The hidden layer includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a second max pooling layer, a fifth convolutional layer, and a global mean pooling layer.

[0023] Furthermore, the first convolutional layer comprises a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 1, and 64 channels, and a ReLU activation function layer; the first max pooling layer and the second max pooling layer are both max pooling layers with a pooling region size of 2 and a stride of 2; the second convolutional layer and the third convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 128 channels, and a ReLU activation function layer; and the fourth convolutional layer and the fifth convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 256 channels, and a ReLU activation function layer.

[0024] Furthermore, the output layer is a classification layer and uses the softmax activation function.

[0025] Furthermore, the improved graph convolutional network structure is specifically as follows:

[0026] In the message passing process of the original graph convolutional network structure, an attention bias mechanism based on node betweenness centrality is introduced, and neighboring nodes are ranked and sampled according to their importance.

[0027] Furthermore, the output of the dynamic knowledge graph and decision-making module includes a visualized fault analysis report, predictive maintenance work orders, and edge task instruction data packets sent to the distributed intelligent acquisition layer.

[0028] In summary, the present invention has at least one of the following beneficial technical effects:

[0029] 1. This invention realizes a unified, efficient, and intelligent data acquisition network. Through a two-tier architecture of HarmonyOS-based intelligent acquisition terminals and edge intelligent clusters, it solves the problem of accessing multi-source heterogeneous devices. The terminals possess local intelligent processing capabilities, and the clusters achieve multi-sensor collaborative fusion, thereby improving data value density from the source.

[0030] 2. It provides advanced and accurate equipment status prediction capabilities. The core spatiotemporal fusion deep learning model, through the parallel and fusion of temporal feature extraction branches, spatial correlation mining branches, and global correlation mining branches, can simultaneously uncover the internal temporal evolution patterns of equipment, the topological correlation effects between equipment, and global implicit dependencies across spatiotemporal dimensions. This enables the system to identify early, subtle fault signs from multidimensional data, achieving a leap from threshold alarms to probabilistic predictions, and significantly improving the proactiveness of operation and maintenance.

[0031] 3. Enhanced depth and interpretability of fault diagnosis: The system combines data-driven conclusions from deep learning models with a knowledge graph containing domain knowledge through a dynamic knowledge graph and decision-making module, enabling automated causal reasoning. This provides a clear logical chain from phenomenon to root cause, greatly improving the scientific rigor and efficiency of decision-making.

[0032] 4. Optimized overall system resources and performance: Adopted an edge-center collaborative computing power allocation strategy. Lightweight processing and diagnostic tasks are offloaded to HarmonyOS intelligent acquisition terminals and edge intelligent clusters, reducing invalid data transmission; complex global models run on the multi-source data fusion and intelligent analysis layer. Simultaneously, the system can dynamically adjust the edge-side acquisition and analysis strategy by issuing edge task instruction data packets based on the analysis results, achieving a dynamic optimal balance between system resources and monitoring accuracy. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0034] Figure 2 A schematic diagram of an improved one-dimensional convolutional neural network structure;

[0035] Figure 3A schematic diagram of the improved graph convolutional network structure. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0037] like Figure 1 As shown, this invention provides a multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of a pumped storage power station based on the open-source HarmonyOS, including:

[0038] A distributed intelligent acquisition layer and a multi-source data fusion and intelligent analysis layer, wherein the distributed intelligent acquisition layer is connected to the multi-source data fusion and intelligent analysis layer based on a preset data interface and control interface;

[0039] The distributed intelligent acquisition layer is used to acquire parameters from multiple sources, perform edge preprocessing, and conduct preliminary diagnosis.

[0040] The multi-source data fusion and intelligent analysis layer is used to fuse and perform deep intelligent analysis on multi-source heterogeneous data from the distributed intelligent acquisition layer. The deep intelligent analysis includes performing deep analysis on the fused multi-source heterogeneous data based on a spatiotemporal fusion deep learning model, and outputting device health status, parameter prediction and anomaly warning.

[0041] The spatiotemporal fusion deep learning model includes a temporal feature extraction branch, a spatial correlation mining branch, and a global correlation mining branch. The temporal feature extraction branch is constructed based on an improved one-dimensional convolutional neural network structure, the spatial correlation mining branch is constructed based on an improved graph convolutional network structure, and the global correlation mining branch is constructed based on a Transformer encoder structure.

[0042] The distributed intelligent acquisition layer includes multiple HarmonyOS intelligent acquisition terminals and one or more edge intelligent clusters;

[0043] 1. First, let's introduce the HarmonyOS intelligent data acquisition terminal:

[0044] HarmonyOS intelligent data acquisition terminal is an embedded intelligent unit deployed on various construction machinery (such as tunnel boring machines and tower cranes) or key facilities (such as substations and pump rooms). Its hardware core is an industrial-grade system-on-a-chip running the open-source HarmonyOS operating system, which integrates dedicated signal conditioning circuits, multi-mode communication chips and necessary AI acceleration units.

[0045] The input to the HarmonyOS intelligent data acquisition terminal consists of three parts. The first part is the raw physical signals from high-precision industrial sensors, such as the analog voltage signal output by a vibration accelerometer, the resistance change signal from a platinum resistance temperature sensor, and the secondary current signal from a current transformer. The second part is a terminal configuration instruction data packet from its respective edge intelligent cluster or directly from the upper-layer system. This data packet is a structured set of digital instructions containing settings for the terminal's operating mode, such as the precise value of the sampling frequency, the type and cutoff frequency parameters of the digital filter, the alarm threshold value for local threshold judgment, the micro-AI model file for lightweight diagnostics, and specific operating mode instructions triggered by construction progress. The third part is a precision clock synchronization signal that ensures time consistency across the entire network. This signal is continuously acquired through the open-source HarmonyOS distributed soft bus protocol.

[0046] The internal processing of the HarmonyOS intelligent data acquisition terminal is a multi-stage pipeline. First, signal conditioning and digitization are performed, involving anti-aliasing filtering and high-speed analog-to-digital conversion of the input raw physical signal to generate a sequence of raw digital signals at equal time intervals. Next, the device micro-feature extraction stage begins. The microprocessor performs real-time calculations on the raw digital signal sequence, generating a set of device micro-feature vectors representing the instantaneous state of the device. These vectors include time-domain statistical features (such as RMS value, peak value, and kurtosis) and frequency-domain transform features (such as the amplitudes of the first few harmonics obtained through Fast Fourier Transform). Subsequently, a local lightweight intelligent diagnostic module is activated, executing in parallel rapid judgment based on a rule engine and inference from a micro-AI model loaded within the terminal. The rule engine compares the device micro-feature vector with alarm thresholds in the terminal configuration command data packet, triggering a primary alarm; the AI ​​model analyzes the raw digital signal sequence or device micro-feature vector, outputting a quantified local anomaly confidence level and a qualitative preliminary anomaly type identifier. Finally, all results are encapsulated, especially the local lightweight intelligent diagnostic module, which can be implemented using some lightweight deep learning models. This invention is merely an application and does not limit its specific structure, as long as it can achieve the functions described in this invention.

[0047] The output of the HarmonyOS intelligent acquisition terminal is a uniformly formatted terminal perception data packet. This data packet is a core data unit sent to its corresponding edge intelligent cluster via the HarmonyOS distributed soft bus. The packet precisely contains the following fields: a unique terminal identifier to uniquely identify the terminal; a timestamp generated based on precise clock synchronization; a master device identifier indicating the master device to which the terminal belongs; a device micro-feature vector generated by the feature extraction stage; a local diagnostic result generated by the local diagnostic stage, including local anomaly confidence and preliminary anomaly type identifier; a data quality identifier reflecting signal quality; and an optional high-fidelity data index for tracing back the original data fragments stored locally when needed.

[0048] 2. Next, we will introduce the edge intelligent cluster. Each edge intelligent cluster is connected to a group of HarmonyOS intelligent acquisition terminals that serve the same main device or the same construction area.

[0049] The edge intelligent cluster is not a physical entity, but a collaborative computing group formed by multiple HarmonyOS intelligent acquisition terminals serving the same large host device or located in the same construction physical area through the distributed capabilities of open source HarmonyOS. One of the terminals or dedicated gateways with stronger computing power is usually dynamically elected as the cluster head node, responsible for coordinating communication with the center.

[0050] Its input consists primarily of data packets from its members. It continuously receives terminal perception data packets reported by each HarmonyOS intelligent acquisition terminal within the cluster, forming a set of terminal perception data packets. Additionally, it receives edge task instruction data packets from the upper-layer multi-source data fusion and intelligent analysis layer. These packets contain descriptions of complex analysis tasks to be executed at the cluster level, dedicated algorithm models for multi-sensor information fusion, and resource allocation strategies.

[0051] Its internal processing flow is as follows: The cluster head node's processing of input aims to achieve cross-sensor information synergy. The first step is spatiotemporal alignment, which interpolates and aligns all input data streams based on the high-precision timestamps in the sensing data packets from each terminal, generating cluster-level multi-source data slices with fully synchronized timelines. The second step is the core multi-source feature fusion. The cluster head loads and runs the fusion model issued by the edge task instruction data packet. This model is typically a lightweight neural network used to learn and integrate device micro-feature vectors from multiple sensors such as vibration, temperature, and current, uncovering their intrinsic correlations and outputting a deeper and more representative multi-dimensional fused feature vector. The third step is cluster-level health assessment. Based on the fused features, an overall device cluster health index is calculated, generating a more comprehensive and reliable cluster-level anomaly diagnostic report summary than single-terminal diagnosis. Finally, the data is refined and cached according to a strategy for later uploading.

[0052] Its output is as follows: The core output of the edge intelligent cluster is a cluster-level fusion feature data packet, which is a key data product uploaded to the central analysis layer. This data packet has a clear structure and includes: a unique identifier for the device cluster; an aligned cluster-level timestamp; a device cluster health index reflecting the overall health of the devices; a multi-dimensional fusion feature vector containing deep state information; a summary of the cluster-level anomaly diagnosis report; and a list of associated terminal indexes participating in this data fusion. In addition, the edge intelligent cluster also outputs a terminal configuration instruction data packet, used to distribute new configuration parameters or control commands to each terminal within the cluster.

[0053] The multi-source data fusion and intelligent analysis layer includes a multi-source data fusion center, a spatiotemporal fusion deep learning model, and a dynamic knowledge graph and decision module. The edge intelligent cluster is connected to the multi-source data fusion center, the multi-source data fusion center is connected to the spatiotemporal fusion deep learning model, and the spatiotemporal fusion deep learning model is connected to the dynamic knowledge graph and decision module.

[0054] They will be introduced separately:

[0055] 1. Multi-source data fusion center

[0056] The multi-source data fusion center is the first stop for data after it enters the center from the edge. Its task is to "simplify complexity and unify standards" to prepare high-quality input for subsequent deep learning models.

[0057] Its input is: a set of cluster-level fused feature data packets periodically reported by all edge intelligent clusters at the power plant infrastructure site;

[0058] Its internal processing flow consists of three steps. First, global spatiotemporal alignment: the center uses a higher-precision time base to timestamp-align all input cluster-level fused feature data packets and performs possible data interpolation, ensuring that data from different device clusters are temporally comparable. Second, topology association: based on prior knowledge such as the power plant's electrical wiring diagram, process piping diagram, and equipment layout diagram, the system constructs a global device topology graph. Nodes in the graph represent the devices or groups of devices corresponding to each edge intelligent cluster, while edges represent the physical connections (such as pipes, cables), energy flow, or information flow relationships between them. Finally, input tensor construction: the center organizes and arranges the aligned multidimensional fused feature vectors of each device cluster according to the cluster's node position in the global device topology graph, forming a well-structured, standardized spatiotemporal feature tensor. Simultaneously, the node connection relationships in the global device topology graph are transformed into a mathematical global device topology adjacency matrix.

[0059] Its output consists of two strictly corresponding data structures from the multi-source data fusion center: a standardized spatiotemporal feature tensor and a global device topology adjacency matrix. The former is a three-dimensional data cube, with dimensions representing device nodes, time series, and feature dimensions, respectively; the latter is a two-dimensional matrix that mathematically encodes the spatial and functional relationships between devices. These two outputs together constitute the complete input context for the deep learning model.

[0060] 2. Spatiotemporal Fusion Deep Learning Model

[0061] The spatiotemporal fusion deep learning model is the core algorithm of this system. It is a carefully designed multi-branch hybrid neural network that integrates temporal processing, spatial graph convolution and global attention mechanisms.

[0062] (1) Its input is: the input of the model is directly derived from the output of the multi-source data fusion center, namely the standardized spatiotemporal feature tensor and the global device topology adjacency matrix.

[0063] (2) Its internal processing flow: The model adopts a parallel branch architecture to process information of different dimensions, specifically including:

[0064] Temporal Feature Extraction Branch: This branch focuses on the dynamic evolution within a single device. It takes temporal slices of individual device nodes from a normalized spatiotemporal feature tensor as input. Its core is an improved one-dimensional convolutional neural network. This network abandons traditional large convolutional kernels, employing multiple stacked layers of small-scale (e.g., 1x3) dilated causal convolutional layers. By exponentially increasing the dilation rate layer by layer, the network achieves a broad temporal receptive field covering everything from instantaneous impacts to long-term slow wear without significantly increasing the number of parameters. Residual connections and batch normalization are introduced after each convolutional block to ensure training stability. At the end of the branch, a global average pooling layer is used to average each feature channel across all time steps, thereby compressing variable-length temporal data into a fixed-length deep temporal feature encoding. This effectively reduces model parameters and suppresses overfitting. Its specific structure is as follows: Figure 2 As shown, it includes: an input layer, a hidden layer and an output layer. The hidden layer includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a second max pooling layer, a fifth convolutional layer and a global mean pooling layer.

[0065] The first convolutional layer comprises a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 1, and 64 channels, and a ReLU activation function layer. The first max pooling layer and the second max pooling layer are both max pooling layers with a pooling region size of 2 and a stride of 2. The second convolutional layer and the third convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 128 channels, and a ReLU activation function layer. The fourth convolutional layer and the fifth convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 256 channels, and a ReLU activation function layer.

[0066] The output layer is a classification layer and uses the softmax activation function.

[0067] Spatial Association Mining Branch: This branch focuses on the mutual influence within the device network. It takes the initial features of all nodes (or features after interaction with other branches) and the global device topology adjacency matrix as input. Its core is an improved graph convolutional network. The improvements are mainly reflected in two aspects: First, during message passing, when calculating the attention coefficient of neighboring nodes, a prior bias based on node betweenness centrality is introduced, making the network naturally pay more attention to the state of device nodes in key hub positions when aggregating information; second, the importance of each target node's neighbors is ranked and sampled, prioritizing the aggregation of neighbor information most relevant to the current node's state and with the greatest influence. This simulates the fault propagation path, improving the model's robustness and interpretability. This branch outputs spatial association feature encodings containing spatial context information for each node. The improved graph convolutional network structure will be described in detail below:

[0068] Firstly, the problem of graph matching falls under the conventional techniques of the original graph convolutional network structure, and will not be elaborated upon here. However, the original graph convolutional network structure has some problems. How to ensure the representativeness of the selected key nodes, and how to ensure that the neighboring nodes can completely represent the structure of the graph during the acquisition of the node's neighborhood, are some key issues that urgently need to be addressed. To solve these problems, this invention proposes an improved graph convolutional network structure, such as... Figure 3 As shown, firstly, three methods for calculating the centrality of network nodes are used to measure and sort the nodes, selecting key nodes. Then, the centrality is used to sort the neighboring nodes and obtain the node neighborhoods sequentially. The model mainly includes:

[0069] (1) Representative node selection: Node centrality is calculated using three network node centrality measurement methods in SNA (including degree centrality, proximity centrality and betweenness centrality) to obtain an ordered sequence of nodes, and the optimal representative node is selected by comparison.

[0070] (2) Neighbour node sort (NNS): Sort the neighbor nodes by centrality and select the neighbor nodes in order of centrality.

[0071] (3) Neighborhood node normalization: Normalize the node neighborhood into a grid structure and use it as the input of the convolutional neural network.

[0072] (4) Feature learning: Normalized graph structure data undergoes feature learning through two convolutional layers and a fully connected layer.

[0073] Global Association Mining Branch: This branch focuses on discovering implicit long-range dependencies across devices and time. It flattens the entire normalized spatiotemporal feature tensor and feeds it into a Transformer encoder layer. By adding spatiotemporal location encoding to the data and leveraging its multi-head self-attention mechanism, this branch can automatically calculate and quantify the strength of the state association between any two devices (regardless of spatial distance) at any two moments (regardless of time interval). For example, it can discover the association between abnormal vibration of a water pump and the operation of an upstream valve several hours earlier. The branch outputs a globally associated feature encoding that captures complex global interactions. Since no improvements are made to the Transformer encoder here, but only standard usage, it will not be described in detail.

[0074] Feature Fusion and Multi-Task Decision Making: The deep temporal feature encoding, spatial correlation feature encoding, and global correlation feature encoding output from the three branches mentioned above are fed into a feature fusion unit (usually composed of a cross-attention mechanism or multiple fully connected layers) for deep fusion. The fused features are then fed into a parallel multi-task output layer. This layer contains three sub-headers: a health diagnosis head that outputs a device health status matrix, providing a specific health score and failure mode probability distribution for each device; a trend prediction head that outputs a key parameter prediction sequence, predicting the trajectory of key operating parameters over a future period; and an anomaly warning head that outputs a list of system-level anomaly warning events, listing the predicted anomaly events, their timing, involved devices, and confidence levels.

[0075] (3) Its output is: the output of the spatiotemporal fusion deep learning model, namely the above-mentioned equipment health status matrix, key parameter prediction sequence and system-level abnormal early warning event list.

[0076] 3. Dynamic Knowledge Graph and Decision Module

[0077] This module endows data-driven results with knowledge-driven interpretability and completes the decision-making loop.

[0078] Its inputs are primarily a list of system-level anomaly warning events and a device health status matrix output by a spatiotemporal fusion deep learning model. It also integrates a domain knowledge base containing device principles, typical fault trees, and maintenance procedures.

[0079] Its internal processing flow is as follows: First, the knowledge graph is updated and causal reasoning is performed. The module adds new events from the warning event list as nodes to the pre-built infrastructure-phase equipment knowledge graph in real time. Based on the rules defined in the graph (such as "bearing wear may lead to increased vibration") and entity relationships (such as "main pump-drive-motor"), it performs automated causal tracing reasoning to find the root cause of the anomaly and the potential impact propagation path. Next, decision generation is performed. Combining the reasoned root cause with maintenance plans, spare parts information, and construction windows in the domain knowledge base, detailed and executable predictive maintenance work order suggestions are automatically generated. Finally, based on the new fault modes or optimization strategies identified in this analysis, new edge task instruction data packages are generated.

[0080] Its output consists of three types of results: First, a visualized fault analysis report for operation and maintenance personnel, which clearly displays the cause-and-effect chain of the fault in the form of a graph; second, predictive maintenance work orders that can be directly issued and executed; and third, edge task instruction data packets used to optimize the behavior of lower-level perception. These data packets will be fed back to the corresponding edge intelligent cluster, thereby starting a new round of more targeted intelligent perception cycle.

[0081] To illustrate the implementation methods and technical effects of the present invention, the following is a description of an example of monitoring the status and providing fault warnings for the main drive system of a tunnel boring machine (TBM) during the excavation of an underground powerhouse of a pumped storage power station.

[0082] Scenario: At the excavation face of the underground powerhouse of a power station, a tunnel boring machine (TBM-001) is excavating through rock strata. Its main drive system is the core power component; a failure in this system would lead to a complete shutdown of the entire line.

[0083] System deployment and operation:

[0084] Deployment: Multiple HarmonyOS intelligent acquisition terminals are deployed in key parts of the TBM-001 main drive system (such as the main bearing housing, drive motor, and gearbox) to collect vibration, temperature, and current signals, respectively. These terminals automatically form an edge intelligent cluster (cluster ID: CLUSTER-TBM001) serving the TBM-001.

[0085] Data Acquisition and Edge Processing: Each terminal continuously acquires raw physical signals and receives terminal configuration instruction data packets (containing monitoring parameters specific to the tunneling conditions) from the cluster. Internally, each terminal performs signal conditioning and digitization, extracts micro-features of the equipment (calculating RMS vibration values, kurtosis, current harmonics, etc.), and runs a lightweight model for local, lightweight intelligent diagnostics. Subsequently, all terminals report the encapsulated terminal perception data packets to the edge intelligent cluster.

[0086] Intra-cluster fusion: The cluster head node of the edge intelligent cluster receives all terminal data packets, performs spatiotemporal alignment, and forms a unified cluster-level multi-source data slice. Next, a pre-defined fusion model is run to perform multi-source feature fusion, correlating vibration, temperature, and current features to generate a comprehensive multi-dimensional fused feature vector and calculate the equipment cluster health index. Finally, the cluster uploads the cluster-level fused feature data packet containing this information to the multi-source data fusion center at the field station.

[0087] Centralized In-Depth Analysis: The multi-source data fusion center aggregates cluster-level data packets from TBM-001 and other devices, performs global spatiotemporal alignment, and constructs a global device topology map based on the relationships between power plant devices. The center outputs a standardized spatiotemporal feature tensor and a global device topology adjacency matrix, which are then input into the spatiotemporal fusion deep learning model.

[0088] The model's temporal feature extraction branch conducted an in-depth analysis of the historical sequence of the vibration signal of the TBM-001 main bearing, revealing a deep-seated trend of slow increase in its high-frequency energy.

[0089] By combining the spatial correlation mining branch of the model with the topology graph, it was found that the current characteristics of the drive motor have recently shown abnormal harmonics, and the motor node has a high degree of betweenness centrality in the network, and its state has an amplified weight in the system.

[0090] Further analysis of the model's global correlation mining branch revealed a long-range correlation between the current vibration trend and a minor tool wear event that occurred a week ago under similar rock conditions.

[0091] The model integrates the findings from the three branches and makes a multi-task output layer judgment: the health score of the TBM-001 main drive system has decreased, and a list of system-level abnormal warning events is generated, predicting that "the probability of the TBM-001 main bearing experiencing moderate wear failure is 78% within the next 48-72 hours", and listing abnormal drive motor current as a key related symptom.

[0092] Decision-making and closed-loop: The dynamic knowledge graph and decision-making module receives the early warning event. It creates the early warning node in the infrastructure-phase equipment knowledge graph and automatically associates it with the "bearing wear" fault mode node. Through graph reasoning, it suggests "focusing on checking the lubrication status of the main bearing and the gear meshing condition." The module then generates a detailed predictive maintenance work order, suggesting maintenance to be carried out in the next planned downtime window. At the same time, the module generates a targeted edge task instruction data package and sends it to CLUSTER-TBM001, instructing it to increase the sampling frequency of the main bearing vibration sensor within the next three days and activate a lightweight AI model specifically designed for "early bearing wear identification" for more refined monitoring.

[0093] Implementation Results: Based on the work orders generated by the system, maintenance personnel conducted targeted inspections of the TBM-001 main drive system during planned downtime. They indeed found early signs of insufficient lubrication and slight wear in the main bearing, and addressed the issue promptly. This prevented a potentially serious unplanned downtime accident caused by bearing damage, verifying the authenticity and effectiveness of the system in predictive maintenance.

[0094] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. A multi-source data acquisition and monitoring system for equipment parameters during the infrastructure construction phase of a pumped storage power station based on the open-source HarmonyOS, characterized in that, include: A distributed intelligent acquisition layer and a multi-source data fusion and intelligent analysis layer, wherein the distributed intelligent acquisition layer is connected to the multi-source data fusion and intelligent analysis layer based on a preset data interface and control interface; The distributed intelligent acquisition layer is used to acquire parameters from multiple sources, perform edge preprocessing, and conduct preliminary diagnosis. The multi-source data fusion and intelligent analysis layer is used to fuse and perform deep intelligent analysis on multi-source heterogeneous data from the distributed intelligent acquisition layer. The deep intelligent analysis includes performing deep analysis on the fused multi-source heterogeneous data based on a spatiotemporal fusion deep learning model, and outputting device health status, parameter prediction, and anomaly warning. The multi-source data fusion and intelligent analysis layer also includes a dynamic knowledge graph and decision module, which is used to perform causal reasoning and decision generation on the model output, and generate control commands that are fed back to the distributed intelligent acquisition layer. The spatiotemporal fusion deep learning model includes a temporal feature extraction branch, a spatial correlation mining branch, and a global correlation mining branch. These three branches are executed in parallel, and their outputs are fed into a feature fusion unit for deep fusion. The temporal feature extraction branch is constructed based on an improved one-dimensional convolutional neural network structure, the spatial correlation mining branch is constructed based on an improved graph convolutional network structure, and the global correlation mining branch is constructed based on a Transformer encoder structure. The improved one-dimensional convolutional neural network structure uses multiple layers of small-scale dilated causal convolutional layers stacked together. By increasing the dilation rate exponentially layer by layer, residual connections and batch normalization are introduced after each convolutional block. At the end of the branch, a global average pooling layer is used to average each feature channel across all time steps. During message passing, the improved graph convolutional network structure introduces a prior bias based on node betweenness centrality when calculating the attention coefficient of neighbor nodes. It sorts and samples the neighbors of each target node by importance, prioritizing the aggregation of neighbor information that is most relevant to the current node state and has the greatest influence.

2. The system for multi-source data acquisition and monitoring of equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 1, is characterized in that... The distributed intelligent acquisition layer includes multiple HarmonyOS intelligent acquisition terminals and one or more edge intelligent clusters. The HarmonyOS intelligent acquisition terminals are connected to physical sensors to perform signal acquisition, local feature extraction and lightweight diagnosis, and output terminal perception data packets. Each edge intelligent cluster is connected to a group of HarmonyOS intelligent acquisition terminals serving the same master device or the same construction area, for fusing multiple terminal perception data packets received and outputting cluster-level fused feature data packets.

3. The system for multi-source data acquisition and monitoring of equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 2, is characterized in that... The terminal-sensing data packet includes at least a unique terminal identifier, a timestamp based on precision clock synchronization, a device micro-feature vector, and local diagnostic results, wherein the local diagnostic results include local anomaly confidence and preliminary anomaly type identifier.

4. The system for multi-source data acquisition and monitoring of equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 3, is characterized in that... The cluster-level fusion feature data package includes at least: a unique identifier for the device cluster, a cluster-level timestamp, a device cluster health index, a multi-dimensional fusion feature vector, and a summary of the cluster-level anomaly diagnosis report.

5. The system for multi-source data acquisition and monitoring of equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 4, is characterized in that... The multi-source data fusion and intelligent analysis layer includes a multi-source data fusion center, a spatiotemporal fusion deep learning model, and a dynamic knowledge graph and decision module. The edge intelligent cluster is connected to the multi-source data fusion center, the multi-source data fusion center is connected to the spatiotemporal fusion deep learning model, and the spatiotemporal fusion deep learning model is connected to the dynamic knowledge graph and decision module. The multi-source data fusion center is used to receive and process cluster-level fusion feature data packets from all edge intelligent clusters, perform global spatiotemporal alignment and topology association, and output standardized spatiotemporal feature tensors and global device topology adjacency matrices. The spatiotemporal fusion deep learning model is used to perform deep analysis on the spatiotemporal feature tensor and the global device topology adjacency matrix, and output device health status, parameter prediction and anomaly warning. The dynamic knowledge graph and decision-making module is used to perform causal reasoning and decision generation on the model output, and generate control instructions that are fed back to the distributed intelligent acquisition layer.

6. The multi-source data acquisition and monitoring system for equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 5, is characterized in that... The improved one-dimensional convolutional neural network structure includes an input layer, a hidden layer, and an output layer. The hidden layer includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a second max pooling layer, a fifth convolutional layer, and a global mean pooling layer.

7. The system for multi-source data acquisition and monitoring of equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 6, is characterized in that... The first convolutional layer comprises a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 1, and 64 channels, and a ReLU activation function layer. The first max pooling layer and the second max pooling layer are both max pooling layers with a pooling region size of 2 and a stride of 2. The second convolutional layer and the third convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 128 channels, and a ReLU activation function layer. The fourth convolutional layer and the fifth convolutional layer each comprise a one-dimensional convolutional layer with a kernel size of 1×3, a stride of 2, and 256 channels, and a ReLU activation function layer.

8. The multi-source data acquisition and monitoring system for equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 7, is characterized in that... The output layer is a classification layer and uses the softmax activation function.

9. A multi-source data acquisition and monitoring system for equipment parameters during the construction phase of a pumped storage power station based on open-source HarmonyOS, as described in claim 8, is characterized in that... The output of the dynamic knowledge graph and decision-making module includes a visualized fault analysis report, predictive maintenance work orders, and edge task instruction data packets sent to the distributed intelligent acquisition layer.

Images