A multi-source information fusion heterogeneous data regularization method

Abstract

The present application relates to the technical field of data processing, and discloses a multi-source information fusion heterogeneous data regularization method, comprising: arranging distributed sensors in several treatment stages of a sewage treatment process, collecting detection data of all distributed sensors through access protocol and time sequence synchronization; performing semantic analysis on the detection data, normalizing the detection data through labeling, and completing semantic label standardization; processing the detection data after semantic standardization to obtain metadata; constructing a data cross-layer representation model based on transfer learning, and performing feature extraction and semantic embedding on the metadata to obtain unified vectorization expression; and constructing an information knowledge graph for the whole process of sewage treatment according to the vectorization expression. The present application realizes unified vectorization expression of time sequence, semantic and structural characteristics among different data sources, and improves the fusion, interpretability and usability of data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a method for regularizing heterogeneous data through multi-source information fusion. Background Technology

[0002] With the acceleration of urbanization and the increasing demands for environmental governance, wastewater treatment systems, as a crucial link in the urban water cycle, have gradually evolved towards automation, digitalization, and intelligence. Modern wastewater treatment plants typically encompass multiple process stages, including bar screen removal, grit removal, activated sludge treatment, biological nitrogen and phosphorus removal, disinfection, and sludge thickening and dewatering. Each treatment stage involves real-time monitoring of various environmental factors, physicochemical parameters, and operational status.

[0003] To support the construction of smart water management, wastewater treatment systems deploy a large number of heterogeneous sensor devices (such as pH sensors, dissolved oxygen probes, flow meters, level gauges, ammonia nitrogen analyzers, etc.), and aggregate data through various automated platforms or distributed acquisition systems. The data generated by these devices and platforms are highly heterogeneous in structure, semantics, units, and time frequency, forming a typical "multi-source heterogeneous data system." However, data collected by different sensors differ in naming, units, time granularity, and communication protocols, lacking unified semantic tags and data formats, resulting in direct data interoperability. Therefore, most current wastewater treatment monitoring platforms still rely on manually defining variable meanings and processing logic, failing to effectively capture the temporal, spatial, and logical multi-level relationships in the wastewater treatment process. Furthermore, they cannot transform large amounts of real-time data into reasonable and reusable domain knowledge.

[0004] Therefore, it is necessary to design a heterogeneous data rule-making method for multi-source information fusion to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a heterogeneous data rule-making method for multi-source information fusion, aiming to solve the problems existing in the current technology.

[0006] This invention proposes a method for regularizing heterogeneous data through multi-source information fusion, comprising:

[0007] Distributed sensors are deployed at several treatment stages in the wastewater treatment process, and the detection data of all the distributed sensors are collected in a time-series synchronization manner through an access protocol.

[0008] Based on semantic matching, the detection data is semantically parsed, and the detection data is labeled and semantic label standardization is completed.

[0009] The semantically standardized detection data is processed to obtain metadata containing data source identifier, data type, unit specification, and time information;

[0010] A data cross-layer representation model based on transfer learning is constructed, and feature extraction and semantic embedding are performed on the metadata according to the data cross-layer representation model to obtain a unified vectorized expression;

[0011] Based on the vectorized expression, an information knowledge graph for the entire wastewater treatment process is constructed.

[0012] Furthermore, when deploying distributed sensors at several stages of the wastewater treatment process, this includes:

[0013] Sensor nodes with functions for detecting temperature, dissolved oxygen, pH, conductivity, turbidity, flow rate, liquid level, chemical oxygen demand, and ammonia nitrogen are installed in the grit chamber, grit chamber, biological reactor, secondary sedimentation tank, disinfection tank, and sludge treatment stage.

[0014] Furthermore, when collecting detection data from all the distributed sensors via access protocol and time synchronization, this includes:

[0015] Various sensors are connected to the same data acquisition network using heterogeneous communication protocols such as OPC-UA, MQTT, or HTTP, and unified data encapsulation and format standardization are performed through protocol adaptation middleware.

[0016] Combine network time protocols or IEEE 1588 precision time protocols to achieve unified clock synchronization for distributed sensors.

[0017] Furthermore, when performing semantic parsing on the detection data based on semantic matching, and normalizing and standardizing the semantic labels of the detection data, the process includes:

[0018] The original descriptive information of the detection data is used to extract semantic features through natural language processing. The semantic feature extraction includes keyword recognition, part-of-speech analysis, entity extraction, and semantic nesting parsing.

[0019] Based on vector semantic similarity calculation, semantic distance measurement and hierarchical semantic relationship matching, the detected data is mapped to the corresponding standard label items, and the synonym, polysemous and abbreviation data labels are normalized.

[0020] The normalized labels are associated with their corresponding physical meanings, unit systems, and processing units, resulting in structured semantic label entries.

[0021] Furthermore, when processing the semantically standardized detection data to obtain metadata containing data source identifiers, data types, unit specifications, and time information, the following is included:

[0022] The semantically standardized detection data is decomposed into fields according to a preset structure template, and at least the following are extracted and aggregated:

[0023] The data source identifier field is used to record the sensor device number, the processing unit to which it belongs, and its geographic location code;

[0024] The data type field is used to indicate the physical attribute type of the detected value;

[0025] The unit specification field is used to standardize the physical units corresponding to numerical values.

[0026] The time information field is used to record the original acquisition time, system reception time, and synchronization timestamp;

[0027] All fields are encapsulated into structured metadata entries, represented in JSON, XML, or RDF semantic formats, while preserving the original data mapping index of the metadata.

[0028] Furthermore, when constructing a cross-layer data representation model based on transfer learning, the following steps are included:

[0029] A dual-branch hybrid model architecture consisting of a pre-trained semantic encoder and a graph neural network is adopted. The pre-trained semantic encoder is used to process the structured semantic tags contained in the metadata and generate an initial semantic embedding vector.

[0030] The graph neural network establishes a process-level structure graph using multiple wastewater treatment process units and their upstream and downstream relationships as nodes and edges. Based on the input metadata, it performs feature propagation and aggregation on the graph to obtain structured embedding vectors.

[0031] Furthermore, constructing a cross-layer data representation model based on transfer learning also includes:

[0032] Initial model parameters are obtained by training the source domain in wastewater treatment projects that have already established semantic annotation and process structure mapping.

[0033] Based on a domain-adaptive migration strategy, the model is migrated to the target wastewater treatment scenario and adjusted using labeled samples;

[0034] The domain adaptive migration strategy process includes feature distribution alignment, maximum mean difference regularization, and semantic label consistency constraints.

[0035] Furthermore, when performing feature extraction and semantic embedding on the metadata based on the aforementioned cross-layer data representation model to obtain a unified vectorized representation, the process includes:

[0036] The data source identifier field, data type field, unit specification field, and time information field in the metadata are independently subspace encoded, and word embedding, position embedding, and multi-scale temporal convolutional network are used to extract features and perform temporal modeling for each field.

[0037] After fusing the embedding vectors of each field, dimensional alignment and feature compression are performed to obtain the vectorized expression, which includes information substructures for representing physical semantics, processing stage location, and dynamic change trends.

[0038] Furthermore, when constructing an information knowledge graph for the entire wastewater treatment process based on the vectorized representation, it includes:

[0039] The entity elements in the vectorized representation are mapped to nodes in the knowledge graph. The entity elements include process units, sensor devices, monitoring parameters, time segments, and control commands.

[0040] Based on the wastewater treatment process logic and technological structure, directed edge relationships are established between nodes, and timestamps or operating conditions are added as attributes of the edges.

[0041] Obtain the information knowledge graph for the entire wastewater treatment process.

[0042] Furthermore, after constructing an information knowledge graph for the entire wastewater treatment process based on the vectorized expression, it also includes:

[0043] Early warning is issued based on the information knowledge graph covering the entire wastewater treatment process;

[0044] Dynamically update the state of each node and its associated edge relationships in the graph, and identify the drift trend of parameter nodes;

[0045] The status of nodes in the current graph is scored in real time based on historical operational data. When the score exceeds the threshold, an early warning flag is triggered. The propagation path of the abnormal impact is determined by graph path backtracking, and an early warning is issued based on the propagation path of the abnormal impact.

[0046] Compared with existing technologies, the beneficial effects of this invention are as follows: by deploying distributed sensors at each processing stage and adopting a unified access protocol and time synchronization mechanism, the invention achieves spatiotemporal consistent acquisition of raw data, completes label normalization and standardization by combining semantic matching, extracts metadata features through structured encoding, and constructs a cross-layer data representation model based on transfer learning. This enables a unified vectorized expression of temporal, semantic, and structural features across different data sources. Based on this vector representation, an information knowledge graph oriented towards the entire process is constructed, improving the data's fusionability, interpretability, and usability. This breaks the traditional data silo pattern and provides unified data support for core smart water management functions such as intelligent analysis, anomaly warning, and process optimization. Attached Figure Description

[0047] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0048] Figure 1 A flowchart of a heterogeneous data rule-making method for multi-source information fusion provided in an embodiment of the present invention. Detailed Implementation

[0049] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0050] In traditional wastewater treatment systems, the heterogeneous data architecture from multiple sources leads to poor data interoperability. Due to differences in naming conventions, communication protocols, unit systems, and time synchronization mechanisms among various sensor devices, data tags lack a unified semantic mapping standard, making it impossible to directly associate them with multidimensional parameters at different process stages. Manually defining variable meanings and processing logic makes it difficult to resolve temporal correlations, spatial distributions, and process-level dependencies between data, hindering the transformation of real-time monitoring data into domain knowledge.

[0051] For example, a wastewater treatment plant deployed pH sensors, dissolved oxygen probes, and ammonia nitrogen analyzers in its biological reactor and secondary sedimentation tank. The pH sensors transmitted data using the OPC-UA protocol, the dissolved oxygen probes reported via the MQTT protocol, and the ammonia nitrogen analyzers accessed the system via HTTP. The data values ​​from different devices were labeled as "pH_1," "DO_Value," and "NH3-N_Concentration," respectively, using units of "none," "mg / L," and "ppm." Due to the lack of a unified semantic parsing mechanism, the system could not automatically recognize the synonymy between "DO_Value" and "dissolved oxygen," nor could it convert the ppm unit of "NH3-N_Concentration" to the standard mg / L unit. Furthermore, the lack of precise clock synchronization between devices resulted in timestamp discrepancies of up to seconds for different parameters within the same process stage, making it impossible to accurately establish a dynamic correlation model between dissolved oxygen changes and ammonia nitrogen degradation rates.

[0052] If the above problems are not addressed, semantic ambiguity in multi-source heterogeneous data will make it difficult to model causal relationships between process parameters, increasing the trend prediction error of key water quality indicators. Data silos will limit data fusion analysis across treatment units, making it impossible to achieve global optimization of control parameters such as aeration rate and sludge return ratio. Process adjustment decisions will rely on offline manual experience, failing to respond in real time to influent load fluctuations, potentially leading to operational risks such as instantaneous exceedances of effluent chemical oxygen demand or sludge bulking.

[0053] For this, please refer to Figure 1 This application proposes a method for regularizing heterogeneous data through multi-source information fusion, including:

[0054] S100: Distributed sensors are deployed in several treatment stages of the wastewater treatment process, and the detection data of all distributed sensors are collected in a time-series synchronization manner through an access protocol.

[0055] S200: Based on semantic matching, perform semantic parsing on the detection data, normalize the labels of the detection data, and complete the semantic label standardization.

[0056] S300: Processes the semantically standardized detection data to obtain metadata containing data source identifier, data type, unit specification, and time information.

[0057] S400: Construct a cross-layer data representation model based on transfer learning, and extract features and embed semantics into metadata according to the cross-layer data representation model to obtain a unified vectorized expression.

[0058] S500: Construct an information knowledge graph for the entire wastewater treatment process based on vectorized representation.

[0059] Specifically, distributed sensors refer to multiple sensing devices with independent detection functions deployed at different stages of wastewater treatment. These can be implemented using temperature, dissolved oxygen, pH, and conductivity sensors to collect physicochemical parameters of each process unit in real time, thus solving the problem of dispersed and heterogeneous data sources.

[0060] Semantic matching refers to parsing the original description of detection data using natural language processing technology. Specifically, it can be achieved through keyword recognition, entity extraction, and semantic nesting parsing. This is used to eliminate naming differences between data labels from different sensors and resolve interoperability barriers caused by semantic ambiguity.

[0061] Metadata refers to structured description entries that include data source identifiers, data types, unit specifications, and time information. Specifically, it can be encapsulated in JSON or XML formats to unify the physical attributes and spatiotemporal benchmarks of multi-source data and solve the problem of inconsistencies between data formats and unit systems.

[0062] The data cross-layer representation model of transfer learning refers to a machine learning architecture that can uniformly encode data features at different levels. Specifically, it can be implemented using a dual-branch hybrid model of a pre-trained semantic encoder and a graph neural network to extract related features across process stages and solve the problem of difficulty in jointly analyzing heterogeneous data.

[0063] Vectorization refers to converting metadata into embedded vectors in a unified mathematical space. Specifically, it can be achieved using word embedding, positional embedding, and multi-scale temporal convolutional networks. It is used to represent the physical semantics and dynamic trends of data and solve the problem of the inability to integrate multi-dimensional information.

[0064] Information knowledge graphs refer to knowledge bases that express the entity relationships of the entire wastewater treatment process in a graph structure. Specifically, they can be constructed using process unit nodes and directed edge relationships to reveal the logical connections between parameters and solve the problem that real-time data cannot be transformed into reasonable knowledge.

[0065] This application utilizes multi-source information fusion technology to transform dispersed and heterogeneous sensor data from wastewater treatment processes into a unified vectorized representation and constructs a knowledge graph across process levels. This scheme eliminates data label differences through semantic matching and combines it with a transfer learning model to achieve cross-layer feature extraction, ultimately forming a structured knowledge system that can support intelligent decision-making. This effectively solves the problems of traditional methods that rely on manually set variables and cannot capture multi-level relationships.

[0066] The working process and principle of this application are as follows: Distributed sensors are deployed at multiple treatment stages of the wastewater treatment process, and the detection data of these sensors are collected synchronously with the time sequence via an access protocol. Next, the detection data undergoes semantic parsing and label normalization to achieve semantic label standardization. Then, the standardized data is processed to generate metadata containing data source identifiers, data types, unit specifications, and time information. A cross-layer data representation model based on transfer learning is constructed, and this model is used to extract features and semantically embed the metadata to obtain a unified vectorized expression. Finally, an information knowledge graph covering the entire wastewater treatment process is constructed based on the vectorized expression.

[0067] Raw data is collected through distributed sensors to ensure the comprehensiveness of data sources. Semantic parsing and standardization steps eliminate semantic differences between data, laying the foundation for subsequent processing. Metadata generation further standardizes the data structure. The construction and application of a cross-layer data representation model enables a unified expression of heterogeneous data, and the resulting knowledge graph integrates information from the entire wastewater treatment process.

[0068] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0069] Distributed sensors are deployed in various stages of the wastewater treatment plant, including grit chambers, sedimentation tanks, biological reactors, secondary sedimentation tanks, and disinfection tanks. These sensors include temperature sensors, dissolved oxygen sensors, pH sensors, conductivity sensors, turbidity sensors, flow meters, and level gauges. Multiple protocols such as OPC-UA, MQTT, and HTTP are used to access these sensors, and data encapsulation and format standardization are achieved through protocol adaptation middleware. Network Time Protocol (NTP) or IEEE 1588 Precision Time Protocol (PMT) are used to synchronize the sensors' clocks, ensuring data timing consistency.

[0070] Semantic analysis is performed on the collected detection data. First, semantic features of the data are extracted using natural language processing techniques, including keyword recognition, part-of-speech analysis, and entity extraction. Then, based on vector semantic similarity calculation and semantic distance measurement, the data is mapped to standard labels. Synonyms, polysemous words, and abbreviations are normalized, and the standardized labels are associated with their physical meaning, unit system, and processing unit.

[0071] The standardized data is processed to generate metadata. This metadata includes data source identifiers (e.g., sensor number, processing unit), data type (e.g., physical attribute type), unit specifications, and time information (e.g., acquisition time, reception time, synchronization timestamp). The metadata is encapsulated in JSON or XML format, while retaining the original data's mapping index.

[0072] A cross-layer data representation model based on transfer learning is constructed. This model consists of a pre-trained semantic encoder and a graph neural network. The semantic encoder processes structured semantic tags in metadata to generate initial semantic embedding vectors. The graph neural network constructs a process-level structure graph, using wastewater treatment process units as nodes and process flow relationships as edges. Feature propagation and aggregation are performed on the graph to obtain structured embedding vectors.

[0073] The constructed model is used to extract features and embed semantics into metadata. Each field of the metadata is independently subspace-encoded, and word embedding, positional embedding, and multi-scale temporal convolutional networks are used for feature extraction and temporal modeling. The embedding vectors of each field are merged, and dimensionality alignment and feature compression are performed to obtain a unified vectorized representation.

[0074] A knowledge graph for the entire wastewater treatment process is constructed based on vectorized representations. Entities (such as process units, sensor devices, and monitoring parameters) in the vectorized representations are mapped to nodes in the graph. Directed edge relationships between nodes are established according to the wastewater treatment process logic and structure, with timestamps or operating conditions added as attributes of the edges.

[0075] Through the above-described scheme, this application achieves effective fusion and regularization of multi-source heterogeneous data. Unified semantic parsing and standardized processing eliminate differences in naming, units, and temporal granularity among different sensor data, improving data interoperability. The data cross-layer representation model based on transfer learning can capture the temporal, spatial, and logical multi-level relationships in the wastewater treatment process, transforming real-time monitoring data into reasonable and reusable domain knowledge. The constructed information knowledge graph covering the entire wastewater treatment process provides a foundation for intelligent decision-making, facilitating global optimization of process parameters and real-time response to influent load fluctuations, reducing operational risks such as instantaneous exceedances of effluent chemical oxygen demand or sludge bulking.

[0076] In some of the solutions described above in this application, distributed sensors are deployed at several treatment stages of the wastewater treatment process to collect detection data. However, since the process objectives of different treatment stages vary greatly, if the sensor type and parameter detection function are not adapted to the specific treatment unit, it may lead to the omission or redundancy of key process parameters, thereby affecting the completeness and accuracy of subsequent data regularization processing.

[0077] This application further proposes to install sensor nodes with functions for detecting temperature, dissolved oxygen, pH, conductivity, turbidity, flow rate, liquid level, chemical oxygen demand and ammonia nitrogen in the grit chamber, grit chamber, biological reactor, secondary sedimentation tank, disinfection tank and sludge treatment stage.

[0078] The sensor nodes in the bar screen tank include flow and level detection functions to monitor influent flow and the accumulation of intercepted impurities. The sensor nodes in the grit chamber are equipped with turbidity and conductivity detection functions to quantify grit settling efficiency and water ion concentration. The sensor nodes in the bioreactor integrate dissolved oxygen, chemical oxygen demand (COD), and ammonia nitrogen detection functions to assess the oxidation environment and pollutant removal efficiency during biodegradation. The sensor nodes in the secondary sedimentation tank are equipped with turbidity and level detection functions to monitor sludge settling efficiency and effluent clarity. The sensor nodes in the disinfection tank are equipped with pH and conductivity detection functions to control disinfectant dosage and water quality stability. The sensor nodes in the sludge treatment stage include flow and level detection functions to optimize sludge dewatering and transportation processes.

[0079] Specifically, the functional configuration of sensor nodes in each treatment unit enables targeted data collection based on its process objectives. For example, the bioreactor uses a dissolved oxygen sensor to monitor the oxygen content required for microbial activity in real time, and combines this with an ammonia nitrogen sensor to quantify denitrification efficiency, ensuring that the metabolic environment during the biological treatment stage is within the optimal range. The disinfection tank uses a pH sensor to detect the acidity or alkalinity of the water, and combines this with a conductivity sensor to assess the residual concentration of disinfectant, avoiding secondary pollution caused by excessive addition. By deploying functionally adapted sensor nodes at each treatment stage, key process parameters throughout the entire wastewater treatment process can be fully covered, avoiding data regularization biases caused by missing or redundant parameters. As a result, subsequent semantic parsing, metadata generation, and knowledge graph construction processes can be based on comprehensive and accurate detection data, thereby improving the reliability of the entire wastewater treatment process's status characterization and anomaly early warning.

[0080] As a preferred embodiment, the solution of this application is implemented as follows: Distributed sensors are deployed at each treatment stage of the wastewater treatment plant. Specifically, temperature and pH sensors are installed in the grit chamber; dissolved oxygen and turbidity sensors are installed in the grit chamber; conductivity, chemical oxygen demand, and ammonia nitrogen sensors are installed in the bioreactor; flow sensors are installed in the secondary sedimentation tank; level sensors are installed in the disinfection tank; and temperature and pH sensors are installed in the sludge treatment stage. These sensor nodes each have corresponding detection functions and can monitor key parameters at each treatment stage in real time.

[0081] Through the above technical solution, this application achieves comprehensive monitoring of the entire wastewater treatment process. This allows for timely understanding of the operational status of each treatment unit, providing data support for process control and anomaly early warning. Furthermore, the coordinated deployment of multiple sensors helps capture parameter correlations between different treatment stages, providing a basis for the overall optimization of the wastewater treatment process.

[0082] In some of the solutions mentioned above in this application, the detection data of all distributed sensors are collected through access protocol and timing synchronization. However, the use of heterogeneous communication protocols by different sensors results in the inability to uniformly encapsulate the data format and the lack of a unified clock synchronization mechanism, causing data acquisition timing disorder.

[0083] This application further proposes a method for collecting detection data from all distributed sensors through access protocols and time synchronization, including: using OPC-UA, MQTT, or HTTP protocols to connect various sensors to the same data acquisition network using heterogeneous communication protocols, and performing unified data encapsulation and format standardization through protocol adaptation middleware. It also combines network time protocols or the IEEE 1588 precision time protocol to perform unified clock synchronization for the distributed sensors.

[0084] The protocol adaptation middleware incorporates multiple communication protocol conversion modules, capable of converting data transmitted via different protocols into data packets of a unified format. For example, when sensors transmit data using the Modbus protocol, the middleware converts it to JSON format and adds a timestamp field. The clock synchronization module broadcasts synchronization signals to all sensors via a time server, ensuring that clock errors at each node are controlled within milliseconds. For high-precision scenarios, the IEEE 1588 protocol, supporting sub-microsecond synchronization, is used to achieve precise time synchronization through hardware timestamps.

[0085] Specifically, when dissolved oxygen sensors deployed in bioreactors send detection data via the MQTT protocol, the protocol adaptation middleware parses the message payload into structured data, adds the device number and process unit code, and encapsulates it into standardized data frames. Simultaneously, the clock synchronization module sends a time synchronization command to the sensor network every 5 minutes to ensure consistent time reference across all nodes. For example, a turbidity sensor deployed in a wastewater treatment plant uses the HTTP protocol to transmit data. The middleware converts this data into XML format data packets containing the device ID, value, unit, and timestamp, and calibrates the sensor's built-in clock via an NTP server, ensuring that the time deviation of all 300 sensors in the plant is less than 50 milliseconds. This dual standardization mechanism aligns detection data from different manufacturers and using different protocols in both format and time dimensions, providing structurally consistent and time-accurate input data for subsequent semantic parsing and feature extraction.

[0086] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0087] Temperature, dissolved oxygen, and pH sensors are connected to the data acquisition network using the OPC-UA protocol. Specifically, the temperature sensor is a PT100 model, with a measurement range of 0-100℃ and an accuracy of ±0.1℃. The dissolved oxygen sensor uses a fluorescence method, with a measurement range of 0-20 mg / L and an accuracy of ±0.1 mg / L. The pH sensor uses a glass electrode method, with a measurement range of 0-14 and an accuracy of ±0.01 pH. These sensors are connected to the OPC-UA server via an RS485 interface.

[0088] The flow meter and level gauge are connected to the data acquisition network using the MQTT protocol. Specifically, the flow meter is electromagnetic, with a measurement range of 0-1000 m³ / h and an accuracy of ±0.5%. The level gauge is ultrasonic, with a measurement range of 0-10 m and an accuracy of ±1 mm. These sensors output 4-20mA analog signals, which are converted into MQTT messages by a signal converter and published to the MQTT broker server.

[0089] The chemical oxygen demand (COD) analyzer and ammonia nitrogen analyzer are connected to the data acquisition network using the HTTP protocol. Specifically, the COD analyzer uses the potassium dichromate method, with a measurement range of 0-1000 mg / L and an accuracy of ±3%. The ammonia nitrogen analyzer uses the Nessler's reagent colorimetric method, with a measurement range of 0-50 mg / L and an accuracy of ±3%. These analyzers upload data to the web server via Ethernet interface using the HTTP POST method.

[0090] Furthermore, a protocol adaptation middleware is used to unify data encapsulation and standardize formats. Specifically, the middleware is built on the Node-RED open-source platform and uses JavaScript functions to convert data from different protocols into a unified JSON format, including fields such as device ID, measurement parameters, values, units, and timestamps.

[0091] Therefore, a Network Time Protocol (NTP) is used to synchronize the clocks of distributed sensors. Specifically, an NTP server is deployed in the data acquisition network, and all sensor devices and servers synchronize their time with the NTP server through NTP clients, achieving synchronization accuracy down to the millisecond level.

[0092] Through the above technical solutions, this application achieves unified access and format standardization of heterogeneous sensor data, improving the efficiency and reliability of data acquisition. The unified data format facilitates subsequent processing and analysis, reducing the workload of data conversion and cleaning. Clock synchronization ensures the time consistency of data from different sources, laying the foundation for multi-source data fusion and time-series analysis. Furthermore, the adoption of open standard communication protocols improves the system's interoperability and scalability.

[0093] In some of the solutions described above in this application, when semantic parsing of detection data is performed based on semantic matching, the original data labels have differences in synonymy, polysemy, and abbreviation, which makes it impossible to uniformly map the semantic descriptions of different sensor data to the standard label system, affecting the accuracy of subsequent metadata generation and knowledge graph construction.

[0094] This application further proposes a semantic parsing method for detection data based on semantic matching. When normalizing and standardizing the semantic labels of the detection data, the method includes: extracting semantic features from the original descriptive information of the detection data using natural language processing, including keyword recognition, part-of-speech analysis, entity extraction, and semantic nesting parsing; mapping the detection data to corresponding standard label items based on vector semantic similarity calculation, semantic distance measurement, and hyponymous semantic relationship matching, and normalizing synonymous, polysemous, and abbreviated data labels; and associating the normalized labels with their corresponding physical meanings, unit systems, and processing units to output structured semantic label entries.

[0095] Semantic feature extraction employs word vector models from natural language processing to segment and vectorize the original descriptive text, such as using Word2Vec or BERT models for keyword recognition and entity extraction. Vector semantic similarity calculation measures the semantic association between different tags using cosine similarity or Euclidean distance; for example, a vector similarity threshold of 0.85 or higher is set for "DO" and "dissolved oxygen" to classify them as synonymous tags. Hyper- and hypo-semantic relationship matching establishes hierarchical relationships based on an ontology library in the wastewater treatment field; for example, "COD" is mapped to the hyper-level concept "water quality indicators." Structured semantic tag entries are stored in triplet format, such as <sensor number, standard tag name, processing unit>.

[0096] Specifically, the raw descriptive information of the detection data is first input into a pre-trained natural language processing model. A bidirectional long short-term memory network extracts contextual features to identify key entities such as "pH" and "liquid level." Then, the similarity between the original label vector and candidate labels in a standard label library is calculated. When the semantic distance between "TOC" and "total organic carbon" is below a preset threshold, they are merged into the same standard label. Furthermore, the hierarchical relationship of the labels is verified using a domain knowledge graph, for example, confirming that "NH3-N" belongs to the "nitrogen compounds" category. The final generated semantic label entries include physical unit conversion rules, such as unifying "mg / L" to "g / m³," and are associated with the geocoding of the bioreactor. This process achieves semantic consistency across sensor data by eliminating label ambiguity and redundancy.

[0097] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0098] The raw descriptive information of the detection data is used for semantic feature extraction through natural language processing. Semantic feature extraction includes keyword recognition, part-of-speech tagging, entity extraction, and semantic nesting parsing. For example, for a raw description like "pH value 7.2", keyword recognition extracts "pH" and "7.2", part-of-speech tagging identifies "pH" as a noun and "7.2" as a numerical value, entity extraction identifies "pH" as a water quality parameter and "7.2" as a specific numerical value, and semantic nesting parsing establishes the attribute-value relationship between "pH" and "7.2".

[0099] Furthermore, based on vector semantic similarity calculation, semantic distance measurement, and hyponym semantic relationship matching, the detected data is mapped to corresponding standard label items, and normalization processing is performed on synonymous, polysemous, and abbreviated data labels. Specifically, the cosine similarity between the original description and the standard label is calculated using a pre-trained word vector model, and the standard label with the highest similarity is selected. Hyponym relationships in the semantic tree are also considered; for example, "dissolved oxygen" is a hyponym of "water quality parameter." For abbreviated labels like "DO," they are mapped to the standard "dissolved oxygen" label through a thesaurus.

[0100] Thus, normalized labels are associated with their corresponding physical meanings, unit systems, and processing units, outputting structured semantic label entries. For example, "pH 7.2" is normalized into a structured entry of {"Parameter": "pH value", "Value": 7.2", "Unit": "Dimensionless", "Processing Unit": "Bioreactor"}.

[0101] Through the above technical solutions, this application achieves semantic unification and standardized expression of heterogeneous data. The semantic features of the detection data are accurately extracted, eliminating ambiguity caused by non-standard naming. Standardized semantic tags facilitate subsequent data analysis and knowledge reasoning, improving data interoperability. Structured semantic tag entries contain complete contextual information, laying the foundation for information fusion throughout the entire wastewater treatment process.

[0102] In some of the solutions mentioned above in this application, although the detection data after semantic matching and standardization has achieved label normalization, it lacks a unified structured metadata organization method, which makes it difficult for subsequent cross-layer representation models to effectively integrate device attributes, physical units and time dimension information in multi-source heterogeneous data, affecting the accurate mapping of entity relationships in the knowledge graph construction process.

[0103] This application further proposes to decompose the semantically standardized detection data into fields according to a preset structure template, extract and collect at least the data source identifier field, data type field, unit specification field, and time information field, encapsulate all fields into structured metadata entries, represent them based on JSON, XML, or RDF semantic formats, and retain the original data mapping index of the metadata.

[0104] The data source identifier field records the data source through a combination of device number, processing unit name, and geographic location code. For example, "BIO-01-12.34N / 45.67E" represents the No. 1 sensor in the bioreactor and its latitude and longitude coordinates. The data type field is matched and labeled using a predefined physical attribute type library; for example, "dissolved oxygen concentration" corresponds to the "DO" type code. The unit specification field uses the International System of Units (SI) for forced conversion; for example, "mg / L" is uniformly converted to "kg / m³". The time information field records the acquisition, reception, and synchronization time in Coordinated Universal Time (UTC) format, for example, "2023-08-01T08:00:00Z". When the structured metadata entries are in JSON format, a nested key-value pair structure can be constructed, containing primary keys such as "source_id", "data_type", "unit", and "timestamp", with sub-fields attached to each primary key. The raw data mapping index generates a unique identifier using a hash algorithm; for example, the SHA-256 algorithm is used to calculate the digest of the raw detection values.

[0105] Specifically, when the semantically standardized detection data enters the metadata encapsulation stage, the preset structure template first splits the data stream into independent fields. The data source identifier field generates a unique identifier containing spatial location information by parsing the sensor device number and the process unit topology, ensuring data traceability. The data type field converts semantic tags into standardized codes by querying a pre-set physical attribute mapping table, eliminating naming ambiguity. The unit specification field normalizes numerical values ​​using a unit conversion function library, for example, unifying flow units of different dimensions to "m³ / s". The time information field aligns the time base of multi-source data through a time synchronization service and records the complete timestamp sequence. During encapsulation, the JSON format uses a hierarchical nested structure to organize the fields; for example, under the "timestamp" key, three subkeys, "collect", "receive", and "sync", are set to store different time nodes respectively. The RDF format expresses entity relationships in the form of triples, for example, "..." <sensor001> <hasunit><kg / m³>The original data mapping index establishes a bidirectional association between metadata entries and original detection data through hash values, ensuring data traceability. Through the above structured processing, metadata entries can fully carry multi-dimensional information, providing a unified data interface for subsequent feature extraction and knowledge graph construction.

[0106] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0107] When processing the semantically standardized detection data to obtain metadata containing data source identifiers, data types, unit specifications, and time information, the following are included:

[0108] The semantically standardized detection data is decomposed into fields according to a preset structure template, and at least the following are extracted and aggregated:

[0109] The data source identifier field is used to record the sensor device number, the processing unit to which it belongs, and its geographic location code.

[0110] The data type field is used to indicate the physical attribute type of the detected value.

[0111] The unit specification field is used to standardize the physical units corresponding to numerical values.

[0112] The time information field is used to record the original acquisition time, system reception time, and synchronization timestamp.

[0113] All fields are encapsulated into structured metadata entries, represented in JSON, XML, or RDF semantic formats, while preserving the original data mapping index of the metadata.

[0114] Specifically, the following steps can be used to achieve this:

[0115] First, define a pre-defined structured template containing four main fields: data source identifier, data type, unit specification, and time information.

[0116] Secondly, each piece of semantically standardized detection data is parsed, and relevant information is extracted and filled into the corresponding fields:

[0117] The data source identifier field can record the device number in the form of "DO_sensor_01_bio_reactor_A", the treatment unit name in the form of "biological_treatment", and the geographic coordinates in the form of "30.5425°N, 114.3572°E".

[0118] The data type field can be labeled as a physical quantity type such as "dissolved_oxygen".

[0119] Unit specifications are standardized to the International System of Units (SI), such as dissolved oxygen concentration being specified as "mg / L".

[0120] The time information field records the original acquisition time in the format "2023-05-01 10:30:15.000", the system receiving time, and the Unix timestamp in the format "1682912215000".

[0121] Then, convert the populated structured data into JSON format, for example:

[0122] {

[0123] "data_source": {

[0124] "device_id": "DO_sensor_01_bio_reactor_A",

[0125] "process_unit": "biological_treatment",

[0126] "location": "30.5425°N, 114.3572°E"

[0127] },

[0128] "data_type": "dissolved_oxygen",

[0129] "unit": "mg / L",

[0130] "time_info": {

[0131] "collection_time": "2023-05-01 10:30:15.000",

[0132] "receive_time": "2023-05-01 10:30:15.500",

[0133] "timestamp": 1682912215000

[0134] },

[0135] "value": 6.5,

[0136] "raw_data_index": 1024

[0137] }

[0138] Finally, the metadata in JSON format is stored in a database or file system, and a mapping relationship with the original data is established.

[0139] Through the above technical solutions, this application achieves a unified structured representation of heterogeneous data. This improves data interoperability and parsability, laying the foundation for subsequent data analysis and knowledge graph construction. Furthermore, by preserving the original data mapping, the traceability of the data processing process is ensured. In addition, adopting the standard JSON format facilitates cross-platform data exchange and system integration.

[0140] In some of the solutions described above in this application, there is a problem that structured semantic labels and process structure information are difficult to effectively integrate when constructing cross-layer data representation models. Traditional single-model architectures cannot simultaneously handle semantic features in metadata and topological relationships between process units, resulting in a lack of hierarchy and correlation in feature representation, which affects the accuracy of entity mapping in subsequent knowledge graphs.

[0141] This application further proposes a dual-branch hybrid model architecture consisting of a pre-trained semantic encoder and a graph neural network. The pre-trained semantic encoder is used to process the structured semantic tags contained in the metadata and generate initial semantic embedding vectors. The graph neural network uses multiple wastewater treatment process units and their upstream and downstream relationships as nodes and edges to establish a process-level structure graph, and performs feature propagation and aggregation on the graph based on the input metadata to obtain structured embedding vectors.

[0142] The pre-trained semantic encoder employs a BERT model based on the Transformer architecture. Pre-trained on massive industrial text corpora, it achieves deep parsing of structured information such as entity types and unit specifications within semantic tags. The graph neural network uses graph convolutional networks or graph attention networks to model process units such as bioreactors and secondary sedimentation tanks as graph nodes, with material flow directions and process dependencies as edge connections. The dual-branch architecture achieves the fusion of semantic embedding vectors and graph structure features through a cross-modal attention mechanism, for example, by using a gating fusion module to dynamically adjust the weight ratio of the two types of features.

[0143] Specifically, during the model training phase, the pre-trained semantic encoder first vectorizes the semantic tags in the metadata, generating an initial semantic embedding with a dimension of 768. Simultaneously, the graph neural network constructs an adjacency matrix based on the connection relationships between process units, aggregating the feature information of adjacent nodes through three layers of graph convolution operations. In the feature fusion phase, the semantic embedding vector and the graph structure embedding vector are input into a multilayer perceptron for dimension alignment, and after normalization, they are concatenated into a 1024-dimensional joint vector. This joint vector, through residual connections and layer normalization operations, ultimately outputs a structured embedding containing both semantic features and topological relationships. For example, the embedding vector of the sludge treatment unit contains both the semantic feature of "ammonia nitrogen concentration" and the process association information between this unit and the bioreactor and disinfection tank. This dual-branch architecture effectively solves the limitations of traditional single-modal models in cross-layer feature representation, improving the entity mapping accuracy of the subsequent knowledge graph to 92.3%.

[0144] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0145] When constructing a cross-layer data representation model based on transfer learning, a dual-branch hybrid model architecture consisting of a pre-trained semantic encoder and a graph neural network is adopted. The pre-trained semantic encoder uses the BERT model to process the structured semantic labels contained in the metadata and generate initial semantic embedding vectors. The graph neural network uses a graph attention network (GAT) to establish a process-level structure graph with multiple wastewater treatment process units and their upstream and downstream relationships as nodes and edges. Specifically, the grit chamber, grit chamber, bioreactor, secondary sedimentation tank, disinfection tank, and sludge treatment unit are used as nodes in the graph, and the material transport relationships in the process flow are used as directed edges. Based on the input metadata, feature propagation and aggregation are performed on the graph to obtain structured embedding vectors.

[0146] Through the above technical solution, this application achieves a unified representation of multi-source heterogeneous data in the wastewater treatment process. Thus, data from different sources and in different formats can be mapped to the same semantic space, facilitating subsequent data analysis and knowledge reasoning. Furthermore, by introducing graph neural networks, this solution can effectively capture the relationships between various process units in wastewater treatment, improving the accuracy and completeness of data representation.

[0147] In some of the solutions described above in this application, when constructing a cross-layer data representation model based on transfer learning, the direct application of the initial model parameters in the target wastewater treatment scenario may lead to a mismatch between the semantic labels and the process structure mapping relationship, limiting the model's generalization ability and making it unable to effectively adapt to the heterogeneous data distribution differences of different wastewater treatment projects.

[0148] This application further proposes obtaining initial model parameters by training in a wastewater treatment project with existing semantic annotations and process structure mappings. Based on a domain-adaptive transfer strategy, the model is transferred to the target wastewater treatment scenario and adjusted using labeled samples. The domain-adaptive transfer strategy includes feature distribution alignment, maximum mean difference regularization, and semantic label consistency constraints.

[0149] In this process, source domain training utilizes a wastewater treatment project dataset with complete semantic annotations and process structure mappings. Supervised learning optimizes model parameters, enabling the model to capture the correlation between data features and process logic. In the domain adaptive transfer strategy, feature distribution alignment is achieved between the source and target domain feature spaces through adversarial training or covariance matching. Maximum mean difference regularization uses a kernel function to calculate the difference in data distribution between the two domains and uses it as a regularization term to constrain model optimization. Semantic label consistency constraints ensure that the semantic labels of the target domain samples remain consistent with those of the source domain through contrastive learning or label propagation algorithms.

[0150] Specifically, during the source domain training phase, labeled data from historical wastewater treatment projects are used. The parameters of the pre-trained semantic encoder and graph neural network are optimized using a cross-entropy loss function, enabling the model to accurately extract semantic features from metadata and establish connections between process units. When migrating to the target scenario, a feature distribution alignment module spatially maps the metadata features of the source and target domains. Maximum mean difference regularization minimizes the difference in feature distributions between the two domains, reducing the impact of domain bias on model performance. Simultaneously, a semantic label consistency constraint module ensures the consistency of label semantics during the migration process by comparing the similarity of embedding vectors of identical semantic labels in the source and target domains. For example, when migrating dissolved oxygen monitoring parameters in the sludge treatment unit, the feature distribution alignment module performs covariance matching between the distribution of sensor data in the target domain and the source domain. Maximum mean difference regularization uses a Gaussian kernel function to calculate the distribution distance of the two domains' data in the latent space and incorporates it into the loss function. The semantic label consistency constraint module forces the model to maintain semantic consistency by calculating the cosine similarity between the dissolved oxygen label in the target domain and the corresponding label in the source domain. Therefore, the model only needs a small number of labeled samples to complete parameter fine-tuning in the target scenario, enabling efficient transfer across wastewater treatment projects and improving the accuracy and stability of data representation.

[0151] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0152] When constructing a cross-layer data representation model based on transfer learning, the following are also included:

[0153] Initial model parameters were obtained by training the model in a wastewater treatment project with existing semantic annotations and process structure mappings. Specifically, a stable wastewater treatment plant with well-annotated data was selected as the source domain, and the initial model was trained using the plant's historical operating data and expert annotation information. During training, the batch stochastic gradient descent algorithm was used to optimize the model parameters, with a learning rate of 0.001 and 100 training epochs.

[0154] Based on a domain-adaptive transfer strategy, the model is transferred to the target wastewater treatment scenario and adjusted using labeled samples. The domain-adaptive transfer strategy includes feature distribution alignment, maximum mean difference regularization, and semantic label consistency constraints. Feature distribution alignment uses the maximum mean difference (MMD) loss function to calculate and minimize the distance between the feature distributions of the source and target domains. Maximum mean difference regularization controls the difference in feature distributions between the source and target domains by adding an MMD term to the objective function. Semantic label consistency constraints introduce label prediction consistency loss to ensure that the model produces consistent prediction results for data with the same semantics in both the source and target domains.

[0155] Furthermore, a progressive transfer strategy was employed during the transfer process. First, the parameters of the pre-trained semantic encoder were frozen, and only the graph neural network portion was fine-tuned. Then, the high-level parameters of the semantic encoder were gradually unfrozen for further fine-tuning. Finally, the entire model was fine-tuned end-to-end. During fine-tuning, the learning rate was set to 0.0001, and the number of iterations was 50.

[0156] Therefore, through the above technical solution, this application achieves cross-scenario transfer and adaptation of wastewater treatment data. The model can quickly adapt to new wastewater treatment scenarios, reducing its dependence on large amounts of labeled data. Simultaneously, through feature distribution alignment and semantic consistency constraints, the model's generalization ability and prediction accuracy in the target domain are improved. Furthermore, the progressive transfer strategy effectively prevents catastrophic forgetting during the transfer process, ensuring the stability of model performance.

[0157] In some of the solutions described above in this application, when the cross-layer data representation model performs feature extraction and semantic embedding on metadata, the encoding methods of different fields fail to effectively separate information such as physical semantics, processing stage location, and dynamic change trends, resulting in information mixing and dimensional redundancy problems in vectorized representation.

[0158] This application further proposes to encode the data source identifier field, data type field, unit specification field, and time information field in metadata independently in a subspace. Word embedding, position embedding, and multi-scale temporal convolutional networks are used to extract features and perform temporal modeling for each field. After fusing the embedding vectors of each field, dimensional alignment and feature compression are performed to obtain a vectorized representation. This vectorized representation contains information substructures used to characterize physical semantics, processing stage location, and dynamic change trends.

[0159] Independent subspace encoding is achieved by assigning an independent encoding channel to each field. For example, the data source identifier field uses hash word embedding based on device number and geocoding; the data type field uses pre-trained word vector mapping based on physical attribute labels; the unit specification field uses joint encoding of unit conversion coefficients and dimensional symbols; and the time information field uses location embedding of segmented timestamps. The multi-scale temporal convolutional network consists of multiple temporal convolutional layers with different kernel widths, for example, setting three time windows of 1 minute, 5 minutes, and 15 minutes to capture dynamic change patterns at different time granularities. Dimension alignment and feature compression are achieved by mapping the embedding vectors of each field to a unified dimensional space through fully connected layers, followed by dimensionality reduction through principal component analysis or autoencoders.

[0160] Specifically, the hash word embedding of the data source identifier field converts the device number and geocoding into a fixed-length vector, avoiding encoding conflicts between different devices. The pre-trained word vectors for the data type field are obtained through training on a corpus in the wastewater treatment domain, ensuring that the semantic similarity of physical attribute labels is maintained in the vector space. The joint encoding of the unit standardization field performs numerical normalization and symbol embedding on the unit conversion coefficients and dimensional symbols respectively, eliminating the influence of unit system differences on the numerical values. The segmented position embedding of the time information field divides the original acquisition time, system reception time, and synchronization timestamp into three time segments: hour, minute, and second, and performs sinusoidal position encoding on each segment. The multi-scale temporal convolutional network adopts a parallel convolutional structure, and the outputs of each convolutional layer are concatenated into a temporal feature vector after max pooling. During dimensional alignment, the weight matrix of the fully connected layer is dynamically adjusted according to the embedding dimension of each field, ensuring that the vectors of different fields are comparable in a unified space. Feature compression is achieved through the encoder part of the autoencoder, retaining 95% of the variance contribution rate to eliminate redundant dimensions. In the resulting vectorized representation, physical semantic information is dominated by the embedding vectors of data type and unit fields, processing stage location information is represented by the embedding vectors of data source identifier fields, and dynamic change trend information is described by the temporal convolution features of the time field.

[0161] As a preferred embodiment, the specific implementation of this application is as follows: The data source identifier field in the metadata undergoes word embedding processing. The device number is converted into a 64-dimensional distributed vector representation using a pre-trained word vector model. The geolocation encoding uses a location embedding algorithm based on polar coordinates to generate a 128-dimensional vector. The data type field is mapped using a physical attribute ontology library, and a one-hot encoding is used to generate a category vector. The unit specification field is normalized using an International System of Units (SI) conversion table, and a linear projection layer is used to generate a 32-dimensional embedding vector. The time information field is input into a multi-scale temporal convolutional network containing three parallel dilated convolutional layers with dilation factors set to 1, 3, and 5, respectively, to capture temporal features at the hourly, daily, and weekly levels. After layer normalization, the encoding results of each subspace are fused using a gated attention mechanism. Finally, a fully connected layer compresses the fused 1024-dimensional vector into a 256-dimensional vectorized expression.

[0162] Through the above technical solutions, this application achieves a unified representation of multimodal metadata features, solving the semantic fragmentation problem caused by differences in encoding methods for different attribute fields in traditional methods. The independent subspace encoding mechanism effectively isolates the interference of heterogeneous features such as equipment identification and spatiotemporal information, and combined with multi-scale temporal modeling, accurately captures the dynamic evolution of wastewater treatment parameters. The feature compression process reduces the redundancy of high-dimensional vectors, enabling the subsequent knowledge graph construction process to efficiently identify the relationships between entities.

[0163] In some of the solutions described above in this application, the heterogeneity of different fields in the feature extraction and semantic embedding process of the cross-layer data representation model makes feature fusion difficult, making it difficult to effectively integrate information on physical semantics, processing stage position and dynamic change trend, thus affecting the integrity and consistency of vectorized expression.

[0164] This application further proposes to encode the data source identifier field, data type field, unit specification field, and time information field in metadata independently in a subspace. Word embedding, position embedding, and multi-scale temporal convolutional networks are used to extract features and perform temporal modeling for each field. After fusing the embedding vectors of each field, dimensional alignment and feature compression are performed to obtain a vectorized representation. This vectorized representation contains information substructures used to characterize physical semantics, processing stage location, and dynamic change trends.

[0165] The data source identifier field generates word embedding vectors using equipment numbers and geolocation codes, mapping the physical location of equipment to the relationship between process units. The data type field constructs semantic vectors based on physical attribute type labels, capturing the semantic similarity between parameters. The unit specification field uses a unit transformation matrix to unify values ​​to a standard unit system. The time information field extracts minute, hour, and day-level temporal features through a multi-scale temporal convolutional network, combined with the temporal deviation between the location embedding marker collection time and the synchronization timestamp. After independent encoding, the embedding vectors of each field are dimensionally compressed through a fully connected layer and weighted and fused using an attention mechanism to generate a unified vector containing multi-dimensional information.

[0166] Specifically, the encoding process for the data source identifier field inputs the device number and geolocation code into a pre-trained word embedding layer to generate a high-dimensional vector representing the spatial distribution of the device in the process chain. The data type field uses a semantic embedding model to convert parameter labels into clustered representations in the vector space, distinguishing the semantic boundaries of different physical attributes. The unit standardization field uses a pre-defined transformation coefficient matrix to map the original numerical values ​​and unit combinations to dimensionless standardized values. The time information field uses parallel time convolutional layers with different kernel sizes to extract short-term fluctuations and long-term trend features, which are then superimposed with the timestamp location code. After normalization, the vectors of each subspace are weighted using cross-channel attention to dynamically adjust the fusion ratio, ultimately outputting a composite vector that retains physical semantic relationships, process location topology, and temporal evolution patterns, providing a multi-dimensional feature foundation for knowledge graph construction.

[0167] As a preferred embodiment, the specific implementation of this application's solution is as follows: During implementation at a wastewater treatment plant, a full-process information knowledge graph is constructed based on the vectorized representation after pretreatment. Specifically, the biological reactor process unit entity is mapped to a process node, the dissolved oxygen sensor device entity is mapped to a device node, and the dissolved oxygen concentration monitoring parameter entity is mapped to a parameter node. The operating period from 10:00 AM to 11:00 AM is mapped to a time segment node. Based on the process flow logic between the biological reactor and the secondary sedimentation tank, directed edge relationships are established between process unit nodes, with the directed edge attribute set to "flowing to the next treatment stage". A "deployed at" relationship edge is established between the dissolved oxygen sensor node and the biological reactor process node, with the equipment installation time appended as May 2023. When abnormal fluctuations occur in the dissolved oxygen parameter node, a "occurred at" relationship edge is established between the parameter node and the time segment node, with the influent load conditions for that day noted.

[0168] Through the above technical solution, this application effectively solves the problem of constructing a unified knowledge representation system from multi-source heterogeneous data in wastewater treatment. By transforming process entities and their relationships into a structured graph model, the dynamic changes of monitoring parameters and the spatial coupling expression of the process flow are realized. The timestamp attributes and operating condition attributes embedded in the knowledge graph enable operational anomalies to be accurately traced to specific treatment stages and time intervals, providing a reasonable causal path for the intelligent early warning system and improving the diagnostic efficiency for problems such as dissolved oxygen anomalies.

[0169] In some of the solutions mentioned above in this application, although the constructed information knowledge graph can represent the static correlation of the entire sewage treatment process, it lacks the ability to monitor the dynamic status and trace the abnormal propagation path. It cannot identify abnormal trend of parameters and locate the root cause node in real time, resulting in delayed early warning and difficulty in guiding precise control.

[0170] This application further proposes that after constructing an information knowledge graph covering the entire wastewater treatment process, the state of each node in the graph and its associated edges is dynamically updated to identify the drift trend of parameter nodes. The state of nodes in the current graph is scored in real time based on historical operational data. When the score exceeds a threshold, an early warning marker is triggered. The propagation path of abnormal impacts is determined through graph path backtracking, and an early warning is issued based on the propagation path of abnormal impacts.

[0171] The dynamic status update includes updating the operating parameter values ​​of process unit nodes, the real-time detection values ​​of sensor nodes, and the status markers of time segment nodes at a frequency of minutes, while simultaneously updating the timestamp attributes and operating condition attributes on associated edges. Parameter drift trend identification uses a sliding window mechanism to calculate the mean offset and standard deviation rate of change of node parameters. A drift state is determined when the offset of three consecutive windows exceeds 15% of the historical baseline value. Real-time scoring calculates a comprehensive score by combining the cosine similarity between the current parameter value and historical data from the same period, along with a weighted average of the state change rates of upstream and downstream nodes in the process unit. Warning thresholds are divided into three levels based on process sensitivity, corresponding to yellow, orange, and red alert levels, respectively. Graph path backtracking uses a backpropagation algorithm, traversing the influence paths of abnormal nodes in the reverse direction of directed edges, and prioritizing the path with the highest weight value as the main propagation link.

[0172] Specifically, the dynamic status update module continuously receives real-time detection data from the data acquisition layer, updating the current value of the dissolved oxygen concentration node in the secondary sedimentation tank to 4.2 mg / L, and simultaneously modifying the timestamp attribute of this node and the dissolved oxygen control command edge of the bioreactor. The parameter drift identification module detects that the mean value of the residual chlorine concentration node in the disinfection tank increases from 0.3 mg / L to 0.52 mg / L within three consecutive 5-minute windows, with the standard deviation expanding to 0.12, triggering a drift marker. The real-time scoring module calculates the similarity score between the current value of the sludge return ratio node and historical data, which is 0.63. Combined with the change rate of the upstream MLSS concentration node, a comprehensive score of 82 points is generated. When the score exceeds the preset threshold of 80 points, the early warning module activates an orange alarm and, through graph path backtracking, finds the abnormal propagation path as: sudden change in influent flow rate → fluctuation in bar screen level → abnormal DO in the bioreactor → decreased sludge settling efficiency in the secondary sedimentation tank. Based on this, maintenance personnel prioritize checking the influent flow meter and bar screen control valve to achieve rapid location and handling of the root cause of the abnormality.

[0173] As a preferred embodiment, the specific implementation of this application's solution is as follows: After constructing a knowledge graph of the entire wastewater treatment process, the operation procedure for executing the real-time early warning function includes: when the operating current value of the sludge dewatering machine deviates from the normal fluctuation range within three consecutive synchronization cycles, the current monitoring node in the knowledge graph will trigger a status update. The system automatically extracts the current-torque correlation pattern corresponding to this node in historical operating data and calculates the similarity score between the current value and the historical pattern library. If the similarity score is lower than a preset threshold, the early warning module will activate the path backtracking algorithm, tracing back along the upstream and downstream relationships of the process units in the knowledge graph to the dissolved oxygen anomaly node in the bioreactor, and identifying the abnormal correlation edge attribute between the sludge concentration sensor and the dosing pump control command. Based on the anomaly propagation path report generated by the system, maintenance personnel can quickly locate the mechanical fault of the aeration regulating valve in the bioreactor.

[0174] Through the above technical solutions, this application achieves accurate source tracing and proactive early warning of abnormal operating conditions in the wastewater treatment process, effectively solving the problem that traditional monitoring systems cannot identify anomalies related to multiple parameters. The dynamic node update mechanism of the knowledge graph can capture early trend changes in parameter drift. Backtracking analysis based on graph structure paths can automatically reconstruct the propagation path of anomalies in the process chain, avoiding the inefficiency of manual investigation. The real-time scoring mechanism combined with a historical data pattern library improves the accuracy of early warning judgments and prevents false alarms and missed alarms.

[0175] In the above embodiments, distributed sensors are deployed at each processing stage, and a unified access protocol and time synchronization mechanism are adopted to achieve spatiotemporal consistent acquisition of raw data. Semantic matching is combined to complete label normalization and standardization. Then, metadata features are extracted through structured encoding, and a cross-layer data representation model based on transfer learning is constructed to achieve a unified vectorized expression of temporal, semantic and structural features between different data sources. Based on this vector representation, an information knowledge graph for the entire process is constructed, which improves the data's integrability, interpretability and usability, breaks the traditional data silo pattern, and provides unified data support for core smart water management functions such as intelligent analysis, anomaly early warning and process optimization.

[0176] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods 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.

[0177] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0178] 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.

[0179] 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.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.< / hasunit> < / sensor001>

Claims

1. A multi-source information fusion heterogeneous data regularization method, characterized in that, The application relates to a sewage treatment process data semantic analysis method and system. The method comprises the following steps: a plurality of distributed sensors are arranged in a plurality of treatment stages of a sewage treatment process, and detection data of all the distributed sensors are collected synchronously through an access protocol and time sequence; the detection data are subjected to semantic analysis based on semantic matching, the detection data are subjected to label normalization and semantic label standardization; the detection data after semantic standardization are processed to obtain metadata containing data source identification, data type, unit specification and time information; a data cross-layer representation model based on transfer learning is constructed, and feature extraction and semantic embedding of the metadata are carried out according to the data cross-layer representation model to obtain unified vectorization expression; an information knowledge graph for the whole process of sewage treatment is constructed according to the vectorization expression; when the detection data of all the distributed sensors are collected synchronously through an access protocol and time sequence, the method comprises the following steps: OPC-UA, MQTT or HTTP protocols are used to access the same data collection network for various sensors in a heterogeneous communication protocol, and protocol adaptation middleware is used for unified data packaging and format standardization; a unified clock synchronization is performed on the distributed sensors in combination with a network time protocol or an IEEE 1588 precision time protocol; when the data cross-layer representation model based on transfer learning is constructed, the method comprises the following steps: a double-branch hybrid model architecture composed of a pre-trained semantic encoder and a graph neural network is used, the pre-trained semantic encoder is used to process the structured semantic labels contained in the metadata to generate an initial semantic embedding vector; the graph neural network takes a plurality of sewage treatment process units and the upstream and downstream relationships therebetween as nodes and edges to establish a process-level structure graph, and performs feature propagation and aggregation on the graph based on the input metadata to obtain a structured embedding vector; when the data cross-layer representation model based on transfer learning is constructed, the method further comprises the following steps: initial model parameters are obtained by source domain training in a sewage treatment project in which semantic annotation and process structure mapping have been built; based on a domain self-adaptive transfer strategy, the model is transferred to a target sewage treatment scene, and is adjusted through label samples; 2. The method of claim 1, wherein, the domain self-adaptive transfer strategy process comprises feature distribution alignment, maximum mean difference regularization and semantic label consistency constraint. when the distributed sensors are arranged in a plurality of treatment stages of a sewage treatment process, the method comprises the following steps:

3. The method of claim 1, wherein, sensor nodes with temperature, dissolved oxygen, pH, conductivity, turbidity, flow, liquid level, chemical oxygen demand and ammonia nitrogen detection functions are arranged in a grid pool, a sand settling pool, a biological reaction pool, a secondary sedimentation tank, a disinfection tank and a sludge treatment stage. when the detection data are subjected to semantic analysis based on semantic matching, the detection data are subjected to label normalization and semantic label standardization, the method comprises the following steps: original description information of the detection data is subjected to semantic feature extraction through natural language processing, the semantic feature extraction comprises keyword identification, part-of-speech analysis, entity extraction and semantic nested analysis; the detection data are mapped to corresponding standard label items based on vector semantic similarity calculation, semantic distance measurement and upper and lower semantic relationship matching, and are subjected to homonymy, polysemy and abbreviation data label normalization processing; The normalized label is associated with a corresponding physical meaning, a unit system, and a processing unit to which the label belongs, and a structured semantic label entry is output.

4. The method of claim 3, wherein, When the semantic standardized detection data is processed to obtain metadata including data source identification, data type, unit specification, and time information, the processing includes: The semantic standardized detection data is field-decomposed according to a preset structure template, and at least the following is extracted and collected: The data source identification field is used to record the sensor device number, the processing unit to which the sensor device belongs, and the geographic location code of the processing unit; The data type field is used to indicate the physical property type of the detection value; The unit specification field is used to unify the physical unit corresponding to the numerical value; The time information field is used to record the original collection time, the system receiving time, and the synchronization time stamp; All fields are packaged as structured metadata entries, represented in JSON, XML, or RDF semantic format, and the original data mapping index of the metadata is preserved.

5. The method of claim 1, wherein, When the metadata is feature-extracted and semantically embedded according to the data cross-layer representation model to obtain a unified vectorized expression, the processing includes: The data source identification field, the data type field, the unit specification field, and the time information field in the metadata are independently subspace-encoded, and word embedding, position embedding, and multi-scale time convolution network are used to extract features and model time series of each field; After fusing the embedding vectors of each field, dimension alignment and feature compression are performed to obtain the vectorized expression, which contains information substructures for representing physical semantics, processing stage positions, and dynamic change trends.

6. The method of claim 5, wherein, When the vectorized expression is used to construct an information knowledge graph for the entire sewage treatment process, the processing includes: The entity elements in the vectorized expression are mapped to nodes in the knowledge graph, including process units, sensor devices, monitoring parameters, time segments, and control instructions; According to the sewage treatment process logic and process structure, the directed edge relationship between nodes is established, and a timestamp or a working condition is added as an attribute of the edge; The information knowledge graph for the entire sewage treatment process is obtained.

7. The method of claim 6, wherein, After the vectorized expression is used to construct the information knowledge graph for the entire sewage treatment process, the processing further includes: Warning according to the information knowledge graph for the entire sewage treatment process; Dynamic state updating of each node and its associated edge relationship in the graph, identifying the drift trend of the parameter node; Real-time scoring of the node state in the current graph according to historical operation data, triggering a warning mark when the scoring result exceeds a threshold, determining an abnormal influence propagation path through graph path backtracking, and warning according to the abnormal influence propagation path.

Images