An intelligent industrial big data acquisition and analysis system based on sensor networks

By using clustered sensor networks and edge-side preprocessing, combined with hierarchical storage of hot and cold data and multimodal analysis, the problems of high power consumption, low accuracy, and high cost in industrial data acquisition systems have been solved, achieving efficient and secure data acquisition and analysis.

CN123123381APending Publication Date: 2026-11-17JIANGSU SHIWEI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611258957.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing industrial data acquisition and analysis systems suffer from problems such as high power consumption during acquisition, low accuracy in analysis, high storage costs, and lax security control.

Method used

By employing a clustered sensor network, adaptive cluster head election, lightweight edge-side preprocessing, hierarchical storage of hot and cold data, and multimodal intelligent analysis, combined with a full-link data security management and control mechanism, we can achieve efficient data acquisition, analysis, and secure storage.

Benefits of technology

It significantly reduces the power consumption of data acquisition nodes, improves data retrieval efficiency, enhances analysis accuracy, meets industrial data security compliance requirements, adapts to different industrial scenarios, and supports multi-terminal visual interaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123381A_ABST
    Figure CN123123381A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent industrial big data acquisition and analysis system based on sensor networks, belonging to the field of intelligent industrial big data acquisition and analysis technology. It includes a sensor network acquisition module, an edge-side data preprocessing module, a distributed hierarchical storage module, a multimodal intelligent analysis module, and a visualization interaction module, all connected in sequence. The sensor network acquisition module adopts a clustered topology deployment and has a built-in cluster head adaptive election unit. It collects three types of raw sensing data in industrial scenarios: operating status of various types of equipment, environmental parameters, and process parameters. The edge-side data preprocessing module sequentially performs outlier filtering, time-series alignment, and lightweight compression on the raw data to obtain standardized preprocessed data. This invention significantly reduces the power consumption of the sensor acquisition end, reduces the transmission and storage resources occupied by invalid data, improves industrial data retrieval efficiency, and effectively controls the overall system operating cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of intelligent acquisition and analysis technology of industrial big data, and in particular to an intelligent acquisition and analysis system for industrial big data based on sensor networks. Background Technology

[0002] The industrial sector is currently at a critical stage of digital and intelligent transformation. Industrial big data, as a core production factor supporting production decisions and optimizing operational efficiency, directly determines the effectiveness of digital transformation in industrial scenarios through its precise data collection, efficient storage, and intelligent analysis capabilities. Sensor networks, as the core infrastructure for industrial data acquisition, are widely used in various industrial scenarios such as discrete manufacturing, process chemicals, and energy and power. The core industry demands focus on five dimensions: reducing power consumption at the acquisition end, improving data transmission efficiency, controlling storage costs, increasing analytical accuracy, and ensuring end-to-end data security. This supports production decisions related to equipment maintenance, energy efficiency optimization, and process iteration, reduces the risk of unplanned downtime, and improves overall production efficiency.

[0003] Currently, mainstream industrial data acquisition and analysis solutions are mainly divided into two categories. The first category is a centralized sensor acquisition + single-node storage and analysis solution. This solution uses a star topology to deploy sensors, with all acquisition nodes directly connected to the field host computer. Raw data is uploaded directly to a single-node relational database for storage without preprocessing, and the analysis logic relies on manually preset fixed thresholds to trigger alarms. This solution has a simple deployment logic and low initial investment cost, and is still widely used in small- to medium-scale, low-complexity industrial production scenarios. However, this solution has significant drawbacks: the centralized topology results in high communication power consumption of acquisition nodes, signal blind spots are prone to occur in obstructed areas of the industrial site, a large amount of abnormal and dirty data consumes transmission bandwidth and storage resources, single-node storage scalability is poor, historical data retrieval efficiency is low, fixed threshold alarms can only identify abnormal conditions that have already occurred and cannot predict potential faults in advance, and storage costs increase linearly with the increase in data volume.

[0004] The second category is a unified cloud-based sensor acquisition solution with centralized storage and analysis. This solution directly uploads all acquired data to the cloud platform, where the cloud handles unified cleaning, storage, and full analysis. This solution boasts ample computing power and strong storage scalability, making it suitable for centralized management of large-scale industrial data across multiple plants. However, this solution also has significant drawbacks: the lack of pre-processing nodes at the edge leads to a large amount of invalid data being uploaded, consuming significant public network bandwidth; edge-to-cloud transmission latency cannot meet the millisecond-level real-time early warning requirements of industrial sites; the entire chain lacks a refined security control mechanism, posing a risk of leakage for core process parameters and sensitive equipment data; data access permissions are broadly defined, failing to meet industrial data security compliance requirements; the lack of tiered storage for hot and cold data results in high overall storage costs; and the generalized analysis model is not adapted to the characteristics of different devices, making it difficult to achieve fault prediction accuracy that meets production-ready standards. Summary of the Invention

[0005] This invention proposes an intelligent industrial big data acquisition and analysis system based on sensor networks to solve the problems mentioned in the prior art, such as high power consumption, low analysis accuracy, high storage cost, and loose security control in existing industrial data acquisition and analysis systems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an industrial big data intelligent acquisition and analysis system based on sensor networks, comprising a sensor network acquisition module, an edge data preprocessing module, a distributed hierarchical storage module, a multimodal intelligent analysis module, and a visualization interaction module that are sequentially connected via industrial Ethernet or LoRaWAN communication links; The sensor network acquisition module is deployed in a clustered topology at the equipment end and environmental monitoring points in the industrial site. It has a built-in cluster head adaptive election unit. The operation cycle of the election unit matches the operation condition fluctuation cycle of the industrial equipment. It is used to collect three types of raw sensing data in the industrial scenario: operating status parameters of power equipment, production equipment, and auxiliary equipment; plant environmental parameters; and production line process parameters. It covers the full-dimensional industrial acquisition needs of vibration, temperature, pressure, current, voltage, flow, humidity, and dust concentration. The edge-side data preprocessing module is deployed on the edge gateway in the industrial field. It is used to sequentially perform outlier filtering, multi-source time-series alignment and lightweight lossless compression on the raw sensing data, and output standardized preprocessed data that conforms to industrial data transmission standards. The distributed hierarchical storage module adopts a cloud-edge collaborative storage architecture to store preprocessed data hierarchically according to hot and cold access attributes, and at the same time builds a multi-dimensional metadata index to achieve millisecond-level fast data retrieval. The multimodal intelligent analysis module is deployed on the computing nodes of the cloud platform and is used to perform multi-dimensional feature correlation analysis on the stored structured and unstructured industrial data, and output three types of actionable analysis results: equipment fault warning, plant energy efficiency optimization, and production process adjustment. The visualization interaction module is deployed at the application layer and is used to visualize the analysis results in multiple dimensions, and to receive user data queries, parameter configurations, and strategy adjustment instructions, and then feed them back to the corresponding functional modules to perform the corresponding operations.

[0007] Preferably, the sensor network acquisition module includes multiple end acquisition nodes, multiple candidate cluster head nodes, and multiple relay nodes. All end acquisition nodes have built-in low-power wake-up units, which only start the acquisition link and radio frequency communication link when the preset acquisition period arrives, when a trigger acquisition command is received from the edge side, or when the acquisition point parameters change suddenly. During other times, the acquisition and communication circuits are turned off, and only the clock unit is kept in sleep mode. Candidate cluster head nodes are pre-selected from the end acquisition nodes. Only nodes with hardware configurations including large-capacity backup batteries and high-gain communication antennas can enter the candidate cluster head pool. Relay nodes are deployed in communication obstruction areas of the industrial site to forward communication data between cluster head nodes and the edge side, eliminating signal blind spots in the field.

[0008] Preferably, the election logic of the cluster head adaptive election unit is as follows: real-time acquisition of three parameters of each candidate cluster head node: remaining power, average communication distance with all end nodes in the cluster, and data forwarding load already undertaken in the current period. After normalizing the three parameters to a value range of 0 to 1, the election priority of each candidate cluster head node is calculated by weighting them according to a preset coefficient with the highest weight for remaining power, the second highest weight for average communication distance, and the lowest weight for forwarding load. The candidate node with the highest priority is selected as the working cluster head node in the current period. When the remaining power of the working cluster head node is lower than a preset safety threshold, the re-election process can be automatically triggered without waiting for the election period to arrive. The election results are synchronized to all nodes in the cluster and the edge-side data preprocessing module.

[0009] Preferably, the edge-side data preprocessing module has a built-in sliding window anomaly detection unit. The size of the sliding window is dynamically adjusted according to the acquisition frequency of the corresponding sensor. A small window is used in high-frequency acquisition scenarios and a large window is used in low-frequency acquisition scenarios. The current acquisition data is compared with the average of historical data in the same window. The deviation threshold is dynamically set based on the 3σ principle of historical data to filter out abnormal dirty data that exceeds the deviation threshold. At the same time, based on the unified timestamp generated by the Beidou time synchronization or NTP network time synchronization of the unified access on the edge side, the acquisition data of different types and different sampling frequencies of sensors are time-aligned. After alignment, they are unified to the same time granularity. After the time alignment is completed, the data is compressed using a lightweight lossless compression algorithm dedicated to industrial time series data. The compressed data is uploaded to the distributed hierarchical storage module after being accompanied by a CRC check code.

[0010] Preferably, the distributed hierarchical storage module has a built-in hot and cold data partitioning unit. The access frequency threshold is dynamically updated based on user access data over the past 30 days. According to the preset access frequency threshold, the preprocessed data is divided into hot data and cold data. Hot data is real-time monitoring and fault analysis related data with access frequency higher than the threshold in the past 7 days. It is stored in a high-performance time-series database to meet the requirements of high-frequency and low-latency access. Cold data is historical archived data that has not been accessed for more than 30 days. It is compressed and stored in a low-cost object storage cluster to reduce the overall storage cost. The metadata index includes three levels of search tags: time dimension, device dimension, and parameter type dimension. During retrieval, the search scope is filtered step by step according to the hierarchy.

[0011] Preferably, the multimodal intelligent analysis module includes a fault prediction subunit, an energy efficiency optimization subunit, and a process adjustment subunit that are interconnected. The three subunits respectively call the machine learning model trained on the historical data labeled in the corresponding industrial scenario. The three subunits support parallel operation and individual calling. The input data source can be real-time access streaming data or stored historical data. The analysis results are synchronously stored in the distributed hierarchical storage module as training samples for subsequent model iteration and fine-tuning. The model is fine-tuned with new labeled data at fixed intervals to ensure that the analysis accuracy is stable above the preset threshold.

[0012] Preferably, the fault prediction subunit adopts an improved temporal Transformer model. The improvement is that a static feature embedding layer is added in front of the self-attention layer. The input data is first fed into the static feature embedding layer, and static features such as equipment type, rated operating parameters, and years of service are embedded into the input vector. The weights of the self-attention layer are dynamically adjusted according to the embedded equipment features. The weights of feature vectors related to the current equipment fault are automatically increased, while the weights of irrelevant features are automatically suppressed. The inference process is synchronously matched with the historical fault case library of the same model of equipment, and three types of prediction results are output: fault type, warning level, and predicted occurrence time. The warning level is divided into four levels: alert, general, important, and urgent, corresponding to different response strategies.

[0013] Preferably, it also includes a data end-to-end security management module, which is communicatively connected to all other functional modules. It is used to perform data anonymization, permission verification, and operation evidence storage for the entire lifecycle of data collection, transmission, storage, analysis, and display. During the data anonymization stage, the unique identifier of the device, core process parameters, and user privacy information are processed by pseudonymization and masking. During the transmission stage, the communication link is encrypted using the TLS 1.3 encryption protocol. During the storage stage, sensitive fields are encrypted using national cryptographic algorithms. All operation evidence storage records are written to an immutable distributed evidence storage module, and all operations are traceable.

[0014] Preferably, the data end-to-end security management module has a built-in dual-dimensional permission matching unit, which performs permission matching based on the user's role attributes and the current access scenario attributes. The role attributes include four categories: system administrator, equipment maintenance personnel, process engineer, and visitor. Different roles have different preset basic access permission ranges. The scenario attributes include four categories: daily monitoring, fault diagnosis, process optimization, and system configuration. The scope of data access is different in different scenarios. The user's access request is only responded to when both dimensions of permission verification pass. If the verification fails, access is directly denied and an abnormal access log is recorded.

[0015] Preferably, the visualization interaction module supports multi-terminal adaptation to PC, mobile, and industrial site dashboards. It adopts a responsive layout to automatically adapt to the screen size of different terminals. It has a built-in custom dashboard configuration unit and an anomaly alarm push unit. The custom dashboard configuration unit allows users to add visualization components such as line charts, bar charts, heat maps, and alarm cards through drag-and-drop operations, and customize the data display dimensions. The anomaly alarm push unit supports multiple push methods such as SMS, APP push, email, and on-site audible and visual alarms. The push path can be customized according to the alarm level. High-level alarms are pushed to multiple corresponding responsible persons simultaneously. Parameter adjustment and strategy modification commands issued by users on the interactive terminal are immediately synchronized to the corresponding functional modules to take effect.

[0016] Compared with existing technologies, the beneficial effects of this invention are: This invention significantly reduces the communication power consumption of acquisition nodes and extends the overall lifespan of the sensor network through a collaborative architecture of clustered sensor network deployment, adaptive cluster head election, lightweight edge-side preprocessing, and hierarchical storage of hot and cold data. It effectively reduces the transmission bandwidth and storage resources occupied by invalid and dirty data, while improving the retrieval efficiency of industrial data through a three-level metadata index structure. It also significantly controls the overall system operating cost and effectively solves the defects of existing technologies, such as high power consumption at the acquisition end, large proportion of invalid data, high storage cost, and low data retrieval efficiency.

[0017] This invention utilizes an improved temporal Transformer multimodal analysis module that embeds static features of equipment, combined with a full-link data security management mechanism. It dynamically adjusts the feature weights of the analysis model according to the operating characteristics of different industrial equipment, significantly improving the accuracy of equipment fault early warning, energy efficiency optimization, and process adjustment. It can achieve early identification of potential industrial anomalies. At the same time, through dual-dimensional permission verification and full-link encrypted evidence storage management logic, it meets the security and compliance requirements of industrial data, effectively solving the defects of existing technologies such as poor adaptability of analysis models, low prediction accuracy, delayed early warning, loose data security management, and inability to meet compliance requirements.

[0018] This invention supports access to multiple mainstream industrial sensor protocols and can be adapted to various complex and large-scale industrial scenarios such as discrete manufacturing, process chemicals, and energy and power. The node deployment topology can be flexibly adjusted according to the actual communication environment on site, and it can be quickly launched without reconstructing the enterprise's existing data acquisition infrastructure. At the same time, it supports multi-terminal visual interaction, can match the usage needs of different users, and can seamlessly connect with the enterprise's existing production management system. It has extremely high promotion and application value in the field of industrial digital transformation. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the industrial big data intelligent acquisition and analysis system based on sensor networks proposed in this invention; Figure 2 This is a flowchart of the sensor network data acquisition and adaptive cluster head election process of the present invention; Figure 3 This is a flowchart of the edge-side data preprocessing process of the present invention; Figure 4 This is a flowchart of the distributed cold and hot hierarchical storage and index construction process of the present invention; Figure 5 This is a flowchart of the multimodal intelligent analysis and end-to-end security management process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figures 1 to 5 This invention discloses an intelligent industrial big data acquisition and analysis system based on sensor networks, comprising a sensor network acquisition module, an edge-side data preprocessing module, a distributed hierarchical storage module, a multimodal intelligent analysis module, and a visualization interaction module, which are connected sequentially via industrial dedicated LoRaWAN (470-510MHz frequency band, communication distance up to 1km in industrial scenarios, capable of penetrating 3 layers of concrete walls) or PROFINET industrial Ethernet (transmission latency less than 1ms).

[0022] The sensor network acquisition module is deployed in a clustered topology at the equipment end and environmental monitoring points in the industrial site. It has a built-in cluster head adaptive election unit. The operation cycle of the election unit is matched with the operation condition fluctuation cycle of the industrial equipment. For example, the election cycle of chemical plant is set to 1 hour and the election cycle of discrete manufacturing production line is set to 2 hours. It can be customized according to the industrial scenario.

[0023] The sensor network acquisition module is used to collect three types of raw sensing data in industrial scenarios: operating status parameters of power equipment, production equipment, and auxiliary equipment; plant environment parameters; and production line process parameters. It covers the full-dimensional industrial data acquisition needs of vibration, temperature, pressure, current, voltage, flow, humidity, dust concentration, and toxic and harmful gas concentration, and supports access to mainstream industrial sensor protocols such as Modbus and OPC UA.

[0024] The edge-side data preprocessing module is deployed on the Advantech UNO-2484G edge gateway in the industrial site. It supports wide temperature operation from -20℃ to 60℃, meets the protection requirements of industrial sites, and is used to sequentially perform outlier filtering, multi-source time-series alignment, and lightweight lossless compression on the raw sensing data, outputting standardized preprocessed data that conforms to the GB / T 37393-2019 industrial data transmission standard.

[0025] The distributed hierarchical storage module is used to store preprocessed data hierarchically according to hot and cold access attributes, and at the same time builds a multi-dimensional metadata index to achieve millisecond-level fast data retrieval.

[0026] The multimodal intelligent analysis module is deployed on the GPU computing nodes of the cloud platform. It is used to perform multi-dimensional feature correlation analysis on the stored structured and unstructured industrial data, and output three types of actionable analysis results: equipment fault warning, plant energy efficiency optimization, and production process adjustment.

[0027] The visualization interaction module is deployed at the application layer and is used to visualize the analysis results in multiple dimensions, and to receive user data queries, parameter configurations, and strategy adjustment instructions, and then feed them back to the corresponding functional modules to perform the corresponding operations.

[0028] This invention also discloses the specific structure of the sensor network acquisition module, which includes multiple end acquisition nodes, multiple candidate cluster head nodes, and multiple relay nodes. All end acquisition nodes use STM32L4 series low-power MCUs with built-in low-power wake-up units. The acquisition link and RF communication link are activated only when the preset acquisition period arrives, when a trigger acquisition command is received from the edge side, or when the change amplitude of the acquisition point parameter exceeds the preset threshold. During other periods, the acquisition and communication circuits are turned off, and only the clock unit is kept in sleep mode. The sleep current is less than 1μA and the wake-up delay is less than 10ms.

[0029] Candidate cluster head nodes are pre-selected from the end acquisition nodes. Only nodes with hardware configurations including a 10000mAh high-capacity backup battery and a high-gain omnidirectional communication antenna can enter the candidate cluster head pool to meet the long-term high-load data forwarding requirements. Relay nodes are deployed in communication obstruction areas of the industrial site, with an adjustable transmission power range of 5dBm to 20dBm. They are used to forward communication data between the cluster head nodes and the edge side, eliminating signal blind spots caused by steel structures and equipment walls on site.

[0030] This invention also discloses the specific election logic of the cluster head adaptive election unit. It acquires parameters in real time from three dimensions: the remaining power of each candidate cluster head node, the average communication distance to all end nodes within its cluster, and the data forwarding load already undertaken in the current period. After normalizing these three parameters to a value range of 0 to 1, the election priority of each candidate cluster head node is calculated by weighting them according to a preset coefficient with the highest weight for remaining power, the second highest weight for average communication distance, and the lowest weight for forwarding load. The candidate node with the highest priority is selected as the working cluster head node for the current period. When the remaining power of the working cluster head node is lower than a preset safety threshold of 20%, a re-election process is automatically triggered without waiting for the election period to arrive. The election results are synchronized to all nodes within the cluster and the edge-side data preprocessing module. The formula for calculating the election priority is as follows: ;in, This represents the election priority for candidate cluster leaders; the higher the value, the higher the priority. The remaining battery power is weighted and fixed at 0.5. The communication distance weight is fixed at 0.3; This is the load weight, with a fixed value of 0.2. This is the normalized value of the remaining charge of the candidate cluster head, ranging from 0 to 1, with a larger value for higher remaining charge. This is a normalized value of the average communication distance between the candidate cluster head and all end nodes in the cluster, ranging from 0 to 1, with the value increasing as the distance increases. This is a normalized value of the data forwarding load that the candidate cluster head has undertaken in the current cycle, ranging from 0 to 1, with a larger value indicating higher load. The core election logic code snippet is as follows: def elect_cluster_head(candidate_nodes, alpha=0.5, beta=0.3, gamma=0.2): max_priority = 0 selected_head = None max_dis = max([n.avg_comm_distance for n in candidate_nodes]) max_load = max([n.current_load for n in candidate_nodes]) for node in candidate_nodes: e_norm = node.remaining_battery / node.max_battery d_norm = node.avg_comm_distance / max_dis l_norm = node.current_load / max_load priority = alpha * e_norm + beta * (1 - d_norm) + gamma * (1- l_norm) if priority > max_priority: max_priority = priority selected_head = node return selected_head This invention also discloses the specific implementation logic of the edge-side data preprocessing module, which has a built-in sliding window anomaly detection unit. The size of the sliding window is dynamically adjusted according to the acquisition frequency of the corresponding sensor. In high-frequency acquisition scenarios, such as a 1kHz vibration sensor, a small window containing 100 sampling points is used; in low-frequency acquisition scenarios, such as a temperature and humidity sensor that samples once per minute, a large window containing 30 sampling points is used. The current acquisition data is compared with the average of historical data within the same window to determine the deviation. The deviation threshold is dynamically set based on the 3σ principle of historical data. Abnormal dirty data that exceeds the deviation threshold is filtered out, and the missing values ​​after filtering are filled by linear interpolation of adjacent time points.

[0031] Simultaneously, based on the unified timestamp generated by the BeiDou or NTP network time synchronization through unified edge access, time-series alignment is performed on data collected by sensors of different types and sampling frequencies. After alignment, the time granularity is unified to 1 second. After time-series alignment, the data is compressed using the TSZ lightweight lossless compression algorithm specifically designed for industrial time-series data, achieving a compression ratio of up to 10:1. The compressed data, accompanied by a CRC32 checksum, is then uploaded to the distributed hierarchical storage module. The core anomaly detection logic code snippet is as follows: import numpy as np def is_abnormal(current_data, history_window, sigma=3): mean = np.mean(history_window) std = np.std(history_window) upper_threshold = mean + sigma * std lower_threshold = mean - sigma * std return not (lower_threshold <= current_data <= upper_threshold) This invention also discloses the specific architecture of the distributed hierarchical storage module, which has a built-in hot and cold data partitioning unit. The access frequency threshold is dynamically updated based on user access data over the past 30 days. By default, preprocessed data is divided into hot data and cold data according to the standard of access frequency exceeding 3 times in the past 7 days. Hot data is high-frequency access data related to real-time monitoring and fault analysis, which is stored in the high-performance time-series database InfluxDB to meet the requirements of high-frequency and low-latency access, with a query response latency of less than 50ms. Cold data is historical archived data that has not been accessed for more than 30 days, which is compressed and stored in the low-cost open-source object storage cluster MinIO to reduce the overall storage cost. The metadata index includes three levels of search tags: time dimension, device dimension, and parameter type dimension. During retrieval, the search scope is filtered step by step according to the hierarchy.

[0032] This invention also discloses the specific structure of the multimodal intelligent analysis module, which includes a fault prediction subunit, an energy efficiency optimization subunit, and a process adjustment subunit that are interconnected. The three subunits respectively call the machine learning model trained on the historical data labeled in the corresponding industrial scenario. During training, the training set, validation set, and test set are divided in an 8:1:1 ratio. The three subunits support parallel operation and individual calling. The input data source can be real-time access Kafka streaming data or stored historical data. The analysis results are synchronously stored in the distributed hierarchical storage module as training samples for subsequent model iteration and fine-tuning. The model is fine-tuned with new labeled data every month to ensure that the analysis accuracy is stable above the preset threshold.

[0033] This invention also discloses the specific algorithm structure of the fault prediction subunit, which adopts an improved temporal Transformer model. The improvement is the addition of a static feature embedding layer at the front end of the self-attention layer. Input data is first fed into the static feature embedding layer, embedding static features such as equipment type, rated operating parameters, and years of service into the input vector. The weights of the self-attention layer are dynamically adjusted according to the embedded equipment features. The weights of feature vectors related to the current equipment fault are automatically increased, while the weights of irrelevant features are automatically suppressed. The inference process synchronously matches the historical fault case library of the same model of equipment, and outputs three types of prediction results: fault type, warning level, and predicted occurrence time. The warning level is divided into four levels: alert, general, important, and urgent, corresponding to different response strategies. The calculation formulas for static feature embedding and weight adjustment are as follows: ;in, The input is a time-aware data sequence with dimensions [seq_len, feature_dim]. To obtain the expanded device static feature vector, the one-dimensional static feature is repeated seq_len times, with dimensions [seq_len, static_dim]. The concatenated input vector, with dimensions [seq_len, feature_dim + static_dim], serves as the input to the self-attention layer. ;in, The feature mask matrix has dimensions of ... Consistent with the setting, the position of the relevant fault feature of the corresponding equipment is set to 1, and the position of the other irrelevant feature is set to 0.2, so as to automatically increase the weight of relevant features and suppress irrelevant features, thereby improving the accuracy of fault prediction.

[0034] This invention also discloses that the system further includes a data end-to-end security management module. This module is communicatively connected to all other functional modules and is used to perform data anonymization, permission verification, and operation evidence storage throughout the entire lifecycle of data acquisition, transmission, storage, analysis, and display. During the data anonymization stage, unique equipment identifiers, core process parameters, and user privacy information are pseudonymous or masked. For example, only the last four digits of the equipment number are displayed, and only the numerical range of sensitive process parameters is shown. During the transmission phase, the TLS 1.3 encryption protocol is used to encrypt the communication link to prevent data from being eavesdropped. During the storage phase, sensitive fields are encrypted using the national cryptographic SM4 algorithm. All operation records are written to the tamper-proof consortium blockchain distributed evidence storage module. The evidence storage nodes include enterprise operators and regulators. All operations are traceable and tamper-proof, which meets the Level 3 requirements of Information Security Protection 2.0.

[0035] This invention also discloses a built-in dual-dimensional permission matching unit in the data end-to-end security management module. It performs permission matching based on the user's role attributes and the current access scenario attributes. The role attributes include four categories: system administrator, equipment maintenance personnel, process engineer, and visitor. Different roles have different preset basic access permission ranges. The scenario attributes include four categories: daily monitoring, fault diagnosis, process optimization, and system configuration. The scope of data access is different in different scenarios. The user's access request is only responded to when both dimensions of permission verification pass. If the verification fails, access is directly denied and an abnormal access log is recorded.

[0036] For example, equipment maintenance personnel can access all operating parameters of the equipment in a fault diagnosis scenario, but can only access publicly available status parameters in a routine monitoring scenario. Visitors, regardless of the scenario, can only access publicly available plant environment parameters.

[0037] This invention also discloses that the visualization interaction module supports multi-terminal adaptation to PC, mobile, and industrial site dashboards, adopts a responsive layout to automatically adapt to the screen size of different terminals, and has a built-in custom dashboard configuration unit and an abnormal alarm push unit.

[0038] The customizable dashboard configuration unit allows users to add over 20 commonly used industrial visualization components, such as line charts, bar charts, heatmaps, and alarm cards, through drag-and-drop operations. Users can customize data display dimensions and build dashboards without writing code. The anomaly alarm push unit supports multiple push methods, including SMS, APP push, email, and on-site audible and visual alarms. The push path can be customized according to the alarm level. High-level alarms are pushed to multiple corresponding responsible persons simultaneously, with a push interval of 5 minutes until the user confirms the alarm. Parameter adjustment and strategy modification commands issued by users on the interactive terminal are instantly synchronized to the corresponding functional modules and take effect.

[0039] Reference Figure 1 This flowchart provides a macroscopic overview of the entire operational process of an industrial big data intelligent acquisition and analysis system based on sensor networks. The system begins with a clustered topology sensor network. The cluster head adaptive election unit drives the end nodes to collect raw sensing data from various types of equipment. Subsequently, the edge-side data preprocessing module cleans, aligns, and performs lossless compression and standardization on this heterogeneous data. The standardized data flows to the distributed hierarchical storage module, triggering a hot / cold partitioning mechanism to achieve hierarchical archiving and the construction of a three-level metadata index. Next, the multimodal intelligent analysis module invokes deep learning networks and machine learning models to output analysis results for fault warnings, energy efficiency optimization, and process adjustments in parallel. Finally, the analysis results are presented in real-time through a multi-terminal adaptive visualization dashboard, and user feedback control commands are transmitted back to each execution module in a closed loop.

[0040] Reference Figure 2 This flowchart details the topology operation mechanism of the sensor network data acquisition module and the control logic of adaptive cluster head election. To ensure the network lifecycle of large-scale sensor nodes in industrial sites, all end acquisition nodes are normally kept dormant through built-in low-power wake-up units, and are only activated momentarily when the cycle arrives or an instruction is triggered. During the data aggregation phase, the adaptive cluster head election unit retrieves three core parameters in real time: the remaining power of each candidate cluster head node, the average communication distance with the end nodes, and the current instantaneous forwarding load. Through internal priority weighted calculation, the optimal working cluster head for the current cycle is dynamically selected. The raw sensing data of each end node is safely and efficiently aggregated to the gateway side through multi-level hop transmission via relay nodes and the current working cluster head.

[0041] Reference Figure 3 This flowchart focuses on the industrial big data cleaning and lightweight architecture of the edge data preprocessing module. When the edge gateway receives massive amounts of high-frequency raw industrial time-series sensing data, it first activates the built-in sliding window anomaly detection unit to dynamically compare the current instantaneous input with the historical benchmark average within the same window, decisively filtering and removing outliers and abnormal pulses that exceed the safety threshold. Addressing the pain points of inconsistent acquisition frequencies and clock misalignment of different sensors in the industrial field, the system rigidly aligns the data of each node on the time axis based on a unified time stamp. The aligned multi-source composite time-series data is then deeply compressed using a lightweight lossless compression algorithm, minimizing the network bandwidth overhead for upward transmission while fully preserving the original industrial semantic information.

[0042] Reference Figure 4 This diagram clearly illustrates the physical layout and logical retrieval mechanism of a distributed hierarchical storage module that meets specific data storage architecture requirements. After receiving preprocessed standardized industrial big data at the system input end, the internal hot and cold data partitioning unit immediately reads the system's preset access frequency threshold. Based on the timeliness of the data and the frequency of business queries, the system automatically defines frequently accessed production data as hot data and writes it into a distributed time-series database with extremely high read and write performance. For cold data with extremely low access frequency, it is batch-transferred to a low-cost object storage cluster to achieve scientific control of storage costs. While writing data, the system automatically extracts retrieval tags along three physical dimensions: time, device, and parameter type, and builds a three-level metadata index to ensure that both hot and cold data can achieve second-level retrieval response.

[0043] Reference Figure 5This flowchart deeply illustrates the core algorithm architecture of the multimodal intelligent analysis module and its data security control mechanism covering the entire data lifecycle. The system's full-link data security control module runs through the entire data lifecycle. When an access request is input, the two-dimensional permission matching unit works together to perform multi-dimensional concurrent verification of the user's role attributes and the current access scenario attributes. After passing the verification, necessary data desensitization and operation evidence storage are performed. In the core intelligent analysis layer, the system calls the improved time-series Transformer model as the computing engine. This model can dynamically adjust the weight allocation matrix of the self-attention layer according to the equipment type, and prioritize deep feature inference with historical defect or fault mode maps, thereby outputting highly accurate fault warning types, warning levels, future energy efficiency optimization curves, and process parameter adjustment schemes in parallel.

[0044] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent industrial big data acquisition and analysis system based on sensor networks, characterized in that, It includes a sensor network acquisition module, an edge data preprocessing module, a distributed hierarchical storage module, a multimodal intelligent analysis module, and a visualization interaction module, which are connected in sequence. The sensor network acquisition module adopts a clustered topology deployment and has a built-in cluster head adaptive election unit, which is used to collect raw sensing data of the operating status, environmental parameters, and process parameters of various types of equipment in industrial scenarios. The edge-side data preprocessing module is used to perform outlier filtering, temporal alignment, and lightweight compression on the raw sensing data to obtain standardized preprocessed data. The distributed hierarchical storage module is used to store preprocessed data hierarchically according to hot and cold attributes, and at the same time build a multi-dimensional metadata index to enable fast data retrieval; The multimodal intelligent analysis module is used to perform multi-dimensional correlation analysis on the stored industrial data and output analysis results such as fault warning, energy efficiency optimization, and process adjustment. The visualization interaction module is used to display the analysis results in a multi-dimensional way and to receive user queries and configuration commands and send them back to each module to execute the corresponding operations.

2. The industrial big data intelligent acquisition and analysis system based on sensor networks according to claim 1, characterized in that, The sensor network acquisition module includes multiple end acquisition nodes, multiple candidate cluster head nodes, and multiple relay nodes. All end acquisition nodes have built-in low-power wake-up units, which only start the acquisition link when the acquisition cycle arrives or when a trigger command is received, and remain in sleep mode during other times.

3. The industrial big data intelligent acquisition and analysis system based on sensor networks according to claim 2, characterized in that, The election logic of the cluster head adaptive election unit is as follows: real-time acquisition of parameters in three dimensions for each candidate cluster head node: remaining power, average communication distance with end nodes, and current data forwarding load; weighted calculation to obtain the election priority of each candidate cluster head node; and selection of the candidate node with the highest priority as the working cluster head node for the current period.

4. The intelligent acquisition and analysis system for industrial big data based on sensor networks according to claim 1, characterized in that, The edge-side data preprocessing module has a built-in sliding window anomaly detection unit. It compares the deviation of the currently collected data with the average of historical data in the same window, filters out abnormal data that exceeds the deviation threshold, and performs time-series alignment of the collected data from different types of sensors based on a unified time stamp. After alignment, the data is compressed using a lightweight lossless compression algorithm before being uploaded.

5. The intelligent industrial big data acquisition and analysis system based on sensor networks according to claim 1, characterized in that, The distributed hierarchical storage module has a built-in hot and cold data partitioning unit. According to the preset access frequency threshold, the preprocessed data is divided into hot data and cold data. Hot data is stored in a high-performance time-series database to meet the high-frequency access requirements, while cold data is stored in a low-cost object storage cluster to reduce storage costs. The metadata index includes three levels of search tags: time dimension, device dimension, and parameter type dimension.

6. The intelligent industrial big data acquisition and analysis system based on sensor networks according to claim 1, characterized in that, The multimodal intelligent analysis module includes a fault prediction subunit, an energy efficiency optimization subunit, and a process adjustment subunit that are interconnected. Each of the three subunits calls its corresponding trained machine learning model and outputs corresponding analysis results in combination with the input industrial data.

7. The intelligent industrial big data acquisition and analysis system based on sensor networks according to claim 6, characterized in that, The fault prediction subunit adopts an improved temporal Transformer model, dynamically adjusts the weight allocation of the self-attention layer according to the fault characteristics of different equipment types, prioritizes matching the historical fault characteristics of the corresponding equipment for inference, and outputs the prediction results of fault type, warning level and predicted occurrence time.

8. The intelligent industrial big data acquisition and analysis system based on sensor networks according to claim 1, characterized in that, It also includes a data end-to-end security management module, which is connected to the other modules to perform data anonymization, permission verification, and operation evidence storage throughout the entire data collection, transmission, storage, analysis, and display process.

9. The intelligent industrial big data acquisition and analysis system based on sensor networks according to claim 8, characterized in that, The data end-to-end security management module has a built-in dual-dimensional permission matching unit, which performs permission matching based on the user's role attributes and the current access scenario attributes. It only responds to the user's access request when both dimensions of permission verification pass.

10. The intelligent acquisition and analysis system for industrial big data based on sensor networks according to claim 1, characterized in that, The visualization interaction module supports multi-terminal adaptation to PC, mobile, and industrial site dashboards. It has a built-in custom dashboard configuration unit and an anomaly alarm push unit, allowing users to customize the display dimensions according to their needs. At the same time, it sends the received warning information to the corresponding user terminal according to the preset push path.