Intelligent water conservancy intelligent supervision method and system based on ai multi-modal large model

Abstract

The present application relates to the technical field of intelligent water conservancy supervision, and discloses an intelligent water conservancy intelligent supervision method and system based on an AI multi-modal large model, which comprises the following steps: accessing sensor numerical values, images and videos, voice transcription content, inspection text and device logs, generating multi-modal supervision data packets through data cleaning, time alignment and space correlation; for the lock station provided with a pre-lock water depth monitoring point, a gate opening feedback device and a post-lock flow monitoring point, a directed water potential transmission chain is established, forming an expected change multiple, an actual change multiple and a transmission closure degree; an AI multi-modal large model service generates a supervision event record, and the intelligent water conservancy supervision platform generates a three-point transmission inconsistency mark when the pre-lock water depth, the gate opening and the post-lock flow all do not cross the preset single threshold and the transmission closure degree continuously falls below the lower limit of the closure degree, and determines the supervision level, the primary check object and the water conservancy supervision task.

Description

Technical Field

[0001] This invention relates to the field of smart water conservancy supervision technology, and more specifically, to a smart water conservancy intelligent supervision method and system based on an AI multimodal large model. Background Technology

[0002] Smart water conservancy monitoring platforms typically integrate data from sources such as water level gauges, flow meters, gate opening meters, video surveillance, drone patrols, mobile inspections, operation logs, and work order systems. They utilize rule bases, alarm modules, and mobile terminals to perform daily inspections, flood control duty, river and lake management, water project operation, and on-site project management. Existing platforms often judge the status of individual measuring points or devices in gate monitoring, such as checking water depth, gate opening, and flow rate separately for limits, and then generating alarms based on video footage or inspection records. This approach is suitable for routine limit violations and equipment anomalies, but when the upstream water depth, gate opening, and downstream flow rate are all within permissible ranges, the platform struggles to determine whether the transmission relationship between these three metrics is consistent across adjacent data collection periods. For on-duty personnel, AI multimodal large model services can provide anomaly alerts from images and videos, speech-to-text content, and inspection text. However, if the alerts do not form a unified basis with upstream water depth, gate opening, downstream flow rate, and task flow fields, the alarms may still fall under the general gate check level. When on-site personnel receive such tasks, they often need to check each item among the gate actuators, downstream flow measurement sections, channel status, and upstream water depth monitoring points. Evidence compilation and verification also require manual data linking. Therefore, the smart water conservancy monitoring system needs to connect on-site data collection, multimodal recognition results, monitoring levels, and task assignment targets into the same business chain without changing the existing on-site layout and gate control boundaries. Summary of the Invention

[0003] This invention provides a smart water conservancy intelligent supervision method and system based on an AI multimodal large model, which solves the technical problems mentioned in the background art.

[0004] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0005] Firstly, the intelligent water conservancy supervision method based on an AI multimodal large model is executed by a water conservancy supervision system. This system includes a data access gateway, a water potential calculation unit, an AI multimodal large model service, a smart water conservancy supervision platform, and management and mobile terminals, comprising: The data access gateway accesses the multimodal water conservancy monitoring data, and the smart water conservancy monitoring platform performs data cleaning, time alignment and spatial correlation on the multimodal water conservancy monitoring data to generate multimodal monitoring data packets; When a gate station is equipped with a water depth monitoring point upstream of the gate, a gate opening feedback device, and a flow monitoring point downstream of the gate, the water potential calculation unit establishes a directional water potential transmission chain between the water body upstream of the gate, the gate, and the flow measurement section downstream of the gate. Based on the gate opening change relationship and the water depth change relationship at adjacent acquisition times, it forms the expected change multiple, and based on the flow change relationship downstream of the gate, it forms the actual change multiple. The degree of transmission closure is formed according to the difference between the expected change multiple and the actual change multiple. The AI ​​multimodal large model service generates regulatory event records based on multimodal regulatory data packages, transmission closure, and water conservancy knowledge data; When the water depth in front of the gate, the gate opening, and the flow rate after the gate all do not exceed the corresponding preset individual thresholds, and the transmission closure is lower than the lower limit of closure for two consecutive collection cycles, the smart water conservancy supervision platform generates a three-point transmission inconsistency marker. Combining the supervision event record, transmission closure, three-point transmission inconsistency marker, and data integrity, the supervision level is determined. Based on the order of changes in water depth in front of the gate, gate opening, and flow rate after the gate, the primary verification object is determined, and water conservancy supervision tasks are generated and assigned.

[0006] Secondly, the intelligent water conservancy monitoring system based on AI multimodal big data models includes: The data access gateway is used to provide multimodal water conservancy monitoring data to the smart water conservancy monitoring platform; The intelligent water conservancy supervision platform is used to perform data cleaning, time alignment, and spatial correlation on multimodal water conservancy supervision data and generate multimodal supervision data packages. The water potential calculation unit is used to receive the upstream water depth, gate opening and downstream flow at adjacent collection times provided by the smart water conservancy supervision platform, establish a directed water potential transmission chain, form the expected change multiple based on the gate opening change relationship and the upstream water depth change relationship, form the actual change multiple based on the downstream flow change relationship, and form the transmission closure degree according to the degree of difference between the expected change multiple and the actual change multiple. AI multimodal large model service is used to generate regulatory event records based on multimodal regulatory data packets, transmission closure and water conservancy knowledge data; The intelligent water conservancy supervision platform is also used to generate a three-point transmission inconsistency mark when the water depth in front of the gate, the gate opening and the flow rate after the gate have not exceeded the corresponding preset individual thresholds and the transmission closure is lower than the lower limit of closure for two consecutive collection cycles. The supervision level is determined by combining the supervision event record, transmission closure, three-point transmission inconsistency mark and data integrity. The primary verification object is determined according to the change order of the water depth in front of the gate, the gate opening and the flow rate after the gate and water conservancy supervision tasks are assigned. Management and mobile terminals are used to receive water conservancy supervision tasks dispatched by the smart water conservancy supervision platform and submit handling feedback data to the smart water conservancy supervision platform. The smart water conservancy supervision platform is also used to perform review and cancellation based on the disposal feedback data.

[0007] Compared with the prior art, the present invention has the following substantial features and significant progress: This invention relies on existing gate station measuring points, gate opening feedback devices, and downstream flow measurement sections. Within the water conservancy supervision system, it incorporates upstream water depth, gate opening, and downstream flow into the same multimodal supervision data package. The transmission closure is used to participate in the generation of supervision event records for AI multimodal large model services, the determination of the level of the smart water conservancy supervision platform, and the assignment of water conservancy supervision tasks. This allows the platform to identify inconsistencies in the transmission of three points even when a single data point does not exceed a preset threshold. This directs gate station verification towards primary verification targets such as gate actuators, downstream flow measurement sections or channels, and upstream water depth monitoring points, reducing irrelevant investigations and improving the consistency of evidence organization, task orientation, and review archiving. Simultaneously, it preserves manual review, existing approval processes, and gate control boundaries. Attached Figure Description

[0008] Figure 1 This is a diagram of the architecture mechanism of the intelligent water conservancy supervision system based on AI multimodal large model of the present invention. Detailed Implementation

[0009] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0010] Example 1: This implementation method is applicable to water conservancy supervision operations in reservoirs, rivers, dikes, sluice gates, pumping stations, water intakes, drainage outlets, water conservancy projects under construction, and water conservancy inspection areas. The method is executed by a water conservancy supervision system, which includes a data access gateway, a water potential calculation unit, an AI multimodal large model service, a smart water conservancy supervision platform, and management and mobile terminals.

[0011] This embodiment describes the monitoring and processing procedures set up in a gate station scenario. The gate station already has a water depth monitoring point in front of the gate, a gate opening feedback device, and a flow monitoring point behind the gate. The water in the field flows from the water in front of the gate through the gate to the flow measurement section behind the gate, thus forming a natural front-to-back transmission relationship.

[0012] Specifically, in existing smart water conservancy monitoring, the platform typically checks whether the upstream water depth, gate opening, and downstream flow exceed their respective preset thresholds. This method can handle single-point boundary violations or obvious equipment malfunctions. However, when the upstream water depth, gate opening, and downstream flow all do not exceed their respective preset thresholds, but the changes in these three parameters are inconsistent between adjacent data collection times, relying solely on single-point judgments makes it difficult to promptly formulate clear water conservancy monitoring tasks.

[0013] See Figure 1The upstream water depth monitoring point 100, the gate opening feedback device 110, and the downstream flow monitoring point 120 are located in the on-site data acquisition area of ​​the sluice gate station. They respectively output the upstream water depth, gate opening, and downstream flow. The data access gateway 200 is located between the on-site data acquisition area and the platform processing area, used to access multimodal water conservancy monitoring data. The intelligent water conservancy monitoring platform 300 is located in the platform processing area, and the water potential calculation unit 310 and the AI ​​multimodal large model service 320 work collaboratively with the intelligent water conservancy monitoring platform 300. The management and mobile terminal 400 is used to receive water conservancy monitoring tasks and submit handling feedback data.

[0014] Example 2: The data access gateway 200 provides multimodal water conservancy monitoring data to the smart water conservancy monitoring platform 300. This multimodal monitoring data includes sensor readings, images and videos, voice-transcribed content, inspection text, and equipment logs. Sensor readings include upstream water depth, gate opening, and downstream flow. Images and videos can originate from video surveillance equipment, fixed cameras, or drone inspection equipment; voice-transcribed content can come from on-site personnel's voice recordings; inspection text can come from mobile inspection terminals; and equipment logs can come from gate opening feedback devices, flow meters, pump unit operation terminals, or existing business systems.

[0015] Point 100, used for monitoring water depth upstream of the gate, is used to collect water depth data. The water depth upstream of the gate represents the water depth read at the current time of data collection from the upstream measuring point. The water depth upstream of the gate can be provided by existing water level gauges or equivalent hydraulic sensing terminals, and the data records must retain at least the collection time, point number, and quality marker.

[0016] The gate opening feedback device 110 is used to output the gate opening degree. The gate opening degree represents the degree of opening read by the gate opening feedback device 110 at the current acquisition time. The gate opening degree is provided by the gate opening feedback device 110; the gate opening feedback device 110 can use existing field equipment and be connected to the monitored object file by device number.

[0017] Downstream flow monitoring point 120 is used to collect downstream flow data. Downstream flow represents the flow rate read at the current acquisition time from the downstream flow measurement section. Downstream flow can be provided by existing flow meters or flow measurement section data, and the data record must retain at least the acquisition time, point number, and quality mark.

[0018] The Smart Water Conservancy Supervision Platform 300 is used to perform data cleaning, time alignment, and spatial correlation on multimodal water conservancy supervision data, and generate multimodal supervision data packages. The Smart Water Conservancy Supervision Platform 300 is also used for rule review, supervision level determination, evidence display, water conservancy supervision task assignment, receiving of handling feedback data, review and cancellation, and statistical archiving.

[0019] The water potential calculation unit 310 receives the upstream water depth, gate opening, and downstream flow rate from the smart water conservancy monitoring platform 300 at adjacent acquisition times, establishes a directed water potential transmission chain, and generates the expected change factor, actual change factor, and transmission closure degree. The water potential calculation unit 310 can be configured as a software module within the smart water conservancy monitoring platform 300, or as an independent data processing task or microservice communicating with the smart water conservancy monitoring platform 300. This deployment configuration represents a conventional architecture choice for existing software systems.

[0020] AI Multimodal Large Model Service 320 is used to generate regulatory event records based on multimodal regulatory data packages, transmission closure, and water conservancy knowledge data, and outputs content that cannot be confirmed in a manner that awaits manual verification.

[0021] The management and mobile terminal 400 is used to receive water conservancy supervision tasks dispatched by the smart water conservancy supervision platform 300 and to submit handling feedback data to the smart water conservancy supervision platform 300. The management and mobile terminal 400 can be used by on-duty personnel or carried by on-site personnel. Its functions include task reception, on-site verification, submission of handling text, submission of on-site photos, submission of retest values, submission of equipment status, and submission of completion time. The smart water conservancy supervision platform 300 performs review and cancellation based on the handling feedback data.

[0022] Example 3: The Smart Water Conservancy Supervision Platform 300 first establishes files for supervised objects. These files include object number, geographical location, monitoring point location, and task receiver identifier. The object number distinguishes different supervised objects among reservoirs, rivers, dikes, sluice gates, pumping stations, water intakes, drainage outlets, water conservancy projects under construction, and water conservancy inspection areas. The geographical location determines the spatial affiliation of the supervised object within administrative divisions, river basins, or engineering areas. Monitoring points record the on-site data collection locations, such as the upstream water depth monitoring point 100, the gate opening feedback device 110, and the downstream flow monitoring point 120. The task receiver identifier identifies the recipient of the electronically assigned water conservancy supervision task.

[0023] The regulated object file can be stored in the existing business system database or retrieved by the Smart Water Conservancy Supervision Platform 300 via a data interface. This file is used to bind the data source, spatial location, and task receiver identifier to the same regulated object. Figure 1 The gate station shown in the figure has at least the following information in its monitoring object file: gate station object number, location number of water depth monitoring point 100 in front of the gate, device number of gate opening feedback device 110, location number of flow monitoring point 120 behind the gate, and task receiving terminal identifier of management and mobile terminal 400.

[0024] When the data access gateway 200 receives water conservancy multimodal monitoring data, it can receive data through existing wired communication, wireless communication, water conservancy private network, unified data interface, batch import interface, or message subscription interface. This access method is a conventional data access method for the water conservancy monitoring platform. The data access gateway 200 registers the source of the accessed data and performs protocol conversion, enabling data reported by different devices or business systems to enter the unified processing flow of the smart water conservancy monitoring platform 300.

[0025] When sensor data is input, the smart water conservancy monitoring platform retains the collection time, location number, and quality tag for each of the 300 pairs of data points: water depth upstream of the sluice gate, gate opening, and flow downstream of the sluice gate. The collection time is used to determine the current and previous collection times; the location number is used to correspond to the monitoring point in the monitored object's file; and the quality tag indicates whether the data is missing, out of bounds, has communication anomalies, or is in a maintenance state. Data with abnormal quality tags are not directly included in this round of intelligent analysis.

[0026] When images and videos are received, the smart water conservancy monitoring platform retains the capture time and device number. The capture time is used for time alignment with the sensor data acquisition time, and the device number is used to confirm that the images and videos belong to the same monitored object or adjacent areas. For sluice gates, images and videos can cover the water body upstream of the gate, the gate itself, the flow measurement section downstream of the gate, the channel status, floating objects, or the on-site operation area.

[0027] When voice-to-text content, inspection texts, and equipment logs are accessed, the smart water conservancy monitoring platform 300 retains the source. The source can include the submitting personnel, mobile inspection terminals, fixed business systems, gate opening feedback devices 110, or the source system recorded by the data access gateway 200.

[0028] To avoid data commingling among different regulated entities, the Smart Water Conservancy Monitoring Platform 300 uses the regulated entity archives as a basis for determining the attribution of multimodal water conservancy monitoring data. When a data record contains an object number, the attribution is determined primarily by the object number; when a data record only contains a location number or equipment number, the attribution is determined by the monitoring point in the regulated entity archive; when images and videos only contain the equipment number and shooting time, the attribution is determined by the equipment number, shooting direction, and geographical location. This attribution determination is a standard engineering implementation of multi-source data processing.

[0029] Example 4: The Smart Water Conservancy Monitoring Platform 300 performs data cleaning on multimodal water conservancy monitoring data. For sensor values, data cleaning includes format verification, missing data marking, duplicate record removal, and unit unification. Format verification confirms whether the upstream water depth, gate opening, and downstream flow can be recognized as numerical records by the platform; missing data marking records gaps caused by non-reporting, communication interruptions, or equipment maintenance; duplicate record removal prevents duplicate data from the same acquisition time from being included in calculations; and unit unification ensures that values ​​from the same type of sensor are stored according to the platform's existing unit rules.

[0030] Format validation failures include empty fields, field types that cannot be converted to numeric values, missing data collection time, missing site numbers, or missing quality markers. Missing markers are written by the data access gateway 200 or the smart water conservancy monitoring platform 300 when any required field is not received. Duplicate record removal is based on a combined judgment condition of the monitored object, site number, and data collection time; when multiple duplicate records exist, the smart water conservancy monitoring platform 300 retains the record with a normal quality marker and the earliest received time. Unit uniformity is performed according to the unit rules registered in the monitored object's file; the converted values ​​and the original values ​​retain their source records.

[0031] For images and videos, data cleaning includes sharpness checking, keyframe extraction, and timestamp verification. Sharpness checking eliminates severely blurry, occluded, or unrecognizable images; keyframe extraction extracts image frames from the video that represent the state of the water surface, embankments, sluice gates, floating objects, and channels; timestamp verification confirms that the image and video capture time corresponds to the sensor data acquisition time. Images and videos that fail sharpness checking or timestamp verification can be included in the abnormal data area.

[0032] Sharpness checks can be performed using image edge strength or the platform's existing sharpness score. Images and videos with a sharpness score below the lower limit are excluded from this round of intelligent analysis by the AI ​​multimodal large model service 320. Keyframe extraction prioritizes image frames whose capture time falls within the time alignment window before and after the current acquisition time and whose sharpness score is not lower than the lower limit. Timestamp verification is based on the difference between the image / video capture time and the sensor's numerical acquisition time; images and videos with a difference exceeding the time alignment window are placed in the abnormal data area.

[0033] The clarity score is generated by the Smart Water Conservancy Supervision Platform 300 according to the image and video processing rules in the monitored object files. If the monitored object files already have a clarity scoring algorithm configured, the Smart Water Conservancy Supervision Platform 300 uses that algorithm. If the monitored object files do not have a clarity scoring algorithm configured, the Smart Water Conservancy Supervision Platform 300 converts the image frame into a grayscale image, calculates the cumulative value of the absolute value of the grayscale difference between adjacent pixels, and divides this cumulative value by the number of adjacent pixel pairs involved in the calculation and 255 to obtain a clarity score between 0 and 1. The above clarity score is only used to determine whether the image and video participate in the current round of intelligent analysis by the AI ​​multimodal large model service 320.

[0034] The lower limit of image clarity is maintained by the Smart Water Conservancy Supervision Platform 300 according to the device number in the file of the supervised object. When using historical normal image samples to determine the lower limit of image clarity, the Smart Water Conservancy Supervision Platform 300 selects historical normal image samples that have been manually reviewed and are identifiable. The number of historical normal image samples is no less than 30 frames, and the 5th percentile of the image clarity score is calculated as the lower limit of image clarity. If there are fewer than 30 historical normal image samples, the Smart Water Conservancy Supervision Platform 300 adopts the lower limit of image clarity in the existing image quality rules of the management unit. If no existing image quality rules exist, the image and video will not participate in the current round of intelligent analysis and data integrity verification of the AI ​​multimodal large model service 320. The lower limit of image clarity must at least retain the applicable supervised object, device number, sample source or rule source, effective time, and maintenance personnel identification.

[0035] The time alignment window is determined by the Smart Water Conservancy Supervision Platform 300 according to the collection cycle in the files of the supervised objects. For Figure 1 For the gate station shown, the time alignment window is half the acquisition cycle of the upstream water depth, gate opening, and downstream flow rate in the monitored object's file. When the time interval between adjacent frames of an image or video is greater than half of this acquisition cycle, the time alignment window is taken as the time interval between adjacent frames of the image or video, but not exceeding one acquisition cycle. Voice transcription, inspection text, and equipment logs are matched with sensor values ​​using the same time alignment window; data that does not fall within the time alignment window is entered into the abnormal data area. When the acquisition cycle, video equipment, or data interface changes, the smart water conservancy monitoring platform 300 synchronously updates the corresponding time alignment window for the monitored object.

[0036] For speech-to-text and inspection text, data cleaning includes field segmentation, invalid character removal, and standardization of common terminology. Field segmentation is used to divide the content such as on-site description, discovery time, object number, and handling suggestions into processable fields; invalid character removal is used to remove garbled characters or duplicate characters that cannot express the business meaning.

[0037] Time alignment is used to establish adjacency between the current acquisition time and the previous acquisition time. The current acquisition time is denoted as... This indicates the current data acquisition time; the previous acquisition time is denoted as... This indicates the adjacent collection time before the current collection time. The Smart Water Conservancy Supervision Platform 300 can use the collection time reported by the device or the platform's receiving time as an auxiliary verification basis. When the collection cycles of data from different sources are different, the closest valid record can be found according to the time alignment window configured in the existing business system. This time alignment window matching is only used for data alignment.

[0038] Spatial association is used to associate numerical values, images and videos, speech-to-text content, inspection texts, and equipment logs of the same monitored object or adjacent areas within the same time period into a multimodal monitoring data package. For Figure 1 For the gate station shown, the smart water conservancy supervision platform 300 maps the water depth monitoring point 100 in front of the gate, the gate opening feedback device 110, and the flow monitoring point 120 behind the gate to the same gate station object, and writes the video key frames, equipment logs, and inspection texts within the same time period as the gate station into the same multimodal supervision data package.

[0039] The multimodal regulatory data package is a collection of business data generated by the Smart Water Conservancy Regulatory Platform 300. It includes at least the regulatory object identifier, data package number, collection time range, sensor values, images and videos, speech-to-text content, inspection text, equipment logs, and subsequently written expected change multiples, actual change multiples, and transmission closure. The multimodal regulatory data package provides unified input to the AI ​​multimodal large model service 320 and is also used by the Smart Water Conservancy Regulatory Platform 300 for rule review and evidence package generation.

[0040] Data that fails the verification is entered into the abnormal data area and is not included in this round of intelligent analysis. The abnormal data area can record the cause of the anomaly, the abnormal field, the time of occurrence of the anomaly, and the processing status, so as to facilitate manual investigation of equipment communication, sensor quality, or data interface problems.

[0041] Example 5: For a gate station equipped with an upstream water depth monitoring point 100, a gate opening feedback device 110, and a downstream flow monitoring point 120, the hydrodynamic calculation unit 310 establishes a directed hydrodynamic transmission chain connecting the upstream water body, the gate, and the downstream flow measurement section. The directed hydrodynamic transmission chain is arranged along the actual flow direction, with the upstream water body as the input end, the gate as the regulating end, and the downstream flow measurement section as the response end.

[0042] When establishing a directed water potential transmission chain, the water potential calculation unit 310 only binds three directly collected field data: upstream water depth, gate opening, and downstream flow rate. The upstream water depth is provided by the upstream water depth monitoring point 100, the gate opening is provided by the gate opening feedback device 110, and the downstream flow rate is provided by the downstream flow rate monitoring point 120.

[0043] Under the condition of the same gate, adjacent sampling times, and other conditions changing little over a short period, the general relationship of gate outflow can be expressed as the flow rate being approximately proportional to the gate opening and approximately varying with the square root of the upstream water depth. In order to eliminate dimensions and retain the transmission relationship between adjacent sampling times, the hydrodynamic calculation unit 310 generates an expected change factor based on the gate opening change relationship and the upstream water depth change relationship between adjacent sampling times.

[0044] ; in, Indicates the expected change factor. This indicates the gate opening at the current data acquisition moment. This indicates the gate opening at the previous data collection time. This indicates the water depth in front of the sluice gate at the current data collection time. This indicates the water depth in front of the sluice gate at the previous data collection time. This indicates the gate opening zero-resistance constant. This represents the zero-prevention constant for the water depth upstream of the gate. The zero-prevention constant for the gate opening and the gate opening use the same dimension or the same normalized scale, and the zero-prevention constant for the water depth upstream of the gate and the water depth upstream of the gate use the same dimension. Both are only used for zero-prevention treatment of the corresponding denominators.

[0045] The expected change factor is used to express the reference change that the downstream flow should show according to the conventional gate outflow relationship when adjacent changes occur in gate opening and upstream water depth. The expected change factor is not the downstream flow itself, nor is it a control quantity for future dispatch commands; it is only one of the processing results output by the hydrodynamic calculation unit 310 to the smart water conservancy monitoring platform 300, and is written into the same multimodal monitoring data packet.

[0046] The gate opening zero-prevention constant is determined by the minimum effective resolution of the gate opening feedback device 110, and its value is not less than the minimum effective resolution and not more than ten times the minimum effective resolution. When the gate opening is stored as a percentage or opening level, the gate opening zero-prevention constant is configured according to the same percentage or the same opening level scale. The upstream water depth zero-prevention constant is determined by the minimum effective resolution of the upstream water depth monitoring point 100, and its value is not less than the minimum effective resolution and not more than ten times the minimum effective resolution. The downstream flow zero-prevention constant is determined by the minimum effective resolution of the downstream flow monitoring point 120, and its value is not less than the minimum effective resolution and not more than ten times the minimum effective resolution. The closure zero-prevention constant is a dimensionless positive number, and its value is not less than the minimum display accuracy of the multiple data stored by the smart water conservancy supervision platform 300 and not more than ten times the minimum display accuracy.

[0047] When the upstream water depth monitoring point 100, the gate opening feedback device 110, or the downstream flow monitoring point 120 is replaced, or when the smart water conservancy supervision platform 300 changes the storage unit of the upstream water depth, the gate opening, or the downstream flow, the smart water conservancy supervision platform 300 rereads the minimum effective resolution of the corresponding device and synchronously updates the gate opening zero constant, the upstream water depth zero constant, the downstream flow zero constant, or the closure zero constant.

[0048] The hydrodynamic calculation unit 310 also generates an actual change factor based on the downstream flow rate variation relationship. The actual change factor is calculated from adjacent downstream flow rate acquisition values ​​and is used to reflect the flow response already observed at the downstream flow measurement section. The actual change factor only uses the downstream flow rate.

[0049] ; in, Indicates the actual change factor. This indicates the downstream flow rate at the current data collection moment. This indicates the downstream flow rate at the previous data collection time. This represents the zero-constant for downstream flow. The zero-constant for downstream flow and the downstream flow are measured in the same dimension or using the same normalized scale. The actual and expected changes are derived from different aspects of the directional hydrodynamic transmission chain: the expected change is derived from the upstream water depth and gate opening, while the actual change is derived from the downstream flow.

[0050] The water potential calculation unit 310 forms a transmission closure degree based on the degree of difference between the expected change multiple and the actual change multiple. The transmission closure degree is used to indicate the degree of consistency between the expected change multiple and the actual change multiple. The closer the transmission closure degree is to 1, the more consistent the expected changes reflected by the upstream water depth and gate opening are with the actual changes in the downstream flow; the smaller the transmission closure degree, the less consistent the three-point transmission relationship is.

[0051] ; in, Indicates transitivity closure. Indicates the actual change factor. Indicates the expected change factor. This represents the zero-constant for closure. The zero-constant for closure is a dimensionless positive number and is in the same dimensionless calculation relationship as the actual change factor and the expected change factor. The transfer closure is used to characterize the consistency of the three-point transfer relationship between the upstream water depth, the gate opening, and the downstream flow at adjacent sampling times.

[0052] The Smart Water Conservancy Supervision Platform 300 writes the expected change multiple, the actual change multiple, and the transmission closure degree into the same multimodal supervision data packet, and uses the transmission closure degree as the event review field, evidence field, and task assignment field in the supervision event record. The event review field is used for rule review and supervision level determination, the evidence field is used for evidence package generation, and the task assignment field is used to determine the primary verification object in the water conservancy supervision task.

[0053] When any data collection point is missing, the quality marker is abnormal, or adjacent data collection times are discontinuous, the smart water conservancy monitoring platform 300 does not require the water potential calculation unit 310 to calculate the current cycle's transmission closure. Instead, it generates data integrity verification items. After the valid data is restored, the smart water conservancy monitoring platform 300 then enters the normal calculation process.

[0054] When a gate is under maintenance, fully closed, undergoing flood discharge mode switching, or when the downstream flow measurement section is affected by construction, management personnel can register this special operating status in the existing business system. The Smart Water Conservancy Supervision Platform 300 can retain three data points and calculation results within the validity period of this status, but will mark it as a special operating condition and submit it for manual review to determine whether to generate a water conservancy supervision task.

[0055] Example 6: After generating a multimodal monitoring data package and obtaining its transmission closure, the intelligent water conservancy monitoring platform 300 loads water conservancy knowledge data and organizes the input for the AI ​​multimodal large model service 320. The water conservancy knowledge data may include operating procedures, inspection points, hazard categories, alarm classification rules, historical handling cases, and commonly used Q&A templates corresponding to the current monitored objects. The water conservancy knowledge data is managed using version numbers and effective dates to prevent obsolete entries from continuing to participate in current business judgments.

[0056] The input provided by the intelligent water conservancy supervision platform 300 to the AI ​​multimodal large model service 320 includes multimodal supervision data packages, transmission closure, and water conservancy knowledge data. Figure 1 The input for the gate station shown can include the water depth in front of the gate, the gate opening, the flow rate after the gate, video keyframes, equipment logs, inspection texts, transmission closure information, and water conservancy knowledge related to the operation of the gate station.

[0057] The AI ​​Multimodal Large Model Service 320 reads images and videos to determine the status of water surfaces, embankments, gates, floating objects, and channels. For example, the AI ​​Multimodal Large Model Service 320 can identify whether gates are in positions consistent with their opening degree, whether there is significant obstruction in the channel downstream of the gate, whether the water surface shows inconsistencies with downstream flow changes, and whether there are any conditions on the embankment or on-site work area that require verification. All of the above identifications rely on images and videos as evidence.

[0058] The AI ​​Multimodal Large Model Service 320 reads sensor data on upstream water depth, gate opening, and downstream flow rate, and combines this with speech-to-text and inspection text to form a unified semantic system. This unified semantic system is used to combine numerical changes, on-site images, on-site descriptions, and equipment logs into explanations for water conservancy supervision. The AI ​​Multimodal Large Model Service 320 marks content that cannot be confirmed as requiring manual verification, rather than directly recording it as established fact.

[0059] The AI ​​multimodal large model service 320 generates regulatory event records according to a fixed output format. This fixed output format is carried by the event object field of the smart water conservancy supervision platform 300 and is used for subsequent water conservancy supervision task assignment.

[0060] The fixed output format is implemented using the event object fields of the Smart Water Conservancy Supervision Platform 300. Required fields include object number, event type, location, discovery time, data evidence, and items requiring verification. The object number uses the object number from the supervised object file; the event type uses the name of the existing water conservancy business item; the location uses the monitoring point or site area name from the supervised object file; the discovery time uses the collection time range of the multimodal supervision data package; data evidence includes source identifiers for sensor values, images and videos, speech-to-text content, inspection text, and equipment logs; items requiring verification include the primary verification object and suggested verification content. Fields that the AI ​​multimodal large model service 320 cannot confirm are written as "Pending manual verification," and the corresponding source identifier is retained in the data evidence.

[0061] The generation order of the fixed output format is as follows: The Smart Water Conservancy Supervision Platform 300 first determines the object and time range to which the supervision event record belongs based on the object number and discovery time. Then, the AI ​​Multimodal Large Model Service 320 generates candidate event types based on sensor values, images and videos, equipment logs, inspection texts, and voice transcription content. Subsequently, the location of occurrence, data evidence, and matters requiring verification are written in. When the candidate event type is inconsistent with the name of existing water conservancy business matters, the event type is written as "Pending manual verification"; when the location of occurrence does not correspond to the monitoring point or on-site area name in the supervision object file, the location of occurrence is written as "Pending manual verification"; when the data evidence cannot be traced back to the source identifier, the supervision event record is not used as a regular handling matter.

[0062] Unified semantics are formed according to the order of evidence sources. The Smart Water Conservancy Supervision Platform 300 first reads the sensor values ​​with normal quality markings, then reads the equipment logs and video keyframes corresponding to the same time alignment window, and then reads the inspection text and speech-to-text content. If different evidence sources describe the same matter consistently, the AI ​​multimodal big data service 320 writes the consistent content into the supervision event record; if different evidence sources describe the same matter in conflict, the AI ​​multimodal big data service 320 does not output a definite fact, but instead writes "pending manual verification" in the corresponding field and retains the source identifier of the conflict in the data evidence.

[0063] After the intelligent water conservancy monitoring platform 300 generates a three-point inconsistency marker, it retrieves keyframes from the video, equipment logs, and inspection texts from the same time period. The AI ​​multimodal large model service 320 generates explanatory fields based on these data to explain the cause of the inconsistency marker. These explanatory fields can indicate whether the gate opening feedback matches the gate position identified in the video, whether there are signs of blockage in the downstream channel, and whether there are any obvious abnormalities at the upstream water depth monitoring point 100.

[0064] The explanation field must include at least the explanation object, evidence source, explanation content, confidence level, and manual verification mark. The explanation object corresponds to one of the following: three-point transmission inconsistency mark, gate actuator, downstream flow measurement section or channel, and upstream water depth monitoring point 100; the evidence source corresponds to video keyframes, equipment logs, or inspection text; the explanation content is described using fixed short sentences; the confidence level uses the dimensionless value output by AI multimodal large model service 320; when the confidence level is lower than the lower limit of the explanation field confidence level, the manual verification mark is written as "pending manual verification".

[0065] The set of fixed short phrases for the explanation content is maintained by the Smart Water Conservancy Supervision Platform 300 in the water conservancy knowledge database and is saved along with the version number and effective date of the water conservancy knowledge database. For Figure 1 For the gate station shown, the fixed short phrase set must cover at least the following: gate opening feedback and video recognition gate position are consistent; gate opening feedback and video recognition gate position are inconsistent; there are signs of blockage in the downstream channel; no signs of blockage in the downstream channel; abnormal behavior at upstream water depth monitoring point 100; no abnormal behavior at upstream water depth monitoring point 100; and insufficient evidence requiring manual verification. The AI ​​multimodal large model service 320 can only select explanation content from the fixed short phrase set and cannot write free descriptions not maintained in the fixed short phrase set into the explanation field.

[0066] The lower limit of the confidence level for the explanatory field is determined by the Smart Water Conservancy Supervision Platform 300 based on manually reviewed samples from the same AI Multimodal Large Model Service version 320. The number of manually reviewed samples must be no less than 30. The Smart Water Conservancy Supervision Platform 300 selects samples whose explanations are correctly interpreted through manual review and calculates the 5th percentile of the corresponding confidence level as the lower limit of the explanatory field's confidence level. If there are fewer than 30 manually reviewed samples, the Smart Water Conservancy Supervision Platform 300 uses the lower limit of the explanatory field's confidence level from the management unit's existing model application rules. If no existing model application rules exist, the manual verification mark for the explanatory field is written as "Pending Manual Verification." When the AI ​​Multimodal Large Model Service version 320 is changed, the fixed phrase set is altered, or the version of water conservancy knowledge data is changed, the Smart Water Conservancy Supervision Platform 300 re-determines the lower limit of the explanatory field's confidence level.

[0067] The Smart Water Conservancy Supervision Platform 300 uses explanatory fields to determine the supervision level or generate evidence packages. These explanatory fields are not used alone as the final decision, but rather participate in the review process along with supervision event records, transmission closure, inconsistency markers in the three-point transmission, and data integrity. Therefore, the output of the AI ​​multimodal large model service 320 is incorporated into the traceable water conservancy supervision business process.

[0068] Example 7: After obtaining the transmission closure, the Smart Water Conservancy Supervision Platform 300 compares the recorded supervision events with fixed thresholds, equipment alarms, inspection rules, and current task status in the water conservancy business system. For gate station events with generated transmission closure, the Smart Water Conservancy Supervision Platform 300 also checks whether the transmission closure is below the lower limit. The lower limit is set by the management unit in existing business rules to determine whether a three-point transmission relationship verification is required.

[0069] The preset individual thresholds are read by the Smart Water Conservancy Supervision Platform 300 from the existing water conservancy business rules associated with the supervised object files, and correspond to the upstream water depth, gate opening, and downstream flow respectively. Each preset individual threshold must at least store the threshold value, applicable supervised object, applicable monitoring point, effective time, and maintenance personnel identification. If no preset individual threshold exists in the current supervised object file, the Smart Water Conservancy Supervision Platform 300 will not generate a three-way inconsistency flag, but will instead generate a data integrity verification item.

[0070] The lower limit of closure is configured by the Smart Water Conservancy Supervision Platform 300 based on historical normal operation samples of the same supervised object or existing business rules of the management unit. When using historical normal operation samples, the Smart Water Conservancy Supervision Platform 300 selects a continuous collection period where manual verification is normal, calculates the 5th percentile of the corresponding transmission closure, and uses this 5th percentile or a higher value confirmed by the management unit as the lower limit of closure. When using existing business rules, the lower limit of closure is stored as a dimensionless value between 0 and 1. The lower limit of closure must at least store the applicable supervised object, effective time, sample source or rule source, and maintenance personnel identifier.

[0071] When using historical normal operation samples to determine the lower limit of closure, the historical normal operation samples are taken from the most recent calendar month of a valid data collection period that has been manually verified as normal for the same regulated object. If there are fewer than 30 valid data collection periods in the most recent calendar month, the period is extended backward to the most recent twelve months until there are 30 valid data collection periods. If there are still fewer than 30 valid data collection periods, the Smart Water Conservancy Supervision Platform 300 does not use historical normal operation samples to determine the lower limit of closure, but instead uses the lower limit of closure in the existing business rules of the management unit. If the existing business rules also do not exist, the Smart Water Conservancy Supervision Platform 300 does not generate a three-point transmission inconsistency flag, but instead generates a data integrity verification item.

[0072] The rules for excluding historical normal operation samples are as follows: samples with missing upstream water depth, gate opening, or downstream flow in any collection period are not included in the calculation; samples with any abnormal essential quality marker are not included in the calculation; samples registered as being under maintenance, fully closed, switching of flood discharge mode, affected by construction of downstream flow measurement sections, or abnormally verified by manual review are not included in the calculation. After the exclusion is completed, the Smart Water Conservancy Supervision Platform 300 sorts the samples in ascending order according to the transmission closure degree corresponding to the remaining valid collection periods, and takes the 5th percentile as the lower limit of closure degree; higher values ​​confirmed by the management unit can only raise the lower limit of closure degree, but cannot be lower than the 5th percentile. The number of samples, sample time range, exclusion records, and confirmation records for the lower limit of closure degree are saved together with the lower limit of closure degree.

[0073] Two consecutive data collection cycles refer to two consecutive transmission closure calculation results formed by three adjacent valid data collection moments for the same monitored object. If the upstream water depth, gate opening, or downstream flow is missing in any data collection cycle, or if the quality marker is abnormal, the smart water conservancy monitoring platform will interrupt continuous counting and generate a data integrity verification item. This rule is used to avoid communication interruptions or single data collection anomalies causing inconsistencies in the three-point transmission markers to be falsely triggered.

[0074] When the upstream water depth, gate opening, and downstream flow rate all do not exceed their corresponding preset individual thresholds, and the transmission closure is below the lower limit for two consecutive data collection cycles, the Smart Water Conservancy Supervision Platform 300 generates a three-point transmission inconsistency flag. The two-consecutive-cycle limit is used to reduce the impact of single-data collection errors or short-term communication jitter on regulatory judgment. The preset individual thresholds are still used for single-point boundary judgment according to existing business rules, while the three-point transmission inconsistency flag is used to detect situations where individual data does not exceed the boundaries but the three-point transmission relationship is inconsistent.

[0075] When the Smart Water Conservancy Supervision Platform 300 generates a three-point transmission inconsistency marker, it also saves the closure degree of the two consecutive transmissions that triggered the marker, the lower limit of the closure degree, the corresponding current acquisition time, the previous acquisition time, the water depth in front of the gate, the gate opening degree, and the flow rate after the gate. The above-mentioned saved information is used for subsequent supervision event recording, evidence packages, and water conservancy supervision task review.

[0076] The Smart Water Conservancy Supervision Platform 300 determines the supervision level by combining supervision event records, transmission closure, inconsistency markers in the three-point transmission, and data integrity. If the data evidence is insufficient or there are abnormal quality markers, a verification item can be generated; if the evidence is complete and meets the general abnormality conditions, a normal handling item can be generated; if the predetermined high-level conditions are met, an emergency handling item can be generated. The specific level name and processing time limit of the supervision level can follow the existing water conservancy supervision business rules.

[0077] The criteria for determining verification-related matters include missing values ​​from any required sensor, abnormal quality markers, images and videos failing clarity checks, explanation fields being marked as "pending manual verification," or missing required fields in the regulatory event record. The criteria for determining ordinary handling matters include normal data integrity, transmission closure being below the lower limit for two consecutive acquisition cycles, and the three-point transmission inconsistency marker being generated without any established high-level conditions. The criteria for determining emergency handling matters follow the existing high-level business rules of the management unit; when existing high-level business rules are triggered, the Smart Water Conservancy Supervision Platform 300 retains the triggering rules, transmission closure, and original evidence. The above-mentioned regulatory level determination rules are used for business review and do not replace on-site protection devices, dispatch orders, and manual emergency command.

[0078] When generating a three-point inconsistency marker, the Smart Water Conservancy Supervision Platform 300 determines the primary verification target based on the order of changes in upstream water depth, gate opening, and downstream flow. This primary verification target helps to pinpoint water conservancy supervision tasks to more specific on-site objects, reducing redundant checks caused by general gate inspections. The determination of the primary verification target only serves the purpose of supervision task allocation and does not directly control the gates.

[0079] Before identifying the primary targets for verification, the Smart Water Conservancy Supervision Platform 300 first calculates the variation ranges of the upstream water depth, gate opening, and downstream flow. The upstream water depth variation range is the absolute value of the difference between the upstream water depth at the current data collection time and the upstream water depth at the previous data collection time; the gate opening variation range is the absolute value of the difference between the gate opening at the current data collection time and the gate opening at the previous data collection time; and the downstream flow variation range is the absolute value of the difference between the downstream flow at the current data collection time and the downstream flow at the previous data collection time.

[0080] The discrepancy between the gate opening degree and the gate position identified by video recognition refers to a situation where the difference between the gate opening degree output by the gate opening degree feedback device 110 (converted to the on-site gate position) and the gate position identified by the AI ​​multimodal large model service 320 from keyframes of the video within the same time period exceeds the gate opening degree inconsistency threshold, and the video recognition confidence level is not lower than the lower limit of the gate position recognition confidence level. The gate opening degree inconsistency threshold is configured by the management unit according to the allowable error of the gate opening degree feedback device 110, and the lower limit of the gate position recognition confidence level is configured by the smart water conservancy supervision platform 300 according to the existing model application rules.

[0081] No corresponding change in downstream flow rate means that after the gate opening change reaches the effective gate opening change threshold, the downstream flow rate change is less than the downstream flow rate response threshold within the same acquisition cycle or the next effective acquisition cycle, and the transmission closure is lower than the lower limit of closure. The effective gate opening change threshold is determined by the minimum effective resolution of the gate opening feedback device 110 or existing business rules, and the downstream flow rate response threshold is determined by the minimum effective resolution of the downstream flow rate monitoring point 120 or existing business rules.

[0082] Abnormal changes in upstream water depth while stable gate opening and downstream flow rate refer to situations where the upstream water depth change reaches the abnormal upstream water depth change threshold, while the gate opening change is less than the stable gate opening threshold and the downstream flow rate change is less than the stable downstream flow rate threshold. The abnormal upstream water depth change threshold is determined by the minimum effective resolution of the upstream water depth monitoring point 100 or existing operational rules; the stable gate opening threshold is determined by the minimum effective resolution of the gate opening feedback device 110 or existing operational rules; and the stable downstream flow rate threshold is determined by the minimum effective resolution of the downstream flow rate monitoring point 120 or existing operational rules.

[0083] If the gate opening degree output by the gate opening degree feedback device 110 is inconsistent with the gate position identified by the video recognition, the smart water conservancy monitoring platform 300 will prioritize the gate actuator as the primary object of verification. This situation indicates that there is an inconsistency between the gate opening degree feedback device 110, the gate actuator, or the video recognition result, requiring on-site personnel to verify the gate actuator, the gate position feedback status, and the corresponding equipment logs.

[0084] If the downstream flow does not change accordingly after the gate opening changes, the Smart Water Conservancy Monitoring Platform 300 will prioritize checking the downstream flow measurement section or channel. This situation indicates that the gate opening has changed, but the downstream flow has not responded reasonably, and it may be necessary to check the downstream flow monitoring point 120, the downstream flow measurement section, the channel status, and the presence of floating objects or blockages.

[0085] If the water depth upstream of the sluice gate changes abnormally while the gate opening and downstream flow remain stable, the intelligent water conservancy monitoring platform 300 will prioritize the upstream water depth monitoring point 100 as the primary target for verification. This situation indicates that the change in upstream water depth is inconsistent with the stable state of the gate opening and downstream flow, and it may be necessary to verify the upstream water depth monitoring point 100, the local state of the upstream water body, and the quality markings of the corresponding data acquisition equipment.

[0086] After identifying the primary verification target, the Smart Water Conservancy Supervision Platform generates and dispatches water conservancy supervision tasks. These tasks may include on-site verification, remote confirmation, equipment inspection, on-site handling, retesting, or data supplementation. Each task may include the monitored object, event type, supervision level, discovery time, location, recommended verification content, processing timeframe, primary verification target, and access to the evidence package. These fields are standard water conservancy supervision work order content, used to support on-site personnel in performing verification.

[0087] Example 8: The Smart Water Conservancy Supervision Platform 300 generates regulatory alerts for regulatory events that require processing. These alerts include the monitored object, event category, regulatory level, discovery time, location, suggested verification content, and processing timeframe. The generation of regulatory alerts is based on regulatory event records, transmission closure, three-point transmission inconsistency markers, explanatory fields, fixed rule hit results, and data integrity, relying not only on the text output of the AI ​​multimodal large model service 320.

[0088] The Smart Water Conservancy Supervision Platform 300 compiles time-series data of sensor values, images and videos, voice-transcribed content, inspection texts, and supervision event records corresponding to supervision alarms into evidence packages. These evidence packages can be associated with unique event numbers and store data sources, collection times, processing records, and model versions. The purpose of these evidence packages is to allow on-duty personnel and on-site personnel to simultaneously view event summaries and original evidence, avoiding reliance solely on model-generated text for action.

[0089] for Figure 1For the gate station shown, when a three-point inconsistency marker has been generated, the evidence package should at least include the upstream water depth, gate opening, and downstream flow rate at the current and previous acquisition times; the expected change factor, actual change factor, and transmission closure degree generated by the water potential calculation unit 310; the regulatory event record generated by the AI ​​multimodal large model service 320; and video keyframes, equipment logs, and inspection texts from the same time period. All of the above content collectively supports the regulatory level and the primary verification target.

[0090] The management and mobile terminal 400 receives water conservancy supervision tasks dispatched by the smart water conservancy supervision platform 300. Task reception can be achieved through platform to-do lists, mobile terminal push notifications, application messages, or existing notification methods. The smart water conservancy supervision platform 300 can record delivery and receipt statuses. This notification method is a standard function of the work order system.

[0091] On-site personnel submit handling feedback data via management and mobile terminals 400. The handling feedback data can include handling text, on-site photos, remeasured values, equipment status, and completion time. The handling text describes the on-site verification process and results; on-site photos supplement visual evidence; remeasured values ​​describe changes in upstream water depth, gate opening, or downstream flow rate after the handling; equipment status describes the status of the gate opening feedback device 110, downstream flow monitoring point 120, or upstream water depth monitoring point 100; and the completion time is used for verification and cancellation.

[0092] The Smart Water Conservancy Supervision Platform 300 adds the feedback data to the original monitoring event record and performs a review and closure. For matters requiring remote review, on-duty personnel review the before-and-after data and on-site evidence to confirm whether the anomaly has been eliminated. Upon successful review, the Smart Water Conservancy Supervision Platform 300 updates the task status to "completed" and records the closure person and closure time; if the review fails, a supplementary task is reassigned. Even after emergency matters are completed, subsequent observation records can still be retained according to existing business procedures.

[0093] The intelligent water conservancy supervision platform archives the entire supervision process (300 records) to create a searchable supervision ledger. Archived content can include raw data, multimodal supervision data packages, model input summaries, supervision event records, rule hit results, supervision alarms, water conservancy supervision tasks, handling feedback data, and review conclusions. Statistical results serve water conservancy industry supervision operations, used to schedule routine inspections, equipment maintenance, and personnel on-duty arrangements.

[0094] As an optional implementation, the upstream water depth, gate opening, and downstream flow rate are all within their respective preset individual thresholds, and the downstream flow rate changes in accordance with the expected change factor after the gate opening changes. The difference between the expected change factor and the actual change factor generated by the hydrodynamic calculation unit 310 is small, and the transmission closure degree does not fall below the lower limit of closure degree for two consecutive acquisition cycles. The smart water conservancy supervision platform 300 can continue to archive multimodal supervision data packets, without generating a three-point transmission inconsistency marker, and without assigning water conservancy supervision tasks based on the transmission closure degree.

[0095] As an optional implementation, the upstream water depth, gate opening, and downstream flow rate do not exceed their respective preset individual thresholds, but the downstream flow rate does not change after the gate opening changes within adjacent data collection times. The smart water conservancy monitoring platform 300 generates a three-point transmission inconsistency marker according to the aforementioned rules and prioritizes the downstream flow measurement section or channel as the primary verification object.

[0096] As an optional implementation, the gate position identified by the AI ​​multimodal large model service 320 from the video keyframes is inconsistent with the gate opening output by the gate opening feedback device 110. The smart water conservancy supervision platform 300 writes this inconsistency into the supervision event record and explanation field, and points it to the gate actuator in the water conservancy supervision task; on-site personnel submit the gate actuator inspection results, on-site photos and equipment status through the management and mobile terminal 400.

[0097] As an optional implementation, abnormal changes in water depth upstream of the sluice gate occur while the gate opening and downstream flow remain stable. The Smart Water Conservancy Monitoring Platform 300 displays time-series data on water depth upstream of the sluice gate, time-series data on gate opening, time-series data on downstream flow, and images and videos from the same time period in the evidence package. If the interpretation field of the AI ​​multimodal large model service 320 indicates that the water body image upstream of the sluice gate is inconsistent with the changes in water depth upstream of the sluice gate, the Smart Water Conservancy Monitoring Platform 300 prioritizes the water depth monitoring point 100 upstream of the sluice gate as the primary object of verification.

[0098] As an optional implementation, when there are missing data on the water depth in front of the gate, the gate opening, or the flow rate after the gate, abnormal quality markers, or discontinuous data collection times, the smart water conservancy supervision platform 300 generates data integrity verification items according to the aforementioned data integrity rules, and can distribute them to the management and mobile terminal 400 for verification of equipment communication, sensor status, or data interface.

[0099] It should be noted that this implementation method can be deployed within an existing smart water conservancy monitoring platform. The data access gateway 200 can be implemented by a gateway device or server software with protocol conversion, source registration, and time verification capabilities. The smart water conservancy monitoring platform 300 can consist of a server, database, business applications, and a monitoring interface. The water potential calculation unit 310 can be implemented by a software module within the smart water conservancy monitoring platform 300, or by an independent data processing service. The AI ​​multimodal large model service 320 can be implemented by an already deployed model service interface. The management and mobile terminal 400 can be implemented through a management terminal page, a mobile inspection terminal, or a mobile application.

[0100] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A smart water conservancy intelligent supervision method based on an AI multimodal large model, executed by a water conservancy supervision system, which includes a data access gateway, a water potential calculation unit, an AI multimodal large model service, a smart water conservancy supervision platform, and management and mobile terminals, characterized in that... include: The data access gateway accesses the multimodal water conservancy monitoring data, and the smart water conservancy monitoring platform performs data cleaning, time alignment and spatial correlation on the multimodal water conservancy monitoring data to generate multimodal monitoring data packets; When a gate station is equipped with a water depth monitoring point upstream of the gate, a gate opening feedback device, and a flow monitoring point downstream of the gate, the water potential calculation unit establishes a directional water potential transmission chain between the water body upstream of the gate, the gate, and the flow measurement section downstream of the gate. Based on the gate opening change relationship and the water depth change relationship at adjacent acquisition times, it forms the expected change multiple, and based on the flow change relationship downstream of the gate, it forms the actual change multiple. The degree of transmission closure is formed according to the difference between the expected change multiple and the actual change multiple. The AI ​​multimodal large model service generates regulatory event records based on multimodal regulatory data packages, transmission closure, and water conservancy knowledge data; When the water depth in front of the gate, the gate opening, and the flow rate after the gate all do not exceed the corresponding preset individual thresholds, and the transmission closure is lower than the lower limit of closure for two consecutive collection cycles, the smart water conservancy supervision platform generates a three-point transmission inconsistency marker. Combining the supervision event record, transmission closure, three-point transmission inconsistency marker, and data integrity, the supervision level is determined. Based on the order of changes in water depth in front of the gate, gate opening, and flow rate after the gate, the primary verification object is determined, and water conservancy supervision tasks are generated and assigned.

2. The intelligent water conservancy supervision method based on AI multimodal large model according to claim 1, characterized in that, The smart water conservancy supervision platform establishes files for supervised objects, which include object number, geographical location, monitoring point location, and task receiving terminal identifier. The files are used to determine the attribution of water conservancy multimodal supervision data and the recipients of electronic water conservancy supervision tasks.

3. The intelligent water conservancy supervision method based on an AI multimodal large model according to claim 1, characterized in that, The multimodal monitoring data for water conservancy includes sensor values, images and videos, voice-transcribed content, inspection texts, and equipment logs; sensor values ​​include water depth in front of the gate, gate opening, and flow rate behind the gate; the smart water conservancy monitoring platform retains the collection time, location number, and quality mark for sensor values, the shooting time and equipment number for images and videos, and the source for voice-transcribed content, inspection texts, and equipment logs.

4. The intelligent water conservancy supervision method based on AI multimodal large model according to claim 1, characterized in that, When the smart water conservancy supervision platform cleans data, it performs format verification, missing mark removal, duplicate record removal and unit unification on sensor values, clarity checks, key frame extraction and timestamp verification on images and videos, and field segmentation, invalid character cleanup and common terminology unification on speech transcription content and inspection text. Data that fails the verification enters the abnormal data area and does not participate in this round of intelligent analysis.

5. The intelligent water conservancy supervision method based on an AI multimodal large model according to claim 1, characterized in that, The intelligent water conservancy supervision platform writes the expected change multiple, the actual change multiple, and the transmission closure into the same multimodal supervision data packet, and uses the transmission closure as the event review field, evidence field, and task assignment field of the supervision event record. The transmission closure is used for supervision event discovery, evidence organization, and task assignment, and does not send control commands to the gate controller.

6. The intelligent water conservancy supervision method based on AI multimodal large model according to claim 1, characterized in that, The AI ​​multimodal large model service reads the water surface, embankment, gate, floating objects and channel status from images and videos, reads the water depth in front of the gate, gate opening and flow rate behind the gate from sensor values, and combines the speech transcription content and inspection text to form a unified semantics. It generates a regulatory event record according to a fixed output format, including object number, event type, location of occurrence, discovery time, data evidence and matters that need to be verified.

7. The intelligent water conservancy supervision method based on an AI multimodal large model according to claim 1, characterized in that, After the smart water conservancy supervision platform generates a three-point inconsistency marker, it retrieves video keyframes, equipment logs, and inspection texts from the same time period. The AI ​​multimodal big data service generates explanatory fields based on the video keyframes, equipment logs, and inspection texts to explain the reasons for the three-point inconsistency marker. The smart water conservancy supervision platform uses these explanatory fields to determine the supervision level or generate evidence packages.

8. The intelligent water conservancy supervision method based on AI multimodal large model according to claim 1, characterized in that, When the smart water conservancy monitoring platform generates a three-point transmission inconsistency marker, if the gate opening output by the gate opening feedback device is inconsistent with the gate position identified by the video, the gate actuator will be the primary object of verification. If the flow rate after the gate opening changes does not change accordingly, the flow measurement section or channel after the gate will be the primary object of verification. If the water depth before the gate changes abnormally but the gate opening and the flow rate after the gate are stable, the water depth monitoring point before the gate will be the primary object of verification.

9. The intelligent water conservancy supervision method based on AI multimodal large model according to claim 1, characterized in that, The smart water conservancy supervision platform generates supervision alarms for the records of supervision events that need to be processed, and compiles the time-series data of sensor values, images and videos, voice transcription content, inspection text and supervision event records corresponding to the supervision alarms into an evidence package; The management and mobile terminals receive water conservancy supervision tasks dispatched by the smart water conservancy supervision platform and submit handling feedback data to the smart water conservancy supervision platform. The smart water conservancy supervision platform adds the handling feedback data to the original supervision event record and performs review and cancellation.

10. A smart water conservancy intelligent monitoring system based on an AI multimodal large model, characterized in that: include: The data access gateway is used to provide multimodal water conservancy monitoring data to the smart water conservancy monitoring platform; The intelligent water conservancy supervision platform is used to perform data cleaning, time alignment, and spatial correlation on multimodal water conservancy supervision data and generate multimodal supervision data packages. The water potential calculation unit is used to receive the upstream water depth, gate opening and downstream flow at adjacent collection times provided by the smart water conservancy supervision platform, establish a directed water potential transmission chain, form the expected change multiple based on the gate opening change relationship and the upstream water depth change relationship, form the actual change multiple based on the downstream flow change relationship, and form the transmission closure degree according to the degree of difference between the expected change multiple and the actual change multiple. AI multimodal large model service is used to generate regulatory event records based on multimodal regulatory data packets, transmission closure and water conservancy knowledge data; The intelligent water conservancy supervision platform is also used to generate a three-point transmission inconsistency mark when the water depth in front of the gate, the gate opening and the flow rate after the gate have not exceeded the corresponding preset individual thresholds and the transmission closure is lower than the lower limit of closure for two consecutive collection cycles. The supervision level is determined by combining the supervision event record, transmission closure, three-point transmission inconsistency mark and data integrity. The primary verification object is determined according to the change order of the water depth in front of the gate, the gate opening and the flow rate after the gate and water conservancy supervision tasks are assigned. Management and mobile terminals are used to receive water conservancy supervision tasks dispatched by the smart water conservancy supervision platform and submit handling feedback data to the smart water conservancy supervision platform. The smart water conservancy supervision platform is also used to perform review and cancellation based on the disposal feedback data.

Images