An unmanned underwater vehicle black box release decision method and system

Abstract

The application discloses an unmanned underwater vehicle black box release judgment method and system, comprising the following steps: S1, acquiring multi-source operation observation data of the unmanned underwater vehicle, preprocessing and constructing a time sequence characteristic vector F1; S2, based on the time sequence characteristic vector F1, through an explicit physical formula and online recursive parameter identification, a mechanism characteristic vector F2 containing a net negative buoyancy coefficient, an equivalent bound drag coefficient, a self-rescue execution efficiency and a communication failure coefficient is calculated; S3, the time sequence characteristic vector F1 and the mechanism characteristic vector F2 are fused into a hybrid feature set, input into a fault classification model, and a fault classification result and a fault confidence are output; S4, the learning rate of the online recursive parameter identification is dynamically adjusted according to the fault confidence, and the adjusted learning rate is fed back to S2 to control the identification speed of the equivalent bound drag coefficient; and S5, the fault classification result is subjected to time sequence steady state verification, and an emergency black box release instruction is output when the verification is passed.

Description

Technical Field

[0001] This invention belongs to the field of underwater unmanned submersible fault diagnosis and safety control technology, specifically relating to a method and system for determining the release of the black box of an unmanned submersible. Background Technology

[0002] Unmanned underwater vehicles (UUVs) are widely used in marine resource exploration, underwater topographic mapping, underwater security inspection, and marine scientific experiments. The black box (the UUV's critical data storage, release, and recovery system) is a core component for storing mission data, recording operational status, tracing faults, and locating accidents. Its emergency release is irreversible. If the black box is accidentally triggered and ejected under normal operating conditions, it will directly disrupt the UUV's original buoyancy balance, causing an increase in passive negative buoyancy, continuous sinking and instability, ultimately leading to the UUV losing contact, equipment damage, and permanent loss of mission data, resulting in significant economic losses and mission failure. Therefore, the black box release condition determination system must meet stringent design requirements for high reliability, accurate fault identification and triggering, and strong resistance to environmental interference.

[0003] During UUV underwater operations, the emergency release of the black box is only permitted in the event of one of four extreme malfunctions that prevent the vessel from regaining its surfacing capability through self-rescue. These four extreme malfunctions are as follows: 1. Physical restraint type failure: When a UUV runs aground on the seabed, gets entangled in seaweed or fishing nets, gets stuck in underwater structures, or has heavy objects attached to its exterior, the hull structure is intact and the self-rescue mechanism is functioning normally, but it is unable to break free from the restraints and float up for recovery due to external physical limitations or additional gravity / resistance. 2. Buoyancy imbalance type of failure: When water enters and leaks into the hull, oil bladders rupture and leak, or ultra-deep pressure-resistant structures are damaged, the overall mass of the machine increases or the buoyancy decreases. The self-rescue mechanism can perform buoyancy adjustment and jettisoning actions normally, but the overall weight of the machine continues to be greater than the buoyancy, the buoyancy balance is completely destroyed, and the trend of continuous sinking cannot be reversed. 3. Failure of self-rescue mechanism: If there is an electrical circuit fault, power supply line breakage, damage to drive motor or electromagnetic actuator, or abnormal control signal in the buoyancy regulating oil pump or jettisoning actuator, the host computer can issue a self-rescue start command normally, but the actuator has no actual effective action, and the UUV completely loses its ability to actively float and rescue itself. 4. Main control failure and communication interruption: The UUV main control module fails, is paralyzed by impact, or the program crashes. The communication link between the black box and the UUV body is completely broken, and no real-time operating data can be obtained. The UUV is completely out of control, unable to start the self-rescue process, and is in a passive and continuous sinking state.

[0004] The macroscopic manifestations of the four types of extreme faults mentioned above are all the same: UUVs are unable to float and continue to sink. However, their physical causes are completely different, and the required multi-source observation characteristics and identification criteria are also fundamentally different. Currently, the industry's detection methods for these faults generally employ single-parameter thresholds or multi-parameter independent threshold judgment schemes, simply comparing real-time acquired data with fixed empirical thresholds. These methods lack logical connections between parameters, temporal change verification, and physical mechanism constraints, resulting in significant technical deficiencies. 1. High risk of misjudgment: Normal operating conditions such as UUV hovering, near-seabed constant altitude cruise, active constant speed descent, underwater short-term communication interference, and instantaneous electrical fluctuations can easily trigger fixed thresholds, causing the black box to be released falsely, directly leading to the loss of the entire UUV. 2. Severe failure to detect faults: For hidden faults such as slow water seepage, minor oil bladder leakage, normal command but mechanism jamming, and local circuit loose connection, all observation parameters are within the normal threshold range, which traditional threshold methods cannot identify, resulting in a large number of fault blind spots. 3. Faults cannot be distinguished: The macroscopic manifestations of the four types of extreme faults are that the UUV cannot float and continues to sink to the bottom. Traditional thresholds rely only on single changing features such as depth, which cannot distinguish the root cause of the fault and do not meet the needs of post-fault review and tracing. 4. Weak resistance to environmental interference: Complex marine environments such as underwater turbulence, natural changes in water depth and pressure, and electromagnetic attenuation interference can easily cause parameter fluctuations and false alarms, further amplifying the instability of threshold detection.

[0005] Another approach is the purely data-driven method, which relies solely on multi-source sensor data to train a classification model for fault determination. This method avoids the limitations of fixed thresholds, but it is difficult to obtain real underwater fault samples, and normal samples constitute the vast majority in the training set. The model's ability to identify subtle, gradual faults is clearly insufficient. Furthermore, the decision-making process lacks physical mechanism support, making the results difficult to trace and interpret, and thus failing to meet the reliability requirements of unmanned underwater vehicles.

[0006] In summary, traditional fixed threshold detection technology suffers from high false positive rates, serious false negatives, and poor reliability. Pure data-driven solutions also cannot balance fault identification accuracy and physical interpretability in scenarios with scarce samples, and cannot meet the high reliability control requirements for irreversible release of black boxes. There is an urgent need to provide a set of intelligent judgment and anti-false triggering technology solutions for release conditions that integrate mechanism and data-driven approaches, have a two-way coupling interaction mechanism, and have complete and implementable quantitative formulas. Summary of the Invention

[0007] The purpose of this invention is to overcome the common industry defects of existing fixed threshold methods, such as serious misjudgments and omissions, scarce fault samples in pure data-driven classification methods, lack of physical mechanism support in the decision-making process, and insufficient traceability and interpretability. This invention provides a method and system for determining the release of the black box of an unmanned underwater vehicle (UUV). It comprehensively characterizes various fault features through explicit physical quantification formulas; dynamically adjusts the recursive identification learning rate using fault classification confidence, forming a two-way feedback loop where mechanistic parameters define the decision boundary and classification results optimize identification accuracy; expands and balances the training dataset using mechanistic Monte Carlo simulation; and combines this with a time-series steady-state decision mechanism to prevent false black box releases, taking into account physical interpretability, operational stability in complex sea conditions, and the need for embedded low-computing-power deployment.

[0008] To achieve the above-mentioned objectives, the first objective of this invention is to provide a method for determining the release of an unmanned underwater vehicle's black box, comprising: S1. Acquire multi-source operational observation data of the unmanned underwater vehicle, perform preprocessing, and construct a time-series feature vector F1; S2. Based on the time-series feature vector F1, the mechanism feature vector F2, which includes net negative buoyancy coefficient, equivalent restraint resistance coefficient, self-rescue execution efficiency and communication failure coefficient, is calculated by identifying explicit physical formulas and online recursive parameters. S3. The temporal feature vector F1 and the mechanism feature vector F2 are fused into a hybrid feature set, which is then input into the fault classification model to output the fault classification result and fault confidence. S4. Dynamically adjust the learning rate of the online recursive parameter identification based on the fault confidence level, and feed back the adjusted learning rate to S2 to control the identification speed of the equivalent restraint coefficient. S5. Perform a time-series steady-state verification on the fault classification results, and output the black box emergency release command when the verification passes.

[0009] Preferably, in S2, the online recursive parameter identification of the equivalent binding drag coefficient adopts a first-order recursive smoothing formula;

[0010] in: This is the smoothed equivalent restraint drag coefficient at the current moment; This is the equivalent restraint drag coefficient after smoothing from the previous moment; The learning rate controls the magnitude of the correction between the current instantaneous value and the estimated value. It is the instantaneous equivalent binding resistance coefficient obtained directly from the force balance equation at the current moment.

[0011] Preferably, S4 is as follows: When the fault confidence level is lower than the preset lower threshold, the learning rate is set to the minimum value; When the fault confidence level is higher than the preset upper limit threshold, the learning rate is set to the maximum value; When the fault confidence level is between the preset lower threshold and the upper threshold, the learning rate is linearly interpolated between the lower threshold and the maximum threshold.

[0012] Preferably, the fault classification model is a lightweight random forest model, and the training set of the lightweight random forest model consists of a mixture of real normal samples and fault samples generated by Monte Carlo simulation based on the mechanism model.

[0013] Preferably, the timing steady-state verification specifically involves: within a sliding window of a preset length, counting the cumulative number of fault classification results; when the cumulative number is greater than or equal to a preset steady-state threshold, the verification is deemed successful.

[0014] Preferably, the mechanism feature vector F2 also includes the oil pump execution efficiency and load dumping execution effectiveness calculated by sliding window integration.

[0015] Preferably, the multi-source operational observation data includes depth, fuel bladder fuel level, fuel pump on / off status, jettisoning power-on status, cabin pressure, jettisoning circuit current, and attitude angle.

[0016] A second objective of this invention is to provide a black box release decision system for an unmanned underwater vehicle, comprising: The data acquisition module is used to acquire multi-source operational observation data of the unmanned underwater vehicle; An embedded computing module is used to execute the unmanned underwater vehicle black box release decision method as described above and output the decision result; The black box drive release module is used to receive the judgment result and drive the black box to perform an emergency release action when the judgment result is release; The data storage module is used to record system operation parameters, identify intermediate results, classify decision logs, and release trigger records; The communication interaction module is used to provide a bidirectional data link between the unmanned underwater vehicle's main control system and the embedded computing module.

[0017] A third objective of this invention is to provide a computer program product, comprising a computer program that is executed by a processor to determine the release of the unmanned underwater vehicle black box as described above.

[0018] A fourth objective of this invention is to provide a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the aforementioned method for determining the release of the unmanned underwater vehicle's black box.

[0019] Compared with the prior art, the present invention has the following technical effects: This invention constructs a closed-loop architecture that couples mechanism identification and machine learning. Through the bidirectional interaction of forward inputting mechanism parameters into a random forest and backward dynamic adjustment of the recursive learning rate based on classification confidence, it directly controls the mechanism. The high update speed enables the mechanism identification module to sense environmental changes and adapt accordingly.

[0020] This invention provides complete explicit mathematical expressions for all fault characteristics, including and Explicit physical formulas Online recursive parameter identification formula, The formula for window integral.

[0021] This invention uses a Monte Carlo simulation balanced dataset with limited parameter ranges, clearly defining the physical disturbance range, sample size, and training / validation splitting rules (8:2 split) for each fault parameter. This effectively solves the engineering pain points of scarce underwater fault samples and imbalanced training datasets, and significantly improves the ability to identify weak and latent faults.

[0022] The sliding timing window continuous counting mechanism effectively shields against misjudgments caused by instantaneous interference and environmental fluctuations, avoids the accidental release of the black box from the algorithm's underlying layer, and ensures the timely identification of real faults.

[0023] This invention is based on lightweight random forest and single-step recursive parameter identification. It does not involve large-scale matrix operations or complex gradient calculations, resulting in low computational overhead, low power consumption, and strong real-time performance. It is compatible with miniaturized embedded hardware platforms and has high engineering feasibility. Attached Figure Description

[0024] Figure 1 The flowchart provided is for a preferred embodiment of the present invention; Figure 2 This is a flowchart of multi-source data preprocessing and time-series feature generation provided in a preferred embodiment of the present invention; Figure 3 This is a flowchart illustrating the explicit mechanism modeling and online recursive parameter identification provided in a preferred embodiment of the present invention; (Note: The last line appears to be a label or annotation and is left untranslated.) Feedback path from S4 back to the parameter identifier; Figure 4 This is a block diagram of hybrid feature random forest classification and confidence output provided in a preferred embodiment of the present invention; Figure 5 The following is a block diagram of the closed-loop logic for confidence-based reverse control of recursive learning rate provided in a preferred embodiment of the present invention; Figure 6 This is a block diagram for the timing steady-state anti-false triggering decision provided in a preferred embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention, and not all of them. Generally, the embodiments of the present invention described and shown in the accompanying drawings are characteristic technologies and solutions. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] For the first embodiment, please refer to... Figure 1 A method for determining the release of the black box of an unmanned underwater vehicle, comprising: S1. Acquire multi-source operational observation data of the unmanned underwater vehicle, perform preprocessing, and construct a time-series feature vector F1; S2. Based on the time-series feature vector F1, the mechanism feature vector F2, which includes net negative buoyancy coefficient, equivalent restraint resistance coefficient, self-rescue execution efficiency and communication failure coefficient, is calculated by identifying explicit physical formulas and online recursive parameters. S3. The temporal feature vector F1 and the mechanism feature vector F2 are fused into a hybrid feature set, which is then input into the fault classification model to output the fault classification result and fault confidence. S4. Dynamically adjust the learning rate of the online recursive parameter identification based on the fault confidence level, and feed back the adjusted learning rate to S2 to control the identification speed of the equivalent restraint coefficient. S5. Perform a time-series steady-state verification on the fault classification results, and output the black box emergency release command when the verification passes.

[0027] To better understand the technical solution of the present invention, the following non-limiting explanation is provided: Please see Figures 2 to 6 The invention is divided into five interconnected closed-loop steps, S1 to S5. S1 receives raw data from multiple sensors, performs filtering and time-series feature calculation, and outputs a time-series feature vector F1. This vector serves as input for both S2 mechanism modeling and S3 intelligent classification. S2 utilizes data from F1, such as oil quantity, pump status, depth change rate, and jettison loop current, to solve for the net negative buoyancy coefficient using explicit physical formulas. Communication failure coefficient The efficiency of the oil pump is calculated by integrating through a sliding window. With the effectiveness of loadout execution The equivalent restraint coefficient is solved by online recursive parameter identification. The five mechanistic parameters are combined into a mechanistic feature vector F2, which is then output to layer S3. Layer S3 merges F1 and F2 into a hybrid feature set F, which is input into the random forest and outputs the classification result. Up to S5, fault confidence Up to S4. S4 layer according to... Dynamically adjust the learning rate for recursive identification The parameters are returned to the parameter identifier in the S2 layer for control. Convergence speed. S5 layer receiver. After the sliding window steady-state check is performed, the final release enable signal is output. .

[0028] Of the five layers mentioned above, S1, S2, S3, and S5 constitute the forward decision pathway, while S3, S4, and S2 constitute the reverse control pathway. Together, they form a two-way closed loop. The core innovation lies in adding an interactive channel that uses classification confidence to reversely control the recognition learning rate, thus dynamically coupling mechanism recognition and intelligent classification.

[0029] In S1, multi-source operational observation data is first acquired using sensors. This data includes depth, fuel tank volume, fuel pump on / off status, jettisoning power-on status, tank pressure, jettisoning circuit current, roll angle, and pitch angle. Then, the multi-source operational observation data is filtered for noise reduction and time-series feature extraction, outputting a time-series feature vector F1, which is used for mechanism modeling (mechanism feature vector F2) and intelligent classification (S3).

[0030] The system uses a fixed sampling period Run, discrete sampling time .

[0031] The original observation vector is :

[0032] in: Real-time depth, in meters; This refers to the oil storage capacity of the oil bladder, in m³. The oil pump is in on / off state. It is in the load-discharge power-on state; The pressure inside the chamber is expressed in Pa. This is the discharge loop current, in amperes (A). The roll angle, The pitch angle is expressed in degrees (°).

[0033] A first-order low-pass filter is applied to continuous analog signals to remove high-frequency noise. , Direct pass:

[0034] in, Represents the original observation vector The filtered state values ​​of continuous analog quantities (depth, oil quantity, tank pressure, current, attitude angle); The filter coefficient is set independently based on the noise characteristics and dynamic response requirements of each sensor. These are the measured values. This is the value from the previous filter. .

[0035] Three time-series derived features are further calculated from the filtered state variables:

[0036]

[0037]

[0038] in: , These are the filtered depth values ​​for the current time and the previous time, respectively, in meters (m). , These are the filtered oil volume in the oil sac at the current time and the previous time, respectively, in m³. This is the rate of change of depth, in m / s; a positive value indicates diving and a negative value indicates surfacing. The oil quantity change rate is expressed in m³ / s. Its absolute value reflects the intensity of the mechanism's action, and the positive or negative sign indicates whether oil is discharged or returned. The data update interval, measured in seconds, represents the real-time synchronization status of the data link.

[0039] The above-mentioned derived features and the filtered steady-state amplitude Together, they constitute a complete phase space description of the system. The temporal eigenvector F1 is defined as:

[0040] The F1 vector is simultaneously output to S2 (mechanism modeling) and S3 (intelligent classification).

[0041] S2. Multi-domain explicit mechanism modeling and online recursive parameter identification: This step receives F1 from S1 and uses data such as oil quantity, pump status, and depth change rate to calculate the net negative buoyancy coefficient using explicit physical formulas. Communication failure coefficient The efficiency of the oil pump is calculated by integral through a sliding window. Loading execution validity The equivalent restraint drag coefficient is solved by online recursive parameter identification. The above five mechanistic parameters are combined into a mechanistic feature vector F2, which is then output to S3. The recursive update speed is controlled by the learning rate fed back by S4.

[0042] S2.1, Buoyancy Balance Model and Net Negative Buoyancy Coefficient

[0043] This step calculates... Required amount of oil and cabin pressure All are derived from the original observation vectors of S1.

[0044] The vertical force balance of the unmanned underwater vehicle is: gravity. Subtract buoyancy This equals the net sinking force. Buoyancy is determined by the displacement volume, where the displacement volume of structural components such as the hull is a fixed constant and cancels each other out when calculating changes in buoyancy; the only variable component that actually contributes to the change in buoyancy is the fuel bladder volume, hence buoyancy is written as... .

[0045] Explicit formula for net negative buoyancy coefficient:

[0046] in: This is the net negative buoyancy coefficient, dimensionless; The density of seawater is expressed in kg / m³. This is the acceleration due to gravity, expressed in m / s². This refers to the nominal oil bladder volume of the UUV under neutral buoyancy, determined by laboratory buoyancy-weight balancing tests, in m³. The current oil bladder capacity after S1 filtering, in m³. This is the chamber pressure-intake conversion factor, in m². This factor was obtained through laboratory calibration: a known weight of water was injected into the sealed chamber. Water, measuring chamber pressure increment ,but ; The current chamber pressure is expressed in Pa. The reference chamber pressure under perfect sealing condition, in Pa; The total weight of the machine is expressed in N.

[0047] In the formula, This characterizes the buoyancy loss caused by the oil bladder volume deviating from its nominal value. The equivalent value of water ingress weight gain corresponding to the increase in cabin pressure.

[0048] single frame The numerical value represents the instantaneous buoyancy deviation at the current moment, which can occur during normal descent and buoyancy-weight imbalance. However, the former's depth change rate Controlled constancy, the latter Continuously increasing and The absolute value continues to increase. Accurate differentiation can be achieved by fusing the S1 temporal features with the full set of S2 mechanistic parameters through a lightweight random forest in S3.

[0049] After the calculation is complete, proceed to the F2 vector.

[0050] S2.2, Underwater Binding Resistance Model and Online recursive parameter identification This step utilizes the oil quantity provided by S1. Oil pump status Discharge state and the rate of change of depth Depth second derivative obtained by difference The equivalent restraint drag coefficient is solved by online recursive parameter identification. .

[0051] Self-rescue in full swing At that time, the dynamic equation in the vertical direction is:

[0052] in: The current buoyancy includes the buoyancy generated by the adjustment of the oil bladder volume, in N; Additional lift provided for jettisoning and release in self-rescue mechanisms, in N; Total weight of the machine, in N; External equivalent restraint resistance, in N; The mass of the entire machine is expressed in kg. The second derivative of depth (acceleration), derived from S1. Obtained by difference, unit m / s².

[0053] In the above equation, , , , , All quantities are known or can be directly calculated from measured values; the only unknown quantity is the external equivalent restraint resistance. Solve by rearranging terms directly:

[0054] Further dimensionless transformation yields the instantaneous equivalent binding drag coefficient:

[0055] Among them, take The purpose is to eliminate small negative values ​​that may be caused by sensor noise and numerical calculation errors, and to ensure that the equivalent restraint drag coefficient is always non-negative, in accordance with the physical definition of the parameter.

[0056] Due to sensor noise and ocean current disturbances, High-frequency fluctuations exist, and direct use would lead to instability in the classifier's input features; therefore, smoothing is necessary. Furthermore, the required degree of smoothing varies under different operating conditions—for example, when a suspected fault is detected. The changing trend itself carries fault information, requiring rapid tracking to capture its dynamic characteristics; under normal operating conditions The fluctuations mainly stem from environmental disturbances, requiring deep smoothing to suppress noise. Therefore, the learning rate cannot be a fixed value and must be dynamically adjusted according to the current operating conditions.

[0057] This invention employs first-order recursive smoothing:

[0058] in: : Equivalent binding drag coefficient after smoothing at the current moment (dimensionless); : Equivalent binding drag coefficient after smoothing at the previous moment (dimensionless); The learning rate controls the magnitude of the correction between the current instantaneous value and the estimated value. The instantaneous equivalent binding resistance coefficient is directly calculated from the force balance equation at the current moment.

[0059] The meaning of this formula is: the current estimated value is based on the estimated value at the previous time step, and is partially corrected to the current instantaneous observation value according to the learning rate. The larger the value, the greater the correction range and the faster the tracking. The smaller the value, the smaller the correction range, and the smoother and more stable the result.

[0060] for The value of the fault confidence level output by S3 is used as the basis for the determination of the fault confidence level in this invention. As a quantitative indicator for determining the current operating condition: A higher value indicates that the current operating condition is closer to a fault state, requiring rapid tracking. A lower value indicates that the current operating condition is closer to normal, requiring deeper smoothing. The S4 layer calculates this in real time. This data is then fed back to the recursive formula, forming a complete control chain for classification confidence, learning rate, and tracking speed of mechanistic parameters. During initialization... .

[0061] This comprehensively characterizes all external resistance and additional gravitational effects that hinder UUV buoyancy, including physical constraints such as grounding, entanglement, and jamming, as well as the additional downward force caused by externally attached weights (equivalent to a continuously downward binding resistance component). When, it indicates that there are no external constraints; when When this occurs, it indicates that there is external resistance hindering the movement of the unmanned underwater vehicle; The higher the value, the more severe the external restraint. In the event of a physical restraint malfunction, even with the self-rescue mechanisms fully activated, the unmanned underwater vehicle (UUV) may still be unable to surface. It will increase significantly, and this parameter provides a quantitative feature that directly represents the external constraint state for subsequent classifiers.

[0062] After the calculation is complete, proceed to vector.

[0063] S2.3 Self-rescue points execution efficiency This step comprehensively evaluates the actual performance of the self-rescue organization from two dimensions: oil pump efficiency and load shedding effectiveness.

[0064] (1) Oil pump efficiency Calculation This sublayer utilizes the instantaneous change rate of oil volume in the oil sac output by S1. Execution efficiency is calculated using definite integrals through a sliding window. Eliminate the interference of instantaneous shock during motor startup and sensor spikes on transient efficiency calculation:

[0065] in: For integral execution efficiency, dimensionless, it represents the ratio of the actual oil adjustment amount of the mechanism to the theoretical oil adjustment amount within the window; The sliding window length is the number of sampling points per unit, compared to the sliding window length for S5 fault statistics. Maintain consistency; The instantaneous change rate of oil volume in the oil bladder after filtering is expressed in m³ / s. The absolute value is used to accumulate the work done by the forward and reverse bidirectional oil adjustment. The nominal volumetric flow rate of the oil pump at its rated operating voltage is obtained through laboratory calibration, and the unit is m³ / s. This is the function to indicate when the oil pump is turned on; it takes the value 1 when the oil pump is turned on and 0 when it is turned off.

[0066] The representative agency failed to take any effective oil adjustment action, which was determined to be a failure of the self-rescue mechanism; under normal working conditions... .

[0067] After the calculation is complete, proceed to the F2 vector.

[0068] (2) Load Execution Validity Calculation A sliding window integration mechanism consistent with the oil pump's efficiency is adopted. The actual energized work level of the load shedding mechanism is characterized by current integration over time, eliminating the influence of transient disturbances such as instantaneous current surges, line contact jitter, and electromagnetic interference spikes on effectiveness determination. The calculation formula is as follows:

[0069] in: The integral load release execution effectiveness is dimensionless and represents the ratio of the actual charge on the load release loop within the window to the theoretical nominal charge. The length of the sliding window, the number of unit sampling points, is consistent with the length of the oil pump execution efficiency integral window and the fault statistics sliding window. The instantaneous operating current of the filtered load dumping circuit is expressed in amperes (A). The nominal operating current of the load release mechanism under rated operating voltage, in amperes (A), is obtained from laboratory calibration. For the load dump power-on indication function, the load dump command is issued ( The value is 1 when there is no instruction and 0 when there is no instruction.

[0070] when When the power supply to the ejection mechanism within the window remains stable, it indicates that the expected ejection action has been reliably executed; when If the value is significantly lower than the nominal threshold, it indicates that there is an open circuit or persistent poor contact in the load release circuit within the window, and the load release action is not executed normally and effectively.

[0071] After the calculation is complete, proceed to the F2 vector.

[0072] S2.4, Communication Failure Factor

[0073] This step uses the data update interval output by S1. , with standard period The ratio is used to calculate the communication failure factor. :

[0074] in: The standard data update cycle is the same as the system sampling cycle by default. .

[0075] This is a dimensionless coefficient that characterizes the health of the communication link. During normal communication... When the main control unit crashes or the link is disconnected It continues to accumulate and increase. It then increases monotonically; Decision logic: When hour( (Based on the maximum allowed number of consecutive packet losses), it is determined to be a master control communication interruption type fault.

[0076] After the calculation is complete, proceed to the F2 vector.

[0077] S2.5, Mechanism Feature Vector The five mechanistic parameters obtained from S2.1 to S2.4 are combined into a mechanistic eigenvector F2:

[0078] The F2 vector is output to S3.

[0079] S3, Lightweight Random Forest Multi-Classification with Hybrid Features This step receives the F1 output from S1 and the F2 output from S2, merges them into a mixed feature set, and then inputs it into the lightweight random forest. The random forest outputs the classification results. (Sent to S5 for final decision) and fault confidence (Sent to S4 for feedback control).

[0080] S3.1 Hybrid Feature Fusion Merge F1 and F2 into a mixed feature set F:

[0081] S3.2 Fault Label Definition : Normal operating conditions; Physical constraint type fault; : Faults related to buoyancy imbalance; Faults related to the failure of self-rescue mechanisms; Faults such as main controller downtime and communication interruption.

[0082] In real-world marine environments, composite faults involving two types of failures may occur simultaneously. In the multi-classification framework of this invention, the feature vector of the composite fault lies in the boundary region of the feature space of the single fault category. Random forest makes a comprehensive judgment based on voting confidence: when one class is dominant, it is assigned to the main fault category and uncertainty is reflected in the confidence score; when the contributions of the two classes are close, a low confidence score is output.

[0083] S3.3 Monte Carlo Simulation Sample Generation and Training Set Configuration The scarcity of real underwater fault samples is the core engineering bottleneck restricting the implementation of UUV intelligent fault diagnosis technology: on the one hand, the cost and risk of actual sea fault tests are high, and it is difficult to obtain sufficient fault data through sea trials; on the other hand, samples generated by pure data augmentation methods are prone to deviating from physical boundaries, causing the model to learn false features.

[0084] To address the scarcity of real underwater fault samples, this invention employs Monte Carlo simulation based on a mechanistic model to generate fault samples. For each type of fault, a disturbance range is set using the corresponding mechanistic characteristic parameters as the core variables. For buoyancy-weight imbalance faults, the net negative buoyancy coefficient is used. The core disturbance parameter; physical constraint faults are represented by the equivalent constraint drag coefficient. The core disturbance parameter; failures of the self-rescue mechanism are respectively represented by the oil pump execution efficiency. Loading execution validity The core disturbance parameter; main control system crashes and communication interruptions are represented by the communication failure coefficient. These are the core perturbation parameters. The perturbation ranges for the above parameters are set based on the physical boundaries of typical UUV fault modes and a limited number of real fault acquisition samples. 2000 time-series samples are generated independently for each fault type, with a single sample time-series length of... The training set is composed of four types of simulated fault samples and an equal number of normal flight samples in a 1:1 ratio. The training set and the validation set are randomly split in an 8:2 ratio.

[0085] This solution has the following advantages: the generated mechanism simulation samples can be used to supplement the volume of fault samples in batches, and a balanced training set can be constructed by matching them with normal samples from actual sea trials, thus solving the problem of class imbalance in the dataset; the generation process is strictly constrained by explicit physical models such as buoyancy balance and restraint drag, ensuring that the samples conform to the laws of underwater motion mechanics; various types of faults and compound faults are covered by continuous parameter perturbation, filling the application scenarios that cannot be covered by sea trials; at the same time, there is no need to carry out high-risk actual sea trial fault tests, which can significantly reduce R&D costs and equipment damage risks.

[0086] S3.4, Random Forest Output Include A lightweight random forest with multiple decision trees outputs the optimal fault category through multi-tree voting:

[0087] Synchronous output of fault confidence , defined as the proportion of the highest number of votes in the voting results to the total number of votes.

[0088] Output to S5, Output to S4.

[0089] S4, Confidence-based reverse regulation of recursive learning rate closed loop This layer receives the fault confidence level output by S3. The learning rate is dynamically adjusted based on a threshold. The recursive update formula returned to S2.2 controls... The convergence speed. This is the reverse regulation pathway, which together with the forward pathway from S2 to S3 forms a bidirectional closed loop.

[0090] The formula for linear and continuous adjustment of the recursive learning rate is:

[0091] In the formula: The recursive learning rate at the current moment is dimensionless and its value is always constrained to... Within the safe zone; The current fault classification confidence level output by S3 is dimensionless. The lower confidence threshold is a dimensionless value that corresponds to the upper boundary of the low confidence smoothing interval. The preset confidence upper limit threshold is dimensionless and corresponds to the lower boundary of the high-confidence fast tracking interval; , These are the maximum and minimum learning rates calibrated by the system, respectively, and are dimensionless.

[0092] After regulation The first-order recursive smoothing formula is fed back to S2.2 in real time. When the fault confidence is in the low range, the learning rate is kept at its minimum value to perform deep smoothing on instantaneous observations, minimizing the coupling of environmental disturbances such as ocean turbulence and sensor noise into the mechanism model and avoiding meaningless drift of the equivalent binding drag coefficient. When the fault confidence is in the high range, the learning rate is kept at its maximum value to accelerate the convergence of mechanism parameters to the real fault state and ensure that the dynamic evolution characteristics of the fault can be captured quickly. When the confidence is in the intermediate transition range, the learning rate changes linearly and smoothly with the confidence, eliminating the step jump caused by threshold switching, making... The recognition process is continuous and stable, avoiding jitter in recognition results caused by sudden changes in the learning rate.

[0093] This linear continuous control mechanism retains the hard constraint boundary of the extreme value range, ensuring that the learning rate is always within the safe range of the engineering calibration, while realizing smooth adaptive adjustment of the transition range, taking into account both anti-interference stability and fault tracking speed.

[0094] S5, Timing Steady-State Decision This step receives the classification results output by S3. After the sliding window steady-state check is performed, the final release enable signal is output. This is the final execution layer of the entire method.

[0095] To address the irreversible nature of black box release, a sliding timing window steady-state verification safety mechanism is implemented to prohibit instantaneous determination-triggered release.

[0096] Sliding window failure count:

[0097] in: This represents the cumulative number of times a particular fault has been identified within the window. For the first The classification results of the second sampling; Label specific fault categories (1-physical restraint, 2-buoyancy imbalance, 3-self-rescue mechanism failure, 4-master control communication interruption). This represents the length of the sliding window.

[0098] To preset the steady-state threshold, only when The initial assessment was that it was a true steady-state fault.

[0099] Final release enable:

[0100] If and only if At that time, the emergency release command for the black box is output.

[0101] A black box release decision system for an unmanned underwater vehicle, used to implement the method of the first embodiment described above, includes: The data acquisition module is used to acquire multi-source operational observation data of the unmanned underwater vehicle; An embedded computing module is used to execute the unmanned underwater vehicle black box release decision method as described above and output the decision result; The black box drive release module is used to receive the judgment result and drive the black box to perform an emergency release action when the judgment result is release; The data storage module is used to record system operation parameters, identify intermediate results, classify decision logs, and release trigger records; The communication interaction module is used to provide a bidirectional data link between the unmanned underwater vehicle's main control system and the embedded computing module.

[0102] The data acquisition module connects to the depth sensor, oil level sensor, tank pressure sensor, current sensor, and attitude sensor. After performing signal conditioning and analog-to-digital conversion on the acquired signals, it transmits them to the embedded computing module. The embedded computing module includes a microprocessor, program memory and data memory, and deploys and runs the complete five-layer closed-loop algorithm from S1 to S5 to complete feature calculation, parameter identification, intelligent classification and timing decision. The data storage module records system operation parameters, identifies intermediate results, classifies judgment logs, and release trigger records, supporting data backtracking and accident analysis. The communication interaction module provides a two-way data link between the UUV main control system and this system, receives parameter configuration commands issued by the main control, reports fault identification status and releases pre-alarms; The black box driver release module receives the release enable signal output by the embedded computing module. The actuator is driven to complete the emergency release action of the black box.

[0103] A computer program product includes a computer program that is executed by a processor to determine the release of the unmanned underwater vehicle black box as described above.

[0104] A computer-readable storage medium includes instructions that, when executed on a computer, cause the computer to perform the aforementioned method for determining the release of the unmanned underwater vehicle's black box.

[0105] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented, in whole or in part, as a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0106] The above description is only a preferred embodiment of the present invention. It should be noted that any improvements, modifications, substitutions or variations made by those skilled in the art without departing from the principle of the present invention should be considered as being included within the protection scope of the present invention.

Claims

1. An unmanned underwater vehicle black box release decision method, characterized by, include: S1. Acquire multi-source operational observation data of the unmanned underwater vehicle, perform preprocessing, and construct a time-series feature vector F1; S2. Based on the time-series feature vector F1, the mechanism feature vector F2, which includes net negative buoyancy coefficient, equivalent restraint resistance coefficient, self-rescue execution efficiency and communication failure coefficient, is calculated by identifying explicit physical formulas and online recursive parameters. S3. The temporal feature vector F1 and the mechanism feature vector F2 are fused into a hybrid feature set, which is then input into the fault classification model to output the fault classification result and fault confidence. S4. Dynamically adjust the learning rate of the online recursive parameter identification based on the fault confidence level, and feed back the adjusted learning rate to S2 to control the identification speed of the equivalent restraint coefficient. S5. Perform a time-series steady-state verification on the fault classification results, and output the black box emergency release command when the verification passes.

2. The unmanned underwater vehicle black box release decision method of claim 1, wherein, In S2, the online recursive parameter identification of the equivalent restraint drag coefficient adopts a first-order recursive smoothing formula; in: Cp,smoothed is the smoothed equivalent drag coefficient for the current time instant; Cp, k-1is the smoothed equivalent drag coefficient at the previous time step; : learning rate, controls the magnitude of the correction of the current instantaneous value to the estimated value; Cp is the instantaneous equivalent drag coefficient calculated directly from the force balance equation for the current time instant.

3. The unmanned underwater vehicle black box release decision method of claim 1, wherein, S4 specifically refers to: When the fault confidence level is lower than the preset lower threshold, the learning rate is set to the minimum value; When the fault confidence level is higher than the preset upper limit threshold, the learning rate is set to the maximum value; When the fault confidence level is between the preset lower threshold and the upper threshold, the learning rate is linearly interpolated between the lower threshold and the maximum threshold.

4. The method for determining the release of the black box of an unmanned underwater vehicle according to claim 1, characterized in that, The fault classification model is a lightweight random forest model. The training set of the lightweight random forest model consists of a mixture of real normal samples and fault samples generated by Monte Carlo simulation based on the mechanism model.

5. The method for determining the release of the black box of an unmanned underwater vehicle according to claim 1, characterized in that, The timing steady-state verification specifically involves: within a sliding window of a preset length, counting the cumulative number of fault classification results; when the cumulative number is greater than or equal to a preset steady-state threshold, the verification is deemed successful.

6. The method for determining the release of the black box of an unmanned underwater vehicle according to claim 1, characterized in that, The mechanistic feature vector F2 also includes the oil pump execution efficiency and load dumping execution effectiveness calculated by sliding window integration.

7. The method for determining the release of the black box of an unmanned underwater vehicle according to claim 1, characterized in that, The multi-source operational observation data includes depth, oil bladder volume, oil pump switch status, jettison power-on status, cabin pressure, jettison circuit current, and attitude angle.

8. A black box release decision system for an unmanned underwater vehicle, characterized in that, include: The data acquisition module is used to acquire multi-source operational observation data of the unmanned underwater vehicle; An embedded computing module is used to execute the unmanned underwater vehicle black box release decision method as described in any one of claims 1 to 7, and output the decision result; The black box drive release module is used to receive the judgment result and drive the black box to perform an emergency release action when the judgment result is release; The data storage module is used to record system operation parameters, identify intermediate results, classify decision logs, and release trigger records; The communication interaction module is used to provide a bidirectional data link between the unmanned underwater vehicle's main control system and the embedded computing module.

9. A computer program product, comprising a computer program, characterized in that, The computer program is executed by the processor to determine the release method of the unmanned underwater vehicle black box as described in any one of claims 1-7.

10. A computer-readable storage medium comprising instructions, characterized in that, When it is run on a computer, it causes the computer to execute the unmanned underwater vehicle black box release decision method as described in any one of claims 1-7.

Images