Intelligent water environment anomaly inspection method based on unmanned ship and deep learning

Abstract

The application discloses an intelligent water environment anomaly inspection method based on an unmanned ship and deep learning, and relates to the technical field of water environment monitoring.The application determines an action side sampling position and a control side sampling position on both sides of a suspected emission point according to a current water flow direction, and forms a paired data unit in the order of the control side, the action side and the control side in a same flow state period.When the water flow direction is reversed, the action side sampling position and the control side sampling position are determined again according to the reversed water flow direction, and the paired water quality feature sequences before and after the reversal are associated with corresponding water flow state data to form a flow direction rearranged feature sequence.The water quality data entering an emission point contribution determination model maintains the relationship between the action side and the control side corresponding to the actual water flow direction during sampling, so that the short-time emission influence of the suspected emission point can be determined according to water quality change data having a corresponding relationship.

Description

Technical Field

[0001] This invention relates to the field of water environment monitoring technology, and in particular to an intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning. Background Technology

[0002] Rainwater pumping stations and riverside outfalls are often set up along urban tributaries in tidal river sections. During the drainage phase after rainfall, the pollution caused by the discharge from pumping stations or short-term discharges from outfalls is affected by tidal hydrodynamics and can migrate along the river channel with changes in water flow direction, resulting in different impact relationships on the water bodies on both sides of the same outfall at different sampling times.

[0003] Existing methods for detecting aquatic anomalies can employ unmanned surface vessels (USVs) equipped with water quality monitoring equipment to collect data on parameters such as ammonia nitrogen, chemical oxygen demand (COD), conductivity, turbidity, and dissolved oxygen. This data can then be combined with data analysis models or deep learning models to identify abnormal areas. Some methods can also perform follow-up checks on the vicinity of suspected discharge points after anomalies are detected, taking into account water flow direction.

[0004] However, in tidal tributaries where short-term discharges following rainfall coexist with tidal changes, the flow direction may reverse during anomaly verification. In such cases, the anomalous water quality data obtained sequentially and the background water quality data may correspond to different flow directions and affected locations, resulting in a lack of effective correlation between the sampled data. Consequently, water quality changes caused by localized discharges are easily confused with fluctuations in river water quality, and anomalous results are difficult to correlate with short-term discharge processes at specific suspected discharge points.

[0005] Therefore, a key technical problem to be solved is how to generate water quality sampling data with corresponding relationships to changes in water flow direction during the investigation of short-term abnormal discharges in tidal tributaries, and how to determine the abnormal contribution of suspected discharge points based on this data. Summary of the Invention

[0006] This application provides an intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning, which solves the problem that it is difficult to accurately verify short-term discharge anomalies when the flow direction of tidal tributaries reverses.

[0007] This invention provides an intelligent inspection method for water environment anomalies based on unmanned surface vessels (USVs) and deep learning. An USV equipped with a water quality parameter acquisition unit, a positioning unit, and a water flow status acquisition unit performs short-term post-rainfall discharge anomaly verification on at least one suspected discharge point among pump station outlets and riverside outfalls in tidal urban tributary waters. The method includes:

[0008] Obtain the identification information of the suspected discharge point and water flow status data, including the direction of water flow;

[0009] Based on the water flow direction, determine the sampling location on the downstream side of the suspected discharge point and the control side sampling location on the upstream side;

[0010] During the same flow period in which the water flow direction remains consistent, the unmanned vessel is controlled to collect water quality parameter data in the order of the control side sampling position, the action side sampling position, and the control side sampling position, with the same water quality parameter category. The action side water quality data is associated with the control side water quality data obtained before and after it as paired data units, and paired water quality feature sequences are formed according to the sampling time.

[0011] The water flow direction is determined based on the continuously acquired water flow state data. When the water flow direction is reversed, the sampling positions on the action side and the control side are re-determined according to the reversed water flow direction. The paired water quality characteristic sequences before and after the reversal are associated with the corresponding water flow state data to form a flow direction rearrangement characteristic sequence.

[0012] Based on the judgment result, the paired water quality feature sequence or the flow direction rearrangement feature sequence is input into the discharge point contribution judgment model to obtain the discharge point contribution judgment result. The discharge point contribution judgment model is a deep learning model.

[0013] The identification information, the contribution determination result of the emission point, and the water quality parameter data and water flow state data that form the result are associated and stored as an emission anomaly verification record.

[0014] In some embodiments, the water flow state data includes water flow direction and flow velocity;

[0015] When the flow velocity reaches the effective flow velocity threshold and the continuously obtained water flow direction remains in the same river channel axis, the same flow state period is determined.

[0016] When the flow velocity reaches the effective flow velocity threshold and the continuously obtained flow direction turns to the opposite river channel axis, it is determined that a flow direction reversal has occurred.

[0017] In some embodiments, the identification information includes the coordinates of the suspected emission point location;

[0018] Based on the location coordinates and the current water flow direction, the sampling positions on the action side and the control side are configured in the projection direction of the river centerline, and the sampling positions on the action side and the control side are respectively separated from the suspected discharge point by a preset distance.

[0019] In some embodiments, when a reversal of the water flow direction is detected during the formation of a pairing data unit, the incomplete pairing data unit is marked as a non-determined data unit.

[0020] New paired data units are collected based on the reversed water flow direction, and the flow direction rearrangement feature sequence is formed according to the sampling time and the corresponding water flow state data.

[0021] In some embodiments, the water quality parameter categories include at least one of ammonia nitrogen characterization parameters or chemical oxygen demand characterization parameters, and at least two of conductivity, turbidity, and dissolved oxygen;

[0022] The background data for the control side is formed from the water quality data of the two consecutive control sides in the paired data unit;

[0023] The paired water quality feature sequence includes corresponding change data of the water quality data on the action side relative to the background data on the control side, sampling time sequence data, and water flow state data.

[0024] In some embodiments, the emission point contribution determination model includes a pairing relationship feature extraction branch, a flow direction change feature extraction branch, and a determination output layer;

[0025] The sampling role status data characterizes the correspondence between the sampling position on the action side, the sampling position on the control side, and the direction of water flow, and whether the correspondence has changed;

[0026] The pairing relationship feature extraction branch extracts the associated change features of the corresponding change data. The flow direction change feature extraction branch extracts the time-series change features based on one of the paired water quality feature sequences and the flow direction rearrangement feature sequences, as well as the associated sampling role status data. The judgment output layer outputs one of the following: discharge point contribution anomaly results, river segment common change results, and results to be verified.

[0027] In some embodiments, training samples include model input fields, category label fields, and traceability fields; the model input fields include one of paired water quality feature sequences and flow direction rearrangement feature sequences, as well as associated water flow state data and sampling role state data;

[0028] The category labeling field includes category labels corresponding to emission point contribution anomaly results, river section co-change results, and results pending verification, respectively; the tracing field includes sample identifier, suspected emission point identifier, sampling location, sampling time, and data quality label;

[0029] The category labels are formed based on on-site sampling records, verification sampling results, and emission operation records.

[0030] In some embodiments, during a period of unidirectional flow, the unmanned vessel is controlled to collect corroborating water quality data downstream of the current flow direction, and the interval between the sampling location of the corroborating water quality data and the suspected discharge point is greater than the interval between the sampling location on the action side and the suspected discharge point.

[0031] The corresponding changes in the water quality data of the action side and the supporting water quality data relative to the background data of the control side are used to form a three-sided water quality feature sequence;

[0032] The discharge point contribution determination model is input with one of the three-sided water quality characteristic sequences, the paired water quality characteristic sequences, and the flow direction rearrangement characteristic sequences, along with the water flow state data and sampling role state data associated with the above sequences, to obtain the discharge point contribution determination result.

[0033] The model input field of the training samples includes one of the three-sided water quality feature sequences, the paired water quality feature sequences, and the flow direction rearrangement feature sequences, and includes water flow state data and sampling role state data associated with the above sequences. The category label field includes a category label corresponding to the contribution determination result of the discharge point.

[0034] In some embodiments, when outputting abnormal results or results to be reviewed for emission point contribution, a review sampling task is generated, and the emission point contribution determination result that triggers the review sampling task is recorded as the first determination result.

[0035] When performing the verification sampling task, a new paired water quality feature sequence is formed based on the current water flow direction; when the current water flow direction is opposite to the water flow direction corresponding to the first judgment result, the paired water quality feature sequence of the first judgment result and the new paired water quality feature sequence are associated with their respective water flow state data to form a flow direction rearrangement feature sequence;

[0036] The new paired water quality characteristic sequence or the flow direction rearrangement characteristic sequence is input into the discharge point contribution determination model to obtain the verification determination result.

[0037] In some embodiments, the emission anomaly verification record includes an event identifier, a suspected emission point identifier, inspection trigger information, water quality parameter data, sampling location, sampling time, water flow status data, sampling role status data, initial judgment result, and review judgment result;

[0038] When both the initial judgment result and the review judgment result are emission point contribution anomaly results, the emission anomaly verification record will be marked as an emission point contribution confirmation record;

[0039] When the initial determination result or the review determination result is a common change result of the river section, the emission anomaly verification record is marked as a common change record of the river section; the remaining records are marked as records to be reviewed again.

[0040] Through the above technical solution, the present invention can achieve at least the following beneficial effects:

[0041] This invention determines the sampling positions on the action side and control side on both sides of a suspected discharge point based on the current water flow direction, and forms paired data units in the order of control side, action side, and control side during the same flow period. When the water flow direction reverses, the sampling positions on the action side and control side are re-determined based on the reversed water flow direction, and the paired water quality characteristic sequences before and after the reversal are associated with the corresponding water flow state data to form a flow direction rearrangement characteristic sequence. The water quality data entering the discharge point contribution determination model maintains the action side and control side relationship corresponding to the actual water flow direction during the sampling period, so that the short-term discharge impact of the suspected discharge point can be determined based on the corresponding water quality change data.

[0042] By determining the same-direction flow period when the flow velocity reaches the effective flow velocity threshold and the flow direction remains in the same channel axis, and marking the unfinished paired data units when the flow direction reverses as non-decision data units, it is possible to prevent data obtained under weak flow conditions and data that have not completed the predetermined paired collection order from entering the judgment sequence, thereby reducing the impact of uncertainty in the interaction between sampling positions on the judgment input.

[0043] By extracting the correlation characteristics of the corresponding changes in water quality data on the action side relative to the background data on the control side, and combining the temporal change characteristics of the water flow direction and the sampling role status data, the judgment result is output. This allows the water quality changes before and after the reversal of the water flow direction to participate in the judgment under their respective sampling location relationships. By forming a three-sided water quality feature sequence through corroborating water quality data, the change relationship between the water quality data on the action side and the corroborating water quality data relative to the same background data on the control side participates in the judgment, thereby distinguishing the local changes in the neighborhood of the suspected discharge point from the changes in the downstream river section that occur simultaneously.

[0044] By performing verification sampling on the initial judgment result and storing the initial judgment result, verification judgment result, sampling location, sampling time and water flow status data as a discharge anomaly verification record, it is possible to maintain the event correlation between the judgment data obtained from the same suspected discharge point under different water flow directions, thus forming a verification record for short-term discharge anomalies. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0046] Figure 1 This is a flowchart of the intelligent water environment anomaly inspection method based on unmanned vessels and deep learning in the embodiments. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0048] All terms used in this application (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0049] To facilitate understanding, the relevant terms and concepts involved in the embodiments of this application will be introduced below.

[0050] Suspected discharge points refer to pump station outlets or shoreline discharge points identified as potential targets for short-term abnormal discharge during inspection missions. Water flow status data refers to data acquired by unmanned surface vessels during sampling that characterizes the water flow status at the sampling location, including at least the flow direction.

[0051] The sampling location on the action side refers to the sampling location downstream of the suspected discharge point, determined based on the current water flow direction. The sampling location on the control side refers to the sampling location upstream of the suspected discharge point, determined based on the current water flow direction. A period of continuous flow in the same direction refers to a continuous sampling period during which the water flow direction remains consistent. Background data on the control side refers to the background water quality parameters corresponding to the action side water quality data, formed by combining two sets of control side water quality data obtained before and after a single action side water quality data collection within the same period of continuous flow.

[0052] Paired water quality feature sequences refer to data sequences formed within a unidirectional flow period, consisting of corresponding changes in water quality data on the active side of a paired data unit relative to background data on the control side, along with sampling time sequence data and flow state data. Flow direction rearrangement feature sequences refer to data sequences formed by associating paired water quality feature sequences before and after a flow direction reversal with corresponding flow state data. Discharge point contribution determination models are deep learning models that take paired water quality feature sequences or flow direction rearrangement feature sequences as input and output discharge point contribution determination results. Discharge anomaly verification records are data records that associate and store suspected discharge points, sampling data, flow state data, and discharge point contribution determination results according to the same verification event.

[0053] Sampling role status data refers to the data characterizing the correspondence between the sampling position on the action side, the sampling position on the control side, and the direction of water flow, as well as whether this correspondence changes due to a reversal of the water flow direction. It includes at least the water flow direction, the sampling position on the action side, the sampling position on the control side, and a role interchange marker. The role interchange marker is a field in the sampling role status data used to characterize whether the sampling positions on the action side and the control side change due to a reversal of the water flow direction.

[0054] On-site sampling records refer to the records of water quality parameters, sampling locations, sampling times, and water flow status data associated with the suspected discharge point, used to form the initial judgment result. Verification sampling results refer to the records of water quality parameters, water flow status data, and corresponding verification judgment results obtained after the unmanned surface vessel (USV) performs verification sampling tasks. Discharge operation records refer to data records associated with the suspected discharge point, recording the suspected discharge point identifier, discharge start time, discharge end time, and discharge operation status. On-site verification sampling conclusions refer to the category assignment records formed based on the on-site sampling records, verification sampling results, and discharge operation records at the time of existence. Verification conclusions refer to the category assignment results recorded in the on-site verification sampling conclusions.

[0055] Example 1:

[0056] like Figure 1 As shown, this embodiment employs an intelligent water environment anomaly inspection method based on unmanned surface vessels and deep learning. It is applied to verify short-term discharge anomalies at pumping station outlets or riverside outfalls after rainfall in tidal urban tributary waters. The pumping station outlets or riverside outfalls are suspected discharge points. The unmanned surface vessel is equipped with a water quality parameter acquisition unit, a positioning unit, and a water flow status acquisition unit, including:

[0057] Step S1: Obtain the identification information of the suspected discharge point and the water flow status data including the direction of water flow;

[0058] Step S2: Based on the current water flow direction, determine the sampling location downstream of the suspected discharge point as the sampling location on the action side, and determine the sampling location upstream of the suspected discharge point as the sampling location on the control side.

[0059] Step S3: During the same flow period when the water flow direction remains consistent, control the unmanned vessel to collect water quality parameter data in the order of the control side sampling position, the action side sampling position, and the control side sampling position, using the same water quality parameter category. Associate one action side water quality data with the two control side water quality data obtained before and after its collection, as well as the corresponding sampling time and water flow state data, to form a paired data unit. Then, organize the continuously formed paired data units into a paired water quality feature sequence according to the sampling time.

[0060] Step S4: Before generating the emission point contribution determination result, determine whether the water flow direction reversal has occurred based on the continuously acquired water flow state data; when the water flow direction reversal is detected, redetermine the sampling position on the action side and the sampling position on the control side according to the reversed water flow direction, generate the reversed paired water quality feature sequence, and associate the paired water quality feature sequences before and after the reversal with their respective corresponding water flow state data as the flow direction rearrangement feature sequence.

[0061] Step S5: Input the paired water quality feature sequence or the flow direction rearrangement feature sequence into the discharge point contribution determination model to obtain the discharge point contribution determination result. The discharge point contribution determination model is trained by the paired water quality feature sequence and the flow direction rearrangement feature sequence with category labels.

[0062] Step S6: Associate and store the identification information, the contribution determination result of the emission point, and the water quality parameter data, sampling location, sampling time and water flow state data that formed the result, and generate an emission anomaly verification record.

[0063] In one implementation, a paired data unit refers to a data unit formed by associating the first set of water quality data (control side, action side, and second control side) obtained sequentially after an unmanned surface vessel completes a closed sampling process within the same unidirectional flow period, following the sequence of control side sampling location, action side sampling location, and control side sampling location. The three samplings use the same water quality parameter category and are associated with their corresponding sampling locations, effective sampling times, and synchronously acquired flow state data.

[0064] When the closed sampling process meets the sampling stability condition, time validity condition, control side background pairing condition, and flow pattern validity condition, the paired data unit formed based on the closed sampling process will be marked as a definite data unit; if any condition is not met, the corresponding data unit will be marked as a non-definite data unit and retained in the discharge anomaly verification record, and will not be included in the paired water quality characteristic sequence.

[0065] For paired data units marked as determinate data units, background data corresponding to the effective sampling time on the action side is generated based on the first and second control side water quality data, and corresponding change data of the action side water quality data relative to the control side background data is generated. The continuously generated determinate data units are organized from earliest to latest according to the effective sampling time corresponding to the action side sampling to form a paired water quality feature sequence, which is used to characterize the water quality parameter change process of the water body on the action side of the suspected discharge point relative to the control side background data under the current water flow direction.

[0066] In one embodiment, the flow state data includes flow direction and flow velocity; when the flow velocity reaches an effective flow velocity threshold and the continuously obtained flow direction remains in the same channel axis, a period of unidirectional flow is determined; when the flow velocity reaches an effective flow velocity threshold and the continuously obtained flow direction changes from one channel axis to the opposite channel axis, a flow direction reversal is determined.

[0067] The effective velocity threshold is a velocity judgment criterion used to distinguish between flow states and weak flow states in which executable sampling roles are configured. Water quality parameter data obtained during periods when the flow velocity does not reach the effective velocity threshold are stored in association with the corresponding water flow state data and are not included in the current input of the discharge point contribution determination model.

[0068] In one embodiment, the identification information includes the location coordinates of the suspected discharge point; based on the location coordinates and the current water flow direction, a sampling position on the action side and a sampling position on the control side are configured in the projection direction of the river centerline, and the sampling positions on the action side and the control side are respectively separated from the suspected discharge point by a preset distance.

[0069] The coordinates of a suspected discharge point refer to the spatial location information of the suspected discharge point recorded during the inspection mission. The river centerline refers to the linear spatial information used to characterize the direction of river extension, determined based on the spatial information of the riverbank of the inspected river section. The neighborhood of a suspected discharge point refers to the spatial range of the river channel determined based on the suspected discharge point, used to configure the sampling positions on the action side and the control side. The unmanned surface vessel (USV) determines the downstream and upstream sides of the suspected discharge point along the current flow direction based on the coordinates of the suspected discharge point, the river centerline, and the current water flow direction, and configures the sampling positions on the action side and the control side respectively. The preset distance is determined based on the neighborhood range of the suspected discharge point, the navigation and positioning status of the USV, and the navigation conditions of the river during the sampling period, ensuring that the sampling positions on the action side and the control side are within the verification range of the same suspected discharge point.

[0070] In one implementation, when a reversal of the water flow direction is detected during the formation of paired data units, the incomplete paired data units are marked as non-decision data units; the sampling positions on the action side and the control side are reconfigured according to the reversed water flow direction, new paired data units are collected, and a flow direction rearrangement feature sequence is formed according to the sampling time and the corresponding water flow state data.

[0071] Non-decision data units refer to data units acquired during a closed sampling process but not included in the paired water quality characteristic sequence. These include data units that do not meet sampling stability conditions, time validity conditions, control-side background pairing conditions, or flow regime validity conditions, as well as data units that did not complete the predetermined paired sampling order due to flow direction reversal. When flow direction reversal is detected, incomplete paired data units are written to the non-decision state, retaining their acquired water quality parameter data, sampling location, sampling time, and flow state data. This data unit does not participate in the current model determination. The flow direction rearrangement characteristic sequence includes paired data units formed before the reversal, paired data units formed after the reversal, the flow state data corresponding to each paired unit, sampling time, and sampling role status data.

[0072] In one embodiment, the water quality data on the action side and the water quality data on the control side are obtained using the same water quality parameter category, which includes at least one of ammonia nitrogen characterization parameter or chemical oxygen demand characterization parameter, and at least two of conductivity, turbidity and dissolved oxygen; the paired water quality feature sequence includes change data, sampling time sequence data and water flow state data formed according to the water quality parameter category.

[0073] Corresponding change data refers to the difference characterization data formed by the water quality data on the action side relative to the background data on the control side under the same water quality parameter category. Water quality parameter data must at least record the water quality parameter category, parameter acquisition value, sampling time, sampling location, and water flow state data associated with that sampling. Data quality markers are record fields used to indicate whether the water quality parameter data meets the model input conditions; when water quality parameter data has missing measurements, exceeds the acquisition range, or the acquisition state does not meet the sampling conditions, the corresponding record carries a data quality marker indicating that the input conditions are not met, and it does not participate in the effective feature generation of the paired water quality feature sequence.

[0074] In one implementation, when outputting anomaly results or results to be reviewed for emission point contribution, a review sampling task is generated. When executing the review sampling task, if the current water flow direction is consistent with the water flow direction corresponding to the emission point contribution judgment result that triggered the review sampling task, a new paired water quality feature sequence is collected according to the current water flow direction, and the new paired water quality feature sequence is input into the emission point contribution judgment model. If the current water flow direction is opposite to the water flow direction corresponding to the emission point contribution judgment result that triggered the review sampling task, the paired water quality feature sequence that forms the emission point contribution judgment result and the paired water quality feature sequence obtained from the review are associated with their respective corresponding water flow state data to form a flow direction rearrangement feature sequence, and the flow direction rearrangement feature sequence is input into the emission point contribution judgment model. The review judgment result is obtained, and the emission point contribution judgment result that triggered the review sampling task is recorded as the initial judgment result.

[0075] The verification sampling task refers to the task of performing paired sampling again for the same suspected emission point when the emission point contribution determination result is an abnormal result of emission point contribution or a result pending verification. In the verification sampling task, when the paired water quality feature sequence used in the initial determination and the paired water quality feature sequence obtained in the verification are opposite in corresponding water flow direction, they are correlated with their respective water flow state data and sampling time to form a flow direction rearrangement feature sequence. When performing the verification sampling task, the unmanned surface vessel determines the sampling position on the active side and the sampling position on the control side based on the current water flow direction obtained during verification, and obtains a new paired water quality feature sequence. If the water flow direction is reversed again during verification, the complete paired data unit is re-obtained according to the sampling role configuration after the reversal, and the complete data unit is included in the flow direction rearrangement feature sequence.

[0076] If the sampling review task fails to generate a data sequence that meets the model input conditions, the corresponding data quality marker will be written into the emission anomaly verification record, and the record status will be recorded as a continued review record. If the water flow direction reverses again during the review period, the flow direction rearrangement feature sequence will be organized separately according to each water flow direction reversal event. Data formed between adjacent water flow direction reversal events will not be merged across events into the same flow direction rearrangement feature sequence.

[0077] In one implementation, the emission anomaly verification record includes an event identifier, a suspected emission point identifier, inspection trigger information, water quality parameter data, sampling location, sampling time, water flow status data, sampling role status data, initial judgment result, and review judgment result. When both the initial judgment result and the review judgment result are emission point contribution anomaly results, the emission anomaly verification record is marked as an emission point contribution confirmation record. When either the initial judgment result or the review judgment result is a common change result of the river section, the emission anomaly verification record is marked as a common change record of the river section. The remaining records are marked as continued review records.

[0078] The emission anomaly verification record stores the event identifier, suspected emission point identifier, inspection trigger information, water quality parameter data, sampling location, sampling time, water flow status data, sampling role status data, data quality marker, initial judgment result, review judgment result, and record status for the same verification event. Inspection trigger information includes at least one of the following: rainfall end information, pump station drainage operation information, or riverside outfall inspection task information. Record status includes emission point contribution confirmation record, river section common change record, and continued review record. The emission anomaly verification record links the event identifier with the data obtained from the initial sampling, the data obtained from the review sampling, and the corresponding judgment result to form an emission anomaly verification record for the same suspected emission point.

[0079] The judgment configuration version refers to the version identifier associated with the model input field structure, data validity judgment criteria, model training parameters, and category output criteria used to form the emission point contribution judgment result. The record fields in the emission anomaly verification record also include the judgment configuration version used when forming the judgment result. The initial judgment result and the review judgment result are stored and associated with their respective judgment configuration versions; when the initial judgment result and the review judgment result are formed within the same inspection task, the same judgment configuration version determined when the inspection task was started is used.

[0080] In a preferred embodiment of Example 1, the effective sampling time, data unit state, and corresponding changed data during the closed sampling process are processed as follows:

[0081] After the unmanned surface vessel (USV) arrives at each sampling location, it maintains a preset stationary state. Following a response delay from the water quality parameter acquisition unit, a common stabilization sampling window is initiated. The median of consecutive readings within this window is used as the current water quality data for the corresponding water quality parameter category. The effective sampling time for each sampling role is determined using the following formula:

[0082] ,

[0083] in, For sampling roles The corresponding valid sampling time; The sampling role in a single closed sampling process; For unmanned ships to reach the sampling role The time corresponding to the sampling position; The duration of stabilization required to suppress hull disturbance after the unmanned vessel reaches the sampling location; This is the set of water quality parameter categories used in this data collection; The water quality parameter categories in the water quality parameter category set; Water quality parameter categories The response delay required for the sensor to enter a stable response range from the moment it comes into contact with the water at the current location; The duration of the common stable sampling window; This indicates that the maximum value is taken from the response delays corresponding to each water quality parameter category.

[0084] The stabilization duration is calibrated based on the time required for water sample disturbance to decay to an allowable range after the unmanned vessel stops propulsion. The response delay is calibrated based on the response time obtained from the sensor standard solution switching test. The duration of the common stabilization sampling window is calibrated based on the shortest acquisition time that can form a stable median. The allowable range of stable acquisition is determined based on the upper limit of fluctuation of the median of continuous readings when the corresponding sensor is repeatedly measured. The above calibration contents remain fixed within a single inspection mission. When the reading fluctuation of any water quality parameter category within the common stabilization sampling window exceeds the corresponding allowable range of stable acquisition, the data unit formed based on this closed sampling process is marked as a non-judgment data unit with insufficient sampling stability. Its original water quality parameter data, sampling location, sampling time, and water flow state data are retained in the discharge anomaly verification record and are not included in the paired water quality characteristic sequence.

[0085] After the water quality data from the first control side, the action side, and the second control side all met the stable acquisition conditions, the time validity of the closed sampling process was determined based on the effective sampling times of the three samplings. The maximum duration of two adjacent sampling processes was determined according to the following formula:

[0086] ,

[0087] in, The maximum duration of two adjacent sampling processes during a single closed sampling process; This is the role of the first control group in the sampling process; For the side sampling role; The role is for the second control side sampling; , and These refer to the effective sampling times for the first control side, the treatment side, and the second control side, respectively, and the meaning of the effective sampling times is the same as that mentioned above. This represents the maximum permissible adjacent sampling time interval for a closed sampling process.

[0088] The maximum permissible adjacent sampling time interval is calibrated based on the trial navigation time, stabilization duration, sensor response delay, and common stabilization sampling window duration of the unmanned surface vessel between preset sampling locations, and remains fixed within a single inspection mission, not extended by any abnormal water quality changes observed during mission execution. When the maximum duration of the adjacent sampling process is within the maximum permissible adjacent sampling time interval, the water quality data on the action side is ready for time pairing with the water quality data on the two control sides; when it exceeds the maximum permissible adjacent sampling time interval, the data unit formed based on this closed sampling process is marked as a non-deterministic data unit with an excessive time span.

[0089] For closed sampling processes that meet the time validity requirement, the degree of change between the first and second control-side water quality data is used to determine whether the control-side background is within the pairable range. Dimensionally consistent background change gating is applied to each water quality parameter category, and the background change index is determined according to the following formula:

[0090] ,

[0091] in, This refers to the background change index corresponding to one closed sampling process; Water quality parameters obtained from the first control side sampling Corresponding water quality data; Water quality parameters obtained from the second control side sampling Corresponding water quality data; Water quality parameter categories The corresponding control side background variation threshold is positive. and The meaning remains the same as before; This indicates taking the absolute difference.

[0092] Rainfall trigger type refers to the historical closed sampling data grouping category determined based on at least one of the following recorded in the inspection trigger information: rainfall end information, pump station drainage operation information, or riverside outfall inspection task information. Tidal current direction type refers to the historical closed sampling data grouping category determined based on the river channel axis represented by the flow state data.

[0093] The allowable background change threshold for the control side is calibrated using historical closed-loop sampling data that matches the rainfall triggering type and tidal current direction type of the current inspection task and has been verified as being contributed by non-emission points. This threshold is also set based on the repeated measurement fluctuation range of the corresponding sensor and remains fixed within a single inspection task. When the background change index is less than or equal to 1, the water quality data on the action side enters the paired data unit; when the background change index is greater than 1, the data unit formed based on this closed-loop sampling process is marked as a non-decision data unit with background change exceeding the limit; if no matching calibration data exists, it is marked as a non-decision data unit with insufficient calibration conditions. Non-decision data units are retained in the emission anomaly verification record and are not entered into the paired water quality characteristic sequence.

[0094] Before generating background data for the control side, the order of the effective sampling times of the three samples was also checked, only when... Background data calculations are performed on the control side. If the time sequence condition is not met, the data unit formed based on this closed sampling process is marked as a non-determined data unit with an abnormal sampling time sequence, and its original water quality parameter data, sampling location, sampling time and water flow state data are retained in the discharge anomaly verification record.

[0095] When the water quality data on the action side meets the conditions for entering the paired data unit, based on the temporal adjacency of the first and second control side water quality data, background data on the control side corresponding to the effective sampling time on the action side is formed, and corresponding change data of the action side relative to the background data on the control side is formed:

[0096] ,

[0097] in, The category of water quality parameters corresponding to the effective sampling time on the action side. The control side background data; Category of water quality parameters obtained from side sampling Corresponding water quality data; This refers to the corresponding changes in water quality data on the action side relative to background data on the control side. , , , , and The meaning remains the same as before.

[0098] The effective sampling time on the action side lies between two effective sampling times on the control side. The background data on the control side is used to characterize the water state on the control side at the sampling time on the action side, and the corresponding change data is used to characterize the change in the water quality data on the action side relative to the background data on the control side. The background data on the control side and the corresponding change data are used to form the model input features; paired data units are only included in the paired water quality feature sequence when the conditions of sampling stability, time validity, background pairability, and flow regime validity are met.

[0099] The effective velocity threshold is the lower limit of the velocity required to maintain the correspondence between the sampling positions on the active side and the control side. It is calibrated based on the measurement fluctuation range of the flow state acquisition unit and the unidirectional flow velocity data obtained from the trial navigation in the vicinity of the suspected discharge point, and remains fixed within a single inspection mission. During each closed sampling process, flow state data is continuously acquired from the first effective sampling time on the control side to the second effective sampling time on the control side. When all velocity readings reach the effective velocity threshold and the flow direction remains in line with the river axis corresponding to the current unidirectional flow period, the median of the continuous velocity readings is used as the representative velocity value of the corresponding paired data unit, and the paired data unit is allowed to be marked as a definite data unit. Otherwise, the data unit formed based on the closed sampling process is marked as a non-definite data unit with insufficient flow state effectiveness, and it is not included in the paired water quality characteristic sequence, nor is a corresponding flow direction time series input element formed.

[0100] The paired data unit includes water quality data obtained from three samplings, background data from the control side, corresponding change data, corresponding sampling location, effective sampling time, water flow state data, and data unit label. When all consecutively obtained paired data units are labeled as identifiable data units, a paired water quality characteristic sequence is formed according to the effective sampling time on the active side from earliest to latest. If a reversal of the water flow direction is detected before the completion of the closed sampling process, the sampling locations on the active side and the control side are reconfigured according to the reversed water flow direction, and the data records obtained before the completion of the closed sampling process before the reversal are organized into non-identifiable data units.

[0101] In one implementation, during the same-direction flow period, the unmanned vessel is controlled to collect circumstantial water quality data at a circumstantial sampling location located downstream of the current water flow direction and outside the neighborhood of the suspected discharge point. The interval between the circumstantial sampling location and the suspected discharge point is greater than the interval between the sampling location on the action side and the suspected discharge point.

[0102] The corresponding changes in the water quality data from the action side and the supporting water quality data relative to the background data from the control side are used to form a three-sided water quality feature sequence. This three-sided water quality feature sequence, along with one of the paired water quality feature sequence and the flow direction rearrangement feature sequence, and the associated water flow state data and sampling role state data, are input into the discharge point contribution determination model to obtain the discharge point contribution determination result. The model input fields for the training samples include one of the three-sided water quality feature sequence, the paired water quality feature sequence, and the flow direction rearrangement feature sequence, and also include the associated water flow state data and sampling role state data. The category label field includes the category label corresponding to the discharge point contribution determination result.

[0103] Evidence-providing sampling locations refer to sampling locations downstream of the current flow direction and outside the vicinity of the suspected discharge point. Evidence-providing water quality data refers to data obtained by unmanned vessels at the evidence-providing sampling locations according to the same water quality parameter categories as the data on the impact side and the control side. The three-sided water quality characteristic sequence refers to a data sequence organized from the corresponding change data of the impact side water quality data relative to the control side background data, the corresponding change data of the evidence-providing water quality data relative to the control side background data, and the corresponding flow state data. When the corresponding change data of the impact side water quality data and the corresponding change data of the evidence-providing water quality data simultaneously show the same direction of change, the three-sided water quality characteristic sequence is used as the basis for determining the common change category of the river section; when the corresponding change data of the impact side water quality data and the corresponding change data of the evidence-providing water quality data show different change characteristics, the three-sided water quality characteristic sequence is used as the basis for determining the local impact of the suspected discharge point.

[0104] When forming the three-sided water quality characteristic sequence, corroborating water quality data are obtained within the same flow period corresponding to the action-side water quality data, using the same water quality parameter category, and associated with the effective sampling time and flow state data. A correspondence is established between the corroborating water quality data and the identifiable data unit closest to the effective sampling time. If the sampling time interval between the corroborating water quality data and the action-side water quality data in that identifiable data unit is not greater than the maximum allowable adjacent sampling time interval, and the flow direction remains along the same river axis during this interval, the corroborating water quality data is included in the three-sided water quality characteristic sequence. If these conditions are not met, the corroborating water quality data is associated with a data quality marker and retained in the discharge anomaly verification record, but is not included in the three-sided water quality characteristic sequence.

[0105] Example 2:

[0106] Based on Example 1, this example provides a specific method for extracting temporal change features associated with the state data of the sampled role in the intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning.

[0107] The emission point contribution determination model includes a pairing relationship feature extraction branch, a flow direction change feature extraction branch, and a determination output layer;

[0108] The sampling role status data represents the correspondence between the sampling position on the action side, the sampling position on the control side, and the direction of water flow, as well as whether the correspondence changes due to the reversal of the direction of water flow;

[0109] The pairing relationship feature extraction branch receives paired data units from paired water quality feature sequences or flow direction rearrangement feature sequences, and extracts the correlation change features of the water quality data on the action side relative to the background data on the control side according to the sampling time order;

[0110] The flow direction change feature extraction branch receives paired water quality feature sequences or flow direction rearrangement feature sequences, as well as sampling role status data associated with the corresponding sequences, and extracts the time-series change features corresponding to the flow direction and sampling role status according to the sampling time order;

[0111] The output layer determines the correlation and temporal change characteristics received, and outputs one of the following: emission point contribution anomaly results, river segment co-change results, or results pending verification.

[0112] In one implementation, the training samples used to train the emission point contribution determination model include a model input field, a category label field, and a tracing field;

[0113] The model input fields include paired water quality feature sequences or flow direction rearrangement feature sequences, as well as water flow state data and sampling role state data associated with the corresponding sequences; the category label field includes category labels corresponding to the emission point contribution anomaly results, the river section co-change results, and the results to be verified, respectively; the traceability field includes sample identifier, suspected emission point identifier, sampling location, sampling time, and data quality label, and the traceability field is stored in association with the corresponding model input field and category label field;

[0114] The category labels are formed based on on-site sampling records, verification sampling results, and emission operation records associated with suspected emission points. During the training phase, the model input fields are input into the emission point contribution determination model, and the corresponding category labels are used as the supervision targets to make the model output categories correspond to the category labels.

[0115] After the model is deployed, the emission point contribution determination model receives paired water quality feature sequences or flow direction rearrangement feature sequences that are consistent with the model input field structure during the training phase, and outputs the emission point contribution determination results; new records with category labels formed after verification form new samples according to the field structure of the training samples. The new samples participate in subsequent model training and verification to form the determination configuration version for subsequent inspection tasks.

[0116] The pairing relationship feature extraction branch receives the identifiable data units contained in the paired water quality feature sequence or the flow direction rearrangement feature sequence, and extracts the correlation change features of the water quality data on the action side relative to the background data on the control side according to the sampling time order; in the implementation method using circumstantial sampling location, the pairing relationship feature extraction branch also receives the water quality feature sequences on three sides to extract the corresponding change relationship between the water quality data on the action side and the circumstantial water quality data relative to the background data on the control side.

[0117] When using three-sided water quality feature sequences for judgment, the model input field structure recorded in the judgment configuration version includes one of the three-sided water quality feature sequences, paired water quality feature sequences, and flow direction rearrangement feature sequences, and includes water flow state data and sampling role state data associated with the above sequences. Specifically, the three-sided water quality feature sequences are input into the paired relationship feature extraction branch; one of the paired water quality feature sequences and the flow direction rearrangement feature sequences, along with the associated water flow state data and sampling role state data, are input into the flow direction change feature extraction branch. The same model input field structure is used in the training and deployment phases. When no three-sided water quality feature sequence meeting the input conditions is formed, the corresponding supporting water quality data does not participate in the current judgment based on the three-sided water quality feature sequences.

[0118] The flow direction change feature extraction branch receives the flow direction rearrangement feature sequence and its corresponding sampling role status data. It extracts the temporal change features corresponding to the changes in pairing relationships before and after the flow direction reversal, and the interchange of sampling positions on the active and control sides, according to the sampling time order. For paired water quality feature sequences where the flow direction has not reversed, the role interchange marker in the sampling role status data is recorded as unchanged, and the flow direction change feature extraction branch forms the temporal change features under the same flow state. The output layer receives the associated change features and temporal change features, and outputs one of the following: discharge point contribution anomaly results, river segment co-change results, or results pending verification.

[0119] The traceability field is stored in association with the corresponding model input field and category label field, and is not used as an independent basis for determining the contribution of suspected emission points.

[0120] The category label is formed based on the verification conclusions corresponding to the emission anomaly investigation records. For samples whose anomalies are confirmed to belong to the corresponding suspected emission points through emission operation records and on-site review sampling conclusions, and which form paired water quality characteristic sequences or flow direction rearrangement characteristic sequences that can be determined by data unit organization, their category label is set to the category label corresponding to the emission point contribution anomaly result. For samples verified to belong to rainfall runoff, overall river section water quality changes, or tidal background changes, their category label is set to the category label corresponding to the common changes in the river section. For samples lacking attribution criteria, only showing unilateral short-term anomaly changes, or whose emission point contribution and river section background changes cannot be distinguished after review, their category label is set to the category label corresponding to the result to be reviewed.

[0121] When generating category labels, the sampling time of the training samples is correlated with the emission time in the emission operation record and the sampling time in the on-site verification sampling conclusion according to the event identifier. For coastal discharge outlets without emission operation records, the emission point attribution result recorded in the on-site verification sampling conclusion is used as the basis for category labeling; for samples without emission point attribution results, their category label is set to the result to be verified.

[0122] In a preferred embodiment of Example 2, the flow direction change feature extraction branch receives determinable data units and sampling role status data organized according to sampling time. A fixed positive axis is set based on the projection point of the suspected discharge point on the river centerline, and the preset sampling positions at both ends of the axis are recorded as the fixed positive side sampling position and the fixed negative side sampling position, respectively. When the water flow direction is along the fixed positive axis, the fixed positive side sampling position is configured as the action side sampling position, and the fixed negative side sampling position is configured as the control side sampling position; when the water flow direction is opposite to the fixed positive axis, the fixed negative side sampling position is configured as the action side sampling position, and the fixed positive side sampling position is configured as the control side sampling position. When a transition between the above two configurations is detected, sampling role status data is formed. The sampling role status data includes the time of water flow direction reversal, the water flow direction before and after the reversal, the action side sampling position before and after the reversal, the control side sampling position before and after the reversal, and a role interchange marker.

[0123] For each paired data unit marked as a definite data unit, the effective sampling time of the sample taken from the active side is used as the time positioning time of that paired data unit, and the paired data units before and after the reversal are uniformly sorted according to the time positioning time from earliest to latest. Incomplete paired data units formed during the reversal process follow the same processing method as the non-definite data units mentioned above; their sampling data is retained in the discharge anomaly verification record, and their corresponding water flow direction reversal events continue to be used to generate role interchange markers. The sorted... Each identifiable data unit corresponds to a flow-direction timing input element, which is formed according to the following formula:

[0124] ,

[0125] in, For the first Each identifiable data unit corresponds to a flow direction timing input element; The sequence number of the data unit after being sorted according to the time of location; For the first Water quality parameter categories in each identifiable data unit The corresponding change data are calculated using the same method as the aforementioned change data of the water quality data on the action side relative to the background data on the control side. Water quality parameter categories The corresponding water quality change scale value, and it is positive; and The meanings of the aforementioned water quality parameter categories and sets of water quality parameter categories shall be retained; For the first The water flow axis code corresponding to each identifiable data unit takes a value of 1 when the water flow direction is along the fixed positive axis and a value of -1 when the water flow direction is opposite to the fixed positive axis. For the first The flow velocity representative value corresponding to each identifiable data unit is obtained by following the aforementioned rule of forming a flow velocity representative value based on the median of continuous flow velocity readings. This is a velocity scale value, and it is positive. For the first The interval between the time positioning time of each definite data unit and the previous definite data unit is 0, and the first definite data unit in the sequence has a corresponding value of 0. This is a time interval scale value, and it is a positive value; For the first The role swap flag corresponding to each identifiable data unit is set to 1 when the data unit is the first identifiable data unit formed after the water flow direction is reversed, and 0 in other cases. To fix the positive side sampling position at the 1st The role code of the sampling position on the action side in each determinable data unit is 1 when the fixed positive side sampling position is used as the sampling position on the action side, and 0 when it is used as the sampling position on the control side. To fix the sampling position on the reverse side at the first The role code of the sampling position on the action side in each determinable data unit is set to 1 when the sampling position on the reverse side is used as the sampling position on the action side, and 0 when it is used as the sampling position on the control side.

[0126] The corresponding change data for each water quality parameter category are written into the flow direction time series input element according to a fixed arrangement order in the preset water quality parameter category list. The fixed arrangement order remains consistent during the model training, model validation, and model deployment phases.

[0127] The water quality change scale value is calibrated based on the distribution of corresponding change data for the same water quality parameter category in the training samples, with the change scale corresponding to repeated sensor measurements as the lower limit; the flow velocity scale value is calibrated based on the flow velocity data of the training samples that reach the effective flow velocity threshold; the time interval scale value is calibrated based on the interval duration of data units that meet the time validity condition. The above scale values ​​correspond to the judgment configuration version of the emission point contribution determination model and remain fixed within a single inspection task.

[0128] When identifiable data units are formed both before and after the water flow direction reverses, the identifiable data unit whose time before the reversal is closest to the moment of role reversal is taken as the boundary data unit before the reversal. Continuous data units before the reversal are selected forward from this boundary data unit, and continuous data units after the reversal are selected backward from the moment of role reversal, forming a fixed-length input window for the flow direction change.

[0129] ,

[0130] in, The flow direction change input window is formed based on the positioning of the reversed front boundary data unit; This involves processing features for connection according to the time sequence of location; This is to reverse the sort order of the data cells at the front boundary; The number of data units that can be determined before reversal is selected in the input window; The number of data units can be determined by reversing the selected data in the input window. The total number of identifiable data units contained in the flow direction change input window; to This is a flow direction timing input element that is arranged according to the time location and enters the flow direction change input window.

[0131] The flow direction change input window, formed by arranging and connecting the identifiable data units before and after the flow direction reversal and their corresponding flow state data according to the time positioning time, serves as the model input representation for the flow direction rearrangement feature sequence. For paired water quality feature sequences that have not undergone flow direction reversal, a sequence of consecutively arranged data according to the time positioning time with a quantity of [number missing] is selected. The determinable data units are connected in the same order as the flow direction change input window to form a unidirectional flow state input window, wherein the role interchange markers in each flow direction time sequence input element are used. All values ​​are 0. This indicates the sorting number of the latest identifiable data unit in the same-direction flow input window. The flow direction change input window and the same-direction flow input window are collectively referred to as the decision input window, and both use... Represent and input the emission point contribution determination model.

[0132] Each flow direction change input window corresponds to only one flow direction reversal event, and and These are positive integers. If a flow direction reversal occurs again during a data unit period after a reversal, the same input window is not formed across two reversal events, and subsequent data is reorganized based on the latest reversal event. If the number of data units is insufficient before or after a reversal, no flow direction change input window is formed; if the number of data units is continuously insufficient within the same flow direction period... At that time, no unidirectional flow input window is formed.

[0133] If a judgment input window cannot be formed by the end of the inspection task, the data quality will be marked as insufficient valid data in the emission anomaly verification record, and no current emission point contribution judgment result will be formed, and a review sampling task will be generated.

[0134] The flow direction change feature extraction branch performs joint temporal coding on the corresponding change data, flow axis coding, flow velocity representative value, time interval, and role interchange marker in the decision input window. When the decision input window is a flow direction change input window, the joint temporal coding is used to characterize the continuity of the corresponding change data of the water quality data on the affected side before and after the sampling role interchange; when the decision input window is a same-direction flow state input window, the joint temporal coding is used to characterize the temporal continuity of the corresponding change data of the water quality data on the affected side under the same flow axis.

[0135] This branch forms an attention coefficient for each identifiable data unit in the decision input window, and when the decision input window contains a role reversal marker with a value of 1, it applies a restricted attention bias to the first identifiable data unit after the corresponding water flow direction is reversed, thereby converging to obtain the time-series variation characteristics:

[0136] ,

[0137] ,

[0138] ,

[0139] in, Extract the temporal variation features of the branch output for the decision input window to identify flow direction change features; To determine the set of sorting numbers for each deducible data unit within the input window; To determine the first input window The attention coefficient corresponding to each identifiable data unit; For the first A timing coding feature is formed by combining a determinable data unit with the timing relationship between the preceding and following input windows; For the temporal coding transformation in the flow direction change feature extraction branch, the temporal coding transformation is implemented by a gated recurrent coding layer that receives the flow direction temporal input elements in the order of time positioning, and outputs the corresponding temporal coding features for each identifiable data unit in the input window; the input feature order and output feature dimension of the gated recurrent coding layer remain consistent during the model training and model deployment phases; This is the training parameter vector used to generate the attention coefficient; This is a vector transpose operation. The training parameters corresponding to the role swap label are set to have values ​​greater than or equal to 0 and less than or equal to the preset upper limit of the role swap bias. The meaning of the aforementioned role-swapping markers will be retained; The summation index is the index of the sorted index set; For sorting sequence number Corresponding temporal coding features; For sorting sequence number Corresponding role swap markers; This is for exponential operations.

[0140] Each attention coefficient ranges from 0 to 1, and the sum of all attention coefficients within the same decision input window is 1. During training, the preset upper limit of the role swap bias is calibrated based on the allowable proportion of category output change caused by the attention bias of the first identifiable data unit after inversion in the validation samples. After model deployment, the preset upper limit of the role swap bias and the training parameters corresponding to the role swap markers remain fixed, allowing the role swap position to participate in feature convergence while limiting the dominance of a single inversion boundary data unit on the decision result. After model deployment, the training parameters in the flow change feature extraction branch remain fixed, and a new decision input window is formed only based on newly acquired data units during a single inspection.

[0141] The pairing relationship feature extraction branch receives the corresponding change data for each water quality parameter category according to the temporal location sequence of each identifiable data unit within the judgment input window, and organizes the corresponding change data according to a fixed arrangement order in the preset water quality parameter category list. When using three-sided water quality feature sequences, the corresponding change data corresponding to the supporting water quality data is organized according to the same temporal location sequence and water quality parameter category arrangement order. The pairing relationship feature extraction branch performs feature encoding on the organized corresponding change data to form associated change features. The fixed arrangement order, input field structure, and output feature dimension remain consistent during the model training and deployment phases. The temporal change features and associated change features are concatenated before the judgment output layer, enabling the discharge point contribution judgment model to utilize the corresponding change data within the same flow period and, when the water flow direction reverses, combine it with the sampled role status data to form a judgment result. The fusion process is performed according to the following formula:

[0142] ,

[0143] in, This determines the category probability vector output by the output layer in response to the input window. This involves normalization to convert the outputs of each category into category probabilities; To determine the training parameter matrix of the output layer; The feature extraction branch for pairing relationships outputs the correlation change features of the data unit corresponding to the decision input window; The meaning of the aforementioned temporal change characteristics is retained; To determine the training parameter vector of the output layer.

[0144] The category probability vector includes the probability of the emission point contribution anomaly category, the probability of the river segment co-change category, and the probability of the category to be verified. To ensure the model output corresponds to the verification task, when the emission point contribution anomaly category or the river segment co-change category reaches its respective output threshold and maintains a calibration interval with the other categories, the corresponding result is output; otherwise, the result to be verified is output, specifically:

[0145] when ,and Furthermore, when the common change categories of the river segment do not simultaneously meet their corresponding output threshold and category interval threshold conditions, ,

[0146] when ,and Furthermore, when the emission point contribution anomaly category does not simultaneously meet its corresponding output threshold and category interval threshold conditions, ,

[0147] When both the emission point contribution anomaly category and the river segment common change category simultaneously meet their respective output thresholds and category interval thresholds, or when neither meets their respective output thresholds and category interval thresholds... ;

[0148] in, To contribute judgment results to the emission points formed by the judgment input window; Anomalies contributing to emission points; This is a result of common changes across the river sections; The result is pending review. The emission point contributes to the anomaly category probability in the category probability vector; The common change category probability of river segments in the category probability vector; The class probability to be verified is in the class probability vector; The output threshold corresponding to the emission point contribution anomaly results; The output threshold corresponding to the common changes in the river section; The category interval threshold corresponding to the emission point contribution anomaly results; The category interval threshold corresponding to the common changes in river segments; To select the larger value among the listed category probabilities.

[0149] The output thresholds and category interval thresholds corresponding to emission point contribution anomalies are calibrated based on the allowable proportion of misclassified river segment co-change results as emission point contribution anomalies in the verification samples with verification conclusions. The output thresholds and category interval thresholds corresponding to river segment co-change results are also calibrated based on the allowable proportion of misclassified emission point contribution anomalies as river segment co-change results in the verification samples with verification conclusions. The range of each output threshold is greater than 0 and less than 1, and the range of each category interval threshold is greater than or equal to 0 and less than 1. The output thresholds and category interval thresholds remain fixed after model deployment. The stabilization duration, response delay, co-stabilization sampling window duration, allowable stabilization acquisition range, allowable background change threshold for the control side, effective flow velocity threshold, water quality change scale value, flow velocity scale value, time interval scale value, number of identifiable data units included in the flow direction change input window, output thresholds, category interval thresholds, and model training parameters used in this inspection task are associated with the same judgment configuration version, and their version identifier is written into the emission anomaly verification record. The judgment configuration version remains unchanged during a single inspection task; if the configuration is missing or the version is inconsistent, no contribution judgment result is generated for the current emission point, and a review sampling task is generated; the updated judgment configuration version is only used for subsequent inspection tasks.

[0150] During the training phase, one of the paired water quality feature sequences with category labels and the flow direction rearrangement feature sequences is used as training samples input into the discharge point contribution determination model. When using three-sided water quality feature sequences, one of the three-sided water quality feature sequences with category labels, the paired water quality feature sequences, and the flow direction rearrangement feature sequences is used as training samples input into the discharge point contribution determination model. Using the corresponding category label as the supervision target, the training parameters of the paired relationship feature extraction branch, the flow direction change feature extraction branch, and the determination output layer are updated based on the classification loss between the category probability vector and the category label. The training parameters determined by the verified samples are associated with the corresponding output thresholds to the same determination configuration version. The verified discharge feature direction is the abnormal change direction of the water quality data on the action side relative to the background data on the control side, recorded according to the water quality parameter category in the on-site verification sampling conclusion. Samples lacking this direction record or whose anomalies cannot be attributed to the corresponding suspected discharge points are not used as confirmation training samples for discharge point contribution anomaly results.

[0151] For samples verified as belonging to suspected discharge points and forming paired water quality characteristic sequences or flow direction rearrangement characteristic sequences organized by identifiable data units, their category labels are set to the category labels corresponding to the discharge point's contribution-type anomaly results. Among these, samples forming flow direction rearrangement characteristic sequences should exhibit anomaly changes on the current impact side consistent with the direction of the verified discharge characteristics, both before and after the reversal. Anomalies in ammonia nitrogen characterization parameters, chemical oxygen demand characterization parameters, conductivity, and turbidity may manifest as increases, while anomalies in dissolved oxygen may manifest as decreases.

[0152] For samples verified to belong to rainfall runoff, overall river section changes, or tidal background changes, their category labels are set to the category labels corresponding to the common changes in the river section. When using circumstantial sampling locations, records where the water quality data on the action side and the circumstantial water quality data show synchronous and identical changes in the same water quality parameter category relative to the background data on the control side can be used as training samples corresponding to the common changes in the river section. For samples that only show short-term abnormal changes on the action side before or after the reversal, or lack evidence to distinguish the contribution of the discharge point from the background changes in the river section, their category labels are set to the category labels corresponding to the results to be verified. The category labels redefined after verification are only used for subsequent new judgment configuration versions, and the judgment results that have already been formed are stored in association with the judgment configuration version used when forming the judgment results.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0154] Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments. For example, all the embodiments above can be used in any combination. The information disclosed in this background section is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that such information constitutes prior art known to those skilled in the art.

Claims

1. A method for intelligent inspection of aquatic environment anomalies based on unmanned surface vessels and deep learning, characterized in that: An unmanned surface vessel (USV), equipped with a water quality parameter acquisition unit, a positioning unit, and a water flow status acquisition unit, will conduct short-term post-rainfall discharge anomaly verification at at least one suspected discharge point among pump station outlets and riverside outfalls in tidal urban tributary waters, including: Obtain the identification information of the suspected discharge point and water flow status data, including the direction of water flow; Based on the water flow direction, determine the sampling location on the downstream side of the suspected discharge point and the control side sampling location on the upstream side; During the same flow period in which the water flow direction remains consistent, the unmanned vessel is controlled to collect water quality parameter data in the order of the control side sampling position, the action side sampling position, and the control side sampling position, with the same water quality parameter category. The action side water quality data is associated with the control side water quality data obtained before and after it as paired data units, and paired water quality feature sequences are formed according to the sampling time. The water flow direction is determined based on the continuously acquired water flow state data. When the water flow direction is reversed, the sampling positions on the action side and the control side are re-determined according to the reversed water flow direction. The paired water quality characteristic sequences before and after the reversal are associated with the corresponding water flow state data to form a flow direction rearrangement characteristic sequence. Based on the judgment result, the paired water quality feature sequence or the flow direction rearrangement feature sequence is input into the discharge point contribution judgment model to obtain the discharge point contribution judgment result. The discharge point contribution judgment model is a deep learning model. The identification information, the contribution determination result of the emission point, and the water quality parameter data and water flow state data that form the result are associated and stored as an emission anomaly verification record.

2. The intelligent water environment anomaly inspection method based on unmanned surface vessels and deep learning according to claim 1, characterized in that, The water flow status data includes the water flow direction and flow velocity; When the flow velocity reaches the effective flow velocity threshold and the continuously obtained water flow direction remains in the same river channel axis, the same flow state period is determined. When the flow velocity reaches the effective flow velocity threshold and the continuously obtained flow direction turns to the opposite river channel axis, it is determined that a flow direction reversal has occurred.

3. The intelligent water environment anomaly inspection method based on unmanned surface vessels and deep learning according to claim 1, characterized in that, The identification information includes the coordinates of the suspected emission point location; Based on the location coordinates and the current water flow direction, the sampling positions on the action side and the control side are configured in the projection direction of the river centerline, and the sampling positions on the action side and the control side are respectively separated from the suspected discharge point by a preset distance.

4. The intelligent water environment anomaly inspection method based on unmanned surface vessels and deep learning according to claim 1, characterized in that, If a reversal of the water flow direction is detected during the formation of paired data units, the incomplete paired data units are marked as non-decision data units; New paired data units are collected based on the reversed water flow direction, and the flow direction rearrangement feature sequence is formed according to the sampling time and the corresponding water flow state data.

5. The intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning according to claim 1, characterized in that, The water quality parameter categories include at least one of ammonia nitrogen characterization parameters or chemical oxygen demand characterization parameters, and at least two of conductivity, turbidity and dissolved oxygen; The background data for the control side is formed from the water quality data of the two consecutive control sides in the paired data unit; The paired water quality feature sequence includes corresponding change data of the water quality data on the action side relative to the background data on the control side, sampling time sequence data, and water flow state data.

6. The intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning according to claim 5, characterized in that, The emission point contribution determination model includes a pairing relationship feature extraction branch, a flow direction change feature extraction branch, and a determination output layer; The sampling role status data characterizes the correspondence between the sampling position on the action side, the sampling position on the control side, and the direction of water flow, and whether the correspondence has changed; The pairing relationship feature extraction branch extracts the associated change features of the corresponding change data. The flow direction change feature extraction branch extracts the time-series change features based on one of the paired water quality feature sequences and the flow direction rearrangement feature sequences, as well as the associated sampling role status data. The judgment output layer outputs one of the following: discharge point contribution anomaly results, river segment common change results, and results to be verified.

7. The intelligent water environment anomaly inspection method based on unmanned surface vessels and deep learning according to claim 6, characterized in that, The training samples include model input fields, category label fields, and traceability fields; the model input fields include one of paired water quality feature sequences and flow direction rearrangement feature sequences, as well as associated water flow state data and sampling role state data; The category labeling field includes category labels corresponding to emission point contribution anomaly results, river section co-change results, and results pending verification, respectively; the tracing field includes sample identifier, suspected emission point identifier, sampling location, sampling time, and data quality label; The category labels are formed based on on-site sampling records, verification sampling results, and emission operation records.

8. The intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning according to claim 7, characterized in that, During the same flow period, the unmanned vessel is controlled to collect corroborating water quality data downstream of the current water flow direction. The interval between the sampling location of the corroborating water quality data and the suspected discharge point is greater than the interval between the sampling location on the action side and the suspected discharge point. The corresponding changes in the water quality data of the action side and the supporting water quality data relative to the background data of the control side are used to form a three-sided water quality feature sequence; The discharge point contribution determination model is input with one of the three-sided water quality characteristic sequences, the paired water quality characteristic sequences, and the flow direction rearrangement characteristic sequences, along with the water flow state data and sampling role state data associated with the above sequences, to obtain the discharge point contribution determination result. The model input field of the training samples includes one of the three-sided water quality feature sequences, the paired water quality feature sequences, and the flow direction rearrangement feature sequences, and includes water flow state data and sampling role state data associated with the above sequences. The category label field includes a category label corresponding to the contribution determination result of the discharge point.

9. The intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning according to claim 7, characterized in that, When outputting abnormal results or results pending verification for emission point contribution, a verification sampling task is generated, and the emission point contribution determination result that triggers the verification sampling task is recorded as the initial determination result. When performing the aforementioned verification sampling task, a new paired water quality characteristic sequence is formed based on the current water flow direction; When the current water flow direction is opposite to the water flow direction corresponding to the initial judgment result, the paired water quality feature sequence of the initial judgment result and the new paired water quality feature sequence will be associated with their respective water flow state data to form a flow direction rearrangement feature sequence; The new paired water quality characteristic sequence or the flow direction rearrangement characteristic sequence is input into the discharge point contribution determination model to obtain the verification determination result.

10. The intelligent inspection method for water environment anomalies based on unmanned vessels and deep learning according to claim 9, characterized in that, The emission anomaly verification record includes event identifier, suspected emission point identifier, inspection trigger information, water quality parameter data, sampling location, sampling time, water flow status data, sampling role status data, initial judgment result, and review judgment result; When both the initial judgment result and the review judgment result are emission point contribution anomaly results, the emission anomaly verification record will be marked as an emission point contribution confirmation record; When the initial determination result or the review determination result is a common change result of the river section, the emission anomaly verification record is marked as a common change record of the river section; the remaining records are marked as records to be reviewed again.

Images