An ai intelligent diagnosis and strategy linkage method for sewage treatment multi-modal perception

Abstract

The application belongs to the technical field of natural language processing, and discloses an AI intelligent diagnosis and strategy linkage method for multi-modal perception of sewage treatment, which realizes standardized text conversion of structured and unstructured data in the whole process through text generation, visual semantic description, audio semantic transcription and other technologies, combines with a domain knowledge graph to complete multi-modal semantic feature fusion and weight distribution, fully excavates complementary information of data, makes an AI model establish global semantic understanding of the whole process of sewage treatment, avoids information one-sidedness of single data modeling, and improves comprehensive utilization efficiency of multi-modal perception data; a hierarchical reasoning diagnosis model combined with the domain knowledge graph is built, abnormal working condition recognition and four-level classification are completed based on a RoBERTa semantic classification model, a reverse chain semantic reasoning algorithm is used to trace core root causes from abnormal phenomena and record reasoning paths, and an interpretable report containing complete diagnosis logic is generated.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically an AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment. Background Technology

[0002] The stable and compliant operation of wastewater treatment processes relies on real-time perception and precise decision-making based on multi-dimensional information. Currently, the industry has gradually adopted multi-modal perception methods such as water quality sensors, visual inspection, voiceprint detection, and SCADA data acquisition. It is also exploring the introduction of AI technology to achieve fault early warning and intelligent control, which has improved the automation level of wastewater treatment plant operation and maintenance to some extent. Meanwhile, natural language processing technology, with semantic understanding, knowledge graphs, and pre-trained language models at its core, has achieved mature applications in fields such as industrial operation and maintenance and medical diagnosis, including unstructured knowledge accumulation, interpretable reasoning, and intelligent decision generation. However, existing technologies still have the following technical challenges: Existing multimodal sensing technologies for wastewater treatment often model structured numerical data, image and audio data, and operation and maintenance text data in a fragmented and independent manner, without using natural language processing technology to achieve unified semantic representation of heterogeneous data. This results in the inability to fully extract complementary information from multiple sources of data, such as sensor data, industry standards, fault cases, and expert experience, and the inability of AI models to establish a global semantic understanding of the entire wastewater treatment process.

[0003] Existing AI diagnostic models are mostly based on numerical deep learning algorithms, which can only classify and identify faults. They cannot combine the process mechanisms, standards and specifications, and historical cases in the field of wastewater treatment to achieve semantic-level reasoning through semantic reasoning technology. The diagnostic results cannot output interpretable root cause analysis, are disconnected from the experience judgment of on-site operation and maintenance personnel, and have poor generalization ability for rare faults.

[0004] Existing technologies mostly adopt a passive operation and maintenance mode of threshold alarms, manual decision-making, and manual execution. Some technologies can only trigger preset fixed strategies and cannot output personalized and implementable handling strategies based on the semantic features of diagnostic results, combined with real-time operating conditions, equipment capability boundaries, and industry standard requirements, through natural language generation technology. Summary of the Invention

[0005] The purpose of this invention is to provide an AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment, in order to solve one or more of the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment, comprising the following specific steps: Preferably, the text representation stage first collects multimodal heterogeneous data of the entire wastewater treatment process, including structured water quality sensor values, equipment operating parameters, SCADA data, as well as unstructured inspection audio and video, equipment voiceprints, maintenance logs, fault reports, industry standards, process documents, expert experience and other full-dimensional data. A domain-customized text generation model is adopted to convert structured numerical data into standard descriptive text with working condition semantic tags according to the domain semantic template; visual features in image and video data are converted into structured visual descriptive text containing abnormal parts, abnormality degree and occurrence time by combining a large visual model with image semantic description generation technology; and voiceprint recognition is combined with audio semantic transcription technology to convert equipment operation voiceprint data into text descriptions containing operating status, abnormal features and suspected fault types. For raw text data, standardization processes including word segmentation, stop word removal, named entity recognition, and semantic normalization are completed through text preprocessing techniques to generate a multimodal semantic text dataset with a unified format and unified semantic dimension. At the same time, an index is built for the text dataset through semantic hashing technology.

[0007] Preferably, the graph construction stage is based on a multimodal semantic text dataset with a unified format and unified semantic dimension generated in the text representation stage. Through wastewater treatment industry standards, process specifications, failure cases, equipment descriptions and expert experience texts, and through Prompt engineering-driven named entity recognition, semantic dependency analysis-based relation extraction and attribute extraction technology, seven core semantic entities are extracted in a targeted manner: water quality indicators, process units, equipment, failure phenomena, failure root causes, treatment strategies and normative constraints. Furthermore, by using hierarchical semantic relationship definition technology, we clarify the multi-level semantic relationships, including phenomena and root causes, faults and strategies, equipment and indicators, processes and constraints. Combined with scenario-based event information from the entire wastewater treatment process, we construct a four-dimensional vertical domain semantic knowledge graph of entities, relationships, events, and constraints. Finally, using context-based semantic matching disambiguation technology, multimodal semantic text data entities are semantically aligned with standard entities in the knowledge graph, eliminating ambiguity between entities with the same name but different meanings, while also building a dynamic update interface for the knowledge graph.

[0008] Preferably, the semantic fusion stage is based on the four-dimensional vertical domain semantic knowledge graph built in the graph construction stage and the multimodal semantic text data with completed semantic alignment. Based on the general pre-trained language model, domain incremental pre-training and initial low-rank adaptation fine-tuning are carried out. Simultaneously, a special semantic dictionary for the wastewater treatment domain is matched to optimize the domain semantic understanding and adaptation capabilities of the model, so as to obtain a domain language model specifically for the wastewater treatment scenario. Then, the real-time multimodal semantic text data is input into the domain language model, combined with the knowledge graph entity embedding vector, and the fusion and extraction of multimodal semantic features are completed through the semantic attention mechanism to obtain a high-dimensional semantic fusion feature vector containing full-dimensional information including working conditions, equipment, water quality anomalies, and regulatory constraints. Meanwhile, through a semantic weight dynamic allocation algorithm, weights are assigned to various semantic features to adapt to different operating conditions, thereby strengthening key semantic features related to fault diagnosis and process anomalies and actively suppressing redundant noise features.

[0009] Preferably, the reasoning and diagnosis stage is based on the high-dimensional semantic fusion feature vector extracted in the semantic fusion stage and the generated domain language model specifically for the sewage treatment scenario. Combined with the semantic reasoning rules of the knowledge graph, a hierarchical semantic reasoning and diagnosis model is built. Based on the RoBERTa semantic classification model fine-tuned and optimized by the domain language model, the abnormal operating condition types, including water quality exceeding standards, equipment failure, process abnormality, and environmental disturbance, are identified, and a four-level semantic classification of severity, namely, early warning, general abnormality, serious abnormality, and emergency failure, is completed. Based on the semantic association rules between phenomena and root causes in the knowledge graph, a reverse chain semantic reasoning algorithm is used to trace the core root cause entity from the abnormal phenomenon entity, locate the root cause of the fault and eliminate indirect interference factors, and record the semantic association reasoning path throughout the process. By using natural language generation technology, an interpretable diagnostic text report is generated that includes the location, type, severity, root cause analysis, and reasoning path of the anomaly, while also outputting standardized diagnostic semantic tags.

[0010] Preferably, the strategy generation stage is based on the standardized diagnostic semantic tags and interpretable diagnostic text reports output by the reasoning and diagnosis stage, and pre-constructs a disposal strategy semantic library bound to the domain knowledge graph. Each strategy in the disposal strategy semantic library is semantically annotated, clearly annotating the core semantic tags including applicable scenarios, execution steps, constraints, and expected effects, covering the disposal needs of faults and abnormal working conditions in the entire sewage treatment scenario, and simultaneously building a strategy semantic index library. A semantic similarity calculation algorithm based on BERT cosine similarity is used to select the candidate strategy set with the highest matching degree with the diagnostic results and real-time operating conditions from the strategy library. Combining multi-dimensional constraints such as industry standards, equipment operating boundaries, water quality compliance, and energy consumption optimization, the candidate strategies are verified for compliance through semantic rule verification technology, and non-compliant and unfeasible solutions are automatically eliminated. By using Prompt-driven natural language generation technology, the optimal candidate strategy is optimized for operational conditions, generating standardized natural language processing strategy texts that fit the current real-time operational conditions and can be implemented step by step, clearly defining the processing objectives, step-by-step operations, precautions, and expected results.

[0011] Preferably, the parsing and execution stage is based on the standardized natural language processing strategy text generated in the strategy generation stage. Through dependency parsing and semantic role labeling technology, a standardized semantic parsing mapping rule for natural language strategy text and equipment control instructions is established, and a dedicated semantic and instruction mapping dictionary is built. Then, through a professional semantic parsing model, the standardized natural language processing strategy text is decomposed into a set of standardized control instructions corresponding to each process unit and equipment, clarifying the execution equipment, timing logic, operating parameters and threshold range of each instruction, and simultaneously completing instruction semantic conflict verification. The dedicated semantic and instruction mapping dictionary is constructed according to four dimensions: process unit, equipment type, operation action, and control parameter, and corresponds one-to-one with the instruction library of the actual equipment control system of the wastewater treatment plant. Each semantic entry in the dictionary corresponds to a unique equipment control instruction code. The entry contains five core pieces of information: standard semantic description, synonymous semantic expression, instruction code, executing equipment, and parameter range. For dedicated operations of core process units such as aeration tanks and sedimentation tanks, separate process-specific semantic entries are added. The dictionary adopts a key-value pair storage format and supports multi-dimensional retrieval by semantic description, equipment type, and instruction code. At the same time, a dictionary maintenance terminal is built, which can add or modify entries according to equipment upgrades and process adjustments. The updated dictionary is synchronized to the semantic parsing model to ensure accurate mapping between semantics and instructions.

[0012] Based on the timing logic and linkage constraints of the instruction set, instructions are issued to the equipment control system; during the execution process, real-time equipment operation feedback data and water quality parameter change data are collected and converted into semantic feedback text of the execution process according to unified textualization rules; After execution, an execution feedback report containing the execution process, execution results, and performance evaluation is automatically generated using text summarization technology.

[0013] Preferably, the incremental learning phase collects new fault handling cases, supplementary texts of operation and maintenance expert experience, and new version of industry standard texts based on the execution feedback text report generated in the parsing and execution phase to build an incremental semantic learning dataset; then, through incremental entity extraction and relation extraction technology, it extracts new semantic entities, new semantic relationships and optimization strategy rules from the incremental data, and dynamically updates the knowledge graph in the field of wastewater treatment. Meanwhile, based on the incremental semantic learning dataset, the domain pre-trained language model is lightly fine-tuned using low-rank adaptation incremental fine-tuning technology to optimize the model's semantic feature extraction, diagnostic reasoning, and policy generation capabilities. Based on feedback from actual implementation results, semantic reasoning rules, semantic similarity calculation algorithms, and constraint verification rules are simultaneously optimized through semantic rule optimization algorithms.

[0014] The beneficial effects of this invention are as follows: 1. This invention achieves standardized text conversion of structured and unstructured data throughout the entire process through technologies such as text generation, visual semantic description, and audio semantic transcription. It combines domain knowledge graphs to complete multimodal semantic feature fusion and weight allocation, fully explores complementary information in the data, and enables the AI ​​model to establish a global semantic understanding of the entire wastewater treatment process. This avoids the information bias of single data modeling and improves the comprehensive utilization efficiency of multimodal perception data.

[0015] 2. This invention constructs a hierarchical reasoning diagnostic model that combines domain knowledge graphs. Based on the RoBERTa semantic classification model, it completes the identification of abnormal operating conditions and four-level classification. Through the reverse chain semantic reasoning algorithm, it traces the core root cause from the abnormal phenomenon and records the reasoning path, generating an interpretable report containing complete diagnostic logic. It combines the sewage treatment process mechanism, industry standards and historical cases, so that the diagnostic results are connected with the experience judgment of on-site operation and maintenance personnel, thereby improving the model's ability to generalize the diagnosis of rare faults.

[0016] 3. Based on the diagnostic results, this invention matches candidate strategies using a semantic similarity algorithm, eliminates non-compliant solutions through multi-dimensional constraint verification, and optimizes and generates personalized implementation strategies. It can also parse natural language strategies into standardized equipment control commands to achieve coordinated execution. At the same time, through incremental learning, it dynamically updates the knowledge graph, fine-tunes the domain model, and optimizes various rules, realizing dynamic iterative optimization of diagnosis, strategy, and execution, thereby improving the automation and intelligence level of sewage treatment operation and maintenance. Attached Figure Description

[0017] Fig. 1 This is a full flowchart of the AI ​​intelligent diagnosis and strategy linkage for multimodal perception in wastewater treatment according to the present invention. Fig. 2 This is a flowchart of the reasoning and diagnosis stage of the present invention; Fig. 3 This is a flowchart of the execution phase of the present invention. Detailed Implementation

[0018] 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.

[0019] like Figs. 1 to 3 As shown in the figure, this embodiment of the invention provides an AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment, including the following specific steps: The text representation stage first collects multimodal heterogeneous data from the entire wastewater treatment process, including structured water quality sensor values, equipment operating parameters, SCADA data, as well as unstructured inspection audio and video, equipment voiceprints, maintenance logs, fault reports, industry standards, process documents, expert experience, and other full-dimensional data. Structured data is collected in real time at a frequency of 5 minutes per collection, covering the entire process of wastewater treatment, including influent, biological treatment, sedimentation, and effluent. Missing data is supplemented based on the average value of the same period. Data exceeding the range is marked as abnormal and triggers a second collection. If the second collection still exceeds the range or is missing, it is marked as invalid data and pushed to the operation and maintenance terminal for manual review. Unstructured text data is stored uniformly in electronic format. Inspection audio and video and equipment voiceprint data are collected at a duration of 1 minute per segment, stored in MP4 and WAV formats. The collection points cover key equipment in core process units such as bar screens, pumps, aeration tanks, and sedimentation tanks.

[0020] A domain-customized text generation model (based on a fine-tuned T5 mini-model) is adopted to convert structured numerical data into standard descriptive text with working condition semantic labels according to the domain semantic template, thereby clarifying the semantic connotations corresponding to numerical anomalies. The T5 small model fine-tuning uses the AdamW optimizer with a learning rate of 5e-5, a batch size of 16, and 10 training epochs. The training set consists of structured numerical-semantic description paired data for wastewater treatment, divided into training, validation, and test sets in an 8:1:1 ratio. The model input is a structured water quality / equipment numerical matrix with dimensions N×128, where N is the number of data points collected in a single batch and 128 represents the dimensions of the core monitoring indicators for wastewater treatment. The output is a domain-standard semantic description text with a length not exceeding 256 characters, labeled with a set of semantic labels for operating conditions. Training stops when the cross-entropy loss on the validation set converges to 0.08.

[0021] By combining visual big data models with image semantic description generation technology, visual features such as sludge status and equipment abnormalities in image and video data are transformed into structured and visualized descriptive text containing abnormal locations, abnormality levels, and occurrence times. By using voiceprint recognition combined with audio semantic transcription technology, the voiceprint data of equipment operation is transformed into text descriptions containing operating status, abnormal features, and suspected fault types. The large-scale visual model is adapted and trained for wastewater treatment scenarios. It focuses on four core visual features for identifying targets: sludge expansion, surface scum, equipment leakage, and instrument malfunction. The image acquisition resolution is no less than 1080P, the frame rate is 25 frames per second, the accuracy of abnormal feature recognition is no less than 95%, the accuracy of abnormal recognition of voiceprint features is no less than 93%, and the accuracy of fault type determination is no less than 90%. The voiceprint recognition technology has a sampling rate of 44.1kHz and a sampling bit depth of 16bit. It identifies four types of voiceprint features for core equipment such as pumps, fans, and mixers: normal operation, abnormal noise, vibration, and overload. The equipment malfunction type is determined by the spectral changes of the voiceprint features.

[0022] For raw text data, standardization processes such as word segmentation, stop word removal, named entity recognition, and semantic normalization are completed through text preprocessing techniques to generate a multimodal semantic text dataset with a unified format and unified semantic dimension. At the same time, an index is built for the text dataset through semantic hashing technology.

[0023] Semantic normalization strictly follows the standard terminology system of knowledge graphs in the wastewater treatment field. Synonymous professional terms with different names are uniformly replaced with standard terms. Vague descriptions of fault phenomena and water quality status are transformed into standardized quantitative descriptions. Process unit names and equipment models with different expressions are standardized. For numerical descriptions, units and precision are unified according to a fixed format. Water quality index values ​​are retained to two decimal places, and equipment operating parameters are unified in units according to the equipment's rated standards. The normalized text data must be consistent with the semantics of entities in the knowledge graph. Inconsistent content is marked and manually calibrated.

[0024] Semantic hashing technology performs hash encoding on the features of multimodal semantic text in wastewater treatment. Using the core semantic tags, process units, and equipment names of the text as hash keys, a locality-sensitive hashing algorithm is employed to map high-dimensional semantic features into fixed-length 64-bit hash values. Texts with the same process unit or similar fault type are assigned similar hash values. A hash index table is constructed, containing three main fields: hash value, text storage address, and core semantic tags. This supports multi-dimensional retrieval by hash value, semantic tags, and process unit, with a retrieval response time controlled within 1 second, enabling fast and accurate querying of multimodal semantic text datasets.

[0025] The graph construction stage is based on a multimodal semantic text dataset with a unified format and unified semantic dimension generated in the text representation stage. Through wastewater treatment industry standards, process specifications, failure cases, equipment descriptions and expert experience texts, and through Prompt engineering-driven named entity recognition, semantic dependency analysis-based relation extraction and attribute extraction technology, seven core semantic entities are extracted in a targeted manner: water quality indicators, process units, equipment, failure phenomena, failure root causes, treatment strategies and normative constraints. A dedicated Prompt template for named entity recognition in the wastewater treatment field is designed. The template includes three parts: a domain background description, entity type hints, and example annotations. The background description is text based on wastewater treatment processes, fault diagnosis, and operation and maintenance specifications, from which relevant entities are extracted. The entity type hints clearly define seven core entity categories. The example annotations select typical wastewater treatment text to demonstrate entity annotation. The Prompt input is combined in the format of "text content + entity extraction requirements," and the output is the entities extracted from the text and their corresponding categories. For entities with ambiguous boundaries, the category is determined in conjunction with the wastewater treatment process mechanism. The entity extraction accuracy is no less than 92%. After automatic deduplication, the extraction results proceed to the relation extraction stage.

[0026] Furthermore, by using hierarchical semantic relationship definition technology, we clarify the multi-level semantic relationships such as phenomena and root causes, faults and strategies, equipment and indicators, processes and constraints. Combined with scenario-based event information such as operating conditions and fault handling events throughout the entire wastewater treatment process, we construct a four-dimensional vertical domain semantic knowledge graph of entities, relationships, events, and constraints. The entity dimension of the knowledge graph includes seven core entities and their attribute information. Water quality indicator entities are labeled with attributes such as range, compliance threshold, and associated processes. Equipment entities are labeled with attributes such as model, operating parameters, and associated process units. The relationship dimension represents the hierarchical semantic relationships between entities, categorized as strong or weak associations. The event dimension records information such as the time, location, phenomenon, and handling results of all process operation events and fault handling events. The constraint dimension includes constraints such as industry standards, process requirements, and equipment operating boundaries. The knowledge graph is stored using the Neo4j graph database, with nodes representing entities and edges representing semantic relationships. It supports dynamic addition, deletion, modification, and querying of nodes and edges. The knowledge graph's dynamic update interface achieves real-time connection with the graph database through Neo4j's REST API.

[0027] Finally, through context-based semantic matching disambiguation technology, multimodal semantic text data entities are semantically aligned with standard entities in the knowledge graph, eliminating ambiguity between entities with the same name but different meanings, completing the semantic mapping between real-time perception data and domain knowledge, and building a dynamic update interface for the knowledge graph.

[0028] Entity semantic alignment follows a process of "literal terminology matching → contextual semantic similarity matching → manual review." First, text entities are precisely matched against standard entities in the knowledge graph; successful matches result in direct alignment. For those failing literal matching, the contextual semantic similarity between the text entity and candidate standard entities is calculated; similarity ≥ 0.85 completes alignment. Entities with similarity below 0.85 undergo manual review, where domain experts determine if they are synonymous and complete the alignment. Entity ambiguity resolution primarily targets polysemous entities. Considering the context of the wastewater treatment scenario, semantic meanings related to processes, faults, and water quality are filtered out, while irrelevant meanings are eliminated to ensure the uniqueness and accuracy of entity semantics.

[0029] The knowledge graph dynamic update interface is a RESTful interface with two levels of access permissions: administrator and operations personnel. Administrators have full CRUD (Create, Read, Update, Delete) permissions, while operations personnel only have the permission to submit new entities / relationships. The conditions for triggering the interface update are divided into automatic and manual triggering. Automatic triggering is triggered synchronously after the dataset is cleaned during the incremental learning phase, while manual triggering is initiated by domain experts based on new industry standards and failure cases. The data exchange format of the interface is JSON, which must include core information such as the name of the added or modified entity, attributes, semantic relationships, data source, and submitter. After submission, the system automatically performs format validation. If the validation passes, the data is temporarily stored and officially updated to the knowledge graph after manual review.

[0030] The semantic fusion stage is based on the four-dimensional vertical domain semantic knowledge graph built in the graph construction stage and the multimodal semantic text data with completed semantic alignment. It uses a general pre-trained language model (such as BERT-base) as a basis to perform domain incremental pre-training and initial low-rank adaptation fine-tuning. It is also combined with a special semantic dictionary for the wastewater treatment domain to optimize the domain semantic understanding and adaptation capabilities of the model, resulting in a domain language model specifically for wastewater treatment scenarios. The incremental pre-training of the BERT-base model uses 100,000 entries of wastewater treatment corpus, including operation and maintenance logs, fault reports, industry standards, and process documents. Low-rank adaptation fine-tuning uses a LoRA adapter with a rank of 8, paired with the AdamW optimizer, a learning rate of 2e-5, a batch size of 32, and 8 training epochs. The model input is word segmentation vectors of multimodal semantic text, and the output is domain semantic feature vectors. The knowledge graph entity embedding vectors are unified to 1×768 dimensions, then added element-wise to the model output vector and layer normalized to complete the fusion of multimodal semantic features. The LoRA adapter is deployed between the attention layer and the fully connected layer of the BERT-base model, using a parallel access method without changing the network structure and parameters of the original model. The adapter's weight matrix consists of two low-rank matrices, A and B. Matrix A is a randomly initialized reduced-dimensional matrix, and matrix B is a randomly initialized increased-dimensional matrix. During training, only the parameters of the two low-rank matrices are updated. After training, the LoRA adapter is fused with the original model and deployed. During the inference phase, the fused model is directly called to ensure the model's inference efficiency and semantic feature extraction capabilities. The adapter's parameters are stored in a separate file, supporting on-demand loading and updating.

[0031] The dedicated semantic dictionary for wastewater treatment is constructed according to seven categories: water quality indicators, process units, equipment, fault phenomena, root causes of faults, treatment strategies, and regulatory constraints, corresponding one-to-one with the seven core entities of the knowledge graph. Each entry contains five core pieces of information: standard name, synonyms, definition, related attributes, and application scenarios. For entries related to equipment and process units, additional related water quality indicators and common fault types are provided. An online maintenance platform is built for the semantic dictionary, allowing domain experts to add, modify, and delete entries based on industry technology updates, process upgrades, and the release of new regulations. The updated semantic dictionary is synchronized to the fine-tuning training of the domain language model to ensure that the model's semantic understanding is consistent with the latest terminology in the domain.

[0032] Then, the real-time multimodal semantic text data is input into the domain language model, combined with the knowledge graph entity embedding vector, and the fusion and extraction of multimodal semantic features are completed through the semantic attention mechanism to obtain a high-dimensional semantic fusion feature vector containing full-dimensional information such as working conditions, equipment, water quality anomalies, and regulatory constraints. Meanwhile, by using a semantic weight dynamic allocation algorithm, weights are assigned to various semantic features to suit different operating conditions, thereby strengthening key semantic features related to fault diagnosis and process anomalies, actively suppressing redundant noise features, and improving the accuracy and stability of subsequent diagnostic reasoning.

[0033] The semantic weight dynamic allocation algorithm is based on the core requirements of wastewater treatment fault diagnosis. It identifies three types of semantic features as key features: abnormal water quality indicators, deviations in equipment operating parameters, and abnormal operating conditions of process units. The initial weight of each of these three types is set to 0.25. Environmentally irrelevant data and descriptions of normal equipment operation are identified as redundant noise features and have an initial weight of 0.05. The remaining 0.20 weight is allocated to semantic features related to environmental disturbances, with an initial total weight of 1. When dynamically adjusting according to operating conditions, the sum of all feature weights remains unchanged at 1. The algorithm dynamically adjusts the weights based on the type of anomaly in real-time operating conditions. If an anomaly of exceeding water quality standards is identified, the weight of semantic features related to water quality indicators is increased by 0.1-0.2. If an anomaly of equipment failure is identified, the weight of semantic features related to equipment operation is increased by 0.1-0.2. After the weight adjustment, the sum of all feature weights is 1, and the adjustment results are synchronized to the feature input layer in real time.

[0034] The reasoning and diagnosis stage is based on the high-dimensional semantic fusion feature vector extracted in the semantic fusion stage and the domain language model generated specifically for the sewage treatment scenario. Combined with the semantic reasoning rules of the knowledge graph, a hierarchical semantic reasoning and diagnosis model is built. Based on the RoBERTa semantic classification model fine-tuned and optimized by the domain language model, the abnormal operating condition types such as water quality exceeding the standard, equipment failure, process abnormality, and environmental disturbance are identified, and a four-level semantic classification of severity is completed: early warning, general abnormality, serious abnormality, and emergency failure. The hierarchical reasoning diagnostic model consists of four layers: a feature input layer, a semantic classification layer, a root cause reasoning layer, and a report generation layer. These layers are connected sequentially according to the data transmission order, and data between layers is exchanged in tensor format. The feature input layer receives high-dimensional semantic fusion feature vectors and reduces them to 1×768 using PCA. The semantic classification layer is a fine-tuned RoBERTa model, responsible for classifying the working condition type and severity. The root cause reasoning layer calls the semantic reasoning rules of the knowledge graph to perform reverse chain reasoning. The report generation layer generates natural language diagnostic reports based on the reasoning results using a template-based generation method. The final interaction format between layers is a combination of 1×768 feature vectors and standardized semantic labels.

[0035] The RoBERTa model uses the base version, with the AdamW optimizer, learning rate 3e-5, batch size 16, and 12 training epochs during the fine-tuning phase. The training set is based on labeled data of abnormal operating conditions in wastewater treatment, and the training, validation, and test sets are divided in an 8:1:1 ratio. The model input is a fused high-dimensional semantic feature vector, and the output is a two-dimensional semantic classification result. The first dimension is the label of 4 types of abnormal operating conditions, and the second dimension is the label of 4 levels of severity. The classification confidence threshold is set to 0.85. If the score is lower than the threshold, it is judged as a suspected abnormality and marked, and a manual review process is triggered. The review results are added to the model training set.

[0036] Based on the semantic association rules between phenomena and root causes in the knowledge graph, a reverse chain semantic reasoning algorithm is used to trace the core root cause entity layer by layer from the abnormal phenomenon entity, locate the root cause of the fault and eliminate indirect interference factors, record the semantic association reasoning path throughout the process, and retain the complete reasoning logic. The specific reasoning steps of the reverse chain semantic reasoning algorithm in the field of wastewater treatment are as follows: 1. Using the identified abnormal entities as the starting point for reasoning, retrieve the set of root cause entities directly associated with them in the four-dimensional vertical domain knowledge graph; 2. Root cause entities with a weight ≥ 0.7 are selected based on the entity association weights in the knowledge graph to exclude low-weight indirect interference factors; the value range of entity association weights is 0-1, and the weights are assigned based on the association frequency of historical failure cases and the experience of industry experts. 3. For the screened root cause entities, the threshold range corresponding to the root cause entities is matched with the real-time collected water quality and equipment operating parameters to verify their rationality. If the verification is successful, they are determined to be the core root cause. 4. Record the reasoning path in the order of phenomenon, intermediate related entities, and core root cause. The path data is stored in the format of triples (subject-relation-object). Simultaneously generate a visual code of the reasoning path for easy viewing by operation and maintenance personnel.

[0037] The reasoning path visualization coding adopts a tree structure coding method, with the identified anomaly as the root node, the intermediate related entities as intermediate nodes, and the core root cause as the leaf node. Nodes are distinguished by different colors to distinguish entity types: water quality indicator entities are blue, equipment entities are red, fault phenomenon entities are yellow, and fault root cause entities are orange. The lines between nodes indicate the semantic relationship and weight between entities, and the thickness of the line is positively correlated with the relationship weight. The visualization coding output is in SVG vector format, which can be zoomed in, zoomed out, and dragged to view on the sewage treatment plant operation and maintenance terminal. At the same time, jump links are embedded in the coding. Clicking on a node can view the detailed attributes of the entity, historical related fault cases, and other information.

[0038] By using natural language generation technology, an interpretable diagnostic text report is generated that includes the location, type, severity, root cause analysis, and reasoning path of the anomaly, while also outputting standardized diagnostic semantic tags.

[0039] The strategy generation stage is based on the standardized diagnostic semantic tags and interpretable diagnostic text reports output by the reasoning and diagnosis stage. It pre-builds a disposal strategy semantic library bound to the domain knowledge graph. Each strategy in the disposal strategy semantic library is semantically annotated, clearly annotating the core semantic tags such as applicable scenarios, execution steps, constraints, and expected effects, covering the disposal needs of faults and abnormal working conditions in the whole scenario of sewage treatment. At the same time, a strategy semantic index library is built to improve the efficiency of strategy matching and retrieval. A semantic similarity calculation algorithm based on BERT cosine similarity is used to select the candidate strategy set with the highest matching degree with the diagnostic results and real-time operating conditions from the strategy library; The semantic similarity calculation algorithm uses the standardized semantic label vector of the diagnostic results (obtained by encoding the labels using a domain language model) and the semantic label vector of the strategies in the strategy library as the calculation objects. Before calculation, both types of vectors are subjected to L2 normalization. The cosine similarity calculation formula is cosθ=(A·B) / (|A|×|B|). The matching degree threshold is set to 0.8. The number of candidate strategies selected does not exceed 5, and they are sorted from high to low similarity. If the matching degree of all strategies is lower than the threshold, the industry expert strategy recommendation process is triggered. The expert recommended strategy is simultaneously entered into the strategy semantic library, and the working condition is recorded as a new working condition case.

[0040] Combining industry standards, equipment operating boundaries, water quality compliance, energy consumption optimization and other multi-dimensional constraints, the candidate strategies are verified for compliance through semantic rule verification technology, and non-compliant and unfeasible solutions are automatically eliminated. Multi-dimensional constraint verification is performed in the following order: "Water quality compliance first → Equipment operation safety → Industry standard compliance → Energy consumption optimization". Water quality compliance verification checks whether the strategy meets the national / local standards for effluent water quality. Equipment operation safety verification checks whether the control parameters of the strategy are within the rated operating limits of the equipment. Industry standard compliance verification checks whether the strategy complies with the operation and maintenance management specifications of the wastewater treatment plant. Energy consumption optimization verification checks whether the unit water treatment energy consumption of the strategy is within a reasonable range. If a single dimension verification fails, it is judged as overall non-compliance. The specific dimension of failure and the reason for failure will be marked and simultaneously fed back to the strategy screening stage for re-screening of candidate strategies.

[0041] By using Prompt-driven natural language generation technology, the optimal candidate strategy is optimized for operational conditions, generating standardized natural language processing strategy texts that fit the current real-time operational conditions and can be implemented step by step, clearly defining the processing objectives, step-by-step operations, precautions, and expected results.

[0042] The Prompt for wastewater treatment scenarios adopts a fixed structure design: basic instructions + real-time operating information + diagnostic results + strategy optimization requirements. The basic instructions are standardized, actionable treatment strategies generated based on the wastewater treatment conditions, including treatment objectives, step-by-step operations, precautions, and expected results. The real-time operating information includes abnormal process units, core water quality index values, and equipment operating status. The diagnostic results include the anomaly type, severity, and root cause. The strategy optimization requirements include constraints such as water quality compliance, equipment safety, and energy consumption control. The generation process combines templates and personalization. The text structure of the treatment strategy is generated according to a fixed template, while the core content such as step-by-step operations and control parameters are personalized according to the real-time operating conditions. The generated text must be concise, clear, and step-by-step, conforming to the reading habits of on-site operation and maintenance personnel. The text length of a single strategy is controlled within 500 characters.

[0043] The parsing and execution phase is based on the standardized natural language processing strategy text generated in the strategy generation phase. Through dependency parsing and semantic role labeling technology, it establishes standardized semantic parsing mapping rules between the natural language strategy text and equipment control instructions, and builds a dedicated semantic and instruction mapping dictionary. Then, through a professional semantic parsing model, the standardized natural language processing strategy text is decomposed into standardized control instruction sets corresponding to each process unit and equipment, clarifying the execution equipment, timing logic, operating parameters and threshold range of each instruction, and simultaneously completing instruction semantic conflict verification to avoid multi-device instruction conflicts. The specialized semantic parsing model consists of a text segmentation layer, a semantic role labeling layer, and an instruction mapping layer. These three modules are connected serially, and the output between layers is a structured semantic information table in JSON format. The text segmentation layer uses jieba word segmentation combined with a custom dictionary in the wastewater treatment field to complete word segmentation and part-of-speech tagging. The semantic role labeling layer uses a BERT-based SRL model to extract core semantic information from the text, such as the executing subject, action type, control parameters, timing requirements, and constraints. The instruction mapping layer calls a dedicated semantic and instruction mapping dictionary to convert the structured semantic information into control instructions that conform to the equipment communication specifications. The semantic information table contains six core fields, including the executing subject and action type.

[0044] The semantic conflict verification of instructions is performed in the order of equipment dimension → process sequence dimension. First, parameter conflicts are verified by grouping by single equipment dimension. The judgment rule is that the same control parameter of the same equipment is only allowed to have one instruction value in the same sequence. If multiple instruction values ​​appear, it is judged as a parameter conflict. Then, the sequence conflict of linkage instructions is verified by wastewater treatment process flow. The judgment rule is that the linkage instructions of different equipment must conform to the process sequence logic. If the sequence logic is contradictory, it is judged as a sequence conflict. Conflicting instructions are eliminated according to the principle of priority for emergency faults and priority for core process equipment. After eliminating conflicting instructions, the missing adaptable instructions are regenerated based on the process logic of the remaining instructions. The second conflict verification is completed simultaneously. The conflict judgment result is recorded in text form, including three core information: conflict type, involved instructions, and conflict reason. It is synchronously fed back to the strategy generation stage for strategy optimization.

[0045] Based on the timing logic and linkage constraints of the instruction set, instructions are issued to the equipment control system to achieve coordinated execution of multiple devices and multiple process units; during the execution process, real-time equipment operation feedback data and water quality parameter change data are collected and converted into semantic feedback text of the execution process according to unified textualization rules; Equipment operation feedback data is collected every 1 minute, with core indicators including equipment operating power, speed, inlet and outlet pressure, and operating status. Water quality parameter change data is collected every 5 minutes, with core indicators including COD, BOD5, ammonia nitrogen, total phosphorus, and SS. The collected raw data must be validated to remove invalid data such as jumps and exceeding the range. Valid data is converted into semantic feedback text in the format of "collection time-process unit-equipment / indicator name-value-status". Abnormal values ​​in the text are marked in red, and normal values ​​are marked in black. The status is divided into three categories: normal, warning, and abnormal.

[0046] The multi-device collaborative linkage timing logic strictly follows the process sequence of wastewater treatment: influent → bar screen → grit chamber → biological treatment tank → sedimentation tank → disinfection tank → effluent. Upstream process unit equipment commands are issued and executed first, and downstream process unit equipment adjusts commands based on the feedback from upstream execution. Command issuance is set with three priority levels: emergency fault, process control, and parameter fine-tuning. Commands within the same priority are executed according to the principle of core process equipment priority, and upstream process unit priority. Emergency fault commands have the highest priority and can interrupt the currently executing low-priority commands. Process control commands are next, and parameter fine-tuning commands have the lowest priority. The multi-device execution mode is divided into serial and parallel. Related devices within the same process unit are executed serially, while non-related devices in different process units are executed in parallel. During execution, the operating status of the equipment is monitored in real time. If any device malfunctions, the execution of commands of its related devices is suspended, triggering a fault warning and feedback to the strategy generation stage.

[0047] After execution, an execution feedback report containing the execution process, execution results, and performance evaluation is automatically generated using text summarization technology.

[0048] The execution feedback report uses a standardized, fixed format and includes six core modules: basic information, execution process record, equipment operation feedback, water quality parameter changes, operation effect evaluation, and problems and suggestions. The basic information module records the process unit, execution time, and triggered abnormal operating conditions of the strategy execution. The operation effect evaluation module evaluates the water quality compliance, equipment operation stability, and energy consumption from three dimensions, and rates it on a four-level scale: excellent, good, average, and poor. When the rating is poor, the problem and suggestion module should provide targeted strategy optimization directions. The report is stored in PDF format and uploaded to the wastewater treatment plant operation and maintenance management system simultaneously.

[0049] The incremental learning phase, based on the execution feedback text report generated in the parsing and execution phase, collects newly added fault handling cases, supplementary texts of operation and maintenance expert experience, and new version of industry standard texts to build an incremental semantic learning dataset. The incremental semantic learning dataset needs to undergo standardized screening and cleaning. The screening criteria are that the data is strongly related to sewage treatment fault diagnosis, process optimization, and equipment operation and maintenance, and the data information is complete and logically clear. Invalid data that lacks specific content, has incomplete information, or is irrelevant to the scenario is removed. The data cleaning steps are deduplication, format unification, and semantic normalization. First, duplicate data is removed, then text data of different formats are unified into TXT format, and finally, semantic normalization is performed on synonymous professional terms with different names to ensure consistency with the standard entities of the domain knowledge graph. The cleaned dataset is divided into fault case set, expert experience set, and industry standard set according to data type, and data labels are marked respectively to facilitate model fine-tuning training by type.

[0050] Then, by using incremental entity extraction and relation extraction techniques, we extract new semantic entities, new semantic relationships and optimization strategy rules from the incremental data, dynamically update the knowledge graph in the wastewater treatment field, and complete the incremental iteration of the knowledge graph and the automated accumulation of domain knowledge. Meanwhile, based on the incremental semantic learning dataset, the domain pre-trained language model is lightweightly fine-tuned using low-rank adaptation incremental fine-tuning technology to continuously optimize the model's semantic feature extraction, diagnostic reasoning, and policy generation capabilities. Low-rank adaptation incremental fine-tuning uses a LoRA adapter with a rank of 4, paired with the AdamW optimizer, a learning rate of 1e-5, a batch size of 8, and 5 training epochs. Fine-tuning is only performed on the attention and fully connected layers of the domain pre-trained language model, while the parameters of the other layers are fixed to reduce computational cost. The confidence threshold for entity extraction in the incremental entity and relation extraction technology is set to 0.9, and the confidence threshold for relation extraction is set to 0.8. Entities and relations below the threshold need to undergo a manual verification process, which is reviewed by senior experts in the wastewater treatment field. Only those that pass the review are updated to the domain knowledge graph. Those that fail the review are annotated with the reasons and added to the training set of the extraction model. The training set for incremental fine-tuning is the incremental semantic learning dataset, which is divided into training, validation, and test sets in a 7:2:1 ratio.

[0051] Based on feedback from actual implementation, semantic reasoning rules, semantic similarity calculation algorithms (strategy matching), and constraint verification rules were simultaneously optimized through semantic rule optimization algorithms, thereby improving the accuracy of fault diagnosis and the adaptability of handling strategies.

[0052] The semantic rule optimization algorithm is based on the performance evaluation results in the execution feedback report and follows the steps of "rule validity determination → rule parameter adjustment → rule verification and implementation". First, the accuracy and fit rate of each semantic rule in actual application are statistically analyzed. Rules with an accuracy rate of less than 85% are judged to be optimized. For inference rules, the association weight threshold between entities and the inference level limit are adjusted. For strategy matching rules, the similarity threshold and feature weight allocation are adjusted. For constraint verification rules, new equipment operation boundaries and industry standard requirements are added. After the rule adjustment, recent fault handling cases are used for verification. If the verification pass rate is not less than 90%, it is officially implemented. For those that fail the verification, the parameters are readjusted. At the same time, the rule optimization process and results are recorded in the incremental learning log for subsequent traceability and analysis.

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal sensing AI-based intelligent diagnosis and strategy linkage method for wastewater treatment, characterized in that, The specific steps include the following: In the text representation stage, multimodal heterogeneous data from the entire wastewater treatment process are collected. Customized text transformation and standardized preprocessing are performed on different modal data to generate a multimodal semantic text dataset. In the graph construction phase, based on the multimodal semantic text dataset and related texts in the field of wastewater treatment, seven core semantic entities are extracted, multi-level semantic relationships are defined to construct a domain semantic knowledge graph, entity semantic alignment and ambiguity elimination are completed, and a dynamic update interface for the knowledge graph is built. In the semantic fusion stage, the general pre-trained language model is fine-tuned and optimized for the domain to obtain a special language model adapted to the sewage treatment scenario. Real-time text data is input and combined with knowledge graph entity vectors to complete the multimodal semantic feature fusion extraction and weight allocation. In the reasoning and diagnosis phase, based on the fused feature vectors and dedicated language models, combined with the semantic reasoning rules of the knowledge graph, a hierarchical reasoning and diagnosis model based on the domain language model is built to identify abnormal working condition types and complete the severity classification, trace the root cause of the fault and record the reasoning path, and generate an interpretable diagnostic text report and standardized diagnostic semantic tags. In the strategy generation stage, a semantic library of disposal strategies bound to the knowledge graph is constructed. Candidate strategy sets are screened through semantic similarity algorithms. Non-compliant solutions are eliminated after multi-dimensional constraint verification. Standardized natural language disposal strategy text is generated after optimization. In the parsing and execution phase, a parsing mapping rule between natural language and equipment control instructions is established. The strategy is decomposed into a standardized set of control instructions corresponding to each process unit and equipment, and instruction semantic conflict verification is completed. Instructions are issued to realize the coordinated execution of multiple devices. Execution feedback data is collected and execution feedback reports are generated. In the incremental learning phase, an incremental semantic learning dataset is built based on the execution feedback report and newly added domain text. The knowledge graph is updated by extracting the semantic relationships between new entities, fine-tuning the domain pre-trained model, and optimizing semantic reasoning, semantic similarity matching, and constraint verification rules.

2. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 1, characterized in that, The multimodal heterogeneous data collected in the text representation stage includes structured water quality sensor values, equipment operating parameters, and SCADA data, as well as unstructured inspection audio and video, equipment voiceprints, maintenance logs, fault reports, industry standards, process documents, and expert experience data across all dimensions. The text conversion of different modal data adopts domain adaptation technology. The preprocessing of unstructured raw text data includes word segmentation, stop word removal, named entity recognition, and semantic normalization, ultimately generating a multimodal semantic text dataset with a unified format and unified semantic dimensions.

3. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 2, characterized in that, The graph construction phase extracts seven core semantic entities—water quality indicators, process units, equipment, fault phenomena, root causes of faults, treatment strategies, and normative constraints—from domain-related text and multimodal semantic text datasets. By defining hierarchical semantic relationships and combining them with scenario-based event information from the entire wastewater treatment process, a four-dimensional vertical domain semantic knowledge graph of entities, relationships, events, and constraints is constructed. Contextual semantic matching technology is used to achieve entity semantic alignment and ambiguity resolution, and a dynamic update interface for the knowledge graph is established.

4. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 3, characterized in that, The semantic fusion stage is based on a general pre-trained language model, combined with a domain-specific semantic dictionary for wastewater treatment for incremental pre-training and fine-tuning, to obtain a domain-specific language model adapted to the wastewater treatment scenario. Real-time multimodal semantic text data and knowledge graph entity embedding vectors are fused through a semantic attention mechanism to extract full-dimensional semantic fusion feature vectors. The feature weights are optimized through a semantic weight dynamic allocation algorithm to strengthen key semantic features related to wastewater treatment fault diagnosis and process anomalies, and suppress redundant noise features.

5. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 4, characterized in that, The hierarchical semantic reasoning diagnostic model in the reasoning and diagnosis stage is based on a semantic classification model fine-tuned and optimized from a domain language model. It identifies four types of abnormal operating conditions: water quality exceeding standards, equipment failure, process abnormality, and environmental disturbance. It also completes a four-level semantic classification of severity: early warning, general abnormality, severe abnormality, and emergency failure. Through a reverse chain semantic reasoning algorithm, it traces the root cause of the fault, eliminates interfering factors, and records the reasoning path. The generated interpretable diagnostic text report contains the core information of the abnormality, including the location, type, severity, root cause analysis, and reasoning path, and simultaneously outputs standardized diagnostic semantic tags.

6. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 5, characterized in that, The strategy generation stage binds the semantic library of disposal strategies to the domain knowledge graph. Each strategy is labeled with core semantic tags including applicable scenarios, execution steps, constraints, and expected effects, covering the disposal needs of faults and abnormal operating conditions in all scenarios of sewage treatment, and a strategy semantic index library is built. Candidate strategy sets are screened through semantic similarity calculation algorithms, and compliance verification is completed by combining multi-dimensional constraints. After eliminating non-compliant solutions, the optimal candidate strategy is optimized for real-time operating condition adaptability to generate standardized disposal strategy text.

7. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 6, characterized in that, The parsing and execution phase utilizes dependency parsing and semantic role labeling techniques to establish parsing mapping rules and a dedicated semantic and instruction mapping dictionary for standardized natural language processing strategy text and equipment control instructions. It decomposes the standardized natural language processing strategy text into standardized control instruction sets corresponding to each process unit and equipment, clarifying the execution equipment, core parameters, timing logic, and threshold range for each instruction, and completing instruction semantic conflict verification. Instructions are issued according to timing logic and linkage constraints, execution feedback data is collected in real time and converted into semantic feedback text of the execution process. After execution is completed, an execution feedback report containing the execution process, execution results, and operational effect evaluation is generated.

8. The AI-powered intelligent diagnosis and strategy linkage method for multimodal perception in wastewater treatment according to claim 7, characterized in that, The incremental semantic learning dataset in the incremental learning phase consists of execution feedback reports, newly added fault handling cases, expert experience supplementary text, and new version of industry standard text. The domain knowledge graph is updated through incremental entity and relation extraction technology, and the domain pre-trained language model is optimized using low-rank adaptation incremental fine-tuning technology. The semantic reasoning rules, semantic similarity calculation algorithm, and constraint verification rules are optimized in combination with actual execution effect feedback.

Images