A Method and System for Anomaly Detection of Unmanned Underwater Vehicle Thrusters Based on Vector Hydrophone Signals and Deep Learning
By combining vector hydrophones and deep learning, precise online detection of the thruster status of unmanned underwater vehicles has been achieved, solving the problem of poor detection stability in existing technologies and improving the reliability and accuracy of detection.
Patent Information
- Application Number
- CN202611018496.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-09
- Publication Date
- 2026-11-10
AI Technical Summary
Current unmanned underwater vehicle (UUV) thruster status detection relies on manual observation, making it difficult to achieve continuous, objective, and quantitative status assessment. The installation of internal sensors is limited, external radiation characteristics are not fully perceived, and multivariate correlation features under complex navigation conditions are difficult to extract effectively and perform time-series modeling, resulting in poor detection stability.
Vector hydrophones are used to simultaneously acquire sound pressure and triaxial vibration velocity signals. Combined with attitude correction and steady-state screening, a deep learning two-layer long short-term memory network (LSTM) model is used to model multivariate time-series samples. The deviation is calculated using Mahalanobis distance and weighted fusion is used to achieve thruster anomaly detection.
It improves the observability and stability of thruster status detection, enhances the ability to detect thruster radiated noise, reduces false alarm and false alarm rates, and has good anti-interference capabilities and environmental adaptability.
Smart Images

Figure CN123065792A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of underwater vehicle status monitoring and intelligent sensing technology, and particularly relates to a method and system for detecting anomalies in the thrusters of unmanned underwater vehicles based on vector hydrophone signals and deep learning. Background Technology
[0002] Autonomous underwater vehicles (UUVs) possess characteristics such as autonomous navigation, strong stealth capabilities, and adaptability to complex marine environments, and have been widely applied in missions such as marine topographic mapping, underwater target detection, environmental monitoring, and seabed resource surveys. During UUV missions, the propulsion system is responsible for providing continuous propulsion and maneuverability, and is the key execution unit for the entire platform to complete track tracking and mission operations. Its operational status directly affects the UUV's speed control accuracy, track maintenance capability, endurance, and mission safety.
[0003] In actual missions, unmanned underwater vehicles (UUVs) typically need to operate continuously over long distances and in complex hydrodynamic environments, with their thrusters under prolonged high loads or frequent speed adjustments. Factors such as water flow disturbances, thruster blade wear, localized adhesion, structural loosening, and changes in the working medium can gradually cause the thruster's operating condition to deviate from normal. These anomalies are often not easily detected visually in their early stages, but they gradually lead to decreased propulsion efficiency, increased energy consumption, increased attitude disturbances, and a heavier control burden. For long-range missions, failure to identify thruster anomalies early on can result in mission failure.
[0004] Current thruster condition monitoring primarily relies on manual observation, typically assessing operational status by checking thruster rotation speed and motor current. This method is highly dependent on operator experience and struggles to achieve continuous, objective, and quantitative condition assessment. Furthermore, existing technologies can employ internal measurement methods such as motor current and speed feedback, shaft vibration, or internally mounted acceleration sensors. However, for compact, pressure-resistant unmanned underwater vehicle platforms, the placement of internal sensors is severely limited by space, power consumption, and sealing conditions. Moreover, internal measurement signals primarily reflect the local state of the drive system, failing to adequately characterize the external radiation characteristics generated by the interaction between the thruster and surrounding fluids, resulting in incomplete condition perception.
[0005] Thrusters generate significant radiated noise and localized fluid disturbances when operating in water. Unlike traditional scalar hydrophones that only measure sound pressure, vector hydrophones can simultaneously acquire sound pressure and particle velocity information, providing richer feature data. However, vector hydrophone signals during the navigation of unmanned underwater vehicles are simultaneously affected by flow noise, attitude changes, rotational speed fluctuations, and environmental background noise. Without condition compensation and time-series modeling, direct comparisons between different time windows are difficult. Furthermore, thruster anomalies often manifest as temporal joint deviations of multiple variables (such as rotational speed, blade frequency energy, vibrational velocity response, and attitude disturbances) over continuous time periods. Using only single-moment threshold comparisons or static feature classification methods fails to fully utilize temporal correlation information, resulting in poor stability and robustness of detection results under complex conditions. Therefore, how to introduce modeling methods that can effectively describe the evolution of multivariate time series (such as Long Short-Term Memory networks, LSTM) to achieve accurate, online thruster anomaly monitoring is a pressing issue in this field. Summary of the Invention
[0006] To address the problems in existing technologies for monitoring the status of unmanned underwater vehicle (UUV) thrusters, such as strong reliance on manual intervention, limited installation of internal sensors, incomplete perception of external radiation characteristics, and poor detection stability due to difficulties in effectively extracting multivariate correlation features and time-series modeling under complex navigation conditions, this invention provides an anomaly detection method and system for UUV thrusters based on vector hydrophone signals and deep learning.
[0007] This invention is implemented as follows: a method for detecting anomalies in the thrusters of unmanned underwater vehicles based on vector hydrophone signals and deep learning, characterized by comprising the following steps: Step S1: Signal acquisition. Acquire sound pressure signals and three-axis vibration velocity signals simultaneously through the vector hydrophone installed at the bow of the unmanned underwater vehicle, and simultaneously acquire the propulsion speed, speed and attitude angle signals of the unmanned underwater vehicle. Step S2: Attitude correction. A rotation matrix is constructed using the attitude angle signal to transform the three-axis vibration velocity signal from the sensor coordinate system to the hull coordinate system, thereby obtaining a vector acoustic signal in a consistent reference system. Step S3: Steady-state screening. After preprocessing the signal, the steady-state condition is segmented according to the fluctuation of the propulsion speed, speed and attitude angle signals, and the steady-state condition segments are retained. Step S4: Feature transformation, taking the propulsion speed as a reference, synchronously resample the steady-state operating condition segment, convert the time domain signal into an order domain sequence, and extract the blade frequency energy sequence from it; Step S5: Sample construction. The sound pressure sequence, the converted triaxial vibration velocity sequence, the blade frequency energy sequence, and the time series sequences corresponding to propulsion speed, speed, and attitude angle are spliced together according to a fixed time window to construct a multivariable time series sample. Step S6: Probability determination. Input the multivariate time series samples into the pre-trained two-layer long short-term memory network LSTM model and output the thruster anomalous probability. Step S7: Fusion decision: Calculate the deviation of the multivariate time series sample from the preset health baseline, normalize the deviation to obtain the normalized deviation, and weight and fuse the anomaly probability with the normalized deviation to calculate the final anomaly score. Step S8: Output the result. When the final abnormal score exceeds the preset threshold, output the abnormal result of the thruster.
[0008] In the above technical solution, preferably, in step S3, the preprocessing includes bandpass filtering and high-pass flow noise suppression of the sound pressure and triaxial vibration velocity signals.
[0009] In the above technical solution, preferably, in step S3, the criterion for the steady-state operating condition segmentation is: within a continuous fixed time window, the propulsion speed fluctuation rate does not exceed 5%, the speed fluctuation rate does not exceed 8%, and the pitch angle change rate does not exceed 3 degrees per second.
[0010] In the above technical solution, preferably, in step S4, the leaf frequency... Satisfy the following formula:
[0011] in, For the number of propeller blades, To increase the rotational speed.
[0012] In the above technical solution, preferably, in step S5, the splicing time window length of the multivariate time series samples is fixed at 5 seconds.
[0013] In the above technical solution, preferably, in step S6, the dual-layer long short-term memory network LSTM model includes two cascaded LSTM hidden layers and one fully connected output layer, and the output layer uses the Sigmoid function to output the anomaly probability.
[0014] In the above technical solution, preferably, the training process of the two-layer long short-term memory network LSTM model uses binary cross-entropy as the loss function and the Adam optimizer.
[0015] In the above technical solution, preferably, in step S7, the health baseline is obtained by extracting feature vectors from samples of the unmanned underwater vehicle thruster under normal conditions and calculating the mean and covariance matrix; the deviation is calculated using Mahalanobis distance.
[0016] In the above technical solution, preferably, in step S7, the weight allocation of the weighted fusion is: the abnormal probability weight is 0.6, and the healthy baseline deviation weight is 0.4; in step S8, the preset threshold is 0.65.
[0017] The unmanned underwater vehicle thruster anomaly detection method based on vector hydrophone signals and deep learning proposed in this invention has the following advantages and effects: This invention addresses the need for anomaly detection in the thrusters of unmanned underwater vehicles (UUVs). By combining vector signal processing and deep learning technologies, it offers significant technical advantages and excellent application results. Firstly, this invention provides enhanced observability in terms of information perception. By employing a vector hydrophone to simultaneously acquire sound pressure and triaxial vibration velocity information, compared to traditional detection methods that rely solely on scalar sound pressure, it significantly enhances the perception of thruster radiated noise and its spatial directional characteristics from a multi-dimensional acoustic feature perspective. This provides a richer and more discriminative raw data foundation for subsequent accurate identification.
[0018] This invention ensures the stability of test results through a standardized processing procedure. By employing fixed operating condition screening rules, fixed time window lengths, and fixed synchronous resampling strategies, it effectively unifies data specifications across different navigation stages and dynamic environments, greatly improving sample consistency. This standardized processing approach makes the thruster characteristics under different navigation states scientifically comparable, thereby enhancing the system's operational reliability in long-cycle missions.
[0019] Regarding the core judgment mechanism, this invention fully leverages the modeling advantages of deep learning for complex temporal features. By employing a two-layer Long Short-Term Memory (LSTM) network to judge multivariate temporal samples of the thruster, it can deeply mine the temporal correlation between thruster radiated noise and motion parameters, effectively capturing the dynamic features of evolution from normal state to slight anomaly. Simultaneously, the system uses a fixed-weight fusion approach, combining the probability output of the LSTM with the deviation from the health baseline based on Mahalanobis distance. This leverages the efficient judgment capability of the data-driven model while introducing a physically meaningful deviation assessment, balancing detection sensitivity and engineering interpretability.
[0020] This invention possesses strong anti-interference capabilities and environmental adaptability. It eliminates deviations caused by changes in the sensor coordinate system through attitude correction, filters transient interference during unsteady navigation using steady-state operating conditions, and eliminates frequency drift caused by speed fluctuations by combining propulsion speed-based order analysis. These targeted technical measures ensure that the propulsion's abnormal characteristics remain highly distinguishable in complex and variable underwater environments, significantly reducing the system's false alarm and false negative rates.
[0021] This invention proposes an anomaly detection system for unmanned underwater vehicle (UUV) thrusters based on vector hydrophone signals and deep learning. The system implements the aforementioned UUV thruster anomaly detection method and includes a signal acquisition module, a preprocessing module, a sample construction module, an anomaly determination module, a baseline deviation calculation module, a fusion decision module, and a result output module. The signal acquisition module simultaneously acquires sound pressure signals, three-axis vibration velocity signals, thruster speed, speed, and attitude angles. The preprocessing module performs attitude correction, flow noise suppression, steady-state condition segmentation, and synchronous resampling on the acquired signals. The sample construction module constructs fixed-length multivariate time-series samples. The anomaly determination module executes a two-layer LSTM model inference and outputs anomaly probabilities. The baseline deviation calculation module calculates the deviation of the test sample from the healthy baseline. The fusion decision module fuses the anomaly probabilities and deviations with fixed weights to obtain a final anomaly score. The result output module outputs the thruster anomaly detection result based on the final anomaly score. Attached Figure Description
[0022] Figure 1 This is a schematic diagram of the process for detecting abnormalities in the thruster of an unmanned underwater vehicle according to the present invention; Figure 2 This is a schematic diagram of the principle structure of the LSTM model in this invention; Figure 3 This is a schematic diagram showing the fixed installation position of the vector hydrophone on the hull of the unmanned underwater vehicle in this invention.
[0023] Figure 4 This is a structural block diagram of the unmanned underwater vehicle thruster anomaly detection system of the present invention. Detailed Implementation
[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0025] This invention provides a method and system for detecting thruster anomalies in unmanned underwater vehicles based on vector hydrophone signals and deep learning. To further illustrate the structure of this invention, a detailed description is provided below in conjunction with the accompanying drawings: Please see Figure 1 and Figure 2 An anomaly detection method for the thrusters of an unmanned underwater vehicle based on vector hydrophone signals and deep learning includes the following steps: Step S1: Signal acquisition. Sound pressure signals and three-axis vibration velocity signals are simultaneously acquired by a vector hydrophone installed at the bow of the unmanned underwater vehicle, and the propulsion speed, speed and attitude angle signals of the unmanned underwater vehicle are also acquired simultaneously.
[0026] Step S2: Attitude correction. A rotation matrix is constructed using the attitude angle signal to transform the three-axis vibration velocity signal from the sensor coordinate system to the hull coordinate system, thereby obtaining a vector acoustic signal in a consistent reference system.
[0027] Step S3: Steady-state screening. After preprocessing the signal, the steady-state condition segment is segmented according to the fluctuation of the propulsion speed, speed and attitude angle signals, and the steady-state condition segment is retained. The preprocessing includes bandpass filtering and high-pass flow noise suppression of the sound pressure and triaxial vibration velocity signals.
[0028] Step S4: Feature transformation: Using the propulsion speed as a reference, synchronously resample the steady-state operating condition segment to convert the time domain signal into an order domain sequence and extract the blade frequency energy sequence from it.
[0029] Step S5: Sample construction. The sound pressure sequence, the converted triaxial vibration velocity sequence, the blade frequency energy sequence, and the time series sequences corresponding to propulsion speed, speed, and attitude angle are spliced together according to a fixed time window to construct a multivariable time series sample.
[0030] Step S6: Probability determination. Input the multivariate time series samples into a pre-trained two-layer long short-term memory (LSTM) network model and output the propeller anomaly probability. The two-layer LSTM network model includes two cascaded LSTM hidden layers and a fully connected output layer. The output layer uses the Sigmoid function to output the anomaly probability. The training process of the two-layer LSTM network model uses binary cross-entropy as the loss function and the Adam optimizer.
[0031] Step S7: Fusion decision-making. Calculate the deviation of the multivariate time-series sample from the preset healthy baseline, normalize the deviation to obtain the normalized deviation, and weight and fuse the anomaly probability with the normalized deviation to calculate the final anomaly score. The healthy baseline is obtained by extracting feature vectors from samples of the unmanned underwater vehicle's thrusters under normal conditions and calculating their mean and covariance matrices. The deviation is calculated using Mahalanobis distance.
[0032] Step S8: Output the result. When the final abnormal score exceeds the preset threshold, output the abnormal result of the thruster.
[0033] It should be noted that steps S1 to S8 constitute the core process framework of the method of this invention. To further clarify the implementation principles and internal logical connections of the key technical points in each step, several core links will be explained in detail below.
[0034] Firstly, regarding time synchronization and attitude correction, the sound pressure channel, three-axis vibration velocity channel, propulsion speed, speed, and attitude angle data are uniformly aligned to the same time axis. Because the unmanned underwater vehicle (UUV) experiences pitch, roll, and yaw changes during navigation, the three-axis vibration velocity components output by the vector hydrophone are always in the sensor coordinate system that changes with attitude, resulting in a lack of direct comparability of directional features acquired at different time points. Therefore, a rotation matrix is constructed using attitude angles to transform the three-axis vibration velocity signals from the sensor coordinate system to the hull coordinate system, obtaining vector acoustic signals with consistent reference coordinates, thereby reducing the impact of attitude changes on subsequent anomaly detection results.
[0035] Secondly, during signal preprocessing and steady-state screening, fixed-bandband bandpass filtering is applied to the sound pressure and triaxial vibration velocity signals, and high-pass flow noise suppression is performed on the low-frequency components. Then, a continuous 5-second sliding time window is used for steady-state condition screening, retaining only sample segments with propulsion speed fluctuations not exceeding 5%, airspeed fluctuations not exceeding 8%, and pitch angle changes not exceeding 3 degrees per second. Without steady-state screening, non-stationary spectral distortions caused by rapid propulsion speed adjustments, large turns, or rapid attitude changes are easily misidentified as propulsion anomalies, significantly increasing the false alarm rate. Therefore, steady-state condition screening ensures a relatively stable correspondence between the input samples and the actual operating state of the propulsion.
[0036] Furthermore, regarding synchronous resampling and order analysis, propulsion speed is used as the sole synchronous reference quantity. Synchronous resampling is performed on steady-state operating segments, converting the time-domain signal into an order-domain sequence. Since propellers are rotating machinery, their anomalies often manifest first as changes in blade frequency and its associated harmonic energy. If time-domain spectrum analysis is directly used, these characteristic frequencies will drift under propulsion speed fluctuations, making direct comparison between different time windows difficult. After synchronous resampling, the blade frequency and its harmonics can be stably mapped to a fixed order position. If the number of propeller blades is Z, the blade frequency position satisfies this requirement. This embodiment extracts the blade frequency energy sequence from the order-domain sequence and combines it with other time-series quantities to form the input sample, highlighting the propeller anomaly-related characteristics.
[0037] Then, for the construction of multivariate time series samples, the sound pressure sequence, triaxial vibration velocity sequence, blade frequency energy sequence, propeller speed sequence, air speed sequence, and pitch angle sequence are spliced together according to a uniform time step to form a multivariate time series sample with a fixed length of 5 seconds. Each sample contains a data structure of the same dimension and length to meet the stable input requirements of the two-layer LSTM.
[0038] Subsequently, regarding the establishment of the health baseline, multiple sets of samples were collected under normal operating conditions of the unmanned underwater vehicle's thrusters. A baseline sample set was generated using the same preprocessing and sample construction methods as in the online detection phase. The mean and covariance matrix of the manually generated eigenvectors in the baseline samples were calculated, and Mahalanobis distance was used as the method for calculating the deviation from the health baseline.
[0039] An anomaly detection network is constructed, consisting of two cascaded LSTM layers and a fully connected output layer. The constructed multivariate temporal samples are input into the network sequentially, with the feature vector at each time step serving as the current input x(t). The first LSTM layer incorporates the hidden state h(t) from the previous time step. 1) and cell state c(t) 1) Complete temporal feature extraction and pass the output to the second LSTM layer. The second LSTM layer further models the extracted temporal features and outputs the hidden state of the last time step. This hidden state is input into the fully connected output layer, and the probability value of the current sample belonging to an abnormal state is output through the Sigmoid function. During the training phase, supervised training is performed using labeled normal samples and abnormal samples. The loss function is fixed as binary cross-entropy, and the optimizer is fixed as Adam.
[0040] Finally, for LSTM model construction and online anomaly detection, the same preprocessing and sample construction process as in the training phase is performed on the online test samples. The generated multivariate time-series samples are input into the LSTM model to obtain the anomaly probability. Simultaneously, the Mahalanobis distance deviation of the test sample relative to the healthy baseline is calculated. After normalizing the anomaly probability and deviation, they are linearly fused with fixed weights of 0.6 and 0.4 respectively to obtain the final anomaly score.
[0041] When the final anomaly score is greater than or equal to 0.65, the system directly determines that the thruster is in an abnormal state and outputs a thruster anomaly alarm message; when the final anomaly score is less than 0.65, the system determines that the thruster is in a normal state.
[0042] To further clarify the specific implementation of the present invention, a preferred hardware deployment example is described below.
[0043] like Figure 3 As shown, a vector hydrophone 1 is fixedly installed at the bow of the unmanned underwater vehicle. The vector hydrophone is used to synchronously collect sound pressure signal p(t) and three-axis vibration velocity signals vx(t), vy(t), vz(t), and simultaneously collect propulsion speed n(t), speed u(t) and attitude angle θ(t).
[0044] Specifically, the sound pressure signal, three-axis vibration velocity signal, propulsion speed, speed, and attitude angle are time-aligned. Because the unmanned underwater vehicle (UUV) undergoes attitude changes during navigation, the three-axis vibration velocity signal resides in a time-varying sensor coordinate system, making signals from different time periods incomparable. Therefore, a coordinate transformation matrix is constructed based on the attitude angle to uniformly transform the three-axis vibration velocity signal to the hull coordinate system, eliminating the influence of attitude changes on the sound field direction distribution and obtaining a vector acoustic signal in a consistent reference frame.
[0045] The corrected vector acoustic signal is subjected to bandpass filtering and high-pass flow noise suppression. During rapid speed regulation or drastic attitude changes of the thruster, the radiated noise spectrum structure will undergo non-stationary distortion, and the spectrum changes caused by operating condition fluctuations will be misidentified as abnormal characteristics of the thruster. Therefore, a steady-state criterion is set based on the propulsion speed fluctuation rate, speed fluctuation rate, and attitude change rate, and the signal is processed by steady-state operating condition segmentation, retaining only the steady-state data. Then, synchronous resampling is performed with the propulsion speed as the only reference quantity to convert the time domain signal into an order domain sequence. Since the thruster is a rotating machine, its fault characteristics are mainly reflected in the order frequencies related to the speed. If the time domain signal is directly used for analysis, the characteristic frequency will drift under the condition of speed fluctuation, making it difficult to extract weak abnormal characteristics stably. Therefore, the blade frequency and its harmonics are stably mapped to a fixed order position through order domain transformation, thereby enhancing the distinguishability of fault characteristics and reducing the false alarm rate.
[0046] The steady-state operating condition segment adopts the following fixed criteria: within a continuous 5-second time window, the propulsion speed fluctuation rate does not exceed 5%, the speed fluctuation rate does not exceed 8%, and the pitch angle change rate does not exceed 3 degrees per second.
[0047] Synchronous resampling uses propulsion speed as the sole reference quantity to convert the time-domain signal into an order-domain signal; when the number of propeller blades is... Z The propulsion speed is n At that time, the leaf frequency satisfies:
[0048] The sound pressure level (SPL) sequence, triaxial velocity sequence, blade frequency energy sequence, propeller speed sequence, air speed sequence, and attitude angle sequence are aligned along a unified time axis and spliced together according to a fixed time window to construct a multivariate time-series sample with a unified dimension and length. This allows various acoustic features and motion parameters to be coupled and expressed in the time dimension. In this embodiment, the multivariate time-series sample consists of the SPL signal, triaxial velocity signal, propeller speed, air speed, pitch angle, and blade frequency energy sequence, with a fixed time window length of 5 seconds.
[0049] Thruster anomalies typically manifest as coordinated changes in multiple physical quantities over time, rather than abrupt changes at a single moment. Therefore, a two-layer Long Short-Term Memory (LSTM) network is used to model multivariate time-series samples to extract the evolutionary characteristics of the thruster's operating state over time and output the thruster anomaly probability. In this embodiment, the two-layer LSTM model consists of two cascaded LSTM hidden layers and a fully connected output layer. The output layer uses a sigmoid function to output the thruster anomaly probability.
[0050] The thruster anomaly probability and the pre-established health baseline deviation are weighted and fused according to a fixed weight. This allows for the simultaneous use of data-driven time-series judgment results and statistical deviation information based on normal sample distribution, achieving a final thruster anomaly score calculation that balances anomaly identification capability and engineering interpretability. In this embodiment, the health baseline model is established based on samples collected under normal conditions of the unmanned underwater vehicle's thrusters. The health baseline deviation is calculated using Mahalanobis distance, and the final anomaly score is fused using an LSTM anomaly probability weight of 0.6 and a health baseline deviation weight of 0.4.
[0051] When the final anomaly score exceeds a preset threshold, the unmanned underwater vehicle's thruster is determined to be in an abnormal state, and an anomaly detection result is output. The preset threshold is fixed at 0.65. When the final anomaly score is greater than or equal to 0.65, a thruster anomaly alarm message is output.
[0052] This technical solution, in low signal-to-noise ratio (SNR) underwater environments, weightedly fuses the temporal anomaly probability output by LSTM with the healthy baseline deviation based on Mahalanobis distance, resulting in a synergistic SNR gain. Specifically, when neither LSTM nor Mahalanobis distance alone can effectively distinguish between abnormal and normal states (e.g., the signal energy output by both is close to the background noise fluctuation level), the weighted fusion result significantly improves the contrast of the anomaly score. This effect is not expected from a simple superposition of the two methods, but rather stems from the complementary nature of LSTM's sensitivity to the time dimension and Mahalanobis distance's sensitivity to the covariance structure between variables, ensuring that the integrated decision remains stable even under interference.
[0053] In this embodiment, to further improve the accuracy of identifying early, subtle anomalies caused by localized blade wear, minor adhesions, or slight deformation, the present invention also introduces an empirical judgment index that reflects the nonlinear coupling relationship between acoustic energy flow and physical power consumption. This index, by capturing the deviation between the mechanical vibration energy sensed by the vector hydrophone and the actual operating conditions of the submersible, can detect propulsion efficiency degradation earlier than conventional features. This empirical judgment index The specific calculation formula is as follows:
[0054] in, To obtain triaxial vibration velocity signals through synchronous acquisition ( The vector total kinetic energy eigenvalue obtained by performing the root mean square integral directly reflects the intensity of the mechanical energy flow generated by the propeller disturbing the water. (Formula denominator term) An ideal power consumption benchmark based on the fundamental principles of fluid mechanics was constructed, in which... The density of seawater, For the speed of the underwater vehicle, This value represents the propulsion energy required for the submersible to maintain steady-state navigation at its current speed, where the thruster diameter is the reference value. Furthermore, The leaf frequency energy center after order transformation The standard deviation of the vibration velocity signal is denoted by , and the ratio of to constitutes the frequency domain stability coefficient, used to characterize the signal coherence. In the formula... , , The coupling coefficient for a specific platform is set based on the aerodynamic structure of the submarine hull and the hydrodynamic characteristics of the propeller. In this embodiment, The value is 1.35. The value is 0.88. The value is 0.08. Using this empirical formula, the system can detect minute load fluctuations that were originally difficult to perceive in the original time-domain signal. The exponential transformation into a significant numerical transition significantly enhances the discernibility of early anomalous signals, providing crucial data support for predictive maintenance of the thruster.
[0055] Example 2 like Figure 2 As shown, this embodiment proposes an anomaly detection system for the thruster of an unmanned underwater vehicle based on vector hydrophone signals and deep learning. The system includes a signal acquisition module, a preprocessing module, a sample construction module, an LSTM anomaly detection module, a baseline deviation calculation module, a fusion decision module, and a result output module. These modules are connected and executed in a fixed order.
[0056] The system comprises the following modules: a signal acquisition module for acquiring sound pressure, three-axis vibration velocity, propulsion speed, speed, and attitude angle; a preprocessing module for time synchronization, attitude correction, bandpass filtering, high-pass flow noise suppression, steady-state condition segmentation, and synchronous resampling; a sample construction module for generating fixed-length multivariate time-series samples; an LSTM anomaly detection module for performing two-layer LSTM inference and outputting anomaly probabilities; a baseline deviation calculation module for calculating Mahalanobis distance deviation; a fusion decision module for calculating the final anomaly score; and a result output module for outputting thruster anomaly results.
[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for detecting thruster anomalies in unmanned underwater vehicles based on vector hydrophone signals and deep learning, characterized in that, Includes the following steps: Step S1: Signal acquisition. Acquire sound pressure signals and three-axis vibration velocity signals simultaneously through the vector hydrophone installed at the bow of the unmanned underwater vehicle, and simultaneously acquire the propulsion speed, speed and attitude angle signals of the unmanned underwater vehicle. Step S2: Attitude correction. A rotation matrix is constructed using the attitude angle signal to transform the three-axis vibration velocity signal from the sensor coordinate system to the hull coordinate system, thereby obtaining a vector acoustic signal in a consistent reference system. Step S3: Steady-state screening. After preprocessing the signal, the steady-state condition is segmented according to the fluctuation of the propulsion speed, speed and attitude angle signals, and the steady-state condition segments are retained. Step S4: Feature transformation, taking the propulsion speed as a reference, synchronously resample the steady-state operating condition segment, convert the time domain signal into an order domain sequence, and extract the blade frequency energy sequence from it; Step S5: Sample construction. The sound pressure sequence, the converted triaxial vibration velocity sequence, the blade frequency energy sequence, and the time series sequences corresponding to propulsion speed, speed, and attitude angle are spliced together according to a fixed time window to construct a multivariable time series sample. Step S6: Probability determination. Input the multivariate time series samples into the pre-trained two-layer long short-term memory network LSTM model and output the thruster anomalous probability. Step S7: Fusion decision: Calculate the deviation of the multivariate time series sample from the preset health baseline, normalize the deviation to obtain the normalized deviation, and weight and fuse the anomaly probability with the normalized deviation to calculate the final anomaly score. Step S8: Output the result. When the final abnormal score exceeds the preset threshold, output the abnormal result of the thruster.
2. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S3, the preprocessing includes bandpass filtering and high-pass flow noise suppression of the sound pressure and triaxial vibration velocity signals.
3. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S3, the criteria for steady-state condition segmentation are: within a continuous fixed time window, the propulsion speed fluctuation rate does not exceed 5%, the speed fluctuation rate does not exceed 8%, and the pitch angle change rate does not exceed 3 degrees per second.
4. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S4, the leaf frequency Satisfy the following formula: in, For the number of propeller blades, To increase the rotational speed.
5. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S5, the splicing time window length of the multivariate time series samples is fixed at 5 seconds.
6. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S6, the dual-layer long short-term memory network (LSTM) model includes two cascaded LSTM hidden layers and a fully connected output layer. The output layer uses the Sigmoid function to output the anomaly probability.
7. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 6, characterized in that, The training process of the two-layer long short-term memory network LSTM model uses binary cross-entropy as the loss function and the Adam optimizer.
8. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S7, the health baseline is obtained by extracting feature vectors from samples of the unmanned underwater vehicle thrusters under normal conditions and calculating the mean and covariance matrix; the deviation is calculated using Mahalanobis distance.
9. The method for detecting abnormalities in the thruster of an unmanned underwater vehicle according to claim 1, characterized in that, In step S7, the weight allocation for the weighted fusion is as follows: the abnormal probability weight is 0.6 and the healthy baseline deviation weight is 0.4; in step S8, the preset threshold is 0.
65.
10. An anomaly detection system for the thruster of an unmanned underwater vehicle based on vector hydrophone signals and deep learning, characterized in that, The system is used to implement the unmanned underwater vehicle thruster anomaly detection method as described in any one of claims 1 to 9, the system comprising: The signal acquisition module is used to synchronously acquire sound pressure signals, three-axis vibration velocity signals, propulsion speed, speed, and attitude angles. The preprocessing module is used to perform attitude correction, flow noise suppression, steady-state segmentation, and synchronous resampling on the acquired signals; The sample construction module is used to construct fixed-length multivariate time series samples. The anomaly detection module is used to perform inference using a two-layer LSTM model and output the anomaly probability. The baseline deviation calculation module is used to calculate the deviation of the test sample from the healthy baseline. The fusion decision module is used to fuse the anomaly probability and deviation degree with fixed weights to obtain the final anomaly score; The results output module is used to output the thruster anomaly detection results based on the final anomaly score.
