Water body evaluation method and equipment based on knowledge distillation and storage medium

Abstract

The invention discloses a water body evaluation method and device based on knowledge distillation and a storage medium, and relates to the technical field of water body evaluation, and the water body evaluation method based on knowledge distillation comprises the steps: determining a multi-modal data set in multiple candidate environment data based on a water body evaluation task; training an original model corresponding to the water body evaluation task according to the multi-modal data set to obtain a teacher model corresponding to the water body evaluation task; determining a sensor data set in the plurality of candidate sensor data according to the water body evaluation task; constructing a student model based on a model structure determined by the sensor data set; and distilling the knowledge of the teacher model to the student model to obtain a distilled student model corresponding to the water evaluation task. The technical effect of improving the timeliness of water body evaluation can be achieved.

Description

Technical Field

[0001] This application relates to the field of water body assessment technology, and in particular to a water body assessment method, device and storage medium based on knowledge distillation. Background Technology

[0002] Currently, comprehensive assessments of surface water conditions, including rivers, lakes, reservoirs, and nearshore waters, primarily rely on fixed monitoring stations. This involves collecting conventional physicochemical parameters such as dissolved oxygen, conductivity, pH, turbidity, nutrients, and organic matter, and combining these with water quality indices, eutrophication indices, ecological health indices, or grading evaluation standards to comprehensively identify the water's condition. To improve the comprehensiveness and precision of water condition identification, some research and engineering practices have introduced high-information-capacity sensing technologies such as hyperspectral remote sensing, infrared imaging, sonar, or radar to acquire multimodal data. This data is used to assist in the analysis of surface optical characteristics, spatial distribution patterns, and sediment morphology, compensating for the insufficient information provided by conventional physicochemical sensors. However, multimodal data is difficult to acquire in real-time and online, making it challenging to meet the application requirements of high-frequency updates or real-time assessments. Summary of the Invention

[0003] The main objective of this application is to provide a water body assessment method, device, and storage medium based on knowledge distillation, aiming to solve the technical problem that existing technologies are unable to meet the application requirements of high-frequency updates or real-time assessment.

[0004] To achieve the above objectives, this application provides a water body assessment method based on knowledge distillation, which includes:

[0005] Based on the water body assessment task, a multimodal data set was determined from multiple candidate environmental data. The original model corresponding to the water body assessment task is trained based on the multimodal dataset, and the teacher model corresponding to the water body assessment task is obtained. Based on the water body assessment task, a set of sensor data is determined from multiple candidate sensor data. A student model is constructed based on the model structure determined by the sensor data set; The knowledge from the teacher model is distilled into the student model to obtain the distilled student model corresponding to the water body assessment task.

[0006] In one embodiment, the step of determining a multimodal data set from multiple candidate environmental data based on a water body assessment task includes: Based on the water body assessment task, target environmental data are identified from the candidate environmental data; Extract data features from the target environment data; The data features are clustered according to the data categories of the target environmental data to obtain at least one multimodal feature, which is used to form a multimodal data set. The multimodal features include water quality parameter features, image features, acoustic features, and semantic features.

[0007] In one embodiment, the step of training the original model corresponding to the water body assessment task based on a multimodal dataset to obtain the teacher model corresponding to the water body assessment task includes: Input the multimodal dataset into the original model corresponding to the water body assessment task; Based on the assessment objectives of the water body assessment task, the model structure of the original model is determined; The original model is trained and optimized based on a multimodal dataset; The original model, after training and optimization, will be used as the teacher model for the water body assessment task.

[0008] In one embodiment, candidate environmental data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, FDOM, spectral data, remote sensing data, infrared data, acoustic detection data, manual inspection records, event annotation information, and semantic description information; candidate sensor data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, and FDOM.

[0009] In one embodiment, the step of determining a set of sensor data from multiple candidate sensor data according to a water body assessment task includes: Determine the correlation between the candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task; Candidate sensor parameters with a correlation degree greater than or equal to the correlation degree threshold are selected as the sensor data set.

[0010] In one embodiment, after determining the correlation between the candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task, the method includes: The candidate sensor data are sorted from largest to smallest according to their correlation to obtain the sorted candidate sensor data. A preset number of candidate sensor data are selected as the sensor data set based on the sorting.

[0011] In one embodiment, the knowledge from the teacher model is distilled into the student model to obtain a distilled student model corresponding to the water assessment task, including: Determine the knowledge distillation method for transferring knowledge from the teacher model to the student model. The knowledge distillation method includes any one of output distillation, feature distillation, and attention distillation. Based on a defined knowledge distillation method, a corresponding knowledge distillation loss function is constructed. Based on multimodal datasets and sensor datasets, the student model is jointly trained using the knowledge distillation loss function; The student model that has completed joint training will be used as the distilled student model for the water assessment task.

[0012] In one embodiment, after the step of distilling the knowledge from the teacher model into the student model to obtain the distilled student model corresponding to the water assessment task, the method includes: Determine the sensor combination appropriate for the water body assessment task; The data set collected by the sensor combination is input into the student model corresponding to the water body assessment task. The student model is obtained by knowledge distillation of the pre-trained teacher model and is associated with the water body assessment task. The water assessment results corresponding to the water assessment task are determined based on the model output of the student model.

[0013] In addition, to achieve the above objectives, this application also provides a water body assessment device based on knowledge distillation. The water body assessment device based on knowledge distillation includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program is configured to implement the steps of the water body assessment method based on knowledge distillation as described above.

[0014] In addition, to achieve the above objectives, this application also provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the knowledge-based distillation-based water body assessment method described above.

[0015] This application provides a water body assessment method based on knowledge distillation. Based on the water body assessment task, a multimodal dataset is determined from multiple candidate environmental data. An original model corresponding to the water body assessment task is trained using the multimodal dataset to obtain a teacher model for the water body assessment task. Based on the water body assessment task, a sensor dataset is determined from multiple candidate sensor data. A student model is constructed based on the model structure determined by the sensor dataset. Knowledge from the teacher model is distilled into the student model to obtain a distilled student model corresponding to the water body assessment task. The student model can be deployed on the patrol device itself or its associated edge computing platform, continuously inputting sensor data and performing real-time inference during the patrol process, realizing online and continuous assessment of water body status, and effectively improving the timeliness of water body assessment. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the knowledge-based distillation-based water body assessment method of this application. Figure 2 This is a flowchart illustrating Embodiment Six of the knowledge-based distillation-based water body assessment method of this application; Figure 3 This is a flowchart illustrating Embodiment 8 of the knowledge-based distillation-based water body assessment method of this application; Figure 4 A flowchart illustrating the knowledge distillation-based water body assessment method of this application; Figure 5 This is a schematic diagram of the structure of a water body assessment device based on knowledge distillation, as described in an embodiment of this application.

[0019] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of this application and are not intended to limit this application.

[0021] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0022] Currently, some methods further incorporate multi-indicator weighting, expert weighting, or fuzzy comprehensive evaluation to generate a single comprehensive score or grade result. However, these methods are highly dependent on fixed-site data, have limited spatial coverage, and cannot meet the application requirements for high-frequency updates or real-time assessments. Furthermore, such comprehensive evaluations are typically conducted offline or semi-offline, relying on periodic data aggregation and manual processing, which also fails to meet the application requirements for high-frequency updates or real-time assessments.

[0023] While some research and engineering practices have incorporated high-information-volume sensing methods such as hyperspectral remote sensing, infrared imaging, sonar, or radar, these devices are bulky, costly, and require significant payload, power, and computing power. This makes them difficult to integrate into lightweight patrol devices like underwater robots or unmanned vessels, limiting their application in large-scale, long-term online patrol scenarios. Furthermore, the acquisition conditions for remote sensing and acoustic data are significantly affected by weather, water transparency, and the operating environment, making it difficult to guarantee data continuity and stability. Related evaluation models typically rely on complete high-modal inputs; once relevant modes are missing, the evaluation capability decreases significantly, making it difficult to adapt to application scenarios with limited sensing conditions in practical engineering.

[0024] Existing machine learning-based multimodal water body assessment methods face significant limitations in engineering applications, primarily due to their high dependence on data acquisition conditions and hardware. Current multimodal assessments typically rely on high-information data sources such as spectroscopy, remote sensing, acoustic detection, or manual semantic annotation. However, such data acquisition equipment is bulky, power-hungry, and complex to deploy, making it difficult to mount on lightweight patrol devices such as underwater robots and unmanned vessels, and also challenging to reliably acquire data during long-term, continuous patrols. Their practical applications are mostly limited to offline analysis and fixed platforms, making direct migration to mobile inspection scenarios that primarily utilize conventional water quality sensors.

[0025] Furthermore, some multimodal machine learning models have complex structures and large parameter scales, requiring high computing power, storage, and data integrity. Real-time operation and long-term stable deployment on edge devices still present certain challenges. When high-information modal data cannot be continuously acquired or is completely missing, existing models often struggle to maintain their original evaluation performance and lack the ability to reliably identify the overall state of water bodies under conditions of limited perception modalities.

[0026] The main solution of this application is as follows: Based on the water body assessment task, a multimodal dataset is determined from multiple candidate environmental data; the original model corresponding to the water body assessment task is trained based on the multimodal dataset to obtain the teacher model corresponding to the water body assessment task; based on the water body assessment task, a sensor dataset is determined from multiple candidate sensor data; a student model is constructed based on the model structure determined by the sensor dataset; the knowledge of the teacher model is distilled into the student model to obtain the distilled student model corresponding to the water body assessment task. The student model can be deployed on the cruise equipment itself or its associated edge computing platform, continuously inputting sensor data and performing real-time inference during the cruise, realizing online and continuous assessment of water body status, and has the technical effect of improving the timeliness of water body assessment.

[0027] It should be noted that the executing entity in this embodiment can be a water body assessment device based on knowledge distillation, or a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a water body assessment device based on knowledge distillation capable of performing the above functions. This embodiment does not specifically limit this. The following uses a water body assessment device based on knowledge distillation as the executing entity as an example to describe this embodiment and the following embodiments.

[0028] The water body assessment method based on knowledge distillation proposed in this application is applicable to application environments that meet one or more of the following conditions: existing routine water body monitoring data, such as dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, and COD (Chemical Oxygen Demand); access to high-information data such as spectral, remote sensing, infrared, and acoustic detection data, or access to historical archived data; access to manual inspection records, event annotations, or semantic description information; and the ability to conduct model training or updates in offline environments or on high-configuration platforms.

[0029] Based on this, Embodiment 1 of this application proposes a water body assessment method based on knowledge distillation. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the knowledge-based distillation-based water assessment method of this application. The knowledge-based distillation-based water assessment method includes steps S10-S50: Step S10: Based on the water body assessment task, determine the multimodal data set from multiple candidate environmental data.

[0030] Water body assessment tasks refer to assessment objectives for specific water body states or problems based on practical application needs, including but not limited to algal bloom risk assessment, pollution source tracing or anomaly source identification, and sediment state or disturbance risk assessment. Water body assessment tasks differ significantly in their focus; for example, algal bloom assessment emphasizes trophic status and optical characteristics, pollution source tracing focuses on pollutant sources and spatiotemporal evolution characteristics, and sediment-related assessments emphasize bottom structure and disturbance information. A multimodal dataset refers to a structured dataset containing multiple types of data from multiple candidate environmental data sources, selected and integrated for a specific water body assessment task, and originating from the same time and location. Multimodal datasets include water quality sensor parameter data, spectral or remote sensing image features, acoustic or radar detection features, and manual inspection or semantic annotation information, used to characterize the comprehensive discrimination rules of water body states.

[0031] In this embodiment, the water body assessment task is determined based on actual needs. Multiple candidate environmental data are selected, and all selected multimodal data are integrated to obtain a multimodal dataset. This adapts to the differentiated requirements of different water body assessment tasks, improving the task adaptability of the multimodal dataset.

[0032] As one implementation method, a multimodal dataset is determined based on the target environmental data of the water body assessment task.

[0033] Specifically, the target environmental data for water body assessment tasks should be analyzed. For example, the target environmental data for algal bloom risk assessment includes chlorophyll concentration, salinity, and spectral or remote sensing images of relevant water areas. The types of environmental data required for the water body's condition should be clearly defined. From candidate environmental data such as water quality sensor parameters, spectral or remote sensing images, acoustic or radar detection, and manual inspection and annotation, data related to the core assessment dimensions should be selected and integrated to form a multimodal dataset specific to this water body assessment task. By selecting multimodal data, redundancy introduced by irrelevant data is avoided, ensuring the relevance and effectiveness of the dataset.

[0034] As another implementation method, weights are assigned to different types of candidate environmental data according to the water body assessment task to determine a multimodal dataset.

[0035] Specifically, for water body assessment tasks, weights are assigned to different types of candidate environmental data based on domain expert experience or historical task data. For example, core water quality parameters have a weight of 0.4, remote sensing data 0.3, acoustic data 0.2, and manual inspection and annotation 0.1. From multiple candidate environmental data, all task-related candidate data types are extracted, and the candidate data are weighted and scored according to the task weight. Candidate data with weights greater than or equal to a preset threshold are selected and integrated to form a multimodal dataset. Through a dynamic weighting mechanism, data with high contribution to the task is retained first, balancing data comprehensiveness with assessment relevance.

[0036] Step S20: Train the original model corresponding to the water body assessment task based on the multimodal dataset to obtain the teacher model corresponding to the water body assessment task.

[0037] The teacher model refers to a high-performance evaluation model trained on multimodal data, directly associated with a specific water body assessment task, and possessing high-precision discrimination capabilities for water body states. This model can be pre-trained in an offline environment or on a high-configuration platform and supports periodic updates or continuous training with newly added multimodal data. The multimodal water body state discrimination rules and evaluation logic learned by the teacher model will be transferred to the student model through a distillation process.

[0038] In this embodiment, the multimodal dataset is preprocessed to construct an original model adapted to the water body assessment task. The preprocessed multimodal data is then input into the original model to obtain the teacher model corresponding to the water body assessment task. Through standardized processing of the multimodal dataset and targeted model training, a high-quality teacher model with multi-dimensional water body state discrimination capabilities is obtained, providing a core knowledge source for subsequent knowledge distillation.

[0039] As one implementation method, the original model is processed once to obtain the teacher model corresponding to the water body assessment task.

[0040] Specifically, an initial model matching the water body assessment task is constructed, and model training objectives are set. The multimodal dataset is divided into training, validation, and test sets, which are input into the initial model for training. The model's assessment accuracy is monitored using the validation set. When the model's accuracy reaches a preset threshold and convergence stabilizes, training is stopped, and the trained model is designated as the teacher model for this water body assessment task. Through the rich information in the multimodal data, the teacher model learns the comprehensive discrimination logic of water body states, possessing high-precision water body assessment capabilities.

[0041] As another implementation method, offline iterative training is performed to obtain the teacher model corresponding to the water assessment task.

[0042] Specifically, offline iterative training runs as a background process, not interfering with the online evaluation task of the cruise equipment. Triggering conditions can include: timed triggering, i.e., automatically starting a round of iterative training at a fixed period; event triggering, such as triggering iterative training when a significant drop in the performance of the evaluation model on the validation set is detected, or when a sufficient number of new multimodal data samples are collected; and version update triggering, such as actively starting iterative training to adapt to new conditions when a new sensor type is added to the system or the evaluation task definition changes. New multimodal water body data is acquired after the previous round of model training. The newly collected data is preprocessed to ensure it maintains the same format as the original historical data, forming a new multimodal dataset. The new multimodal dataset is merged with the original historical multimodal dataset to form a more comprehensive new multimodal dataset. The merged dataset can be reshuffled and divided into training, validation, and test sets to avoid introducing temporal bias. The teacher model obtained after the current training is used as a pre-trained model, and incremental training or fine-tuning is performed on it using the new multimodal dataset. Iterative training objectives are set, and the model performance is monitored using a new validation set during training. After training, the performance of the updated model is evaluated using a new test set. Iteration stops once the set iteration objective is met, and the iterated model is used as the new teacher model. By continuously optimizing the model through accumulating historical data, the problem of insufficient generalization ability of a single-training model is solved, enabling the teacher model to cover more water body scenarios and providing more comprehensive evaluation knowledge for knowledge distillation.

[0043] Step S30: Based on the water body assessment task, determine the sensor data set from multiple candidate sensor data.

[0044] In this embodiment, based on the water body assessment task, sensor data that is associated with the core dimensions of the water body assessment task and can be collected in real time by the patrol device is determined from multiple candidate sensor data, and integrated to form a sensor data set. This achieves alignment between the sensor data and the core dimensions of the water body assessment task and matches the final deployment environment perception capabilities, ensuring that the sensor data set can effectively carry the assessment knowledge of the teacher model and provide a highly adaptable input foundation for knowledge distillation.

[0045] As one implementation method, candidate sensors are pre-defined for water body assessment tasks, and a sensor data set is determined.

[0046] Specifically, a mapping table is predefined, which clearly specifies the candidate sensors corresponding to each water body assessment task. This mapping table is determined based on historical experience, domain expert knowledge, or previous offline research conclusions, and is directly invoked as prior knowledge in this process. Based on the water body assessment task data, the corresponding task category in the mapping table is located, and the corresponding candidate sensors are extracted, resulting in a candidate sensor set. Using the mapping table to determine candidate sensors significantly simplifies the sensor selection process. Simultaneously, the mapping table undergoes expert verification and offline testing to ensure a high degree of compatibility between candidate sensors and tasks, providing accurate and reliable low-dimensional input for subsequent knowledge distillation.

[0047] As another implementation method, the correlation between each candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task is calculated, and the candidate sensor data with high correlation is selected as the sensor dataset.

[0048] Specifically, the correlation between each candidate sensor data in the multimodal dataset corresponding to the water assessment task and the current water assessment task is calculated. The candidate sensor data is then traversed, and highly correlated candidate sensor data is selected and integrated into a sensor dataset. By quantifying the correlation between candidate sensor data and the water assessment task, highly correlated data is selected and integrated to form a sensor dataset that accurately meets the task requirements. This provides high-quality input for knowledge distillation, ensuring the assessment accuracy and practicality of the student model.

[0049] Step S40: Construct a student model based on the model structure determined by the sensor data set.

[0050] The student model refers to a lightweight assessment model specifically built to adapt to the real-time assessment scenarios of cruise equipment. Its input is limited to sensor data that the cruise equipment can reliably acquire. The cruise equipment must be able to reliably acquire at least one or more of the following types of sensor data: dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, and FDOM. Through knowledge distillation, the student model inherits the water body assessment knowledge and discrimination logic formed by the teacher model based on multimodal data. It does not rely on high-information modal data such as spectral, remote sensing, and acoustic data, and can receive sensor data in real time during cruise operations and output assessment results.

[0051] In this embodiment, the input and output characteristics of the sensor data set are determined, such as data dimension, type, and format. Combining the structure of an existing model or a newly designed network framework, a model structure adapted to this data set is constructed as a student model. The resulting student model is highly adapted to the sensor data, ensuring the stability and accuracy of the data input and output.

[0052] As one implementation method, the model structure is determined to be a student model based on the feature dimensions and temporal characteristics of the sensor data set.

[0053] Specifically, based on the feature dimensions and temporal characteristics of the sensor dataset, a suitable model structure is selected and adapted as the basic framework of the student model. For example, for non-time-series instantaneous data evaluation, lightweight fully connected neural networks, support vector machines, or decision tree ensemble models, such as LightGBM, can be used, with inputs being single-sampled multi-dimensional sensor data vectors. For time-series trend evaluation, computationally efficient recurrent neural networks, such as GRU, lightweight variants of LSTM, one-dimensional temporal convolutional neural networks, or Temporal Fusion Transformers, can be used, with inputs being sensor data sequences within a time window. This yields the student model. The model structure is determined based on the sensor data characteristics to ensure that the model can fully extract the effective information from the sensor data.

[0054] As another implementation method, the trained teacher model is extracted and lightweighted to obtain the student model.

[0055] Specifically, the core feature extraction and inference logic of the trained teacher model are extracted, and its key network layers and output dimensions are identified. Based on the input dimensions of the sensor dataset, the teacher model is proportionally lightweighted and pruned, such as by removing redundant hidden layers and reducing the number of neurons, ensuring that the student model's input layer matches the sensor data dimensions and its output layer is fully aligned with the teacher model. For the temporal and numerical characteristics of the sensor data, the inter-layer activation functions and connection methods are fine-tuned to complete the construction of the student model structure. By lightweighting and pruning the teacher model as a benchmark, a student model is constructed that precisely matches the dimensions of the sensor dataset and is fully aligned with the teacher model's inference logic. This ensures the efficient inheritance and transfer of knowledge from the teacher model to the student model, achieving a dual guarantee of model lightweighting and evaluation accuracy.

[0056] Step S50: Distill the knowledge from the teacher model into the student model to obtain the distilled student model corresponding to the water assessment task.

[0057] In this embodiment, the knowledge from the teacher model is distilled into the student model to obtain the student model corresponding to the water body assessment task. Through knowledge distillation, efficient transfer of multimodal assessment knowledge from the teacher model is achieved, enabling the student model to maintain assessment accuracy close to that of the teacher model even when relying solely on sensor data.

[0058] As one implementation method, knowledge distillation can include: output distillation, feature distillation, and attention distillation.

[0059] Specifically, for output distillation, the soft labels of the training samples from the teacher model are used as supervision signals. These soft labels represent the probability distribution of the model's output, for example, low risk of algal bloom: 0.1, medium: 0.2, high: 0.7. They contain more information than hard labels, such as "high". One of the training objectives of the student model is to minimize the difference between its output probability distribution and the soft labels of the teacher model, typically using the KL divergence loss function. This allows the student model to learn the smooth and informative decision boundaries of the teacher model.

[0060] Specifically, for feature distillation, the focus shifts from solely the final output to aligning the intermediate feature representations of the student model with the corresponding layers of the teacher model. Specifically, certain intermediate layers in the teacher model network are selected to make the corresponding layer outputs of the student model as close as possible in the feature space, using mean squared error or cosine similarity as the loss function. This allows the student model to learn how the teacher model extracts and constructs key features from the data, representing a deeper level of knowledge transfer.

[0061] Specifically, for attention distillation, the teacher model does not treat all input information equally when making judgments, but focuses on certain key features or key time points. For attention distillation of multi-dimensional sensor features, when the input is a single sample of multi-dimensional sensor data vector, the teacher model can use an attention mechanism layer or calculate gradient weights to generate a feature importance weight vector. This vector indicates which sensor parameters are the key indicators that the teacher model focuses on in this evaluation. A similar structure or method is introduced into the student model to generate its corresponding feature importance weight vector. A distillation loss function is designed, such as using mean squared error (MSE) or KL divergence, to force the student model's feature importance weight vector to match that of the teacher model. For example, in an algal bloom assessment task, the teacher model might highly focus on chlorophyll and dissolved oxygen. Through attention distillation, the student model will also learn to give these two features higher weights in the limited sensor data, rather than treating all parameters equally. Attention distillation for time-series data refers to the input of sensor data within a time window, such as sensor readings per minute over the past hour. For the teacher model, its temporal attention map can be obtained, which shows the degree of attention the model pays to data from past time points when making a state judgment at a given point in time. The student model learns from the teacher model's temporal attention map; that is, the student model learns to identify which key time segments or events have the greatest impact on the evaluation results. The difference between the student and teacher model temporal attention maps is minimized through loss calculation. For example, in a pollution source tracing task, the teacher model might pay particular attention to abnormal fluctuations in conductivity and COD in the period leading up to a pollution event. Through temporal attention distillation, the student model can also learn to capture these key temporal patterns, thus more accurately identifying pollution events.

[0062] As another implementation method, the knowledge distillation process can include offline and online methods.

[0063] Specifically, the knowledge distillation process can be conducted offline. After the teacher model completes training or offline iterative updates based on multimodal data, the student model is trained in one go using a fixed teacher model as the knowledge source through strategies such as output distillation and feature distillation. After training, the student model is independently deployed on the cruise equipment. Alternatively, the knowledge distillation process can be conducted online. The teacher model continuously undergoes online training and updates with newly added multimodal data. The distillation process is synchronized with the teacher model updates, and the newly added evaluation knowledge from the teacher model is transferred to the student model in real time through incremental distillation. This eliminates the need for the student model to retrain fully, ensuring that the evaluation capabilities of the student model remain dynamically consistent with the teacher model, adapting to the changing water conditions required in long-term cruise monitoring. Both offline and online knowledge distillation methods flexibly adapt to different application scenarios. Offline distillation can quickly build a high-precision student model adapted to the cruise equipment, while online distillation enables dynamic synchronization of the student model's evaluation capabilities with the teacher model, jointly ensuring the optimized implementation of multimodal evaluation knowledge in cruise scenarios.

[0064] This embodiment provides a water body assessment method based on knowledge distillation. First, based on the water body assessment task, a multimodal dataset is determined from multiple candidate environmental data. Then, an original model corresponding to the water body assessment task is trained using the multimodal dataset to obtain a teacher model for the task. Next, a sensor dataset is determined from multiple candidate sensor data based on the water body assessment task. A student model is constructed based on the model structure determined by the sensor dataset. Finally, the knowledge from the teacher model is distilled into the student model to obtain the distilled student model corresponding to the water body assessment task. The student model can be deployed on the patrol device itself or its associated edge computing platform, continuously inputting and inferring sensor data during the patrol process, achieving online and continuous assessment of the water body state, thus improving the timeliness of water body assessment.

[0065] Based on any of the above embodiments of this application, Embodiment 2 of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. Based on this, the step of determining a multimodal data set from multiple candidate environmental data based on the water body assessment task includes: Step S11: Determine the target environmental data from the candidate environmental data based on the water body assessment task.

[0066] Candidate environmental data refers to the collection of all raw data related to the aquatic environment that may be considered and acquired. This includes, but is not limited to, readings from various water quality sensors, satellite remote sensing imagery, UAV spectral data, sonar detection data, meteorological data, manual inspection reports, and records of historical pollution events. Target environmental data refers to a subset selected from the candidate environmental data based on a specific water body assessment task. For example, for an algal bloom assessment task, target environmental data refers to data closely related to algal growth.

[0067] In this embodiment, the current water body assessment task is clearly defined, such as algal bloom risk assessment, pollution source tracing or anomaly identification, and sediment condition or disturbance risk assessment. The types of candidate environmental data are identified, including raw data from water quality sensors, spectral or remote sensing images, acoustic or radar detection data, manual inspection records, or semantic annotations. Candidate environmental data associated with the corresponding water body assessment task are then selected to obtain the target environmental data. For example, the algal bloom risk assessment task will focus on target environmental data including chlorophyll concentration, salinity, and spectral or remote sensing images of the relevant water area; the sediment condition or disturbance risk assessment task will focus on acoustic detection data used to depict sediment topography and structure, water turbidity, flow velocity, and relevant manual inspection records. This achieves precise binding between tasks and data, avoiding the introduction of redundant information from irrelevant data and reducing the computational cost of subsequent model training. It also adapts to the differentiated needs of different water body assessment tasks, such as selecting different data for algal bloom assessment and pollution source tracing, improving the task adaptability of multimodal datasets.

[0068] Step S12: Extract data features from the target environment data.

[0069] Data features refer to the numerical representations of key attributes of the original data obtained after processing, transformation, and extraction from target environment data.

[0070] In this embodiment, feature extraction is performed on different types of target environmental data. For example, for water quality sensor data, sensor readings and statistical features such as mean, variance, and slope are extracted; for spectral or remote sensing image data, texture and spectral features are extracted; for acoustic or radar detection data, time-domain and frequency-domain features are extracted; for manual inspection or semantically labeled data, keyword features such as "black and smelly" and "floating matter" are extracted through natural language processing, as well as semantic classification features such as pollution type labels. Corresponding extraction methods are adopted for the characteristics of different types of data to maximize the retention of effective information in the data.

[0071] Step S13: Cluster the data features according to the data categories of the target environmental data to obtain at least one multimodal feature to form a multimodal data set. The multimodal features include water quality parameter features, image features, acoustic features, and semantic features.

[0072] Data categories refer to modal classifications based on data source, perception method, and physical properties, such as categories like water quality parameters, images, acoustics, and semantics. Multimodal features refer to a collection of multiple data features belonging to the same data category. For example, all features extracted from an image collectively constitute the image feature, a multimodal feature. A multimodal dataset consists of multiple samples, each containing multiple multimodal features from the same time and location. For example, a sample might simultaneously contain water quality parameter features, image features, and semantic features.

[0073] In this embodiment, pre-defined data category classification rules clearly define the feature attribution standards for water quality parameters, images, acoustics, and semantics. Based on the data source and physical meaning, data features are categorized into different modalities. For example, water quality parameter features include features from sensor data such as dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, and COD; image features include all visually relevant features extracted from spectra and remote sensing; acoustic features include sound wave-related features extracted from sonar, radar, and other detection devices; and semantic features include semantic information features extracted from text such as manual records and reports. All data features are clustered, with each cluster corresponding to a multimodal feature, such as water quality parameter features, image features, acoustic features, and semantic features. All features are integrated to form a complete multimodal data set. Clustering by data category makes feature classification clear, facilitating efficient fusion of multi-source information in subsequent teacher models.

[0074] In this embodiment, by accurately screening target environment data, extracting data features, and clustering them by category to form a multimodal data set, high-quality training data support with strong targeting, comprehensive information, and structure is provided for the teacher model.

[0075] Based on any of the above embodiments of this application, Embodiment 3 of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. Based on this, the steps of training the original model corresponding to the water body assessment task according to the multimodal dataset to obtain the teacher model corresponding to the water body assessment task include: Step S21: Input the multimodal dataset into the original model corresponding to the water body assessment task.

[0076] The original model refers to a model that has not been trained or whose parameters are in an initial random state before training begins. It does not have any water body assessment capabilities and is the predecessor of the teacher model.

[0077] In this embodiment, the multimodal dataset is divided into a training set, a validation set, and a test set according to a preset ratio. The training set is used for model parameter learning, the validation set is used for accuracy monitoring during training, and the test set is used for a final unbiased evaluation of the model's generalization ability after the entire training process is completed. The multimodal dataset is then input into the original model corresponding to the water body assessment task. By dividing the dataset into training, validation, and test sets, accuracy monitoring and overfitting prevention during model training are ensured.

[0078] Step S22: Based on the assessment objectives of the water body assessment task, determine the model structure of the original model.

[0079] Model structure refers to the architectural design of a machine learning model. This includes, but is not limited to: model type, such as convolutional neural network, recurrent neural network, Transformer, multilayer perceptron; number of network layers, i.e., how many layers there are; number of neurons or convolutional kernels in each layer; and modality fusion method, i.e., at what stage of the model and in what way data from different modalities are fused, such as by concatenation or attention mechanisms.

[0080] In this embodiment, the core assessment objectives of the water body assessment task are clearly defined, such as the risk level classification for algal bloom risk assessment and the disturbance risk prediction for sediment state assessment. An appropriate model architecture is selected for each assessment objective. For example, algal bloom risk assessment relies on spectral or remote sensing image data; therefore, the original model architecture might be a model with a convolutional neural network (CNN) as its backbone to extract image features, which are then fused with water quality parameter features. Pollution source tracing tasks may require analyzing time-series changes in pollutant concentrations; therefore, the original model architecture might include recurrent neural network (RNN) or Transformer modules to better handle time-series data. Determining the model structure based on the assessment objectives ensures that the original model focuses on the core requirements of the task, avoiding generalization of assessment logic or insufficient specificity caused by a generic model architecture, and improving the model's adaptability to the target task.

[0081] Step S23: Train and optimize the original model based on the multimodal dataset.

[0082] In this embodiment, for each input multimodal sample, the model outputs a prediction result, which is then compared with the true label. This comparison evaluates the actual value corresponding to the target, such as whether algal blooms actually occurred. A loss function is set, and the corresponding loss is calculated. Optimization algorithms such as gradient descent are used to backpropagate based on the loss value, updating the parameters of the original model so that its prediction continuously approximates the true situation. During the optimization iteration process, the model's accuracy, such as precision, recall, and mean squared error, is monitored in real time using a validation set. After completing one training cycle or a fixed number of iterations, the model parameters are temporarily frozen, and data samples from the validation set are input into the current model. Based on the model's performance on the validation set, it is determined whether the model is learning effectively or overfitting. This guides hyperparameter adjustment; for example, the learning rate may be adjusted or early stopping may be enabled based on the validation set performance. The model is then used to perform a final evaluation using a test set to obtain an unbiased estimate of its performance. By training and optimizing the original model, it is able to learn complex multimodal data and accurately map it to the final evaluation result.

[0083] Step S24: Use the trained and optimized original model as the teacher model for the water body assessment task.

[0084] In this embodiment, the core performance metrics of the trained and optimized model are verified to confirm that the metrics meet the preset evaluation criteria, thus obtaining the trained and optimized original model. A binding relationship is established between this original model and the water body assessment task, and the trained and optimized original model is marked as the teacher model corresponding to this task. Performance metric verification ensures the reliability of the teacher model's evaluation, prevents unqualified models from entering subsequent processes, and guarantees the effectiveness of knowledge distillation and the accuracy of subsequent water body assessments.

[0085] In this embodiment, by adapting to input multimodal data, customizing the model structure, and conducting targeted training and optimization, a high-quality teacher model that focuses on the core requirements of water body assessment tasks and has the ability to fuse and discriminate multimodal information is obtained, providing reliable knowledge source support for subsequent knowledge distillation.

[0086] Based on any of the above embodiments of this application, Embodiment 4 of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. In this regard, candidate environmental data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, FDOM, spectral data, remote sensing data, infrared data, acoustic detection data, manual inspection records, event annotation information, and semantic description information; candidate sensor data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, and FDOM.

[0087] In this embodiment, the candidate environmental data covers multiple dimensions of information such as water quality physicochemical properties, spectral remote sensing, acoustic detection, and artificial semantics. The candidate sensor data focuses on the core water quality parameters that the patrol equipment can acquire in real time. Both provide comprehensive information support for teacher model training and provide convenient input for student model reasoning, and together adapt to the needs of the entire process of water body assessment, including offline training and online evaluation.

[0088] Based on any of the above embodiments of this application, Embodiment 5 of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. Based on this, as one implementation method, the step of determining a sensor data set from multiple candidate sensor data based on a multimodal data set includes: Step S31: Determine the correlation between the candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task.

[0089] In this embodiment, candidate sensor data is extracted from the multimodal dataset corresponding to the water assessment task. Correlation calculation methods, such as information gain, Pearson correlation coefficient, and mutual information, are used. Taking the Pearson correlation coefficient method as an example, the Pearson correlation coefficient is calculated between each sensor data sequence, such as continuously collected pH values, and the water assessment task label, such as pollution level scores. For each candidate sensor data, the correlation with the water assessment task is calculated. This ensures that the sensor data input to the student model has strong features highly correlated with the task, avoiding the introduction of irrelevant or redundant features.

[0090] Step S32: Filter candidate sensor parameters with a correlation degree greater than or equal to the correlation degree threshold to form a sensor data set.

[0091] In this embodiment, a correlation threshold is set based on the requirements of the water body assessment task. This threshold can be dynamically adjusted. Candidate sensor data is traversed, and candidate sensor data with a correlation greater than or equal to the threshold are selected as the sensor data set. Sensors that contribute little or no to the current water body assessment task are filtered out, simplifying the structure of the student model and reducing the risk of overfitting.

[0092] As one implementation method, sensor parameters with a correlation degree greater than or equal to the correlation degree threshold for the water body assessment task are extracted to form a reference parameter set; for each candidate sensor data, the feature similarity between it and the corresponding parameter in the reference parameter set is calculated; based on the calculated feature similarity, target parameters with similarity higher than a preset threshold are selected from the candidate sensor data; and the selected target parameters are combined into a sensor data set.

[0093] Specifically, sensor parameters with a correlation degree greater than or equal to a correlation degree threshold are used as a reference parameter set. Core feature vectors are extracted from the corresponding parameters in both the sensor data and the reference parameter set. Quantification methods such as Pearson correlation coefficient and cosine similarity are used to calculate the feature similarity between a single sensor data point and its corresponding parameter in the reference parameter set. Considering data type and other factors, a preset similarity threshold is set, and target parameters with similarity higher than the preset threshold are selected. These selected target parameters are then combined into a sensor dataset. By constructing a sensor dataset that focuses on the core requirements of water body assessment tasks and possesses high data reliability, redundant and low-value data can be eliminated, ensuring the efficiency and accuracy of knowledge distillation, thus laying a data foundation for student models that combine assessment accuracy and engineering practicality.

[0094] In this embodiment, by quantifying the correlation between candidate sensor data and the water assessment task and selecting highly correlated candidate sensor parameters, a set of sensor data that accurately meets the core requirements of the task is formed, providing high-quality, low-dimensional input for subsequent knowledge distillation and ensuring that the student model has both assessment accuracy and engineering practicality.

[0095] Based on any of the above embodiments of this application, Embodiment Six of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. On this basis, Figure 2 This is a flowchart illustrating S31, S33, and S34 of the knowledge distillation-based water assessment method provided in Embodiment Six of this application. Following the step of determining the correlation between each candidate sensor data in the multimodal dataset and the water assessment task, the method includes: Step S33: Sort the candidate sensor data from largest to smallest according to the correlation to obtain the sorted candidate sensor data.

[0096] In this embodiment, all candidate sensor data are sorted in descending order of correlation score. During the sorting process, if candidate sensor data with the same correlation score exist, they can be sorted according to a preset priority to obtain the sorted candidate sensor data. Clearly defining the importance of sensors provides a clear basis for the next step of precise screening.

[0097] Step S34: Select a preset number of candidate sensor data as a sensor data set according to the sorting.

[0098] In this embodiment, considering the sensor capacity of the cruise equipment and the accuracy requirements of the task, a preset number of sensors is selected. From the sorted candidate sensor data, the top N data points are chosen, where N is the preset number. These top N candidate sensor data points form the sensor data set. By prioritizing the selection of the N most important sensors, the student model is ensured to receive the most discriminative information input.

[0099] In this embodiment, by sorting candidate sensor data in descending order of relevance and selecting a preset number of candidate sensor data, efficient and adaptable low-dimensional input is provided for knowledge distillation, ensuring a balance between lightweight deployment and accurate evaluation of student models.

[0100] Based on any of the above embodiments of this application, Embodiment Seven of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. On this basis, the knowledge of the teacher model is distilled into the student model to obtain the distilled student model corresponding to the water body assessment task, including: Step S51: Determine the knowledge distillation method for transferring knowledge from the teacher model to the student model. The knowledge distillation method includes any one of output distillation, feature distillation, and attention distillation.

[0101] In this embodiment, a suitable distillation method is selected based on the characteristics of the water body assessment task and the data features. The knowledge distillation method includes any one of output distillation, feature distillation, and attention distillation. Differentiated selection of distillation methods allows for targeted transfer of assessment knowledge from the teacher model, enhancing the relevance and completeness of knowledge transfer.

[0102] Step S52: Based on the determined knowledge distillation method, construct the corresponding knowledge distillation loss function.

[0103] In this embodiment, for output distillation, KL divergence loss is introduced into the loss function to measure the difference between the student model's output and the teacher model's soft labels; for feature distillation, MSE loss is introduced into the loss function to align the intermediate layer features of the student and teacher models; for attention distillation, cosine similarity loss is introduced into the loss function to constrain the attention response distribution of both models. When combining output distillation, feature distillation, and attention distillation, KL divergence loss, MSE loss, and cosine similarity loss are introduced, and corresponding weights are assigned to each loss term. The loss terms are then integrated through a weighted summation to obtain the corresponding knowledge distillation loss function. By designing the loss function, it is possible to control whether the student model more closely imitates the teacher model's experience or focuses more on fitting the relationship between limited sensor data and real labels, thereby finding the optimal balance point.

[0104] Step S53: Based on the multimodal dataset and the sensor dataset, the student model is jointly trained using the knowledge distillation loss function.

[0105] In this embodiment, a multimodal dataset is input into the teacher model, and a sensor dataset is input into the student model. The student model outputs a prediction result based on the sensor data, while the teacher model obtains an output based on the multimodal data, such as at least one of soft labels, intermediate layer features, and attention maps. For example, when output distillation is used, the teacher model outputs soft labels. The difference between the student model's prediction result and the teacher model's knowledge is calculated using a loss function. The teacher model parameters are fixed, and only the student model's parameters are updated to minimize the total loss. During training, control measures can be implemented, such as early stopping mechanisms and regularization strategies, to prevent overfitting. Iterative training continues until the model converges and stabilizes. Through joint training, the student model not only learns the evaluation results based on sensor data but also learns the more accurate judgment logic made by the teacher model based on richer information. Ultimately, the student model inherits the experience of the teacher model, but its own structure is simple, requiring only sensor data input to operate.

[0106] Step S54: The student model that has completed joint training is used as the distilled student model corresponding to the water assessment task.

[0107] In this embodiment, the student model after joint training is used as the distilled student model corresponding to the water assessment task. Only the student model needs to be deployed in the final cruise system. The student model runs independently, outputting high-quality assessment results relying solely on real-time sensor data.

[0108] This embodiment achieves the goal of compressing and transferring the knowledge or experience of complex multimodal evaluation models into a lightweight model. This enables resource-constrained patrol devices such as underwater robots and unmanned vessels to obtain comprehensive water body assessment capabilities that are close to those achieved using expensive and bulky multimodal equipment, even with only conventional water quality sensors. This significantly improves the real-time performance, convenience, coverage, and engineering feasibility of water body monitoring.

[0109] Based on any of the above embodiments of this application, Embodiment Eight of this application proposes a water body assessment method based on knowledge distillation, which can be referred to the above description and will not be repeated hereafter. Based on this, please refer to... Figure 3 , Figure 3 This is a flowchart illustrating Embodiment 8 of the knowledge distillation-based water assessment method of this application. Steps S50, S501, S502, and S503 are shown. After the step of distilling the knowledge from the teacher model to the student model to obtain the student model corresponding to the water assessment task, the following steps are included: Step S501: Determine the sensor combination corresponding to the water body assessment task.

[0110] In this embodiment, the water body assessment task type is obtained, such as algal bloom risk assessment, pollution source tracing or anomaly source identification, and sediment state or disturbance risk assessment. Based on the water body assessment task, sensors directly related to the water body state characterization are selected from the acquisition dimensions corresponding to the multimodal data. The multimodal data includes water quality sensor parameter data, spectral or remote sensing image features, acoustic or radar detection features, and manual inspection or semantic annotation information, used to characterize the comprehensive discrimination rules of the water body state. Based on the selected sensors, a sensor combination suitable for the task is determined. This achieves precise matching between the water body assessment task and the sensors, avoids blind sensor selection, reduces the cost of invalid data collection, adapts to the differentiated needs of different water body assessment tasks, and improves the versatility and adaptability of the method.

[0111] Step S502: Input the data set collected by the sensor combination into the student model corresponding to the water body assessment task. The student model is obtained by knowledge distillation of the pre-trained teacher model and is associated with the water body assessment task.

[0112] In this embodiment, the corresponding student model is invoked based on the type of water body assessment task currently being performed. The sensor data set is input into the corresponding student model, which performs feature extraction and inference calculations on the input data based on multimodal assessment knowledge obtained from distillation. The student model inherits the multimodal assessment knowledge and discrimination logic of the teacher model, maintaining high assessment accuracy while simplifying the input dimensions, thus balancing assessment convenience and reliability.

[0113] Step S503: Determine the water assessment result corresponding to the water assessment task based on the model output of the student model.

[0114] In this embodiment, the output of the student model is extracted. This output includes water body state category, risk level, or confidence level information, and can be output separately for different assessment tasks. When multiple assessment tasks are configured, each student model can run in parallel and output corresponding assessment results. The core information in the model output is accurately extracted, directly anchoring the core conclusions of the water body assessment, avoiding redundant information interference, and ensuring the relevance and clarity of the assessment results.

[0115] In this embodiment, by determining the sensor combination suitable for the water body assessment task, the data collected by the sensor is input into the student model obtained by knowledge distillation of the teacher model, and the water body assessment results that are accurate and adapted to the task requirements are quickly output, so as to realize efficient and reliable online assessment of water body status in the cruise scenario.

[0116] For example, to help understand the technical concept or principle of the knowledge-based distillation water assessment method in combination with the above embodiments, please refer to... Figure 4 , Figure 4The flowchart illustrating the knowledge distillation-based water assessment method of this application is as follows: First, during the training phase, an input system for both the teacher and student models is constructed. The teacher model receives multimodal data as input, including high-information data such as characteristics of new pollutants, spectral data, remote sensing images, semantic features, infrared detection, and acoustic radar. The student model receives only sensor data that can be acquired in real time by the patrol equipment, such as dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, and FDOM. Based on these inputs, the teacher model learns the water body state discrimination rules under multimodal conditions and transfers its evaluation capabilities to the student model through a knowledge distillation process. Knowledge distillation can employ different knowledge categories, such as output distillation and feature distillation, and supports organizational methods such as cross-modal distillation, multi-teacher distillation, online distillation, or offline distillation. After completing the distillation training, only the student model is retained for practical applications, enabling it to output comprehensive water body evaluation results without relying on real-time acquisition of multimodal data. Ultimately, the student model can be deployed in underwater robots, unmanned boats, and other automated water patrol devices equipped with sensor probes to achieve comprehensive water body assessment functions such as water quality early warning, algal bloom risk assessment, pollution source tracing, and sediment status assessment.

[0117] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the water body assessment method based on knowledge distillation in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0118] This application provides a knowledge-based distillation water assessment device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the knowledge-based distillation water assessment method in the first embodiment described above.

[0119] The following is for reference. Figure 5 The diagram illustrates a structural schematic of a knowledge-based distillation-based water assessment device suitable for implementing embodiments of this application. The knowledge-based distillation-based water assessment device in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, personal digital assistants (PDAs), tablets, and in-vehicle terminals, as well as fixed terminals such as digital TVs and desktop computers. Figure 5 The water assessment device based on knowledge distillation shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0120] like Figure 5As shown, the knowledge-based distillation water assessment device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the knowledge-based distillation water assessment device. The processing unit 1001, the ROM 1002, and the RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the knowledge-based distillation water assessment equipment to communicate wirelessly or wiredly with other devices to exchange data. Although the figure shows a knowledge-based distillation water assessment equipment with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0121] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0122] The knowledge-based distillation water assessment device provided in this application, employing the knowledge-based distillation water assessment method described in the above embodiments, can solve the technical problem that existing technologies struggle to meet the application requirements of high-frequency updates or real-time assessments. Compared with existing technologies, the beneficial effects of the knowledge-based distillation water assessment device provided in this application are the same as those of the knowledge-based distillation water assessment device provided in the above embodiments, and other technical features of this knowledge-based distillation water assessment device are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0123] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0124] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0125] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the knowledge-based distillation water assessment method in the above embodiments.

[0126] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, radio frequency (RF), or any suitable combination thereof.

[0127] The aforementioned computer-readable storage medium may be included in a knowledge-based distillation water assessment device; or it may exist independently and not assembled into a knowledge-based distillation water assessment device.

[0128] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a knowledge-distillation-based water assessment device, the knowledge-distillation-based water assessment device: determines a multimodal data set from multiple candidate environmental data based on the water assessment task; trains an original model corresponding to the water assessment task based on the multimodal data set to obtain a teacher model corresponding to the water assessment task; determines a sensor data set from multiple candidate sensor data based on the water assessment task; constructs a student model based on the model structure determined by the sensor data set; and distills the knowledge from the teacher model into the student model to obtain a distilled student model corresponding to the water assessment task.

[0129] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0130] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0131] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0132] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described knowledge-distillation-based water body assessment method. This solves the technical problem that existing technologies struggle to meet the application requirements of high-frequency updates or real-time assessments. Compared to existing technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the knowledge-distillation-based water body assessment method provided in the above embodiments, and will not be elaborated upon here.

[0133] This application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the knowledge-based distillation-based water assessment method described above.

[0134] The computer program product provided in this application can solve the technical problem that existing technologies are unable to meet the application requirements of high-frequency updates or real-time evaluation. Compared with the prior art, the beneficial effects of the computer program product provided in the embodiments of this application are the same as the beneficial effects of the water body assessment method based on knowledge distillation provided in the above embodiments, and will not be repeated here.

[0135] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent scope of this application.

Claims

1. A water body assessment method based on knowledge distillation, characterized in that, The knowledge-based distillation-based water body assessment method includes: Based on the water body assessment task, a multimodal data set was determined from multiple candidate environmental data. The original model corresponding to the water body assessment task is trained based on the multimodal dataset to obtain the teacher model corresponding to the water body assessment task. Based on the water body assessment task, a set of sensor data is determined from multiple candidate sensor data. A student model is constructed based on the model structure determined by the aforementioned sensor data set; The knowledge from the teacher model is distilled into the student model to obtain the distilled student model corresponding to the water body assessment task.

2. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, The step of determining a multimodal dataset from multiple candidate environmental data based on a water body assessment task includes: Based on the water body assessment task, target environmental data are determined from the candidate environmental data; Extract the data features of the target environment data; The data features are clustered according to the data categories of the target environmental data to obtain at least one multimodal feature to form the multimodal data set. The multimodal features include water quality parameter features, image features, acoustic features, and semantic features.

3. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, The step of training the original model corresponding to the water body assessment task based on the multimodal dataset to obtain the teacher model corresponding to the water body assessment task includes: Input the multimodal dataset into the original model corresponding to the water body assessment task; Based on the assessment objectives of the water body assessment task, the model structure of the original model is determined; Based on the multimodal dataset, the original model is trained and optimized; The original model, after training and optimization, is used as the teacher model for the water body assessment task.

4. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, The candidate environmental data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, FDOM, spectral data, remote sensing data, infrared data, acoustic detection data, manual inspection records, event annotation information, and semantic description information; the candidate sensor data includes at least one of dissolved oxygen, conductivity, pH, turbidity, chlorophyll, ammonia nitrogen, COD, salinity, and FDOM.

5. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, The step of determining the sensor data set from multiple candidate sensor data according to the water body assessment task includes: Determine the correlation between the candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task; Candidate sensor parameters with a correlation degree greater than or equal to the correlation degree threshold are selected as a sensor data set.

6. The water body assessment method based on knowledge distillation as described in claim 5, characterized in that, After determining the correlation between the candidate sensor data in the multimodal dataset corresponding to the water assessment task and the water assessment task, the following steps are included: The candidate sensor data are sorted from largest to smallest according to the correlation degree to obtain the sorted candidate sensor data. A preset number of candidate sensor data are selected as the sensor data set based on the sorting.

7. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, The step of distilling the knowledge from the teacher model into the student model to obtain the distilled student model corresponding to the water assessment task includes: Determine a knowledge distillation method for transferring knowledge from the teacher model to the student model, wherein the knowledge distillation method includes any one of output distillation, feature distillation, and attention distillation; Based on a defined knowledge distillation method, a corresponding knowledge distillation loss function is constructed. Based on the multimodal dataset and the sensor dataset, the student model is jointly trained using the knowledge distillation loss function; The student model that has completed joint training will be used as the distilled student model corresponding to the water assessment task.

8. The water body assessment method based on knowledge distillation as described in claim 1, characterized in that, After the step of distilling the knowledge from the teacher model into the student model to obtain the distilled student model corresponding to the water assessment task, the following steps are included: Determine the sensor combination appropriate for the water body assessment task; The data set collected by the sensor combination is input into the student model corresponding to the water body assessment task. The student model is obtained by knowledge distillation of the pre-trained teacher model, and the teacher model is associated with the water body assessment task. The water assessment result corresponding to the water assessment task is determined based on the model output of the student model.

9. A water body assessment device based on knowledge distillation, characterized in that, The knowledge-based distillation water assessment device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the knowledge-based distillation water assessment method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the knowledge-based distillation water assessment method as described in any one of claims 1 to 8.

Images