Robot sampling depth control antifouling method based on multi-source self-supervised active learning algorithm

Abstract

The invention provides a robot sampling depth control antifouling method based on a multi-source self-supervised active learning algorithm, and the method comprises the steps: obtaining multi-source data, such as environment parameters, postures, propulsion states and sampling port neighborhood fluid, constructing a self-supervised environment feature model, and extracting water source, flow field disturbance and attachment reentry features; calculating three indexes of a target depth water body matching degree, a different depth mixing degree and equipment disturbance brought-in pollution, forming antifouling evaluation, judging a sampling safety level and generating a sampling trigger condition; constraining approaching stage propulsion and attitude by combining the safety level and the evaluation result, determining sampling port opening time, duration and propulsion constraint according to trigger conditions after reaching a target depth range, executing sampling, and outputting sampling completion state data; the mixing risk is reduced, and the cleanliness and stability are improved.

Description

Technical Field

[0001] This invention relates to the field of robot sampling and pollution prevention control in stratified aquatic environments, and particularly to a robot sampling deep control and pollution prevention method based on a multi-source supervised active learning algorithm. Background Technology

[0002] Significant differences in water properties exist at different depths within stratified water bodies. High-purity sampling for hydrological, ecological, and pollution source tracing analyses often relies on sampling robots such as AUVs / ROVs. During descent and approach, propulsion disturbances and attitude changes can easily lead to the entrainment of water bodies at different depths, alterations in the flow field near the sampling port, and the re-entry of adhering substances, resulting in decreased sample representativeness. On-site monitoring often employs sensors such as CTDs, turbidity sensors, and dissolved oxygen sensors, along with flow velocity monitoring near the sampling port. However, environmental responses are time-varying and coupled, requiring coordinated sampling timing and control. Traditional manual or pre-planned triggering methods struggle to balance pollution prevention and efficiency, especially near thermoclines.

[0003] Existing technologies mostly employ preset descent trajectories and target depth points, selecting sampling layers based on CTD profiles and depth thresholds; upon approaching the target depth, sampling is triggered by fixed thresholds for depth error, turbidity, or local flow velocity, as well as short-term stability criteria; during the approach phase, thruster speed is typically limited, a fixed attitude is maintained, or the approach is slow; the control process is mostly a rule-based sequential control, namely "reaching depth—stabilizing—opening the valve—timed closing," and the sampling time and basic monitoring data are recorded; some also verify the consistency between the sampling window and the target layer through data playback after the mission.

[0004] The above scheme lacks a unified quantitative closed loop between judgment, control, and sampling: it cannot jointly characterize the source matching of water bodies at the target depth, mixing of water from different depths, and introduction of water due to equipment disturbances based on multi-source time series of robot actions and environmental responses, and it also lacks stability conditions across time windows to constrain triggering; proximity control and pollution prevention assessment are independent of each other, and propulsion and attitude are difficult to dynamically shrink with risk, making it difficult to coordinate the sampling timing and control actions when reaching the target depth.

[0005] Therefore, a robotic sampling method for deep-control and anti-contamination that can overcome the shortcomings of the existing technology is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a robot sampling depth control anti-fouling method based on a multi-source supervised active learning algorithm. In stratified water body operations, the aim is to utilize multi-source time series to simultaneously address water source, flow field disturbance, and re-entry of attached substances, thereby forming executable anti-fouling judgments and triggering conditions online, and applying closed-loop constraints to approach and control sampling, enabling the robot to reliably trigger within the target depth range and reduce external contamination.

[0007] According to an embodiment of the present invention, a robot sampling deep control anti-fouling method based on a multi-source supervised active learning algorithm includes:

[0008] S1. Operate a sampling robot in a stratified water environment to acquire environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port during the diving and approaching target depth process, and form multi-source state data in time sequence.

[0009] S2. Input the multi-source state data into the self-supervised environmental feature model. Within a limited time interval, the self-supervised environmental feature model simultaneously receives changes in robot actions and environmental responses. Under the constraint of predicting subsequent changes in environmental parameters, the multi-source state data is processed by feature extraction and encoding to obtain sampling environmental feature data reflecting the source of water, flow field disturbances, and re-entry of attached objects.

[0010] S3. Calculate the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value based on the sampling environment characteristic data. Combine the three values ​​to form pollution prevention assessment data. Under the condition of meeting the preset pollution prevention conditions, determine the sampling safety level and sampling trigger condition data based on the pollution prevention assessment data. The sampling trigger condition data is used to limit the relationship between the three values ​​when sampling is allowed.

[0011] S4. Calculate proximity control command data based on the sampling safety level. The proximity control command data is used to constrain the propulsion output and attitude adjustment of the sampling robot during the process of approaching the target depth.

[0012] S5. When the sampling robot reaches the target depth range, a sampling trigger determination is made based on the sampling environment feature data, the anti-fouling assessment data, the sampling trigger condition data, and the proximity control command data. When the anti-fouling assessment data meets the relationship defined by the sampling trigger condition data, sampling control command data is generated.

[0013] S6. Drive the sampling robot to perform sampling port opening, sample introduction and sealing actions according to the sampling control command data, and generate sampling completion status data as the output of the robot sampling deep control and anti-contamination method.

[0014] Optionally, S1 is as follows:

[0015] In a stratified water environment, the sampling robot is controlled to move continuously along the depth direction according to a preset diving trajectory. The time and depth information of the sampling robot at each preset sampling time interval are recorded to limit the time order and depth order of multi-source state data.

[0016] Within each preset sampling time interval, the environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port under the corresponding time and depth information are acquired simultaneously, and the environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port are combined into a single input vector in a fixed arrangement order.

[0017] The input vectors formed within multiple consecutive preset sampling time intervals are arranged sequentially in chronological order. The set of arranged input vectors is used as multi-source state data, and the multi-source state data is matched in terms of temporal resolution with the limited time interval and input layer requirements of the self-supervised environment feature model.

[0018] Optionally, S2 is as follows:

[0019] Multi-source state data are input into the input layer of the self-supervised environment feature model in chronological order within a limited time interval. At each time point, the input neurons in the input layer receive the environmental parameter values, robot posture values, propulsion device working status values, and fluid parameter values ​​near the sampling port from the corresponding input vectors, so as to form a sequence of input vectors arranged in chronological order.

[0020] At each time step, the input vector is weighted and summed and nonlinear function is performed by the feature neurons in the first feature layer. The correlation between environmental parameters, robot posture, propulsion device working state and fluid parameters near the sampling port at the same time step is encoded into an intermediate feature vector to express the instantaneous coupling state of robot action change and environmental response in the feature space.

[0021] The intermediate feature vector is input into the time recursion layer. The time recursion neurons in the time recursion layer simultaneously receive the intermediate feature vector at the current time and the output of the time recursion layer at the previous time. The recursive feature data is obtained through recursive operation, so that the recursive feature data represents the time relationship between the robot's motion trajectory and the continuous change of environmental parameters within a limited time interval.

[0022] The recursive feature data is input into the second feature layer. The feature neurons in the second feature layer perform weighted combination and nonlinear function operations on the recursive feature data to obtain the second feature data. This allows the recursive feature data to be compressed and recombined while maintaining the relationship between robot motion changes and environmental response time in the recursive feature data.

[0023] The second feature data is input into the output layer, and the output neurons in the output layer generate sampling environment feature data through linear transformation. The output neurons in the output layer are divided into three feature sub-vectors corresponding to water source-related features, flow field disturbance-related features, and attached re-entry-related features in a pre-fixed order, so that the sampling environment feature data can simultaneously express water source information, flow field disturbance information, and attached re-entry status at each time point.

[0024] During the training phase of the self-supervised environmental feature model, the model predicts environmental parameter data at a later time point within a defined time interval using sampled environmental feature data as input. The predicted environmental parameter data is then compared with the actual environmental parameter data. Based on the error between the predicted and actual environmental parameter data, the neuron connection weights in the input layer, first feature layer, time recursion layer, second feature layer, and output layer are iteratively optimized. This enables the self-supervised environmental feature model to complete feature extraction and encoding processing based on the supervision signal formed by multi-source state data within a defined time interval, and output sampled environmental feature data that reflects the source of water bodies, flow field disturbances, and re-entry of attached organisms.

[0025] Optionally, S3 specifically refers to:

[0026] The feature sub-vectors of the water source-related features in the sampled environmental feature data are extracted into the water source feature sequence. The typical environmental parameters of the target depth are converted into the target depth feature vector. The feature sub-vectors of the water source feature sequence at the current time are calculated with the target depth feature vector to obtain the target depth water matching degree value at the current time.

[0027] The feature vectors of the flow field disturbance-related features in the sampled environmental feature data are jointly analyzed with the fluid parameters near the sampling port within a limited time interval. Based on the amplitude change of the flow field disturbance-related features and the proportion of time when the target depth water body matching degree value is lower than the different depth judgment threshold, the value of the different depth water body mixing degree at the current time is calculated.

[0028] The feature vectors of the re-entry related features of the attached objects in the sampled environmental feature data are jointly analyzed with the working state of the propulsion device and the robot posture within a limited time interval. Based on the abrupt change amplitude of the re-entry related features of the attached objects and the synchronization relationship with the changes in the working state of the propulsion device and the changes in the robot posture, the degree of pollution brought in by the equipment disturbance at the current time is calculated.

[0029] The target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value calculated at the current time are combined into pollution prevention assessment data according to a preset sorting order.

[0030] Based on the comparison results between the pollution prevention assessment data and the target depth water body matching degree threshold, the different depth water body mixing degree threshold, and the equipment disturbance pollution degree threshold set in the preset pollution prevention conditions, the pollution prevention assessment data at the current time is classified and judged, and the state corresponding to the pollution prevention assessment data is divided into one of the sampling safety levels: prohibited sampling level, observation and adjustment level, and permitted sampling level.

[0031] Within a limited time interval, the anti-fouling assessment data and corresponding sampling safety levels at multiple consecutive time points are statistically analyzed. Based on the duration for which the target depth water body matching degree value is continuously higher than the target depth water body matching degree value threshold, and the duration for which the values ​​of the degree of mixing of water bodies at different depths and the degree of pollution brought in by equipment disturbance are continuously lower than their respective thresholds, the anti-fouling stability judgment result corresponding to the current time point is obtained.

[0032] Based on the antifouling stability assessment results and preset antifouling conditions, sampling trigger condition data is generated. The constraint rules in the sampling trigger condition data used to limit the relationship between the three values ​​when sampling is allowed are set as follows: the target depth water body matching degree value is greater than the target depth water body matching degree value threshold and the condition is continuously met within a preset time length; the values ​​of the degree of mixing of water bodies at different depths and the degree of pollution brought in by equipment disturbance are respectively less than their respective thresholds and the condition is continuously met within a preset time length. This allows the sampling trigger condition data to limit the relationship between the target depth water body matching degree value, the degree of mixing of water bodies at different depths, and the degree of pollution brought in by equipment disturbance when sampling is allowed in subsequent steps.

[0033] Optionally, S4 specifically refers to:

[0034] The sampling safety level is used as input. Based on the sampling safety level, the propulsion output limit parameters and attitude adjustment limit parameters are determined through a preset mapping relationship. The propulsion output limit parameters and attitude adjustment limit parameters are combined into a basic proximity control parameter set, which is used to limit the range of propulsion output amplitude and attitude adjustment amplitude that the sampling robot can generate under different sampling safety levels.

[0035] The difference between the depth where the sampling robot is located and the target depth is used as the state quantity of the approach stage. Based on the state quantity of the approach stage and the set of basic approach control parameters, the target value of the propulsion output and the target value of the attitude adjustment are calculated at each time step. The target value of the propulsion output and the target value of the attitude adjustment are combined into uncorrected approach control command data, so that the uncorrected approach control command data can drive the sampling robot to move towards the target depth under the condition of meeting the corresponding restrictions of the sampling safety level.

[0036] By jointly analyzing the uncorrected proximity control command data with the sampling environment characteristic data and pollution prevention assessment data, and based on the changing trends of the target depth water body matching degree value, the degree of mixing of water bodies at different depths, and the degree of pollution brought in by equipment disturbance relative to the threshold values ​​of the target depth water body matching degree value, the threshold value of mixing of water bodies at different depths, and the threshold value of pollution brought in by equipment disturbance, the propulsion output target value and attitude adjustment target value are corrected, and proximity control command data is generated. This ensures that the proximity control command data constrains the propulsion output and attitude adjustment of the sampling robot during the approach to the target depth while meeting the sampling safety level requirements.

[0037] Optional, S5 specifically includes:

[0038] The current depth of the sampling robot is compared with the target depth. When the difference between the two is less than the target depth judgment threshold, the current time is marked as reaching the target depth range. The propulsion output target value and attitude adjustment target value at the corresponding time in the proximity control command data are extracted as proximity state parameters.

[0039] Within a limited time interval after reaching the target depth range, the sampling environmental characteristic data and pollution prevention assessment data at each time point are compared with the sampling trigger condition data. Based on whether the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value simultaneously meet the threshold relationship set in the sampling trigger condition data, the sampling trigger judgment result at the corresponding time point is generated.

[0040] When the sampling trigger determination result indicates that the pollution prevention assessment data meets the sampling trigger condition data limit relationship, the proximity state parameters at the corresponding time moment are combined with the pollution prevention assessment data to calculate the sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling. The sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling are then combined into sampling control command data.

[0041] Optional, S6 specifically includes:

[0042] The sampling control command data is parsed into execution control parameters, including the sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling. At the sampling port opening time, control signals are sent to the sampling port actuator and propulsion device control module in the sampling robot according to the execution control parameters.

[0043] During the sampling port opening duration, the propulsion output of the propulsion device is limited according to the propulsion output constraint parameters during sampling, and the sampling port actuator is controlled to be in the open state, guiding the water body at the target depth into the sample storage cavity inside the sampling robot through the sampling port. When the sampling port opening duration reaches the sampling port opening duration set in the sampling control command data, the sampling port actuator is controlled to be in the closed state and the restriction on the propulsion output is lifted.

[0044] When the sampling port actuator changes from the open state to the closed state, the sampling port opening time, sampling port opening duration, and sampling port actuator closing state are correlated with the sampling environment characteristic data and anti-fouling assessment data at the corresponding time to generate sampling completion status data as the output of the robot sampling deep control anti-fouling method.

[0045] The beneficial effects of this invention are:

[0046] 1. This proposal puts forward an improved self-supervised environmental feature modeling and sampling determination method, which is oriented towards multi-source time series of environmental parameters, robot posture, propulsion state and sampling port neighborhood fluid. It adopts an input layer-feature layer-time recursive layer-compressed feature layer-output layer structure, and constructs a self-supervised signal by multi-step environmental parameter prediction. Combined with channel weight and time weighted training, the feature sub-vectors are divided at the output end according to water source, flow field disturbance and attachment re-entry. This improved technology encodes the instantaneous coupling and time dependence of action and environment without relying on manual annotation, so that it can stably give the characterization of source and disturbance under different layers and working conditions, supporting subsequent quantitative evaluation and triggering.

[0047] 2. This proposal presents a novel closed-loop method for pollution prevention assessment, triggering, and proximity control. Based on sampling environment characteristics, it calculates three indicators: target depth water body matching degree, degree of mixing with water from different depths, and pollution introduced by equipment disturbance. This forms a pollution prevention assessment and classifies sampling safety levels. Simultaneously, it performs stability checks across time windows to generate sampling trigger condition data. During the proximity phase, the safety level is mapped to amplitude limits for propulsion and attitude, and a safety attenuation coefficient is constructed based on the deviation of the three indicators from their relative thresholds to adjust commands in real time. After reaching the target depth range, the sampling port opening time, duration, and propulsion constraints during sampling are determined only when the triggering relationship is satisfied and the duration meets the criteria. This closed loop enables coordinated judgment and control, helping to reduce water body entrainment from different depths and the introduction of pollution by equipment disturbance, improving triggering reliability and sample representativeness within the target depth range.

[0048] 3. This proposal presents a comprehensive method for deep-contamination prevention sampling in stratified water bodies. On the data side, multi-source state data is constructed through fixed arrangement and time-depth alignment. Missing data substitution and labeling are set to ensure input stability and avoid interference from poor-quality fragments during training. On the execution side, sampling control command data uniformly drives the opening of the sampling port, sample introduction, and sealing, and generates sampling completion status data containing contamination prevention assessment and characteristics at the sampling time for retrospective and quality control. Unlike the rule-based sequential control of "to depth—stabilization—valve opening," this scheme adopts a chain design of "characteristics—assessment—control," transforming executable contamination prevention judgments into specific constraints at the approach and sampling stages, enhancing operability and process traceability in field applications. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 The flowchart shows a robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm proposed in this invention.

[0051] Figure 2 This is a flowchart illustrating the construction of multi-source state data in a stratified water environment for a robot sampling deep-control pollution prevention method based on a multi-source supervised active learning algorithm proposed in this invention.

[0052] Figure 3 This is a flowchart of a self-supervised environmental feature model for a robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm proposed in this invention.

[0053] Figure 4 This is a flowchart illustrating the proximity control command data generation process for a robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm proposed in this invention.

[0054] Figure 5 A sectional view of a layered water body using traditional methods;

[0055] Figure 6 This is a cross-sectional view of a stratified water body for deep-control pollution prevention sampling, based on a multi-source supervised active learning algorithm proposed in this invention.

[0056] Figure 7 This is a network structure diagram of a self-supervised environmental feature model for a robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figures 1 to 7 A robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm, characterized by comprising:

[0059] S1. Operate a sampling robot in a stratified water environment to acquire environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port during the diving and approaching target depth process, and form multi-source state data in time sequence.

[0060] S2. Input the multi-source state data into the self-supervised environmental feature model. Within a limited time interval, the self-supervised environmental feature model simultaneously receives changes in robot actions and environmental responses. Under the constraint of predicting subsequent changes in environmental parameters, the multi-source state data is processed by feature extraction and encoding to obtain sampling environmental feature data reflecting the source of water, flow field disturbances, and re-entry of attached objects.

[0061] S3. Calculate the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value based on the sampling environment characteristic data. Combine the three values ​​to form pollution prevention assessment data. Under the condition of meeting the preset pollution prevention conditions, determine the sampling safety level and sampling trigger condition data based on the pollution prevention assessment data. The sampling trigger condition data is used to limit the relationship between the three values ​​when sampling is allowed.

[0062] S4. Calculate proximity control command data based on the sampling safety level. The proximity control command data is used to constrain the propulsion output and attitude adjustment of the sampling robot during the process of approaching the target depth.

[0063] S5. When the sampling robot reaches the target depth range, a sampling trigger determination is made based on the sampling environment feature data, the anti-fouling assessment data, the sampling trigger condition data, and the proximity control command data. When the anti-fouling assessment data meets the relationship defined by the sampling trigger condition data, sampling control command data is generated.

[0064] S6. Drive the sampling robot to perform sampling port opening, sample introduction and sealing actions according to the sampling control command data, and generate sampling completion status data as the output of the robot sampling deep control and anti-contamination method.

[0065] In this embodiment, step S1 specifically includes:

[0066] The sampling robot is controlled according to a preset descent trajectory, which is generated and stored in advance by the controller. The preset descent trajectory contains multiple depth target points and corresponding descent speed constraints. The controller controls the sampling robot at preset sampling time intervals. The sampling robot moves continuously along the depth direction, driven by a time step. The robot is equipped with a depth sensor and a clock module. At the end of each preset sampling time interval, the clock module outputs the current time information. The current depth information is output by the depth sensor. The controller will send time information With depth information Storage is used to define the temporal order and depth-based correspondence of multi-source state data.

[0067] Preset sampling time interval The value is determined by the time interval specified by the self-supervised environment feature model and the input requirements of the input layer. Specifically, the length of the time interval is divided into several consecutive preset sampling time intervals and several consecutive time indices. Corresponding time information The time window used to input the self-supervised environment feature model is defined. The number of samplings included in the time window is consistent with the number of time steps that the time recursion layer of the self-supervised environment feature model can receive. In this way, the diving control cycle and depth acquisition cycle of the controller are unified with the time series modeling requirements of the self-supervised environment feature model, so that the multi-source state data matches the time resolution of the self-supervised environment feature model within the limited time interval.

[0068] Within each preset sampling time interval, the controller obtains time information. and depth information Subsequently, raw data at corresponding times are read from the environmental parameter acquisition module, attitude measurement module, propulsion device monitoring module, and fluid parameter acquisition module near the sampling port, respectively. The environmental parameter vector output by the environmental parameter acquisition module is denoted as... Environmental parameter vector This includes scalar channel values ​​such as temperature, conductivity, turbidity, and dissolved oxygen concentration to describe the environmental state of stratified water bodies. The robot attitude vector output by the attitude measurement module is denoted as... Robot pose vector This includes pitch angle, roll angle, yaw angle, and angular velocities for each attitude angle. The propulsion device operating state vector output by the propulsion device monitoring module is denoted as... Propulsion device operating state vector This includes the rotational speed command value, actual rotational speed feedback value, and thrust estimate value for each propulsion device. The fluid parameter vector near the sampling port output by the fluid parameter acquisition module near the sampling port is denoted as... Fluid parameter vector near the sampling port This includes the flow velocity, flow direction, local turbidity, and local temperature near the sampling port;

[0069] Controller in time index At the corresponding sampling time, the environmental parameter vector Robot posture vectors Propulsion device operating state vector and the fluid parameter vector near the sampling port By concatenating them sequentially in a fixed order, we obtain the input vector. The fixed arrangement order is set during the system design phase and maintains a correspondence with the order of input neurons in the input layer of the self-supervised environment feature model. The total number of values ​​included is denoted as This number is the same as the number of input neurons set in the first input layer of the self-supervised environment feature model, and the input vector... Each value in the vector corresponds to an input neuron. In this way, environmental parameters, robot posture, propulsion device operating status, and fluid parameters near the sampling port are combined into a single input vector at the same time. This allows the robot's actions and environmental responses to be reflected simultaneously in subsequent time series modeling.

[0070] The input vector formed within multiple consecutive preset sampling time intervals Arranged in ascending order of time index, forming the input vector sequence. The number of time steps contained within a defined time interval is denoted as . The controller extracts a length of [length] within each defined time interval. Continuous input vector sequence The input vector sequence is taken as a complete multi-source state data input segment. The input vector sequences are arranged sequentially in chronological order to form a multi-source state data set covering the entire descent process and the approach to the target depth. The sampling interval of the multi-source state data set on the time axis is determined by a preset sampling time interval. The sampling position on the depth axis is determined by the depth information corresponding to each time index. The decision, together with the other two, defines the temporal and depth order of the multi-source state data;

[0071] When a certain time index When data from a specific type of sensor is missing or significantly deviates from the preset reasonable range, the controller replaces the value with the most recent valid sampled data, or estimates it using a linear interpolation method based on data from adjacent time indices, to ensure the input vector... The dimensions and numerical range are stable. When the substitution method is adopted, the controller marks the replaced values ​​and removes the input vector sequence containing a large number of replaced values ​​during the training phase of the self-supervised environment feature model, thereby avoiding serious missing data from interfering with the self-supervised environment feature model's learning of the temporal and correlation relationships in multi-source state data.

[0072] Through the above implementation method, the multi-source state data constructed in step S1 not only includes time information and depth information, but also uniformly collects environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port within each preset sampling time interval, and combines the four types of parameters into an input vector in a fixed arrangement order. The continuous input vector sequence matches the time interval setting of the self-supervised environmental feature model in terms of time resolution and length, enabling the self-supervised environmental feature model to directly use multi-source state data as input in step S2 to perform time series modeling of robot action changes and environmental responses on the same time axis. This provides structurally complete and time-aligned basic data for subsequent extraction of sampled environmental feature data and calculation of target depth water body matching degree, degree of mixing of water bodies at different depths, and degree of pollution brought in by equipment disturbance.

[0073] In this embodiment, step S2 specifically includes:

[0074] In step S2, a self-supervised environmental feature model is used to perform time series modeling on multi-source state data. The self-supervised environmental feature model is connected in the following order: input layer, first feature layer, time recursion layer, second feature layer, and output layer. The input layer is set with a first number of input neurons, the first feature layer is set with a second number of feature neurons, the time recursion layer is set with a third number of time recursion neurons, the second feature layer is set with a fourth number of feature neurons, and the output layer is set with a fifth number of output neurons. The layers are connected by neuron connection weights. In the structural design of the self-supervised environmental feature model, multi-source state data is used as the only source of input data. The supervision signal is constructed by predicting the time change results of environmental parameter data, which is used to constrain the feature extraction and encoding process during the training phase.

[0075] In this embodiment, the input vector obtained by time sampling from the multi-source state data is denoted as... Time Index An integer, representing the nth time interval within a specified time range. Each preset sampling time interval corresponds to a time point and a depth position, and the input vector... The input vector sequentially lists environmental parameter values, robot posture values, propulsion device operating status values, and fluid parameter values ​​near the sampling port. The dimension is denoted as The number of input neurons set in the input layer is equal to the number of input neurons in the first number of input layers. The same, at every moment. The input vector is received by each input neuron in the input layer. The value at a fixed position is used to form a sequence of input vectors arranged in chronological order. ;

[0076] In the first feature layer, the time index is... input vector The second number of feature neurons in the first feature layer are input, and each feature neuron corresponds to the input vector. All The numerical values ​​are weighted and summed, and a non-linear activation function is applied to the weighted sum to obtain the time index of the feature neuron. The output below will index the second number of feature neurons in time. The corresponding output combination is the intermediate feature vector. intermediate feature vector The dimension is the second quantity, the intermediate feature vector Simultaneously encoding the correlation between environmental parameter values, robot posture values, propulsion device working status values, and fluid parameter values ​​near the sampling port at the same time point is used to express the instantaneous coupling state of robot motion changes and environmental response at that time point in the feature space.

[0077] In the time recursion layer, a third number of time recursion neurons are set up to index the time. Corresponding intermediate feature vector Input time recursion layer, and simultaneously index the time. The corresponding time recursion layer output vector is input to the time recursion layer, used to build memory units in the time direction inside the time recursion layer. The time recursion layer uses time indexing. The recursive feature data vector output below is denoted as Recursive feature data vector The dimension is the third quantity, and each time-recursive neuron has a time index. Lower receiving intermediate feature vector The recursive feature data vector of all components and the previous time index For each corresponding component, the output of the current component is generated through weighted summation and nonlinear activation function operations. By repeating the above process over the entire limited time interval, the feature data vector sequence is recursively calculated. It also includes the intermediate feature vectors of each time point within a defined time interval and the time dependencies between each time point, which are used to represent the time relationship between the robot's motion trajectory and the continuous changes in environmental parameters;

[0078] In the second feature layer, the time index is used. The corresponding recursive feature data vector The input consists of the fourth number of feature neurons in the second feature layer, each of which is associated with a recursive feature data vector. All components are weighted and combined, and a nonlinear activation function is applied to generate the second feature layer at the time index. The output of the fourth number of feature neurons is combined to form the second feature data vector. Second feature data vector The dimension is the fourth quantity, and the second feature layer maintains the recursive feature data vector. Given the already encoded relationship between robot motion changes and environmental response time, the recursive feature data is compressed and recombined to create a second feature data vector. Relatively recursive feature data vector in terms of dimension More compact, while retaining key information related to water sources, flow field disturbances, and reentry of deposits;

[0079] In the output layer, the second feature data vector The fifth output neuron in the input-output layer, each output neuron is connected to the second feature data vector. Perform a linear transformation operation on all components to obtain the output layer at the time index. The output value of the corresponding output neuron is used to determine the time index of the fifth number of output neurons. The output combination is the sampled environment feature data vector. Sampling environment feature data vector The dimension is the fifth quantity. According to a pre-fixed order, the output neurons in the output layer are divided into a first part of output neurons corresponding to water source-related features, a second part of output neurons corresponding to flow field disturbance-related features, and a third part of output neurons corresponding to attachment re-entry-related features. The outputs of the first part of output neurons are combined into a water source-related feature sub-vector. The outputs of the second part of the output neurons are combined into a flow field perturbation-related feature subvector. The outputs of the third part of the output neurons are combined into attachments and then fed into the relevant feature subvectors. In the time index Below, feature subvectors related to water body sources This is used to express the probability characteristics of fluid near the sampling port originating from water at the target depth or other depths; it is a flow field disturbance-related feature subvector. This is used to express the influence of changes in the operating state of the propulsion device on the velocity and direction of the fluid near the sampling port, and the re-entry of the adhering material into the relevant feature sub-vector. Used to describe the abnormal fluctuation characteristics of fluid parameters near the sampling port when the robot's posture changes rapidly;

[0080] During the training phase of the self-supervised environment feature model, the multi-source state data is divided into multiple limited time intervals, and the input vector sequence within each limited time interval is... Input self-supervised environment feature model, where This represents the number of time steps contained within each defined time interval. To avoid the direct superposition of errors with different physical dimensions between different environmental parameter channels, the environmental parameter data vector is channel normalized before calculating the prediction error, resulting in a standardized environmental parameter data vector sequence. The corresponding predicted environmental parameter data vectors are then subjected to the same normalization process to obtain a standardized data vector sequence of predicted environmental parameters. The normalized channel values ​​are dimensionless. The self-supervised environmental feature model outputs a sequence of sampled environmental feature data vectors corresponding to the time index range within the current limited time interval. Then, a multi-step prediction method is adopted, using the continuous predictions after the end of the current time interval. Each time step is a prediction interval. The prediction is performed using a standardized data vector of environmental parameters corresponding to the time index. The prediction result is denoted as... The standardized data vector of the environmental parameters after actual collection and normalization is denoted as... ,in From arrive The integer, denoted as the number of environmental parameter channels contained in the standardized environmental parameter data vector, is denoted as . To highlight the constraining effect of environmental parameter channels such as temperature and conductivity, which are closely related to the source and stratification of water bodies, channel weighting coefficients are set for each environmental parameter channel. Based on the above, a self-supervised prediction error scalar is constructed, and the self-supervised prediction error scalar is denoted as... It can be represented as:

[0081] ;

[0082] in, This represents a scalar of the self-supervised prediction error of the self-supervised environmental feature model within the current time and prediction intervals. Indicates the number of time steps within the prediction interval. This represents the number of environmental parameter channels in the standardized data vector of environmental parameters. This represents the time-weighted decay factor, which is between... and The dimensionless constants between them This represents the index of the predicted time step relative to the end of the current time interval. Indicates the first The channel weight coefficients for each environmental parameter channel are dimensionless constants. Indicates time index and the Standardized values ​​of predicted environmental parameters on each environmental parameter channel Indicates time index and the Standardized values ​​of real environmental parameters on each environmental parameter channel;

[0083] Employing a self-supervised prediction error scalar The gradient descent-like optimization method iteratively updates the neuron connection weights in the input layer, first feature layer, time recursion layer, second feature layer, and output layer, and adjusts the time weight decay factor. and channel weight coefficient The training process is fixed to preset values, and is conducted under the combined constraints of multi-step prediction, channel weighting, and time weighting. By repeating the above training process under different combinations of stratified water environments, different diving conditions, and different propulsion device operating states, the self-supervised environmental feature model adjusts the internal neuron connection weights based on the supervision signal generated by the multi-source state data itself, without relying on manual annotation. This ensures that the sampled environmental feature data vector formed within a limited time interval is optimized. Able to guarantee the standardization of environmental parameter data vectors in multi-step prediction While maintaining accuracy, it provides a stable representation of water sources, flow field disturbances, and re-entry of attached materials. During online operation, the multi-source state data collected in real time is input into the self-supervised environmental feature model in the manner described above. No further updates to neuron connection weights are performed; instead, the input vector sequence is forward-propagated using fixed neuron connection weights, directly outputting the sampled environmental feature data vector at the corresponding time index. This serves as the unified input data basis for calculating the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance in step S3.

[0084] In this embodiment, step S3 specifically includes:

[0085] In one specific implementation, step S3 is executed by the controller after obtaining the sampling environment feature data, and the controller indexes at each time step. The sampled environmental feature data vector is received from the output of the self-supervised environmental feature model. And extract water source-related feature subvectors from them. Flow field disturbance related feature vectors And the attachments then enter the relevant feature sub-vectors Among them, the feature subvectors related to water body sources The dimension is denoted as Flow field disturbance related feature vectors And the attachments then enter the relevant feature sub-vectors The dimensions of each dimension correspond to the number of output neurons in the corresponding part of the output layer. Simultaneously, the controller reads the typical environmental parameters corresponding to the target depth from the configuration storage. During the system calibration phase, these typical environmental parameters are converted into a target depth feature vector, which is denoted as... Target depth feature vector Dimensions and water source-related feature subvectors Dimensions The same, used as a feature template for the source of water at the target depth;

[0086] During the calculation of the target depth water body matching degree, the controller first uses the water body source-related feature sub-vectors on the continuous time index. To construct a water source feature sequence, and in order to weight the contributions of water source-related features across different dimensions and eliminate feature scale differences, the controller sets a channel weight coefficient for each dimension of the water source-related feature sub-vector during the system calibration phase. The channel weight coefficients of dimension are denoted as And determine the difference normalization factor based on the statistical results of historical samples. Differential normalization factor Used to limit the maximum reference value of the sum of feature differences, at the current time index. The controller will then generate feature subvectors related to the water body source. With target depth feature vector Perform a dimension-by-dimensional comparison, calculate the target depth water body matching degree value based on the weighted absolute difference form, and denote the target depth water body matching degree value as... Specifically, it is expressed as:

[0087] ;

[0088] in, Indicates time index The target depth water body matching degree value is a dimensionless value. The normalization factor representing the differences in water source-related characteristics is a positive dimensionless constant. Represents the feature subvectors related to water body sources The dimension is a positive integer. The first subvector representing the water body source-related features. The channel weight coefficients for each dimension are dimensionless constants. Indicates time index The feature subvectors related to the source of the wastewater body are in the th... Values ​​in each dimension The target depth feature vector is represented at the th... Values ​​in each dimension Indicates a time index. This represents the dimension index of the feature sub-vectors related to water body sources. The controller selects an appropriate index during the configuration phase. and the weighting coefficients of each channel The maximum possible value of the weighted absolute difference under typical operating conditions shall not exceed Thus Numerically located to Within the range, and based on the weighted difference between the water source-related features in the sampling environment feature data and the target depth feature vector, it can quantitatively reflect the degree to which the fluid near the sampling port originates from water bodies at the target depth or other depths;

[0089] During the calculation of the degree of mixing of water at different depths, the controller performs calculations for each time index. Select with To terminate the index, a time window with a preset mixing analysis window length is used to jointly analyze the flow field disturbance-related feature sub-vectors and the target depth water body matching degree values ​​within the time window. For each time index within the mixing analysis window... The controller will correlate the flow field disturbance with the relevant feature vectors. Feature vectors related to flow field perturbation at the previous time index Subtracting each dimension sequentially and taking the absolute value, then summing the absolute values ​​of each dimension along the dimensional direction yields the flow field disturbance amplitude indication. Finally, averaging the flow field disturbance amplitude indications across all time indices within the mixed analysis window yields the current time index. The controller simultaneously calculates the flow field disturbance intensity index and the water body matching degree value at the target depth within the mixing analysis window. The number of time indices below the differential depth determination threshold is divided by the length of the blending analysis window to obtain the differential depth determination time ratio. The controller then maps the flow field disturbance intensity index to a value between [a certain threshold] using linear normalization. and The dimensionless disturbance intensity value is then combined with the ratio of the depth-differential determination time according to a preset weighting coefficient to obtain the current time index. The numerical value of the degree of mixing of water bodies at different depths is denoted as , For the middle and The dimensionless values ​​between these values ​​are used to represent the degree to which non-target depth water components are involved in the fluid near the sampling port within a defined time interval.

[0090] During the calculation of the pollution level value brought about by equipment disturbance, the controller performs calculations for each time index. Select with To terminate the index, within a time window of a preset length for contamination analysis, the attachments within that time window are then entered into the relevant feature sub-vectors. The working status of the propulsion device and the robot's posture are jointly analyzed, and the index of each time point within the pollution analysis window is calculated. The controller calculates the attachments and then enters the relevant feature sub-vectors. The attachments from the previous time index re-enter the relevant feature sub-vectors The absolute values ​​of the differences in each dimension are summed to obtain the attachment re-entry mutation amplitude indicator. This indicator is then compared with a preset attachment re-entry mutation judgment threshold. If the attachment re-entry mutation amplitude indicator is greater than the threshold, the current time index is determined. There is a potential re-entry event; the controller indexes the time of the potential re-entry event. The process further calculates the absolute value of the difference between the propulsion device's working state vector and the previous time-indexed propulsion device's working state vector, and the absolute value of the difference between the robot's posture vector and the previous time-indexed robot posture vector. These differences are then compared with preset propulsion device working state change thresholds and preset robot posture change thresholds. When any change exceeds the corresponding threshold, the time index is adjusted. The time index is marked as the equipment disturbance trigger time index. The controller counts the number of equipment disturbance trigger time indices within the pollution analysis window, divides this number by the length of the pollution analysis window to obtain the equipment disturbance trigger ratio, and simultaneously averages and normalizes the re-entry mutation amplitude indicators of all attachments within the pollution analysis window to obtain the attachment re-entry intensity value. Then, the equipment disturbance trigger ratio and the attachment re-entry intensity value are combined according to preset weights to obtain the current time index. The equipment disturbance introduces a pollution level value, which is recorded as... , For the middle and The dimensionless values ​​between these values ​​are used to represent the risk level of attachments and residues entering the sample due to changes in the working state of the propulsion device and changes in the robot's posture.

[0091] During the construction of pollution prevention assessment data, the controller indexes at each time step. The target depth water body matching degree values ​​obtained above Numerical values ​​of the degree of mixing of water bodies at different depths The pollution level is introduced by equipment disturbance. The data are arranged into a three-dimensional vector according to a preset fixed order, and this three-dimensional vector is recorded as the pollution prevention assessment data vector. Pollution prevention assessment data vector The first dimension corresponds to the target depth water body matching degree value, the second dimension corresponds to the degree of mixing with water bodies at different depths, and the third dimension corresponds to the degree of pollution brought in by equipment disturbance. This is the pollution prevention assessment data vector. As the current time index A comprehensive description of the water source and equipment disturbance pollution in the sampling environment is used for subsequent sampling safety level determination and sampling trigger condition data generation;

[0092] During the sampling safety level determination process, the controller reads the target depth water body matching degree threshold, the cross-depth water body mixing degree threshold, and the equipment disturbance-induced pollution degree threshold from the preset anti-pollution conditions, and records these three values ​​as follows: , and For each time index The controller first uses the anti-fouling assessment data vector. A preliminary classification is performed by comparing the relationships between each component and its corresponding threshold. This classification considers the water body matching degree at the target depth. Less than Or the numerical value of the degree of mixing of water bodies at different depths. Greater than Or, equipment disturbance may introduce pollution levels into the numerical values. Greater than At that time, index the current time. The corresponding state is directly classified as a prohibited sampling level. For time indexes that do not meet the above prohibited conditions, The controller marks it as a candidate allowed sampling state, and then the controller indexes it at each time step. A time window of the same length as a preset safe smoothing window is selected around the current time index. The preliminary classification results within the time window are checked. If a candidate is allowed to be sampled and at least one time index is classified as prohibited from sampling within a safe smoothing window, then the current time index is... The corresponding status is divided into observation and adjustment levels, when the current time index When the candidate is allowed to sample and all time indices within the safe smoothing window are not classified as prohibited from sampling, the current time index is... The corresponding states are divided into allowed sampling levels, and the index at each time point is obtained through the above method. The sampling safety level below is denoted as: , The value range is the prohibited sampling level, the observation adjustment level, and the permitted sampling level;

[0093] During the antifouling stability assessment process, the controller evaluates the antifouling assessment data vectors from multiple consecutive time indices within a defined time interval. and sampling safety level To perform time statistics, the controller maintains a continuously satisfied condition counter that increments by time index. When the target depth water body matching degree value The numerical threshold of water body matching degree at or above the target depth And the numerical value of the degree of mixing of water bodies at different depths Less than or equal to the numerical threshold of the degree of mixing of water bodies at different depths Furthermore, equipment disturbance introduces pollution levels. Less than or equal to the numerical threshold of contamination introduced by equipment disturbance When a condition is not met, the continuous condition fulfillment counter is incremented by one. When any of the above conditions is not met, the continuous condition fulfillment counter is reset to zero. The controller compares the current value of the continuous condition fulfillment counter with the preset stability determination time step threshold. If the value of the continuous condition fulfillment counter is greater than or equal to the stability determination time step threshold and the current sampling safety level is met, the controller will proceed with the test. To allow sampling levels, index the current time. The corresponding antifouling stability assessment result is marked as a stable state; under other circumstances, the antifouling stability assessment result is marked as an unstable state, and the controller records the antifouling stability assessment result as... , The value can be either a stable state or an unstable state;

[0094] During the data generation process for the sampling trigger condition, the controller bases its analysis on the anti-fouling stability determination results. The controller generates sampling trigger condition data for subsequent steps based on preset antifouling conditions. During the system configuration phase, a preset time length is set and converted into a corresponding time step threshold. This threshold is used to limit the duration requirement of the relationship between the three values ​​during sampling. When generating the sampling trigger condition data, the controller ensures that the target depth water body matching degree value is greater than the target depth water body matching degree threshold within consecutive time steps. The number of time steps is not less than a preset time step threshold, and the value of the mixing degree of water bodies at different depths is less than the value threshold of the mixing degree of water bodies at different depths within consecutive time steps. The number of time steps is not less than the preset time step threshold, and the value of the contamination level introduced by the equipment disturbance is less than the threshold value of the contamination level introduced by the equipment disturbance within consecutive time steps. The three constraint rules—that the number of time steps is not less than a preset time step threshold—are written into the sampling trigger condition data. These constraint rules are then associated with the criteria that the sampling safety level is the permissible sampling level and the antifouling stability assessment result is a stable state. The sampling trigger condition data generated in this way can be used in subsequent steps to determine the target depth water body matching degree value during permissible sampling. Numerical values ​​of the degree of mixing of water bodies at different depths The pollution level is introduced by equipment disturbance. The relationship between them and their duration are defined to provide clear quantitative constraints for sampling trigger determination.

[0095] In this embodiment, step S4 specifically includes:

[0096] In one specific implementation, step S4 is executed by the controller after obtaining the sampling safety level, sampling environment characteristic data, and pollution prevention assessment data. The controller indexes at each time step. The sampling security level corresponding to the current time index is received below. It receives sampled environmental feature data vectors output by the self-supervised environmental feature model. Receive the antifouling assessment data vector calculated in step three. Pollution prevention assessment data vector The values ​​of water body matching at target depth are included in a fixed order. Numerical values ​​of the degree of mixing of water bodies at different depths The pollution level is introduced by equipment disturbance. The controller simultaneously obtains the current depth information from the depth sensor. Read target depth information from configuration storage The difference between the two is taken as the approximate stage state variable, denoted as... , used to represent the distance and direction of the sampling robot relative to the target depth;

[0097] The controller pre-establishes a mapping relationship between the sampling safety level and the propulsion output limit parameters and attitude adjustment limit parameters, denoted as follows: This parameter is used to limit the absolute value upper limit of the propulsion output under a certain sampling safety level. The attitude adjustment limit parameter is denoted as... This is used to limit the upper limit of the absolute value of attitude adjustment commands under a certain sampling safety level. During the system configuration phase, the controller sets a set of parameters for prohibited sampling levels, observation and adjustment levels, and permitted sampling levels. and The value that corresponds to the prohibited sampling level and Approaching zero is used to suppress propulsion and attitude changes; the corresponding adjustment levels are observed. and A moderate amplitude is chosen for slow approach and limiting attitude changes, allowing for sampling levels corresponding to... and A larger value is selected to improve proximity efficiency while meeting anti-fouling requirements; the controller indexes at each time step. Based on the current sampling safety level Query the mapping relationship to get the corresponding and The two are combined into a set of basic proximity control parameters to constrain the maximum amplitude of the proximity control command obtained in subsequent calculations.

[0098] The controller is based on the state quantities of the approach phase. The controller first determines the uncorrected proximity control command data calculated from the basic proximity control parameter set. The symbol, when If the target depth is determined to be below the current depth, it needs to be moved downwards. When the value is negative, the target depth is determined to be above the current depth, requiring upward movement. The controller employs a proportional depth approach strategy. Multiplying the value by the preset depth scaling factor yields the propulsion output reference value. Then, the absolute value of the propulsion output reference value is multiplied by the propulsion output limit parameter. Comparison, when the absolute value is greater than At that time, the output baseline value will be trimmed to the value from... Within the defined range, the target value for propulsion output is obtained, and this target value for propulsion output is denoted as... Regarding posture adjustment, the controller calculates the posture adjustment reference value based on the posture error between the current posture of the sampled robot and the preset ideal posture, and then compares the absolute value of the posture adjustment reference value with the posture adjustment limit parameter. Compare and crop to obtain the target value for attitude adjustment, and record this target value as . The controller is in the time index The next step is to advance the output target value. and attitude adjustment target value Combined, they form uncorrected proximity control command data, which is used to drive the sampling robot to approach the target depth along the depth direction and maintain attitude stability under the condition that the amplitude limit corresponding to the current sampling safety level is met;

[0099] The controller performs joint analysis of uncorrected proximity control command data, sampled environmental characteristic data, and pollution assessment data, and bases the analysis on the target depth water body matching degree value. Numerical values ​​of the degree of mixing of water bodies at different depths The pollution level is introduced by equipment disturbance. The controller quantitatively corrects the propulsion output target value and attitude adjustment target value based on the deviations from the target depth water body matching degree threshold, the threshold for the degree of mixing with water from different depths, and the threshold for the degree of contamination introduced by equipment disturbance. The controller reads the target depth water body matching degree threshold from the preset anti-fouling conditions. Numerical threshold for the degree of mixing of water bodies at different depths And the numerical threshold of pollution level introduced by equipment disturbance Weighting coefficients are assigned to the three types of deviations, and the weighting coefficients related to the water body matching degree at the target depth are denoted as follows: The weighting coefficients related to the numerical values ​​of the degree of mixing of water bodies at different depths are denoted as... The weighting coefficient related to the pollution level value brought in by equipment disturbance is denoted as... Index at the current time The controller according to , and The safety attenuation coefficient is calculated based on the normalized deviation from each threshold, and this safety attenuation coefficient is denoted as... Specifically, it is expressed as:

[0100] ;

[0101] in, Indicates time index The safety attenuation coefficient is a dimensionless value. This represents the weighting coefficient for the safety attenuation of the target depth water body matching degree value deviation, and is a dimensionless constant. The coefficient representing the deviation of the numerical value of water mixing at different depths from the safety attenuation is a dimensionless constant. This represents the weighting coefficient for the deviation of the pollution level introduced by equipment disturbance from the safety degradation, and is a dimensionless constant. This represents the numerical threshold for the water body matching degree at the target depth, and is a dimensionless value. This represents a numerical threshold indicating the degree of mixing of water bodies at different depths; it is a dimensionless value. This represents a numerical threshold indicating the degree of contamination introduced by equipment disturbance; it is a dimensionless value. Indicates time index The target depth water body matching degree value is a dimensionless value. Indicates time index The numerical values ​​for the degree of mixing of water bodies at different depths are dimensionless. Indicates time index The pollution level values ​​introduced by equipment disturbances are dimensionless. This indicates the operation of taking the larger value of the two real number parameters within the parentheses. Indicates a time index;

[0102] The controller is selected based on typical operating conditions during the configuration phase. , and The sum of the three weighted deviations mentioned above under the most unfavorable operating condition shall not exceed [a certain value]. This results in a safety attenuation coefficient. Always in and Between, when Below ,or Higher than ,or Higher than At that time, the corresponding deviation term passes through the safety attenuation coefficient. The controller applies attenuation to the uncorrected propulsion output target value and attitude adjustment target value, and applies it at the time index. The next step is to advance the output target value. and attitude adjustment target value Multiply by the safety attenuation factor respectively The corrected propulsion output command value and the corrected attitude adjustment command value are obtained, and the two are combined to form proximity control command data, which is output to the propulsion device control module and the attitude control module. In this way, the proximity control command data, under the premise of meeting the basic limit parameters corresponding to the sampling safety level, further dynamically shrinks the propulsion output and attitude adjustment amplitude based on real-time anti-fouling assessment data, so that the sampling robot can take into account both proximity efficiency and anti-fouling constraints when approaching the target depth.

[0103] In this embodiment, step S5 specifically includes:

[0104] In one specific implementation, step S5 is executed by the controller after obtaining proximity control command data, sampling environment characteristic data, anti-fouling assessment data, and sampling trigger condition data. The controller indexes at each time step. The system receives proximity control command data corresponding to the current time index, including propulsion output command values. and attitude adjustment command value Receives the sampled environmental feature data vector output by the self-supervised environmental feature model. Receive the antifouling assessment data vector obtained in step three. Pollution prevention assessment data vector The values ​​of water body matching at target depth are included in a fixed order. Numerical values ​​of the degree of mixing of water bodies at different depths The pollution level is introduced by equipment disturbance. The controller simultaneously obtains the current depth information from the depth sensor and indexes the current depth information in time. The following is recorded as Read the target depth information from the configuration storage and record the target depth information as... This is used to determine the target depth range in step S51;

[0105] The controller pre-sets the target depth determination threshold during the system configuration phase, and this threshold is denoted as... , A positive depth value is used to limit the maximum depth deviation that is allowed to be considered as having reached the target depth range, indexed at each time step. The controller calculates the current depth. Depth of target The difference between the two values ​​is calculated, and the absolute value of this difference is obtained. The absolute value is then compared with the target depth determination threshold. Compare, when the absolute value is not greater than At that time, the controller will index the current time. The time index marked as the time of arrival at the target depth range is denoted as . And extract the time index from the proximity control command data. Corresponding propulsion output command value and attitude adjustment command value The two are combined into proximity state parameters, which are used for subsequent sampling trigger determination and sampling control parameter calculation, for subsequent time indexing. The controller will only operate if the absolute value of the depth difference is not greater than a certain value. Under the premise of maintaining the marker within the target depth range, otherwise cancel the marker and stop sampling to trigger a judgment;

[0106] The controller obtains the time index for the first time. Subsequently, during the system configuration phase, the length of the time interval after reaching the target depth range is pre-defined, and the number of time indices contained in this time interval is recorded as follows: The controller Starting from the time index interval As the sampling trigger determination time interval, during the configuration phase, the controller generates sampling trigger condition data based on preset anti-fouling conditions, and records the lower limit threshold of the target depth water body matching degree value in the sampling trigger condition data. This lower limit threshold is denoted as... Record the upper limit threshold of the degree of mixing of water bodies at different depths, and denote this upper limit threshold as... Record the upper limit threshold of the pollution level introduced by equipment disturbance, and denote this upper limit threshold as... Simultaneously, record the minimum number of consecutive time indices required for all three values ​​to simultaneously satisfy the threshold relationship, and denote this number as... In the time index interval Within, for each time index The controller reads the corresponding anti-fouling assessment data vector. , obtain , and ,Will and Comparison, will and Comparison, will and Comparison, when both conditions are met Not less than , Not greater than and Not greater than When the condition is met, the controller will index the current time. Marking a condition as met for a single-step sampling trigger, the controller maintains a continuously satisfied counter to count the number of consecutively satisfied conditions. The number of time indices at which the single-step sampling trigger condition is continuously met, until the counter value is not less than [a certain value]. When the sampling trigger judgment result corresponding to the current time index is set to allow sampling, the sampling trigger judgment result is recorded as the allow sampling state. In other cases, the sampling trigger judgment result is set to the prohibit sampling state. The controller outputs the corresponding sampling trigger judgment result for each time index within the entire limited time interval.

[0107] When a certain time index When the sampling trigger determination result is that sampling is allowed, the controller records the time index as... index the time The actual physical time is converted to the sampling port opening time, which is used to write the sampling control command data. The controller presets the target sample volume during the configuration phase, and the target sample volume is recorded as... Preset or measure the sampling pump flow rate, and record the sampling pump flow rate as... By increasing the target sample volume Divide by sampling pump flow rate The basic opening duration of the sampling port is calculated. Based on this, the controller uses the time index... Corresponding pollution prevention assessment data vector In , and Adjust the base activation duration when near and or When the input approaches its respective upper limit threshold, the controller will appropriately shorten the duration of the sampling port opening. Significantly higher and and When the time is significantly lower than the corresponding upper limit threshold, the controller allows the sampling port to remain open for a duration close to the basic opening duration, and the controller also adjusts the time index accordingly. The corresponding proximity state parameters include the propulsion output command value. The absolute value of the parameter sets the propulsion output constraint parameter during the sampling period, and the propulsion output constraint parameter during the sampling period is denoted as... ,when or At higher levels, Set to no greater than The smaller value is used to further reduce the propulsion output during sampling, when and At a lower level, Set to near The value is used to ensure basic attitude stability. The controller combines the sampling port opening time, sampling port opening duration and propulsion output constraint parameters during sampling according to a preset format to generate sampling control command data. The sampling control command data is then sent to the sampling port actuator and propulsion device control module, so that the sampling port is opened within the time period that meets the sampling trigger condition data limit relationship, and the propulsion output is constrained during the sampling period, thereby completing the anti-fouling sampling control within the target depth range.

[0108] In this embodiment, step S6 specifically includes:

[0109] Step S6 is executed by the controller after obtaining the sampling control command data. Before executing step S6, the controller has already obtained the sampling control command data from the previous steps. The sampling control command data includes the sampling port opening time, the sampling port opening duration, and the advance output constraint parameters during sampling. The controller maintains the system time variable internally and records the system time variable as follows: , The real-time time under the current control cycle is recorded by the controller as the moment the sampling port is opened. , This indicates the point in time when the sampling port actuator needs to switch from the closed state to the open state, and the duration of the sampling port being open is denoted as... , This represents the time length from when the sampling port starts to open until it is ready to close. The output constraint parameters during the sampling period are denoted as... , The upper limit of the absolute value of the output allowed by the propulsion device during sampling is set. At the same time, the controller obtains the current state of the sampling port actuator from the previous steps and records the state of the sampling port actuator as an internal state flag to indicate whether the sampling port is currently in a closed or open state.

[0110] The controller parses the sampling control command data, and interprets the sampling port opening time as... The duration of the sampling port being open is parsed as The output constraint parameters during the sampling period are parsed as follows: The analysis results are then combined into a set of execution control parameters, and the controller adjusts the system time variable at a fixed control period. Update the control cycle duration as follows: , This represents the time interval between two adjacent control cycles. The controller compares... and The size relationship, when detected Not less than When the sampling port actuator is currently in the closed state, the controller sends an open control signal to the sampling port actuator through the sampling port control interface to set the sampling port actuator to the open state. Simultaneously, it sends a propulsion output constraint enable signal to the propulsion device control module and writes the necessary information into the propulsion device control module. This is used to limit the upper limit of the propulsion output within subsequent control cycles. When the controller switches the sampling port actuator from the closed state to the open state, it will... and Stored in the control cache;

[0111] After detecting that the sampling port actuator is in the open state, the controller starts the sampling port open duration timer, and the controller calculates the time based on the system time variable. With the sampling port opening time The difference determines the length of time the sampling port has been open. and The difference is less than the sampling port opening duration. At this time, the current time is considered to be within the sampling port opening duration. During the sampling port opening duration, the controller receives the propulsion output command value from the upper control module in each control cycle, and records the propulsion output command value as... , This represents the original propulsion output command without considering the propulsion output constraint parameters during sampling. The controller will... absolute value and When comparing, The absolute value is greater than At that time, Adjust to absolute value equal to And the propulsion output command value with the same symbol as the original command will be sent to the propulsion device control module. The absolute value is not greater than At that time, the controller directly The data is sent to the propulsion control module. Using the aforementioned method, the propulsion output of the propulsion device is limited according to the propulsion output constraint parameters during the sampling port's open duration. Simultaneously, the sampling port actuator remains open, allowing water at the target depth to enter the sample storage chamber inside the sampling robot through the sampling port. When the controller detects... and The difference is not less than When the sampling port opening duration has reached the requirement set in the sampling control command data, the controller sends a closing control signal to the sampling port actuator, setting the actuator to the closed state, and sends a propulsion output constraint release signal to the propulsion device control module, ensuring that the propulsion device is no longer subject to constraints in subsequent control cycles. Restrictions;

[0112] In the control loop where the sampling port actuator changes from the open state to the closed state, the controller records the sampling port opening time. Sampling port open duration The closing state of the sampling port actuator is recorded as key control information for the sampling operation. Simultaneously, the controller reads the sampling environment feature data at the time index corresponding to the instant the sampling port closes from the output buffer of the self-supervised environment feature model, and denotes this sampling environment feature data vector as... , This represents the sampling environment characteristic data vector corresponding to the moment the sampling port is closed. It reads the anti-fouling assessment data vector under the same time index from the anti-fouling assessment module and denotes this anti-fouling assessment data vector as... , This represents the anti-fouling assessment data vector corresponding to the moment the sampling port is closed. The controller will... , The closed status flag of the sampling port actuator and the sampling environment feature data vector. and pollution prevention assessment data vector The system performs associated storage to generate sampling completion status data. This data includes at least the sampling time information field, sampling control parameter field, sampling environment characteristic field at the sampling time, and anti-contamination assessment field at the sampling time. This data is used for retrospective analysis of the sampling process after the robot task is completed or uploaded to the upper-level monitoring system. After generating the sampling completion status data, the controller outputs it as the robot's sampling depth control anti-contamination method, marking the end of the current sampling operation. The system can then proceed to the next depth segment's approach control and sampling determination process according to the task plan.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm, characterized in that, include: S1. Operate a sampling robot in a stratified water environment to acquire environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port during the diving and approaching target depth process, and form multi-source state data in time sequence. S2. Input the multi-source state data into the self-supervised environmental feature model. Within a limited time interval, the self-supervised environmental feature model simultaneously receives changes in robot actions and environmental responses. Under the constraint of predicting subsequent changes in environmental parameters, the multi-source state data is processed by feature extraction and encoding to obtain sampling environmental feature data reflecting the source of water, flow field disturbances, and re-entry of attached objects. S3. Calculate the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value based on the sampling environment characteristic data. Combine the three values ​​to form pollution prevention assessment data. Under the condition of meeting the preset pollution prevention conditions, determine the sampling safety level and sampling trigger condition data based on the pollution prevention assessment data. The sampling trigger condition data is used to limit the relationship between the three values ​​when sampling is allowed. S4. Calculate proximity control command data based on the sampling safety level. The proximity control command data is used to constrain the propulsion output and attitude adjustment of the sampling robot during the process of approaching the target depth. S5. When the sampling robot reaches the target depth range, a sampling trigger determination is made based on the sampling environment feature data, the anti-fouling assessment data, the sampling trigger condition data, and the proximity control command data. When the anti-fouling assessment data meets the relationship defined by the sampling trigger condition data, sampling control command data is generated. S6. Drive the sampling robot to perform sampling port opening, sample introduction and sealing actions according to the sampling control command data, and generate sampling completion status data as the output of the robot sampling deep control and anti-contamination method.

2. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S1 specifically refers to: In a stratified water environment, the sampling robot is controlled to move continuously along the depth direction according to a preset diving trajectory. The time and depth information of the sampling robot at each preset sampling time interval are recorded to limit the time order and depth order of multi-source state data. Within each preset sampling time interval, the environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port under the corresponding time and depth information are acquired simultaneously, and the environmental parameters, robot posture, propulsion device working status and fluid parameters near the sampling port are combined into a single input vector in a fixed arrangement order. The input vectors formed within multiple consecutive preset sampling time intervals are arranged sequentially in chronological order. The set of arranged input vectors is used as multi-source state data, and the multi-source state data is matched in terms of temporal resolution with the limited time interval and input layer requirements of the self-supervised environment feature model.

3. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S2 specifically refers to: Multi-source state data are input into the input layer of the self-supervised environment feature model in chronological order within a limited time interval. At each time point, the input neurons in the input layer receive the environmental parameter values, robot posture values, propulsion device working status values, and fluid parameter values ​​near the sampling port from the corresponding input vectors, so as to form a sequence of input vectors arranged in chronological order. At each time step, the input vector is weighted and summed and nonlinear function is performed by the feature neurons in the first feature layer. The correlation between environmental parameters, robot posture, propulsion device working state and fluid parameters near the sampling port at the same time step is encoded into an intermediate feature vector to express the instantaneous coupling state of robot action change and environmental response in the feature space. The intermediate feature vector is input into the time recursion layer. The time recursion neurons in the time recursion layer simultaneously receive the intermediate feature vector at the current time and the output of the time recursion layer at the previous time. The recursive feature data is obtained through recursive operation, so that the recursive feature data represents the time relationship between the robot's motion trajectory and the continuous change of environmental parameters within a limited time interval. The recursive feature data is input into the second feature layer. The feature neurons in the second feature layer perform weighted combination and nonlinear function operations on the recursive feature data to obtain the second feature data. This allows the recursive feature data to be compressed and recombined while maintaining the relationship between robot motion changes and environmental response time in the recursive feature data. The second feature data is input into the output layer, and the output neurons in the output layer generate sampling environment feature data through linear transformation. The output neurons in the output layer are divided into three feature sub-vectors in a pre-fixed order, corresponding to water source-related features, flow field disturbance-related features, and attached re-entry-related features, so that the sampling environment feature data can simultaneously express water source information, flow field disturbance information, and attached re-entry status at each time point. During the training phase of the self-supervised environmental feature model, the model predicts environmental parameter data at a later time point within a defined time interval using sampled environmental feature data as input. The predicted environmental parameter data is then compared with the actual environmental parameter data. Based on the error between the predicted and actual environmental parameter data, the neuron connection weights in the input layer, first feature layer, time recursion layer, second feature layer, and output layer are iteratively optimized. This enables the self-supervised environmental feature model to complete feature extraction and encoding processing based on the supervision signal formed by multi-source state data within a defined time interval, and output sampled environmental feature data that reflects the source of water bodies, flow field disturbances, and re-entry of attached organisms.

4. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S3 specifically refers to: The feature sub-vectors of the water source-related features in the sampled environmental feature data are extracted into the water source feature sequence. The typical environmental parameters of the target depth are converted into the target depth feature vector. The feature sub-vectors of the water source feature sequence at the current time are calculated with the target depth feature vector to obtain the target depth water matching degree value at the current time. The feature vectors of the flow field disturbance-related features in the sampled environmental feature data are jointly analyzed with the fluid parameters near the sampling port within a limited time interval. Based on the amplitude change of the flow field disturbance-related features and the proportion of time when the target depth water body matching degree value is lower than the different depth judgment threshold, the value of the different depth water body mixing degree at the current time is calculated. The feature vectors of the re-entry related features of the attached objects in the sampled environmental feature data are jointly analyzed with the working state of the propulsion device and the robot posture within a limited time interval. Based on the abrupt change amplitude of the re-entry related features of the attached objects and the synchronization relationship with the changes in the working state of the propulsion device and the changes in the robot posture, the degree of pollution brought in by the equipment disturbance at the current time is calculated. The target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value calculated at the current time are combined into pollution prevention assessment data according to a preset sorting order. Based on the comparison results between the pollution prevention assessment data and the target depth water body matching degree threshold, the different depth water body mixing degree threshold, and the equipment disturbance pollution degree threshold set in the preset pollution prevention conditions, the pollution prevention assessment data at the current time is classified and judged, and the state corresponding to the pollution prevention assessment data is divided into one of the sampling safety levels: prohibited sampling level, observation and adjustment level, and permitted sampling level. Within a limited time interval, the anti-fouling assessment data and corresponding sampling safety levels at multiple consecutive time points are statistically analyzed. Based on the duration for which the target depth water body matching degree value is continuously higher than the target depth water body matching degree value threshold, and the duration for which the values ​​of the degree of mixing of water bodies at different depths and the degree of pollution brought in by equipment disturbance are continuously lower than their respective thresholds, the anti-fouling stability judgment result corresponding to the current time point is obtained. Based on the antifouling stability assessment results and preset antifouling conditions, sampling trigger condition data is generated. The constraint rules in the sampling trigger condition data used to limit the relationship between the three values ​​when sampling is allowed are set as follows: the target depth water body matching degree value is greater than the target depth water body matching degree value threshold and the condition is continuously met within a preset time length; the values ​​of the degree of mixing of water bodies at different depths and the degree of pollution brought in by equipment disturbance are respectively less than their respective thresholds and the condition is continuously met within a preset time length. This allows the sampling trigger condition data to limit the relationship between the target depth water body matching degree value, the degree of mixing of water bodies at different depths, and the degree of pollution brought in by equipment disturbance when sampling is allowed in subsequent steps.

5. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S4 specifically refers to: The sampling safety level is used as input. Based on the sampling safety level, the propulsion output limit parameters and attitude adjustment limit parameters are determined through a preset mapping relationship. The propulsion output limit parameters and attitude adjustment limit parameters are combined into a basic proximity control parameter set, which is used to limit the range of propulsion output amplitude and attitude adjustment amplitude that the sampling robot can generate under different sampling safety levels. The difference between the depth where the sampling robot is located and the target depth is used as the state quantity of the approach stage. Based on the state quantity of the approach stage and the set of basic approach control parameters, the target value of the propulsion output and the target value of the attitude adjustment are calculated at each time step. The target value of the propulsion output and the target value of the attitude adjustment are combined into uncorrected approach control command data, so that the uncorrected approach control command data can drive the sampling robot to move towards the target depth under the condition of meeting the corresponding restrictions of the sampling safety level. By jointly analyzing the uncorrected proximity control command data with the sampling environment characteristic data and pollution prevention assessment data, and based on the changing trends of the target depth water body matching degree value, the degree of mixing of water bodies at different depths, and the degree of pollution brought in by equipment disturbance relative to the threshold values ​​of the target depth water body matching degree value, the threshold value of mixing of water bodies at different depths, and the threshold value of pollution brought in by equipment disturbance, the propulsion output target value and attitude adjustment target value are corrected, and proximity control command data is generated. This ensures that the proximity control command data constrains the propulsion output and attitude adjustment of the sampling robot during the approach to the target depth while meeting the sampling safety level requirements.

6. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S5 specifically refers to: The current depth of the sampling robot is compared with the target depth. When the difference between the two is less than the target depth judgment threshold, the current time is marked as reaching the target depth range. The propulsion output target value and attitude adjustment target value at the corresponding time in the proximity control command data are extracted as proximity state parameters. Within a limited time interval after reaching the target depth range, the sampling environmental characteristic data and pollution prevention assessment data at each time point are compared with the sampling trigger condition data. Based on whether the target depth water body matching degree value, the degree of mixing of water bodies at different depths value, and the degree of pollution brought in by equipment disturbance value simultaneously meet the threshold relationship set in the sampling trigger condition data, the sampling trigger judgment result at the corresponding time point is generated. When the sampling trigger determination result indicates that the pollution prevention assessment data meets the sampling trigger condition data limit relationship, the proximity state parameters at the corresponding time moment are combined with the pollution prevention assessment data to calculate the sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling. The sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling are then combined into sampling control command data.

7. The robot sampling deep-control anti-fouling method based on a multi-source supervised active learning algorithm according to claim 1, characterized in that, S6 specifically refers to: The sampling control command data is parsed into execution control parameters, including the sampling port opening time, sampling port opening duration, and propulsion output constraint parameters during sampling. At the sampling port opening time, control signals are sent to the sampling port actuator and propulsion device control module in the sampling robot according to the execution control parameters. During the sampling port opening duration, the propulsion output of the propulsion device is limited according to the propulsion output constraint parameters during sampling, and the sampling port actuator is controlled to be in the open state, guiding the water body at the target depth into the sample storage cavity inside the sampling robot through the sampling port. When the sampling port opening duration reaches the sampling port opening duration set in the sampling control command data, the sampling port actuator is controlled to be in the closed state and the restriction on the propulsion output is lifted. When the sampling port actuator changes from the open state to the closed state, the sampling port opening time, sampling port opening duration, and sampling port actuator closing state are correlated with the sampling environment characteristic data and anti-fouling assessment data at the corresponding time to generate sampling completion status data as the output of the robot sampling deep control anti-fouling method.

Images