Underwater glider energy consumption modeling and residual life prediction method based on digital twinning

Abstract

The invention discloses an underwater glider energy consumption modeling and residual life prediction method based on digital twinning, and belongs to the technical field of underwater robots. Deterministic parameters and time-varying parameters of an energy consumption model are constructed and updated by fusing simulated running test and sea test data, and a self-adaptive updating strategy based on a genetic algorithm is designed for degradation coefficients in the time-varying parameters, so that the model is evolved along with a physical entity. In the residual life prediction application of the underwater glider, the residual life prediction method based on domain adversarial migration is provided for solving the problem that the model generalization ability is weak due to the data distribution difference between a historical model machine and an in-service model machine by taking the residual electricity percentage of an energy system as a health index. And the prediction precision of the residual life of the in-service underwater glider can be effectively improved. The method breaks through a traditional recycling decision mode dominated by human experience, and important technical support is provided for energy efficiency early warning and maintenance plan adjustment of an in-service model machine.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of underwater robots, and particularly relates to a residual life prediction method for underwater gliders based on digital twinning. BACKGROUND

[0002] An underwater glider is a long-endurance, unmanned and autonomous sea sample three-dimensional observation device. With the continuous breakthroughs of China's underwater glider in low-power and long-endurance technology, its endurance capability has realized a leap from "months" to "years". However, for the underwater glider that sails in complex marine environments for a long time, its energy consumption characteristics are affected by nonlinear time-varying factors during service, showing significant individual differences, which increases the difficulty of high-precision residual life prediction. At present, the life prediction of the underwater glider only relies on the subjective experience of the operator, which not only causes the waste of the energy of the underwater glider, but also may even lead to the loss of the prototype.

[0003] The energy consumption model of the underwater glider provides an important reference for optimizing the control parameters and improving the energy utilization rate during operation, and also provides important data support for the residual power assessment. At present, two methods commonly used to construct the energy consumption model of the underwater glider are based on physical models and data driving. The construction method based on the physical model has strong interpretability, but according to the energy consumption monitoring results, the single profile energy consumption characteristics of the underwater glider are affected by multiple uncertain factors such as biofouling and structural wear, showing time-varying nonlinear characteristics. The physical model alone is difficult to accurately model these multiple uncertain factors, limiting the simulation accuracy. The data-driven modeling method does not have sufficient data support for accurate energy consumption estimation because the cost of obtaining sea trial energy consumption data is high and it is difficult to cover the full working condition energy consumption characteristics of all types of underwater gliders.

[0004] Looking back at the research status of residual life prediction technology, in terms of implementation, it is also divided into two types based on physical models and data driving. However, the physical model has insufficient modeling capability for the time-varying nonlinear degradation characteristics of complex systems, and there are too many assumptions and simplifications, which restrict the prediction accuracy of the model. The data-driven method is limited by the lack of sample data throughout the life cycle and the difference in data distribution between the historical prototype (source domain) and the in-service prototype (target domain), which seriously affects the prediction performance of the model for the target domain. In terms of research objects, it mainly focuses on rotating mechanical parts, such as bearings, hydraulic pumps, aircraft engines and lithium-ion batteries, and there is no residual life prediction technology system suitable for marine unmanned equipment for the residual life estimation of underwater gliders sailing in complex marine environments. The existing maintenance and recovery decision-making still remains at the experience-dominated stage, lacks scientific and quantitative prediction methods, and the device loss risk prevention and control capability needs to be improved. SUMMARY

[0005] In view of the problems that the underwater glider has large individual differences in energy consumption characteristics in complex marine environment, modeling and data support are insufficient, resulting in difficulty in high-precision prediction of residual life, maintenance decision relies on experience, and there is risk of energy waste and equipment loss, the application provides an underwater glider energy consumption modeling and residual life prediction method based on digital twinning.

[0006] The application is implemented in the following manner.

[0007] Step 1: Build a hardware-in-the-loop simulation platform to simulate the load and power supply of each energy consumption unit during the underwater glider run, and obtain the power characteristics and flow characteristics parameters of the energy consumption unit;

[0008] Step 2: Extract limited time-varying parameters related to energy consumption on the onboard end, aggregate and then return to the twin database through satellite;

[0009] Step 3: Based on the data of step 1 and step 2, establish a high-fidelity digital twin energy consumption model of the underwater glider;

[0010] Step 4: Periodically identify and update the degradation coefficient in the power unit energy consumption model using a genetic algorithm to compensate for the energy consumption calculation error caused by transmission performance degradation, and realize the adaptive evolution of the twin energy consumption model with the entity energy consumption characteristics;

[0011] Step 5: Construct source domain data set and target domain data set, wherein the source domain contains historical prototype data, and the target domain contains in-service prototype data;

[0012] Step 6: Extract time series degradation features using a long short-term memory network (LSTM), construct a loss function containing domain classification loss and regression prediction loss, optimize the performance of the field discriminator and the source domain regressor through adversarial training, and train the residual life prediction network;

[0013] Step 7: Based on the trained residual life prediction model, accurately predict the residual life of the in-service underwater glider.

[0014] In the above technical solution, preferably, in step 1, the hardware-in-the-loop simulation platform comprises:

[0015] An attitude adjusting unit simulates voltage changes of the energy system through an adjustable voltage stabilizer, simulates pitch angle adjustment through an electronic compass and a clamp self-locking device, and measures the relationship between power and pitch angle and voltage;

[0016] The buoyancy adjustment unit adjusts the energy system voltage through an adjustable voltage regulator, simulates seawater pressure through an adjustable overflow valve, and measures the power and flow characteristics of the oil discharge stage through a flow meter and a pressure gauge.

[0017] The energy-consuming electrical unit adjusts the energy system voltage through an adjustable voltage regulator, measures the function of the communication positioning unit power with voltage, and fixes the static power of the navigation unit and the task detection unit (the power changes slightly with voltage).

[0018] In the above technical solution, preferably, the limited time-varying parameters in step 2 include:

[0019] The time consumption of each stage of single profile operation, the average pitch angle of the diving / ascending gliding stage, the pitch adjustment time, the cumulative roll adjustment time, the average depth of oil discharge in the diving / ascending conversion stage, and the average voltage of the battery in the diving / ascending conversion stage and the non-diving / ascending conversion stage.

[0020] In the above technical solution, preferably, the expression of the high-fidelity digital twin energy consumption model in step 3 is:

[0021]

[0022] In the formula, k bat is the reciprocal of the discharge efficiency of the underwater glider energy system, E gm is the attitude adjustment unit single profile energy consumption, E c is the control navigation unit single profile energy consumption, E b is the buoyancy adjustment unit single profile energy consumption, E d is the task detection unit single profile energy consumption, E com is the communication positioning unit single profile energy consumption, k pm , k rm and k pump are the degradation coefficients of the pitch adjustment module, the roll adjustment module and the oil discharge stage buoyancy adjustment unit respectively, t per is the time interval from the last closing to the next opening of the state detector of the control navigation unit, Δt cm is the single running time of the state monitor, t all is the total duration of each stage, and are the depth intervals of the i-th task sensor from the last closing to the next opening in the diving and ascending stages respectively, and are the single sampling depths of the i-th task sensor after opening in the diving and ascending stages respectively, and δV is the back oil discharge amount of the buoyancy adjustment unit.

[0023] In the above technical solution, preferably, the energy consumption of the power unit of the first n sections is monitored with m working sections as a cycle, the degradation coefficient is optimized by genetic algorithm, and the root mean square error between the measured value and the model prediction value is used as the objective function to periodically update the buoyancy adjustment unit degradation coefficient k pump , attitude control unit degradation coefficient (k pm ,k rm ), the objective function and degradation coefficient constraints of the buoyancy adjustment unit and attitude adjustment unit are as follows:

[0024]

[0025] Where, are the measured values ​​of energy consumption of the i-th section of the buoyancy adjustment unit and the attitude adjustment unit, is the predicted value of the i-th profile model of the attitude adjustment unit.

[0026] In the above technical solution, preferably, the feature spaces of the source domain and the target domain in step 5 include the average battery voltage and the cumulative energy consumption prediction value in the diving and floating conversion stage, the source domain label space includes the remaining power percentage (SOC) and the domain classification label, and the target domain label space only includes the domain classification label; the calculation formula of the SOC is:

[0027]

[0028] Where W sum is the cumulative energy consumed by the underwater glider measured by the power consumption sensor, E eol Indicates the remaining releasable energy at the end of the underwater glider mission.

[0029] In the above technical solution, preferably, in step 6, the domain classification loss Using binary cross entropy loss function, regression prediction loss The mean square error loss function is used, and the domain classification loss of the i-th sample is and regression prediction loss They are as follows:

[0030]

[0031] Where, The estimated remaining power of sample i at time t, Represents the domain label category prediction result for sample i. The network is defined to include domain classification loss and regression prediction loss The overall loss function is as follows:

[0032]

[0033] where n S is the number of source domain samples, n T is the number of target domain samples, and alpha is a hyperparameter balancing the relative importance of the two loss terms.

[0034] In the above technical solution, preferably, the LSTM is used to extract the time sequence degradation features of the source domain and the target domain, the domain adversarial transfer learning is used to reduce the cross-domain data distribution difference, and the model is dynamically updated to enhance the adaptability to the target domain data.

[0035] In the above technical solution, preferably, the remaining life prediction in step 7 is based on the real-time feedback battery average voltage and cumulative energy consumption parameters, the trained LSTM-DANN model is input, and the remaining battery percentage prediction value is output.

[0036] In the above technical solution, preferably, the method further comprises: after the prototype is recycled, discharging test is performed on the energy system, actual working conditions are simulated to verify the prediction accuracy of the remaining battery percentage.

[0037] The underwater glider energy consumption modeling and remaining life prediction method based on digital twinning provided by the application has multiple advantages and significant effects.

[0038] Firstly, by building a hardware-in-the-loop simulation platform, the real load and power supply state of each energy-consuming unit in the underwater glider navigation process can be highly restored in the laboratory environment, providing a reliable experimental basis for obtaining high-precision power characteristics and flow characteristics parameters, and significantly improving the authenticity and accuracy of the energy consumption model. Secondly, for the time-varying nonlinear factors that are difficult to model and affect the energy consumption during the operation of the underwater glider, this method aggregates and extracts key finite parameters on the onboard end, and transmits them back to the digital twin database via the satellite link, realizes the dynamic capture and mapping of unobservable environmental changes and component state degradation factors, and enhances the timeliness and adaptability of the energy consumption modeling.

[0039] In the process of energy consumption modeling, the application innovatively constructs a high-fidelity digital twin energy consumption model, and introduces a genetic algorithm to periodically identify and update the degradation coefficients of the power unit (including the attitude adjusting unit and the buoyancy adjusting unit), so that the model can adapt to the evolution process of the performance of the entity system, thereby maintaining high simulation accuracy for a long time, overcoming the limitations of traditional static modeling that cannot reflect the system degradation process. At the same time, in the life prediction aspect, the application constructs a cross-platform sample set combining source domain and target domain data, and extracts the time series degradation features in the multi-source data based on the LSTM network. By introducing a joint loss function containing domain classification loss and regression prediction loss, and using an unsupervised domain adversarial training mechanism, the model migration difficulty caused by the lack of labeled samples in the target domain and the inconsistency of source-target domain data distribution is effectively overcome, and the generalization ability and prediction accuracy of the life prediction model in the target domain are improved.

[0040] In summary, the method improves the accuracy of single profile energy consumption simulation in the energy consumption modeling stage through multi-source fusion and model adaptive evolution, and effectively alleviates the data distribution problem in the life prediction stage through cross-domain transfer learning, realizing accurate prediction of the remaining life of the in-service underwater glider. The overall scheme breaks through the dependence on experience judgment and the over-reliance on labeled samples in data-driven modeling of existing methods, providing reliable and scientific technical support for intelligent maintenance, energy optimization scheduling and equipment loss risk prevention of underwater gliders, and has good engineering application prospect and popularization value. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 Fig. 1 is a schematic diagram of energy consumption units and their working processes of an underwater glider;

[0042] Figure 2 Fig. 3 is a schematic diagram of the principle of a hardware-in-the-loop simulation platform for power characteristics and flow characteristics of a buoyancy adjusting unit;

[0043] Figure 3 Fig. 5 is a schematic diagram of a technology process of extracting and returning time-varying parameters on board to a twin database;

[0044] Figure 4 Fig. 7 is a prediction result of the energy consumption model outputting single profile energy consumption total and energy consumption distribution;

[0045] Figure 5 Fig. 9 is a flowchart of adaptive identification and update of degradation coefficients in the energy consumption model;

[0046] Figure 6 Fig. 11 is a comparison of simulation accuracy of each energy consumption model in the whole flight cycle;

[0047] Figure 7 Fig. 13 is a flowchart of training and testing of a remaining life prediction model based on domain antagonism;

[0048] Figure 8 The comparison chart of prediction accuracy before and after the field confrontation for the remaining life prediction model;

[0049] Figure 9 The flowchart of the prediction method of the present application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0051] In order to solve the problem that the underwater glider has large individual differences in energy consumption characteristics, insufficient modeling and data support in complex marine environments, resulting in difficulty in high-precision prediction of remaining life, maintenance decision relying on experience, and risk of energy waste and equipment loss, the present application provides an underwater glider energy consumption modeling and remaining life prediction method based on digital twinning. In order to further illustrate the structure of the present application, the detailed description is as follows in combination with the drawings:

[0052] The energy consumption units of the underwater glider include attitude adjustment units, buoyancy adjustment units, control navigation units, task detection units, communication positioning units, and their corresponding start-up stages, as shown in Figure 1 The stages a-e in the figure correspond to the positioning communication stage, the diving-floating conversion stage, the diving gliding stage, the floating-diving conversion stage, and the floating gliding stage, respectively. The control navigation unit is working in each stage of the profile process and can be regarded as a fixed static load of the underwater glider. In addition, due to the influence of ocean current disturbance, seawater pressure and density on the net buoyancy of the underwater glider, the attitude adjustment unit will perform random attitude adjustment in stages b and e to keep the underwater glider in the target attitude and target heading.

[0053] The underwater glider energy consumption modeling and remaining life prediction method based on digital twinning proposed by the present application, as shown in Figure 9 , includes the following steps:

[0054] Step 1: Build a hardware-in-the-loop simulation platform to simulate the load and power supply of each energy consumption unit during the underwater glider run, and obtain the power characteristics and flow characteristics parameters of the energy consumption unit.

[0055] The hardware-in-the-loop simulation platform of the attitude adjustment unit is composed of the attitude adjustment unit, adjustable stabilized power supply, electronic compass, program controller, data storage device, and clamp self-locking device. In terms of adjusting the pitch angle of the attitude adjustment unit, the shaft shoulder of the clamp self-locking device is in contact with the end face of the shell to position, and then the attitude adjustment unit is clamped by applying friction force by the clamp, and then the attitude adjustment unit is rotated to any angle by driving the stepping motor and transmitting torque by the shaft coupling, and the current pitch angle is sensed by the electronic compass, and the reverse self-locking is realized by the self-provided worm and gear reducer when the motor stops. In terms of adjusting the working voltage of the attitude adjustment unit, the adjustable stabilized power supply can adjust the working voltage within the adjustable voltage interval, and the data storage device is used to read the register value of the power supply for storing real-time voltage and current values, and the power characteristics of the attitude adjustment unit are monitored in real time. The specific pitch angle adjustment interval and working voltage adjustment interval are determined by the actual sea trial performance of the underwater glider. According to the sea trial data of the Petrel-UL underwater glider, the pitch angle adjustment interval is set to [0°, 90°] (10° interval sampling), and the working voltage adjustment interval is set to [12V, 19V] (1V interval sampling). Due to the difference in motor power caused by the change in direction of sliding friction in the lifting / descending stage, the input power of the pitch adjustment module of the movable battery pack in the two movement directions is collected, and a total of 160 samples are measured. The measured power data of the movable battery pack in the lifting and descending stages are averaged, and the fitting function of the average power required by the pitch adjustment module and the working voltage and the fitting function of the roll adjustment module input power and the working voltage are obtained.

[0056] The power and flow characteristics of the buoyancy adjustment unit are divided into the oil discharge stage and the oil return stage. The oil return stage is realized by an electromagnetic valve, and the average input power P v of the electromagnetic valve is measured by the adjustable stabilized power supply, and the fitting function between the average input power and the working voltage U is obtained. To measure the power characteristics and flow characteristics of the buoyancy adjustment unit in the oil discharge stage, a hardware-in-the-loop simulation platform is built as shown in Figure 2 . The overflow valve with adjustable opening pressure is used to simulate the seawater pressure at different depths, and the flow meter and pressure gauge are installed in the hydraulic circuit to monitor the real-time flow and pipeline pressure values, respectively. The input power and flow of the buoyancy adjustment unit under different pressures and working voltages are measured, and the power fitting function and the flow fitting function are obtained.

[0057] The energy-consuming electrical unit is used to simulate the output voltage change by the adjustable stabilized power supply, and the fitting function of the communication positioning unit input power and the working voltage is measured. The power of the control navigation unit and the task detection unit varies insignificantly with the voltage, and is taken as a fixed value.

[0058] Specifically, the hardware simulation platform is constructed according to the load of each energy-consuming unit and the time-varying characteristics of the energy system input, and the power function of each energy-consuming unit and the flow function of the buoyancy adjusting unit are fitted. For the energy-consuming electrical unit such as the communication positioning unit, the output voltage of the energy system of the underwater glider is considered to decrease with the decrease of the electric quantity, the adjustable voltage stabilizing power supply is used to simulate the output voltage change, the average power of each energy-consuming electrical unit under different working voltages is measured, and the power P com (U) is as follows:

[0059] P com (U) = -15.8890 + 3.5266U - 0.2397U 2 + 0.0056U 3

[0060] In addition, the power of the measured control navigation unit and the task detection unit changes insignificantly with the voltage, wherein the static power of the control navigation unit is P sl = 0.56W, the additional power required after the state monitor is turned on is P cm = 0.22W; the task detection unit refers to the scientific load carried by the underwater glider, and the average power of the i-th load is denoted as P si .

[0061] For the power unit with a larger energy consumption ratio, including the attitude adjusting unit and the buoyancy adjusting unit, wherein the attitude adjusting unit includes the pitch adjusting module and the roll adjusting module, the body pitch angle is adjusted at any angle through the designed clamp self-locking device, the relationship between the power of the attitude adjusting unit and the body pitch angle θ and the working voltage U can be measured, the test data is fitted by using the Rational2D function, and the power function P pm (U, θ) of the pitch adjusting module is as follows:

[0062]

[0063] The roll adjusting module mainly changes with the input voltage U, and the fitting function between the power P rm of the roll adjusting module and the working voltage U is as follows:

[0064]

[0065] The power and flow characteristics of the buoyancy adjusting unit are divided into the oil discharging stage and the oil returning stage. The oil returning stage is realized by the electromagnetic valve, and the average input power P v of the electromagnetic valve is measured by the adjustable voltage stabilizing power supply, and the fitting function between the power and the working voltage U is as follows:

[0066] P v(U) = -14.6046 + 3.2378U - 0.2015U 2 + 0.0054U 3

[0067] The flow rate of the oil return after the electromagnetic valve is opened is relatively stable, and the average oil return flow rate is q v = 558.6 mL / min.

[0068] In the oil discharge stage, the overflow valve with adjustable opening pressure is used to simulate different depths of seawater pressure, and a flow meter and a pressure gauge are installed in the hydraulic circuit to monitor the real-time flow rate and pipeline pressure value, respectively, to measure the power P of the buoyancy adjusting unit under different pressures and working voltages pump (U, h) and flow rate q pump (U, h) are as follows:

[0069] P pump (U, h) = (-17.1713 + 0.00674h + 5.0292U - 0.3829U 2 + 0.00935U 3 )

[0070] ÷ (1 - 6.9117x10 -4 h + 5.9078x10 -7 h 2 - 1.9210x10 -10 h 3

[0071] - 0.0591U + 0.00194U 2 )

[0072]

[0073] Step 2: Extract limited time-varying parameters related to energy consumption at the on-board end, aggregate and then return to the twin database through the satellite.

[0074] The limited time-varying features extracted and returned by the on-board end are shown in Table 1.

[0075] Table 1

[0076]

[0077] The a-e stages in the table are the positioning communication stage, the diving and floating conversion stage, the diving gliding stage, the floating and diving conversion stage, and the floating gliding stage, respectively.

[0078] For the time-varying parameters in the above table that are difficult to accurately model, closely related to energy consumption and limited, considering the high delay and low bandwidth characteristics of the data transmission process, the data preprocessing and time-varying parameter extraction, dimensionality reduction calculation of the real-time running data of the entity equipment need to be completed by the on-board ARM processor, and only in the positioning communication stage, the time-varying parameters are returned to the twin database of the shore control center through the communication satellite, as shown in Figure 3 .

[0079] Step 3: Based on the data of step 1 and step 2, a high-fidelity underwater glider digital twin energy consumption model is established. Through the automatic program of the simulation system, the time-varying parameters are input into the expression of the high-fidelity digital twin energy consumption model, and the prediction results of the current profile energy consumption total and the energy consumption distribution of each unit are output by the model, as shown in Figure 4 , and the prediction results are combined and stored with the time-varying parameters in the twin database.

[0080] The expression of the high-fidelity digital twin energy consumption model in this step is:

[0081]

[0082] In the formula, k bat is the reciprocal of the energy system discharge efficiency of the underwater glider, E gm is the single profile energy consumption of the attitude adjustment unit, E c is the single profile energy consumption of the control navigation unit, E b is the single profile energy consumption of the buoyancy adjustment unit, E d is the single profile energy consumption of the task detection unit, E com is the single profile energy consumption of the communication positioning unit, k pm , k rm and k pump are the degradation coefficients of the pitch adjustment module, the roll adjustment module and the oil discharge stage of the buoyancy adjustment unit, respectively, t per is the time interval from the last closing to the next opening of the state detector of the control navigation unit, Δt cm is the single running time of the state detector, t all is the total duration of each stage, and are the depth intervals of the i-th task sensor from the last closing to the next opening in the diving and ascending stages, respectively, and are the single sampling depths of the i-th task sensor after opening in the diving and ascending stages, respectively, and δV is the oil discharge amount of the buoyancy adjustment unit.

[0083] Step 4: periodically identify and update the degradation coefficients in the power unit energy consumption model using genetic algorithms to compensate for the energy consumption calculation errors caused by transmission performance degradation, and to achieve adaptive evolution of the twin energy consumption model with the entity energy consumption characteristics.

[0084] With m working profiles as a period, monitor the power unit energy consumption measured values of the first n profiles, optimize the degradation coefficients through genetic algorithms, and periodically update the degradation coefficients k pump , the attitude adjustment unit degradation coefficients (k pm , k rm ) of the buoyancy adjustment unit, and the objective functions and constraint conditions of the buoyancy adjustment unit and the attitude adjustment unit are as follows:

[0085]

[0086] In the formula, are the energy consumption measured values of the i-th profile of the buoyancy adjustment unit and the attitude adjustment unit, is the model prediction value of the i-th profile of the attitude adjustment unit.

[0087] After the underwater glider is just deployed, it will first start the power consumption monitoring function, and measure the energy consumption of the power unit for m consecutive profiles through the current and voltage sensor; the root mean square error between the measured power unit energy consumption and the model predicted energy consumption is used as the optimization objective function. The degradation coefficients k pump of the buoyancy adjustment unit and the degradation coefficients k pm and k rm of the attitude adjustment unit are identified and updated using genetic algorithms, and after updating, the power consumption monitoring function is turned off to save energy; Figure 5 The judgment in the formula whether the number of new profiles is greater than n means that from the profile after the last time the degradation coefficient is updated, it is the first profile, and after completing the sawtooth gliding motion of n consecutive profiles, the power consumption monitoring function is started again to measure the energy consumption of m profiles and update the degradation coefficients of the power unit based on genetic algorithms again. Considering that long-time opening of the power consumption sensor to monitor the real-time power of the underwater glider will increase its own power consumption burden, especially for underwater gliders with limited energy carrying capacity, slight power consumption will affect its endurance time, and the performance degradation process of its power unit occurs in a large time scale, so the number of times the power consumption monitoring function is started in the identification and update process should be minimized. In the model degradation coefficient update process, the power consumption sensor is started to monitor the power unit energy consumption of the first 5 profiles with 100 working profiles as a period, that is, n = 95 and m = 5.

[0088] The full-flight cycle energy consumption monitoring data of Petrel-UL-01 are selected for comparison with the simulation results of the proposed energy consumption model (DT-ECMM) and the traditional energy consumption model based on dynamics derivation (FP-ECMM). Figure 6 As shown in the figure, the comparative results show that by integrating the time-varying characteristic parameters during navigation to construct an energy consumption model and adopting a periodic degradation coefficient identification method driven by measured data, the prediction accuracy of the energy consumption model can be significantly improved, and the robustness can still be maintained on a long time scale.

[0089] See Figure 7 , which is a flow chart of the underwater glider remaining life prediction method based on the domain adversarial transfer learning network of the present invention. The method uses the complete power consumption history data of the underwater glider from normal operation to power exhaustion, and realizes the remaining life prediction of the underwater glider in service through transfer learning technology.

[0090] Step 5: Construct source domain dataset and target domain dataset, where the source domain contains historical prototype data and the target domain contains in-service prototype data.

[0091] The feature spaces of the source and target domains include the average battery voltage and cumulative energy consumption prediction value during the submersible-floating transition phase. The source domain label space contains the remaining charge percentage (SOC) and domain classification labels, while the target domain label space only contains domain classification labels. The SOC is calculated as follows:

[0092]

[0093] Where W sum is the cumulative energy consumed by the underwater glider measured by the power consumption sensor, E eol Indicates the remaining releasable energy at the end of the underwater glider mission.

[0094] In the process of acquiring data in the feature space and label space of the source domain and the target domain, the average battery voltage parameter U in the latent-floating conversion stage in the feature space is S The onboard ARM processor completes the extraction and aggregation operations and transmits them back to the ground system via the communication satellite. S The calculation formula is as follows:

[0095]

[0096] Where, t 4_s is the starting moment of the snorkeling conversion phase in a profile cycle of the underwater glider, and t 4_eU(t) is the output voltage of the battery at time t, and t is the end time of the snorkeling conversion phase. In addition, by introducing the time-varying parameters of each profile (see Table 1) into the expression of the high-fidelity digital twin energy consumption model, a full sequence of single-profile energy consumption prediction value sets from the first deployment profile to the current latest profile is obtained, and the set elements are accumulated to obtain the cumulative energy consumption parameter in the feature space The label space of the source domain and the target domain has a domain classification label d. By manually labeling the domain classification label of the sample, it is specified that when the sample belongs to the source domain, d takes 0, and vice versa, d takes 1. The SOC label unique to the source domain label space is calculated by the calculation formula of SOC. In order to obtain the residual releasable energy E sum Before the sea trial, high-precision voltage and current sensors were installed on each prototype to monitor the battery output energy in real time during navigation at a frequency of 1 Hz. In order to calculate the residual releasable energy E eol After the prototype is recovered, the residual power of the internal energy system is discharged. During the test, the ambient temperature is controlled at 10-15℃, which is similar to the temperature in the underwater glider cabin when performing a sawtooth profile motion. During the measurement process, the periodic current working conditions of the underwater glider "dive gliding-hibernation conversion-float gliding" are simulated, and the specific execution process is as follows: first, discharge at a constant current of 0.06A for 7200s, then discharge at a constant current of 2.3A for 540s, and finally discharge at a constant current of 0.06A for 7200s. If the battery terminal voltage does not reach the set cutoff voltage of 10V, the above discharge process is repeated.

[0097] Step 6: Extract time series degradation features using long short-term memory network (LSTM), build loss function containing domain classification loss and regression prediction loss, optimize the performance of field discriminator and source domain regressor through adversarial training, and train the remaining life prediction network.

[0098] In this step, the domain classification loss The binary cross-entropy loss function is used, and the regression prediction loss The mean square error loss function is used, and the domain classification loss of the i-th sample and the regression prediction loss are as follows:

[0099]

[0100] In the formula, The residual power estimate of sample i at time t is , which represents the domain label category prediction result of sample i. Define the overall loss function of the network containing domain classification loss and regression prediction loss as follows:

[0101]

[0102] where n S is the number of source domain samples, n T is the number of target domain samples, and a is a hyper-parameter balancing the relative importance of the two loss terms.

[0103] Based on the obtained dataset, the training of the remaining life prediction network is carried out, which includes Figure 7 two core stages of feature extraction and model updating. Specifically, in the feature extraction stage, first, the long short-term memory network (LSTM) is used to deeply mine the time sequence degradation features of the source domain data; at the same time, based on the unsupervised domain adversarial mechanism, the invariance features between the source domain and the target domain are learned to alleviate the influence of the cross-domain data distribution difference on the prediction accuracy. In the model updating stage, with the continuous accumulation of the feature data and the energy consumption model simulation results returned online, the target domain feature data is continuously enriched. Based on the new data, the model is dynamically updated, and the extracted cross-domain common feature space is continuously optimized to enhance the adaptability of the model to the target domain data.

[0104] Step 7: Based on the trained remaining life prediction model, the remaining life of the in-service underwater glider is accurately predicted. The real-time feature parameters (battery average voltage parameter U S and cumulative energy consumption parameter ) of the in-service glider are input into the trained remaining life prediction network, and the prediction result of the current remaining battery percentage of the glider is output.

[0105] To quantitatively evaluate the improvement effect of the domain adversarial transfer learning (DANN) on the model prediction performance, a double-dimension comparative experiment is designed: first, comparing LSTM / RNN-DANN with LSTM / RNN model trained only based on source domain data; second, exploring the influence of RNN and LSTM two time sequence feature extraction methods on the model performance. The full flight cycle data of underwater gliders numbered Petrel-UL-01 and Petrel-UL-02 are selected as the source domain, and the remaining battery percentage of the target domain in-service glider Petrel-UL-03 is predicted, as shown in Figure 8 , the baseline model trained only based on the source domain has a significant deviation between the remaining life prediction value in the target domain and the measured value after the glider is recovered. In contrast, RNN-DANN and LSTM-DANN introduced the domain adversarial mechanism can greatly reduce the prediction error with the measured value, and in terms of feature extraction, compared with RNN, the feature extraction method based on LSTM can slightly improve the model prediction accuracy.

[0106] The above merely describes preferred embodiments of the present application, and is not used to limit the present application, any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for energy consumption modeling and remaining life prediction of underwater gliders based on digital twins, characterized in that: The following steps are involved: Step 1: Build a hardware-in-the-loop simulation platform to simulate the load and power supply of each energy-consuming unit during the underwater glider's flight, and obtain the power characteristics and flow characteristic parameters of the energy-consuming unit; Step 2: Extract the finite time-varying parameters related to energy consumption on the airborne side, aggregate them, and transmit them back to the twin database via the satellite; Step 3: Based on the data from steps 1 and 2, a high-fidelity digital twin energy consumption model of the underwater glider is established; Step 4: Use a genetic algorithm to periodically identify and update the degradation coefficient in the power unit energy consumption model to compensate for the energy consumption calculation error caused by transmission performance degradation, and achieve adaptive evolution of the twin energy consumption model with the entity's energy consumption characteristics; Step 5: Construct source domain dataset and target domain dataset, where the source domain contains historical prototype data and the target domain contains in-service prototype data; Step 6: Use a long short-term memory (LSTM) network to extract temporal degradation features, construct a loss function that includes domain classification loss and regression prediction loss, optimize the performance of the domain discriminator and source domain regressor through adversarial training, and train the remaining life prediction network; Step 7: Based on the trained remaining life prediction model, accurately predict the remaining life of the underwater glider in service.

2. The underwater glider energy consumption modeling and remaining life prediction method based on digital twin according to claim 1 is characterized in that: In step 1, the hardware-in-the-loop simulation platform includes: The attitude adjustment unit simulates pitch angle adjustment through an adjustable voltage-regulated power supply, an electronic compass, and a clamp self-locking device, and measures the relationship between power, pitch angle, and voltage; The buoyancy regulating unit simulates seawater pressure through an adjustable overflow valve and combines flowmeters and pressure gauges to measure the power and flow characteristics during the oil discharge phase; The energy consumption electrical unit measures the power of the communication and positioning unit as a function of voltage change, and fixedly controls the static power of the navigation unit and the mission detection unit.

3. The method for underwater glider energy consumption modeling and remaining life prediction based on digital twin according to claim 1 is characterized in that: The finite time-varying parameters in step 2 include: The time consumed in each stage of single-profile operation, the average pitch angle of the diving / surfacing gliding stage, the pitch adjustment time, the cumulative roll adjustment time, the average oil discharge depth of the diving and surfacing transition stage, and the average battery voltage of the diving and surfacing transition stage and the non-submerging and surfacing transition stage.

4. The method for energy consumption modeling and remaining life prediction of underwater glider based on digital twin according to claim 1 is characterized in that: The expression of the high-fidelity digital twin energy consumption model described in step 3 is: Where k bat is the inverse of the discharge efficiency of the underwater glider energy system, E gm is the energy consumption of the single section of the attitude adjustment unit, E c To control the energy consumption of a single profile of the navigation unit, E b is the energy consumption of a single section of the buoyancy adjustment unit, E d is the energy consumption of a single profile of the mission detection unit, E com is the energy consumption of a single cross-section of the communication positioning unit, k pm 、k rm and k pump are the degradation coefficients of the pitch control module, roll control module and buoyancy control unit in the oil discharge phase, t per To control the time interval from the last closing to the next opening of the status detector of the navigation unit, Δt cm is the single operation time of the condition monitor, t all is the sum of the duration of each stage, and are the depth intervals from the last shutdown to the next startup of the i-th mission sensor during the diving and surfacing phases, and are the single sampling depths of the i-th mission sensor after it is turned on during the diving and surfacing stages, and δV is the oil discharge volume of the buoyancy adjustment unit.

5. The method for energy consumption modeling and remaining life prediction of underwater glider based on digital twin according to claim 1 is characterized in that: With m working sections as a cycle, the measured values ​​of the power unit energy consumption of the first n sections are monitored, and the degradation coefficient is optimized by genetic algorithm. The root mean square error between the measured value and the model prediction value is used as the objective function, and the buoyancy adjustment unit degradation coefficient k is periodically updated. pump , attitude control unit degradation coefficient (k pm ,k rm ), the objective function and degradation coefficient constraints of the buoyancy adjustment unit and attitude adjustment unit are as follows: Where, are the measured values ​​of energy consumption of the i-th section of the buoyancy adjustment unit and the attitude adjustment unit, is the predicted value of the i-th profile model of the attitude adjustment unit.

6. The underwater glider energy consumption modeling and remaining life prediction method based on digital twin according to claim 1 is characterized in that: The feature spaces of the source and target domains in step 5 include the average battery voltage and cumulative energy consumption prediction value during the submersible-floating conversion phase. The source domain label space contains the remaining charge percentage (SOC) and domain classification labels, while the target domain label space only contains domain classification labels. The SOC is calculated as follows: Where W sum is the cumulative energy consumed by the underwater glider measured by the power consumption sensor, E eol Indicates the remaining releasable energy at the end of the underwater glider mission.

7. The method for underwater glider energy consumption modeling and remaining life prediction based on digital twin according to claim 1 is characterized in that: In step 6, the domain classification loss Using binary cross entropy loss function, regression prediction loss The mean square error loss function is used, and the domain classification loss of the i-th sample is and regression prediction loss They are as follows: Where, The estimated remaining power of sample i at time t, Represents the domain label category prediction result for sample i. The network is defined to include domain classification loss and regression prediction loss The overall loss function is as follows: Where n S is the number of source domain samples, n T is the number of target domain samples, and α is a hyperparameter that balances the relative importance of these two loss terms.

8. The underwater glider energy consumption modeling and remaining life prediction method based on digital twin according to claim 1 is characterized in that: LSTM is used to extract the temporal degradation features of the source and target domains, domain adversarial transfer learning is used to eliminate the cross-domain data distribution differences, and the model is dynamically updated to enhance its adaptability to the target domain data.

9. The underwater glider energy consumption modeling and remaining life prediction method based on digital twin according to claim 1 is characterized in that: The remaining life prediction described in step 7 is based on the real-time average battery voltage and cumulative energy consumption parameters, which are input into the trained LSTM-DANN model and output as a predicted value of the remaining battery percentage.

10. The method for energy consumption modeling and remaining life prediction of underwater glider based on digital twin according to any one of claims 1 to 9, characterized in that: The method further includes: performing a discharge test on the energy system after the prototype is recovered, simulating actual operating conditions to verify the calculation accuracy of the remaining power percentage.

Images