CN-121981273-A - Water supply network intelligent diagnosis method and system based on large visual language model

CN121981273ACN 121981273 ACN121981273 ACN 121981273ACN-121981273-A

Abstract

The invention discloses a water supply network intelligent diagnosis method and system based on a large visual language model, wherein the method comprises the following steps of collecting multi-source time sequence data, standardizing the multi-source time sequence data, converting the multi-source time sequence data into a visual chart, and constructing a diagnosis map containing context semantics by combining domain knowledge; after interference is filtered by the fixed fluctuation mode library, a general visual language model is input, and abnormal perception, type identification and root cause inference are completed by zero/few sample visual semantic analysis and causal reasoning, so that a structured natural language report is generated. The method has the core that the problem of abnormal diagnosis of the numerical sequence is converted into the task of visual semantic analysis and reasoning which can be understood by a visual language model, the diagnosis with high precision and interpretation can be realized under the condition of few labels, the false alarm is reduced, the reliability is enhanced, the problems that the existing pure numerical method depends on a large amount of label data, the interpretation is poor, the long-term hidden leakage is difficult to detect and the like are effectively solved, and the method is suitable for the long-term hidden leakage monitoring scene.

Inventors

ZHAO YUMENG
WANG RUNZHI
Zhao Zitang
ZHONG BAOYI
PAN XINYU
ZHANG HANQI
MA JUN

Assignees

哈尔滨工业大学

Dates

Publication Date: 20260505
Application Date: 20260210

Claims (10)

1. A water supply network intelligent diagnosis method based on a large visual language model is characterized by comprising the following steps: step 1, multi-source time sequence sensing data of a plurality of monitoring points in a water supply network are collected, and standardized processing is carried out on the time sequence sensing data; Step 2, converting the time sequence sensing data after standardized processing into a visual chart in a preset format, and generating a diagnostic map containing context semantic information by combining with a predefined knowledge prompt in the structured field; Step 3, carrying out pattern recognition on the diagnostic map based on a pre-configured fixed pattern library, and filtering out data conforming to a non-fault fluctuation pattern defined in the fixed pattern library; Step 4, inputting the diagnosis map subjected to mode filtering into a pre-trained universal visual language model, guiding the visual language model to perform visual semantic analysis and causal reasoning on the diagnosis map, and outputting a diagnosis reasoning result comprising abnormal types, possible reasons and confidence; And 5, generating a structured natural language diagnosis report according to the diagnosis reasoning result.
2. The intelligent diagnosis method for the water supply network based on the large visual language model according to claim 1, wherein in the step 1, the multi-source time sequence sensing data comprises at least two of pressure data, flow data, noise data and water quality data, and the normalization processing comprises time alignment, missing value processing, outlier removal and normalization operation.
3. The intelligent diagnosis method for the water supply network based on the large visual language model according to claim 1 is characterized in that in the step 2, the specific step of converting the standardized time sequence data into a visual chart with a preset format is that a line graph, an area graph, a scatter graph or a combined chart is selected and generated according to the data type and the diagnosis requirement, a structured domain knowledge prompt is embedded into the diagnosis chart in a text form, and the structured domain knowledge prompt comprises at least one of monitoring point position information, a sensor type, a data acquisition period, a pipe network topology relation context and a historical maintenance record.
4. The intelligent diagnosis method for water supply network based on large visual language model according to claim 1, wherein in the step 3, the predefined non-fault fluctuation mode in the fixed mode library comprises at least one of periodic water peak mode, pressure gradual change mode caused by starting and stopping of water plant pump and flow rate return-to-zero mode caused by planned maintenance, and the filtering is realized by matching the visual characteristics of the current diagnosis map with the mode characteristics in the fixed mode library.
5. The intelligent diagnosis method for the water supply network based on the large visual language model, which is disclosed by claim 1, is characterized by comprising the specific steps of constructing a hierarchical prompt template, wherein the prompt template comprises task instructions, visual focus guide questions and causal reasoning chain questions, and inputting the diagnosis map and the prompt template subjected to pattern filtering into the visual language model together so as to answer the questions in the prompt template in sequence, thereby cooperatively completing abnormal perception, type identification and root cause inference.
6. A large visual language model based intelligent diagnosis system for water supply network implementing the method according to any one of claims 1-5, characterized in that the system comprises a data acquisition and processing module, a diagnosis map construction module, a fixed pattern filtering module, a visual language model reasoning module and a report generation module, wherein: The data acquisition and processing module is used for acquiring multi-source time sequence sensing data and carrying out standardized processing on the time sequence sensing data; the diagnostic map construction module is used for converting the standardized time sequence sensing data into a visual chart, and fusing knowledge prompts in the structural field to generate a diagnostic map; the fixed pattern filtering module is embedded with a fixed pattern library and is used for identifying and filtering non-fault fluctuation patterns in the diagnostic map; the visual language model reasoning module is integrated with a pre-trained universal visual language model and is used for receiving the filtered diagnostic map, executing visual semantic analysis and causal reasoning and outputting a diagnostic reasoning result; the report generation module is used for automatically generating a structured natural language diagnosis report according to the diagnosis reasoning result.
7. The large visual language model-based water supply network intelligent diagnosis system according to claim 6, wherein the diagnosis map construction module comprises a map generation unit and a prompt embedding unit, the map generation unit supports configuration and generation of multiple visual map types, and the prompt embedding unit is used for carrying out association packaging on structured text prompts from a knowledge base and generated charts.
8. The large visual language model based intelligent diagnosis system for water supply network according to claim 6, wherein the fixed pattern filtering module comprises a pattern feature extractor for extracting key visual features from the diagnosis map and a pattern matcher for calculating the similarity of the key visual features extracted by the pattern feature extractor and the pattern features stored in the fixed pattern library and deciding whether to filter according to a threshold value.
9. The large visual language model-based intelligent diagnosis system for water supply network according to claim 6, wherein the visual language model reasoning module comprises a prompt management unit for managing and invoking hierarchical prompt templates for different diagnosis tasks and a model interface unit for encapsulating interaction protocols with cloud or locally deployed visual language models.
10. A computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, carries out the steps of the method according to any of the preceding claims 1-5.

Description

Water supply network intelligent diagnosis method and system based on large visual language model Technical Field The invention belongs to the technical field of intelligent water affairs and industrial Internet of things, relates to an intelligent monitoring and intelligent diagnosis method and system for key infrastructure such as urban water supply networks, and particularly relates to a water supply network intelligent diagnosis method and system for carrying out data visual analysis and causal reasoning based on a large visual language model. Background Urban water supply networks are used as key infrastructure, and stable operation is crucial. However, the leakage and pipe explosion risks are continuously increased due to factors such as pipe network aging, environmental change and the like, and huge water resource waste and economic loss are caused. Currently, along with popularization of industrial Internet of things technology, a water supply network is provided with a large number of sensors, so that time sequence data such as pressure, flow and the like can be continuously collected, and a data base is provided for intelligent monitoring. The current mainstream abnormality detection methods are mainly divided into two types. One type is based on statistical thresholds, such as setting upper and lower thresholds for pressure, and triggering an alarm when the data exceeds the thresholds. Although the method is simple, the false alarm rate is high, the true fault mode cannot be distinguished from normal water fluctuation (such as the early and late peak), and the slow-development long-term hidden leakage is more difficult to find. Another class is based on deep learning methods, such as modeling time-ordered data using long-short-term memory networks or self-encoders to detect anomalies. The method improves the detection precision to a certain extent, but has two key defects that firstly, the method is severely dependent on a large amount of historical fault data accurately marked by experts to carry out supervision training, and the marked data in actual operation and maintenance are extremely scarce and have high acquisition cost, so that the wide application and iterative optimization of the model are limited. Second, deep learning models are often considered as "black boxes" whose decision process lacks interpretability. When the system gives an alarm, an operation and maintenance person cannot understand the data characteristics based on which model makes a judgment, and the reliability of the alarm is difficult to evaluate, so that the intelligent system is not trusted, and the effective deployment of the technology in an actual scene is blocked. Therefore, an intelligent diagnosis method which can work under the condition of few marked data, has transparent diagnosis process and interpretable result is urgently needed, so that various pipe network anomalies including long-term hidden leakage can be accurately identified, and the operation and maintenance efficiency and decision trust degree are improved. Disclosure of Invention The invention provides a large visual language model-based intelligent diagnosis method and system for a water supply network, which are high in data efficiency, transparent in process and reliable in result, in order to overcome the defects that the conventional intelligent diagnosis method for the water supply network has strong dependence on labeling data and poor interpretability and is difficult to detect long-term hidden leakage. The method has the core that the problem of abnormal diagnosis of the numerical sequence is converted into the task of visual semantic analysis and reasoning which can be understood by a visual language model, the diagnosis with high precision and interpretation can be realized under the condition of few labels, the false alarm is reduced, the reliability is enhanced, the problems that the existing pure numerical method depends on a large amount of label data, the interpretation is poor, the long-term hidden leakage is difficult to detect and the like are effectively solved, and the method is suitable for the long-term hidden leakage monitoring scene. The invention aims at realizing the following technical scheme: a water supply network intelligent diagnosis method based on a large visual language model comprises the following steps: Step 1, multi-source time sequence sensing data of a plurality of monitoring points in a water supply network are collected, and standardized processing is carried out on the time sequence sensing data, wherein the multi-source time sequence sensing data comprises at least two of pressure data, flow data, noise data and water quality data; Step 2, converting the time sequence sensing data after the standardization processing into a visual chart with a preset format, and generating a diagnosis chart containing context semantic information by combining with a predefined structured domain knowledge prompt, wher