A method and system for assessing the health of intelligent ships
By constructing a timing buffer and a dual-channel collaborative architecture in the ship propulsion system, the initial fault classification probability of artificial intelligence is cached and corrected using physical damping characteristic parameters. This solves the problem of timing misalignment between artificial intelligence and physical calculation, and achieves accuracy and reliability in ship health assessment.
Patent Information
- Application Number
- CN202611623878.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
There is a temporal misalignment between the rapid reasoning of artificial intelligence and the slow settlement of physical data, which makes it difficult for physical features to constrain the diagnostic results of artificial intelligence in a timely manner, thus making it difficult to achieve accurate ship health assessment.
By acquiring multimodal operational status data of the ship propulsion system, using a time-series buffer to cache the initial fault classification probability of artificial intelligence, and correcting it when generating physical damping characteristic parameters, a dual-channel collaborative architecture is constructed to achieve alignment and fusion in the time dimension.
It effectively eliminates the impact of timing misalignment, improves the accuracy and reliability of health assessment of ship propulsion systems, ensures timely constraints of physical characteristics on artificial intelligence diagnostic results, and enhances the system's assessment capabilities under complex operating conditions.
Smart Images

Figure CN123115979A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an intelligent ship health assessment method and system, belonging to the field of ship condition monitoring technology. Background Technology
[0002] As ships become larger and more intelligent, the structure of their propulsion systems is becoming increasingly complex, making accurate assessment of their health status crucial for ensuring navigational safety. Currently, in the field of ship health assessment, artificial intelligence diagnostic techniques based on multimodal sensor data or physical diagnostic techniques based on dynamic models are commonly used.
[0003] In existing evaluation methods, AI diagnostic models can quickly extract and finely classify multi-source nonlinear fault features, while physical feature calculation requires waiting for the complete transient response process, which is time-consuming. Due to the significant temporal misalignment between the fast AI reasoning and the slow physical calculation, it is difficult to accurately align the two in the time dimension when they are fused. Physical features cannot constrain the AI diagnostic results in a timely and effective manner, making it difficult to achieve real-time correction of AI evaluation by physical laws.
[0004] Therefore, a new solution is needed to address the technical problem of a temporal misalignment between the fast reasoning of artificial intelligence and the slow settlement of physical systems, which prevents physical features from constraining the diagnostic results of artificial intelligence in a timely manner. Summary of the Invention
[0005] The technical problem to be solved by this invention is to provide an intelligent ship health assessment method and system, which solves the problem that the timing misalignment between the fast reasoning of artificial intelligence and the slow settlement of physical systems makes it impossible to achieve timely physical constraints on artificial intelligence.
[0006] The technical problem to be solved by this invention is achieved by the following technical solution:
[0007] A method and system for assessing the health of intelligent ships, comprising the following steps:
[0008] Acquire multimodal operating status data of the ship propulsion system, wherein the multimodal operating status data includes transient torque signals;
[0009] The multimodal operating state data, excluding the transient torque signal, are input into a preset fault classification model for inference to obtain an initial fault classification probability, and the initial fault classification probability is stored in a time-series buffer.
[0010] The transient excitation event is determined based on the fluctuation characteristics of the transient torque signal, the torque decay envelope corresponding to the transient excitation event is extracted, and the physical damping characteristic parameters are obtained by fitting the torque decay envelope.
[0011] At the time when the physical damping characteristic parameters are generated, the corresponding initial fault classification probability is read from the time-series buffer, and the initial fault classification probability is corrected using the physical damping characteristic parameters to obtain the target health assessment result.
[0012] Optionally, determining the transient excitation event based on the fluctuation characteristics of the transient torque signal includes:
[0013] The natural frequency of shaft torsional vibration of the ship propulsion system is obtained, and the length of the sliding time window is determined based on the natural frequency of shaft torsional vibration.
[0014] The coefficient of variation of the transient torque signal is calculated within the sliding time window, where the coefficient of variation is the ratio of the standard deviation to the mean of the transient torque signal within the sliding time window.
[0015] When the coefficient of variation exceeds a preset threshold, the transient excitation event is determined to have occurred, and the current triggering time is recorded. The triggering time is used as the start time of the transient excitation event.
[0016] Optionally, the step of extracting the torque decay envelope corresponding to the transient excitation event and fitting the physical damping characteristic parameters based on the torque decay envelope includes:
[0017] Starting from the triggering time, the transient torque signal after the transient excitation event is subjected to bandpass filtering to obtain a target frequency band torsional vibration signal that includes the natural frequency of the shaft system torsional vibration.
[0018] Perform Hilbert transform on the torsional vibration signal in the target frequency band to extract the torque attenuation envelope;
[0019] An exponential function is fitted to the torque attenuation envelope to calculate the physical damping coefficient, which characterizes the energy dissipation level of the shaft system, and serves as the physical damping characteristic parameter.
[0020] Optionally, after obtaining the physical damping characteristic parameters and before correcting the initial fault classification probability using the physical damping characteristic parameters, the method further includes:
[0021] Obtain the cooling water temperature parameters of the ship's propulsion system and obtain a preset reference temperature;
[0022] Calculate the difference between the cooling water temperature parameter and the reference temperature;
[0023] Calculate the square of the difference, and sum the product of the difference and a first preset coefficient, the product of the square and a second preset coefficient, and a preset environmental correction benchmark value to obtain the environmental correction coefficient.
[0024] The physical damping characteristic parameters are compensated and corrected using the environmental correction coefficient to eliminate damping reference drift caused by cross-ocean environmental temperature differences.
[0025] Optionally, the step of correcting the initial fault classification probability using the physical damping characteristic parameters to obtain the target health assessment result includes:
[0026] Obtain the deviation between the physical damping characteristic parameters and the preset physical reference damping parameters;
[0027] The deviation is input into a preset mapping function to obtain the penalty weights corresponding to each fault category. A diagonal matrix is constructed with each penalty weight as the diagonal element, which serves as the physical constraint penalty function matrix.
[0028] The physical constraint penalty function matrix and the initial fault classification probability are subjected to matrix operation. The matrix operation is to multiply the physical constraint penalty function matrix and the initial fault classification probability vector element by element, and then divide by the sum of the elements of the vector obtained by the element-wise multiplication to achieve linear normalization. This is to suppress abnormal classification results that violate physical laws in the initial fault classification probability, and to enhance the classification weight of fault categories that are positively correlated with the deviation trend of the physical damping feature parameter relative to the physical reference damping parameter, so as to obtain the corrected target fault classification probability.
[0029] The target health assessment result includes the target failure classification probability.
[0030] Optionally, storing the initial fault classification probability in a time-series buffer includes:
[0031] The fault classification model is continuously run at a preset inference period and an initial fault classification probability with a timestamp is generated. The inference period is from 10 milliseconds to 100 milliseconds.
[0032] Construct a time-series buffer with a preset buffer duration. The time-series buffer is a sliding buffer queue. The buffer duration is 2 to 5 seconds. The buffer duration is greater than or equal to the time window length required to calculate the physical damping characteristic parameters.
[0033] When generating the physical damping characteristic parameters, an initial fault classification probability is selected from the time-series buffer based on the timestamp of the generation time, with a timestamp no later than the generation time and the smallest time difference between the timestamp and the generation time.
[0034] Optionally, it also includes a static water condition assessment step:
[0035] When the coefficient of variation of the transient torque signal is detected to be lower than the preset still water determination threshold for a preset duration, it is determined that the transient excitation event has not occurred, and the system switches to still water disturbance detection mode.
[0036] The perturbation signal of the blade passing frequency torque generated by the propeller rotation is obtained, wherein the blade passing frequency is the product of the number of propeller blades and the shaft rotation frequency;
[0037] The time-domain averaging of the blade passage frequency torque perturbation signal is performed using a multi-cycle synchronous averaging method to amplify the perturbation characteristic amplitude. The average number of cycles of the multi-cycle synchronous averaging method is determined based on the signal-to-noise ratio of the blade passage frequency torque perturbation signal. The average number of cycles is determined by the larger of a preset minimum number of cycles and a ratio k / SNR, where k is a preset constant and SNR is the signal-to-noise ratio.
[0038] The shaft system micro-deformation dynamic characteristic parameters are calculated based on the amplified perturbation characteristic amplitude. These parameters are the equivalent damping coefficients under still water conditions. The shaft system micro-deformation dynamic characteristic parameters are used as alternative physical constraint parameters to correct the initial fault classification probability read from the time-series buffer, thereby obtaining the target health assessment result.
[0039] Optionally, before inputting the multimodal operating state data (excluding the transient torque signal) into a preset fault classification model for inference, the method further includes:
[0040] The multimodal operating status data also includes vibration signals;
[0041] An adaptive Kalman filter is constructed based on a preset shaft inertia matrix and stiffness matrix. The state variables of the adaptive Kalman filter include shaft torsional angular displacement and angular velocity. The gain matrix of the adaptive Kalman filter is dynamically adjusted according to the real-time rotational speed of the host machine.
[0042] The vibration signal is decoupled using the adaptive Kalman filter to separate the coupling interference between the main machine torsional vibration component and the auxiliary machine torsional vibration component, and the decoupled vibration signal is then input into the fault classification model.
[0043] Optionally, after obtaining the target health assessment result, the method further includes:
[0044] When the probability of a preset fault category in the target fault classification probability exceeds the corresponding safety threshold, a load limiting instruction corresponding to the fault category is generated and sent to the ship's main engine control system to dynamically adjust the maximum allowable load of the main engine.
[0045] The ship's energy efficiency operation index parameters are updated based on the target fault classification probability. The energy efficiency operation index parameters include the energy efficiency operation index and the carbon intensity index.
[0046] An intelligent ship health assessment system includes:
[0047] The data acquisition module is used to acquire multimodal operating status data of the ship propulsion system, including transient torque signals.
[0048] The AI inference and caching module is used to input the multimodal operating state data other than the transient torque signal into a preset fault classification model for inference, obtain the initial fault classification probability, and store the initial fault classification probability in the time-series cache area.
[0049] The physical feature calculation module is used to determine the transient excitation event based on the fluctuation characteristics of the transient torque signal, extract the torque attenuation envelope corresponding to the transient excitation event, and obtain physical damping feature parameters based on the torque attenuation envelope.
[0050] The fusion adjudication module is used to read the corresponding initial fault classification probability from the time-series buffer at the time of generation of the physical damping characteristic parameters, and to correct the initial fault classification probability using the physical damping characteristic parameters to obtain the target health assessment result.
[0051] The beneficial effects of this invention are as follows: By constructing a temporal buffer, the initial fault classification probabilities obtained from rapid artificial intelligence inference are cached in the time dimension. At the moment the physical damping feature parameters are calculated, the initial probabilities, aligned with the timestamps, are read from the buffer for correction. This scheme effectively eliminates the temporal misalignment between millisecond-level artificial intelligence inference and second-level physical calculation, achieving timely and accurate constraints of physical damping features on the artificial intelligence diagnostic results. This not only improves the physical rationality and accuracy of the ship propulsion system health assessment results but also avoids the problem of separation between physical diagnosis and artificial intelligence diagnosis, enhancing the reliability of the system's assessment under complex operating conditions. Attached Figure Description
[0052] Figure 1 This is a flowchart of the method of the present invention.
[0053] Figure 2 This is a flowchart of the physical feature calculation and fusion correction process in this invention.
[0054] Figure 3 This is a flowchart of the static water condition assessment process in this invention.
[0055] Figure 4 This is the system architecture diagram of the present invention. Detailed Implementation
[0056] To facilitate a clear understanding of the technical means, creative features, objectives, and effects of this invention, the invention will be further described below in conjunction with specific illustrations.
[0057] To address several technical problems in existing technologies, such as the unclear division of labor between physical diagnosis and artificial intelligence diagnosis leading to their being considered interchangeable, the lack of a triggering mechanism for transient physical feature extraction during normal, stable navigation resulting in its long-term dormancy, and the temporal misalignment between rapid AI inference and slow physical calculation preventing the achievement of truly timely physical constraints on AI, this application provides a ship health assessment method based on natural transient impact attenuation characteristics and bidirectional physical constraint AI. (Refer to...) Figure 1 This includes the following steps.
[0058] In this method, the edge computing gateway runs a dual-channel inference engine. The high-frequency artificial intelligence inference channel performs high-speed feature extraction and fault classification inference on multimodal sensor signals such as vibration, temperature, and pressure. The low-frequency physical calculation channel performs transient excitation detection, damping envelope extraction, and physical feature parameter fitting on torque meter signals. The two channels are aligned and fused in the time dimension through a time-series buffer.
[0059] Overall, this method forms a closed loop consisting of four stages: action triggering, physical feedback, multi-level judgment, and dynamic response. It constructs a clear collaborative architecture between artificial intelligence (AI) and physics: the AI layer is responsible for extracting multi-source nonlinear fault features and outputting refined classification probabilities of composite faults; the physics layer is responsible for extracting low-dimensional but physically clear macroscopic damping features and outputting independent physical arbitration results. The physical constraint penalty function applies physical constraints and incentives to the AI's softmax output, rather than replacing AI in diagnosis. The AI layer is responsible for the fine differentiation of multiple fault categories, while the physics layer is responsible for macroscopic verification based on physical laws; the two are complementary rather than mutually substitutive.
[0060] Step S1: Acquire multimodal operating status data of the ship's propulsion system. This multimodal operating status data includes transient torque signals, as well as vibration signals, temperature signals, and pressure signals. The transient torque signal is acquired in real time by a high-precision torque sensor installed on the intermediate shaft or stern shaft of the ship's propulsion system, with a sampling frequency of 1 kHz. The vibration signal is acquired by an acceleration sensor installed on the main engine cylinder block, bearing housing, and shaft system, with a sampling frequency of 10 kHz. The temperature signals, including the exhaust temperature of each cylinder of the main engine, the cooling water temperature, and the lubricating oil temperature, are acquired by thermocouples or resistance temperature detectors (RTDs). The pressure signals, including the compression pressure, combustion pressure, and scavenging pressure of each cylinder of the main engine, are acquired by pressure sensors. The above multimodal operating status data is synchronously acquired by the data acquisition unit at a unified time reference and transmitted to the edge computing gateway.
[0061] Step S2 involves inputting multimodal operating state data (excluding transient torque signals) into a pre-set fault classification model for inference to obtain an initial fault classification probability, which is then stored in a time-series buffer. Since the AI inference cycle is in the millisecond range while the calculation of physical damping characteristic parameters requires waiting for the complete torque decay process, there is an order-of-magnitude time difference between the two. Therefore, a time-series buffer is used to achieve alignment in the time dimension.
[0062] Before inputting multimodal operating state data into the fault classification model, a physical-level decoupling process for the vibration signal is included. An adaptive Kalman filter is constructed based on a preset shaft inertia and stiffness matrix. The state variables of this adaptive Kalman filter include shaft torsional angular displacement and angular velocity. The state equation and observation equation of this adaptive Kalman filter are constructed using the standard Kalman filter form. The state transition matrix is obtained by discretizing the shaft torsional dynamics equation, and the observation matrix is determined according to the mode shape corresponding to the sensor installation location. The gain matrix of this adaptive Kalman filter is dynamically adjusted according to the real-time rotational speed of the main engine. This adaptive Kalman filter is used to decouple the vibration signal, separating the coupling interference between the main engine torsional vibration component and the auxiliary engine torsional vibration component.
[0063] After completing the physical-level decoupling process described above, the decoupled multimodal operating state data is preprocessed. Preprocessing includes bandpass filtering of the vibration signal to remove high-frequency noise and DC components, and outlier removal and smoothing of the temperature and pressure signals. The preprocessed vibration, temperature, and pressure signals undergo feature extraction to form a multi-channel feature tensor. This feature tensor is then input into a deep learning model based on an attention mechanism. In a feasible example, this deep learning model employs an encoder structure with alternating stacks of multi-head self-attention mechanisms and feedforward neural networks. The encoder has four stacked layers, each with eight self-attention heads, and the feedforward network has a hidden layer dimension of 2048. The dimension of the input tensor is the number of channels multiplied by the time step, where the number of channels corresponds to the number of sensor categories (vibration, temperature, pressure, etc.), and the time step corresponds to the number of sampling points within a preset historical time window. The output layer of this deep learning model uses a softmax function to output a probability vector. Each element of this probability vector corresponds to the probability of a preset fault category, which includes injector wear, abnormal valve clearance, localized main bearing wear, and turbocharger failure. This probability vector is the initial fault classification probability, denoted as […]. The deep learning model is optimized using a cross-entropy loss function during the training phase. The training data for this deep learning model comes from historical fault records of ship propulsion systems and / or simulated fault data generated based on shafting dynamics models.
[0064] The fault classification model runs continuously at a preset inference cycle, which ranges from 10 to 100 milliseconds. The initial fault classification probability generated during each inference, along with its timestamp, is stored in a time-series buffer. This buffer is a sliding queue with a preset buffer duration of 2 to 5 seconds, longer than the time window required to calculate the physical damping characteristic parameters. The time-series buffer is maintained on a first-in, first-out (FIFO) basis; when the buffer is full, the earliest stored initial fault classification probability is removed.
[0065] Step S3: Determine the transient excitation event based on the fluctuation characteristics of the transient torque signal, extract the torque attenuation envelope corresponding to the transient excitation event, and obtain the physical damping characteristic parameters based on the torque attenuation envelope. The transient excitation event can be caused by propeller water ejection due to wind and waves, or by drastic changes in the wake field. This method uniformly captures the event using the following quantitative judgment conditions.
[0066] refer to Figure 2 This step specifically includes the following sub-steps.
[0067] Obtain the natural frequency of the shafting torsional vibration of the ship's propulsion system. The frequency range of the first-order low-frequency torsional vibration of the shafting is 1 Hz to 10 Hz. Determine the length of the sliding time window based on this natural frequency. The natural frequency of the shafting torsional vibration is obtained through torsional vibration calculations using the structural parameters of the shafting, including its moment of inertia, stiffness, length, and diameter. The length of the sliding time window is set to 3 to 5 times the period corresponding to the natural frequency of the shafting torsional vibration.
[0068] The coefficient of variation of the transient torque signal is calculated within a sliding time window. When the coefficient of variation exceeds a preset threshold, a transient excitation event is determined to have occurred, and the current trigger time is recorded as the start time of the transient excitation event. The preset threshold is 5%. When the coefficient of variation exceeds 5%, it is considered a valid load transient impact.
[0069] Starting from the trigger moment, the transient torque signal after the transient excitation event is bandpass filtered to obtain a target frequency band torsional vibration signal containing the shaft system's torsional vibration natural frequency. The passband center frequency of the bandpass filter is set to the shaft system's torsional vibration natural frequency, and the passband bandwidth is set to ±10% of the shaft system's torsional vibration natural frequency. A Hilbert transform is then performed on the target frequency band torsional vibration signal to extract the torque attenuation envelope. This torque attenuation envelope physically follows an exponential decay law, and its form is... ,in The initial amplitude, The physical damping coefficient is obtained by performing an exponential fit on the extracted torque attenuation envelope using the least squares method. This physical damping coefficient serves as the physical damping characteristic parameter. The moment when the exponential fitting is complete is the moment when the physical damping characteristic parameter is generated.
[0070] After obtaining the physical damping characteristic parameters, the process also includes steps to compensate for environmental factors. This includes obtaining the cooling water temperature parameters of the ship's propulsion system. And obtain the preset reference temperature. Calculate the difference between the cooling water temperature parameter and the reference temperature. Calculate the square of this difference. The product of the difference and the first preset coefficient 'a', and the product of the squared value and the second preset coefficient 'b', are summed and then one is added. This sum is then compared with the third preset coefficient. Multiply by this to obtain the environmental correction factor:
[0071]
[0072] in, The environmental correction baseline value is the preset baseline environment, i.e., the third preset coefficient mentioned above; and These are the thermodynamic empirical correction factors preset for the system. The above factors... and The initial universal calibration values are obtained in advance based on the thermal expansion coefficient of the shaft material and the viscosity-temperature characteristics of the lubricating medium, among which... The typical range of values is to per degree Celsius, The typical range of values is to Per square degree Celsius. Using this environmental correction factor. For physical damping characteristic parameters Compensation and correction are performed to obtain the corrected physical damping characteristic parameters. Unless otherwise specified, the physical damping characteristic parameters described below refer to those adjusted for environmental compensation. The above coefficients and After the system's initial cross-sea deployment, it will undergo 30 to 50 hours of operation in a real-world environment, during which the system will make online corrections using a recursive least squares algorithm. and The value of is determined to achieve self-calibration and adaptation to different marine environmental conditions.
[0073] Step S4: At the moment the physical damping characteristic parameter is generated, the corresponding initial fault classification probability is read from the time-series buffer, and the initial fault classification probability is corrected using the physical damping characteristic parameter to obtain the target health assessment result.
[0074] This step specifically includes the following sub-steps.
[0075] Obtain the deviation between the physical damping characteristic parameters and the preset physical reference damping parameters. ,in This serves as the physical damping reference value for the ship's propulsion system under healthy conditions. During the initial commissioning of the system, either during the sea trial phase after new ship construction or a major overhaul of the propulsion system, the physical damping characteristic parameter is measured by continuously triggering transient excitation events after 30 minutes of stable operation under rated conditions. If calm sea conditions during the sea trial prevent the triggering of sufficient transient excitation events, the theoretically calculated shafting damping value from the ship's design phase is used as a temporary reference value, which is then replaced after accumulating sufficient measured data during subsequent voyages. After triggering more than 10 transient excitation events consecutively, the arithmetic mean of the measured physical damping characteristic parameters is taken as the reference value. After long-term operation of the ship, if the long-term sliding mean of the physical damping characteristic parameter deviates significantly and this deviation persists even after environmental compensation correction, the system prompts maintenance personnel to confirm the shafting status and recalibrate the reference value after confirming a healthy state.
[0076] Deviation Input a preset mapping function to obtain the penalty weight corresponding to each fault category. In an implementable example, for the first fault... Class of faults, their penalty weights Calculate using the following formula:
[0077]
[0078] in, For the first The sensitivity coefficient corresponding to a type of fault. When When positive, it indicates increased shaft damping and intensified energy dissipation. In this case, the sensitivity coefficient for fault categories related to increased friction, such as localized wear of the main bearing, takes a positive value, giving that category a larger penalty weight. A negative value indicates decreased shaft damping and reduced energy dissipation. In this case, the sensitivity coefficient for fault categories related to increased clearance or seal failure, such as abnormal valve clearance, takes a negative value. The sensitivity coefficients for each type of fault are pre-calibrated based on the fault's physical mechanism. In a specific example, the sensitivity coefficient corresponding to localized wear of the main bearing... Values Sensitivity coefficient corresponding to abnormal valve clearance Values Sensitivity coefficient corresponding to injector wear Values Sensitivity coefficient corresponding to turbocharger failure Values A diagonal matrix is constructed with each penalty weight as its diagonal element, serving as the physical constraint penalty function matrix. .
[0079] When the physical damping characteristic parameters drift slowly relative to the baseline value, but the probabilities of all fault categories in the initial fault classification probability do not change significantly, an additional incentive term is applied to the physical constraint penalty function matrix. This forcibly increases the weight coefficient of the fault category associated with the physical drift trend, thereby increasing the probability of the corresponding fault category in the fused target fault classification probability. Specifically, when the deviation... When the absolute value of the error exceeds a preset drift threshold and the difference between the highest and second-highest probabilities in the initial fault classification probabilities is less than a preset discrimination threshold, it is determined to be a slow drift and an unresponsive state of artificial intelligence. In an implementable example, the drift threshold is preset to 0.05 and the discrimination threshold is preset to 0.1. In this case, the excitation increment is superimposed on the diagonal elements of the physical constraint penalty function matrix:
[0080]
[0081] in, This is a dimensionless excitation intensity coefficient, which is taken as 0.5 in a feasible example. For physical reference damping parameters, This is a sign function. The above normalization process makes the excitation increment dimensionless, avoiding the problem of inconsistent physical dimensions.
[0082] At the time when the physical damping characteristic parameters are generated, an initial fault classification probability is selected from the time-series buffer based on the timestamp of the generation time. The selection principle is to choose the initial fault classification probability whose timestamp is no later than the generation time and has the smallest time difference with the generation time.
[0083] A matrix operation is performed between the physical constraint penalty function matrix and the initial fault classification probability vector. This matrix operation involves element-wise multiplication of the physical constraint penalty function matrix and the initial fault classification probability vector, followed by division by the sum of the elements of the resulting vector to achieve linear normalization.
[0084] in, This represents element-wise multiplication. This represents summing the elements of a vector. This represents the corrected target fault classification probability vector. Through the above calculations, aberrant classification results that violate physical laws in the initial fault classification probability are suppressed, and the classification weights of fault categories that are positively correlated with the deviation trend of physical damping characteristic parameters relative to physical reference damping parameters are enhanced, resulting in the corrected target fault classification probability. .
[0085] During the initial calibration phase, the system calculates the baseline mean and baseline standard deviation based on the physical damping characteristic parameters measured from multiple consecutive transient excitation events. Following statistical process control methods, the upper and lower control limits are determined by adding or subtracting three times the baseline standard deviation from the baseline mean. The system maintains a long-term moving average of the physical damping characteristic parameters. When this moving average exceeds the control limits, a strong physical constraint mode is activated, and the weight magnitude of the physical constraint penalty function matrix increases. When the moving average is within the control limits, the weight magnitude of the physical constraint penalty function matrix remains neutral.
[0086] Under normal navigation conditions without transient excitation events, taking the most recent generation time of physical damping feature parameters as the starting point, if no new transient excitation event is triggered within the following 60 seconds, the system directly outputs the initial fault classification probability of the current artificial intelligence inference as the final result, and sets the physical constraint flag to "No" in the data structure of the output result to indicate that the result has not been corrected for physical constraints. If a transient excitation event is triggered within this 60-second waiting window, the artificial intelligence inference features in the timing buffer are realigned based on the trigger time, and the previously cached features continue to participate in subsequent fusion calculations.
[0087] In the time dimension of the above process, the system executes the following sequence: At the first moment, multimodal operational status data is input into the system; at the second moment, the fault classification model completes inference and caches the initial fault classification probability in the time-series buffer; at the third moment, physical feature calculation is completed and physical damping feature parameters are generated; at the fourth moment, the fusion adjudication module applies the physical constraint penalty function matrix to the cached initial fault classification probability and outputs the final diagnostic result. The entire system follows a closed-loop processing architecture of caching, adjudication, and output. The intermediate feature vectors generated by AI inference are stored in the time-series buffer; after physical calculation is completed, a physical constraint penalty function matrix is generated; the fusion adjudication module completes matrix operations and generates the final diagnostic result, which is then written to the output buffer, generating a timestamp-aligned event record for querying or pushing to the upper-layer operation and maintenance system.
[0088] After obtaining the target health assessment results, a dynamic response step is also included. When the probability of a preset fault category in the target fault classification probability exceeds the corresponding safety threshold, a load limit command corresponding to that fault category is generated and sent to the ship's main engine control system to dynamically adjust the maximum allowable load of the main engine.
[0089] In one embodiment, the safety threshold corresponding to partial wear of the main bearing is preset to 70%. When the probability of this fault type exceeds 70%, a limiting command is generated to limit the maximum allowable load of the main unit to 80% of the rated load. The safety threshold corresponding to abnormal valve clearance is preset to 75%. When the probability of this fault type exceeds 75%, a limiting command is generated to limit the maximum allowable load of the main unit to 85% of the rated load. The safety threshold corresponding to fuel injector wear is preset to 80%. When the probability of this fault type exceeds 80%, a limiting command is generated to limit the maximum allowable load of the main unit to 90% of the rated load.
[0090] The ship's energy efficiency operation index parameters are updated based on the target failure classification probability. These parameters include the Energy Efficiency Operational Indicator (EEOI) and the Carbon Intensity Indicator (CII). The EEOI is updated by multiplying the current voyage's fuel consumption by the carbon emission factor and then dividing by the product of cargo volume and voyage distance. The CII is updated by dividing the annual cumulative carbon emissions by the annual cumulative transport workload. The specific calculation formulas for the EEOI and CII adopt the standard formulas specified in the current International Maritime Organization (IMO) regulations.
[0091] The method in this application also includes a static water condition assessment step. (Reference) Figure 3 The above step S3 relies on the triggering of natural transient excitation events. When no transient excitation events occur for a long time, this method provides the following alternative evaluation path.
[0092] When the coefficient of variation of the transient torque signal is continuously lower than the preset still water detection threshold within a preset time period, it is determined that no transient excitation event has occurred, and the system switches to the still water disturbance detection mode. The still water detection threshold is preset to 2%, and the preset time period is set to 30 to 60 seconds. If the coefficient of variation is not continuously lower than the still water detection threshold within the preset time period, the system does not switch to the still water disturbance detection mode and returns to continue executing the transient excitation event detection: if a transient excitation event is triggered, steps S3 and S4 are executed sequentially to complete the extraction and fusion correction of physical damping feature parameters; if no transient excitation event is triggered within a 60-second waiting window starting from the time of the most recent physical damping feature parameter generation, the initial fault classification probability of the current artificial intelligence inference is directly output as the final result. The above-mentioned maximum waiting time of 60 seconds is independent of the triggering conditions of the still water disturbance detection mode. When the coefficient of variation remains below the still water threshold for a preset time period, the system is switched to still water perturbation detection mode. The shaft micro-deformation dynamic characteristic parameters obtained in this mode are considered as valid physical constraint parameters and the rule of direct output after a 60-second timeout is no longer applicable.
[0093] In still water perturbation detection mode, the torque perturbation signal generated by the propeller rotation at the blade passage frequency is acquired. The blade passage frequency is the product of the number of propeller blades and the shaft rotation frequency.
[0094] A multi-period synchronous averaging method is employed to perform time-domain averaging on the blade passage frequency torque perturbation signal, thereby amplifying the characteristic amplitude of the perturbation. Using the period of the blade passage frequency as a reference, the torque signal is periodically segmented into multiple data segments, and then these segments are averaged at corresponding points. Through multi-period synchronous averaging, the amplitude of minute periodic torque perturbations can be amplified by more than 10 times. The number of average periods for the multi-period synchronous averaging method is determined based on the signal-to-noise ratio of the blade passage frequency torque perturbation signal. :
[0095]
[0096] in, To preset the minimum number of cycles, This is a dimensionless proportionality coefficient, ranging from 5 to 20, with the specific value determined based on the number of propeller blades and shaft speed. Signal-to-noise ratio. The average number of cycles is calculated by taking the preset minimum number of cycles and the ratio. The larger value in the range is determined.
[0097] The dynamic characteristic parameters of the shaft system's micro-variation are calculated based on the amplified perturbation characteristic amplitude. The calculation method is as follows: extract the perturbation characteristic amplitude within different periods after multi-period synchronous averaging, and then calculate it according to the exponential decay law. For the period number With amplitude Perform fitting, where This is the equivalent damping coefficient under still water conditions. The period of the blade passage frequency is obtained by fitting. The value represents the micro-variable dynamic characteristic parameter of the shafting system. This parameter reflects the damping dissipation level of the shafting system under still water conditions. The baseline value for the micro-variable dynamic characteristic parameter of the shafting system is... During the initial system commissioning, the arithmetic mean of at least five consecutive measurements were taken under still water conditions using a multi-cycle synchronous averaging method. If still water conditions were not available during initial deployment, the theoretical value of the shaft system design damping parameter, converted from the blade passing frequency period, was used as a temporary reference value. Using this slightly variable dynamic characteristic parameter of the shaft system as a substitute physical constraint parameter, the parameter and its reference value were calculated in the same manner as in step S4. The deviation is calculated and the corresponding physical constraint penalty function matrix is constructed. The initial fault classification probability read from the time-series buffer is then corrected to obtain the target health assessment result.
[0098] refer to Figure 4 The present invention also provides an intelligent ship health assessment system for implementing the above-mentioned ship health assessment method based on natural transient impact attenuation characteristics and bidirectional physical constraint artificial intelligence, including a data acquisition module, an artificial intelligence reasoning and caching module, a physical feature settlement module and a fusion adjudication module.
[0099] The data acquisition module is used to acquire multimodal operating status data of the ship's propulsion system. This multimodal operating status data includes transient torque signals, as well as vibration signals, temperature signals, and pressure signals. The data acquisition module is connected to high-precision torque sensors, acceleration sensors, thermocouples, and pressure sensors installed on the ship's propulsion system. It synchronously acquires data using a unified time reference through the data acquisition unit and transmits the acquired data to subsequent modules.
[0100] The AI inference and caching module is used to input multimodal operating state data (excluding transient torque signals) into a pre-set fault classification model for inference, obtain initial fault classification probabilities, and store these probabilities in a time-series buffer. This module also includes a decoupling processing submodule and a preprocessing submodule. The decoupling processing submodule constructs an adaptive Kalman filter based on a pre-set shaft inertia and stiffness matrix. The state variables of this adaptive Kalman filter include shaft torsional angular displacement and angular velocity. The state equation and observation equation adopt the standard Kalman filter form, and the gain matrix is dynamically adjusted according to the real-time rotational speed of the main engine. This adaptive Kalman filter is used to decouple the vibration signal, separating the coupling interference between the main engine torsional vibration component and the auxiliary engine torsional vibration component. The preprocessing submodule performs bandpass filtering on the decoupled vibration signal and outlier removal and smoothing on the temperature and pressure signals. The AI inference and caching module continuously runs the fault classification model with an inference cycle of 10 to 100 milliseconds, generating initial fault classification probabilities with timestamps, and storing them in a sliding buffer queue with a buffer duration of 2 to 5 seconds.
[0101] The physical feature calculation module is used to determine transient excitation events based on the fluctuation characteristics of transient torque signals, extract the torque decay envelope corresponding to the transient excitation event, and obtain physical damping characteristic parameters based on the torque decay envelope. This module includes a transient excitation detection submodule, an envelope extraction submodule, a damping fitting submodule, and an environmental compensation submodule. The transient excitation detection submodule acquires the shaft torsional vibration natural frequency of the ship's propulsion system, determines the length of the sliding time window based on this natural frequency, calculates the coefficient of variation of the transient torque signal within the sliding time window, and determines that a transient excitation event has occurred and records the trigger time when the coefficient of variation exceeds a preset 5% threshold. The envelope extraction submodule uses the trigger time as the starting time, performs bandpass filtering on the transient torque signal after the transient excitation event to obtain the target frequency band torsional vibration signal, and then performs a Hilbert transform on the target frequency band torsional vibration signal to extract the torque decay envelope. The damping fitting submodule performs exponential function fitting on the torque decay envelope to calculate the physical damping coefficient. The environmental compensation submodule acquires the cooling water temperature parameter, calculates its difference from the reference temperature and its square value, and calculates the physical damping coefficient according to the formula... Calculate the environmental correction factor and use it to compensate and correct the physical damping characteristic parameters.
[0102] The fusion adjudication module is used to read the corresponding initial fault classification probability from the time-series buffer at the time of generation of physical damping feature parameters, and correct the initial fault classification probability using the physical damping feature parameters to obtain the target health assessment result. This module includes a deviation calculation submodule, a penalty function construction submodule, a time-series matching submodule, and a fusion operation submodule. The deviation calculation submodule calculates the deviation between the physical damping feature parameters and the preset physical benchmark damping parameters. The penalty function construction submodule inputs this deviation into a preset mapping function to obtain the penalty weights corresponding to each fault category, and constructs a diagonal matrix as the physical constraint penalty function matrix using each penalty weight as its diagonal element. The time-series matching submodule selects the initial fault classification probability with a timestamp no later than the generation time and the smallest time difference from the generation time from the time-series buffer at the time of generation of physical damping feature parameters. The fusion operation submodule performs element-wise multiplication of the physical constraint penalty function matrix with the initial fault classification probability vector, and then divides by the sum of the elements of the resulting vector to achieve linear normalization, obtaining the corrected target fault classification probability.
[0103] The intelligent ship health assessment system also includes a still water condition assessment module. This module determines that no transient excitation event has occurred and switches to still water disturbance detection mode when the coefficient of variation of the transient torque signal remains below a preset still water threshold for a preset duration. This module acquires the blade passage frequency torque disturbance signal generated by the propeller rotation, performs time-domain averaging on the signal using a multi-cycle synchronous averaging method to amplify the disturbance characteristic amplitude, and determines the number of averaging cycles based on the signal-to-noise ratio. Based on the amplified perturbation characteristic amplitude, the shaft system micro-deformation dynamic characteristic parameters are calculated using the exponential decay fitting method. These shaft system micro-deformation dynamic characteristic parameters and their benchmark values are then used as alternative physical constraint parameters to correct the initial fault classification probability.
[0104] The intelligent ship health assessment system also includes a dynamic response module. When the probability of a preset fault category in the target fault classification probability exceeds the corresponding safety threshold, this module generates a load limit command corresponding to that fault category and sends it to the ship's main engine control system to dynamically adjust the maximum allowable load of the main engine. It also updates the ship's energy efficiency operation index parameters according to the standard formula specified in the current International Maritime Organization (IMO) regulations based on the target fault classification probability. These energy efficiency operation index parameters include the energy efficiency operation index and the carbon intensity index.
[0105] In terms of engineering implementation, the deployment of transient torque sensors and hard clock synchronization in the above scheme requires dry-docking construction during the conversion of old ships, resulting in high hardware costs per ship; at the same time, the coefficients in the thermodynamic empirical correction formula... and After the initial cross-sea deployment, the system requires approximately 30 to 50 hours of operation in a real-world environment, during which it completes self-calibration and adaptation using a recursive least squares algorithm. The aforementioned limitations are engineering boundaries that can be recognized by those skilled in the art during implementation and do not affect the normal implementation of this technical solution on newly built vessels or existing vessels with suitable retrofitting capabilities.
[0106] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments, and various changes and modifications can be made without departing from the spirit and scope of the invention, all of which fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for assessing the health of intelligent ships, characterized in that, Includes the following steps: Acquire multimodal operating status data of the ship propulsion system, wherein the multimodal operating status data includes transient torque signals; The multimodal operating state data, excluding the transient torque signal, are input into a preset fault classification model for inference to obtain an initial fault classification probability, and the initial fault classification probability is stored in a time-series buffer. The transient excitation event is determined based on the fluctuation characteristics of the transient torque signal, the torque decay envelope corresponding to the transient excitation event is extracted, and the physical damping characteristic parameters are obtained by fitting the torque decay envelope. At the time when the physical damping characteristic parameters are generated, the corresponding initial fault classification probability is read from the time-series buffer, and the initial fault classification probability is corrected using the physical damping characteristic parameters to obtain the target health assessment result.
2. The method according to claim 1, characterized in that, The step of determining the transient excitation event based on the fluctuation characteristics of the transient torque signal includes: The natural frequency of shaft torsional vibration of the ship propulsion system is obtained, and the length of the sliding time window is determined based on the natural frequency of shaft torsional vibration. The coefficient of variation of the transient torque signal is calculated within the sliding time window, where the coefficient of variation is the ratio of the standard deviation to the mean of the transient torque signal within the sliding time window. When the coefficient of variation exceeds a preset threshold, the transient excitation event is determined to have occurred, and the current triggering time is recorded. The triggering time is used as the start time of the transient excitation event.
3. The method according to claim 2, characterized in that, The step of extracting the torque decay envelope corresponding to the transient excitation event and fitting the physical damping characteristic parameters based on the torque decay envelope includes: Starting from the triggering time, the transient torque signal after the transient excitation event is subjected to bandpass filtering to obtain a target frequency band torsional vibration signal that includes the natural frequency of the shaft system torsional vibration. Perform Hilbert transform on the torsional vibration signal in the target frequency band to extract the torque attenuation envelope; An exponential function is fitted to the torque attenuation envelope to calculate the physical damping coefficient, which characterizes the energy dissipation level of the shaft system, and serves as the physical damping characteristic parameter.
4. The method according to claim 3, characterized in that, After obtaining the physical damping characteristic parameters and before correcting the initial fault classification probability using the physical damping characteristic parameters, the method further includes: Obtain the cooling water temperature parameters of the ship's propulsion system and obtain a preset reference temperature; Calculate the difference between the cooling water temperature parameter and the reference temperature; Calculate the square of the difference, and sum the product of the difference and a first preset coefficient, the product of the square and a second preset coefficient, and a preset environmental correction benchmark value to obtain the environmental correction coefficient. The physical damping characteristic parameters are compensated and corrected using the environmental correction coefficient to eliminate damping reference drift caused by cross-ocean environmental temperature differences.
5. The method according to claim 1, characterized in that, The step of correcting the initial fault classification probability using the physical damping characteristic parameters to obtain the target health assessment result includes: Obtain the deviation between the physical damping characteristic parameters and the preset physical reference damping parameters; The deviation is input into a preset mapping function to obtain the penalty weights corresponding to each fault category. A diagonal matrix is constructed with each penalty weight as the diagonal element, which serves as the physical constraint penalty function matrix. The physical constraint penalty function matrix and the initial fault classification probability are subjected to matrix operation. The matrix operation is to multiply the physical constraint penalty function matrix and the initial fault classification probability vector element by element, and then divide by the sum of the elements of the vector obtained by the element-wise multiplication to achieve linear normalization. This is to suppress abnormal classification results that violate physical laws in the initial fault classification probability, and to enhance the classification weight of fault categories that are positively correlated with the deviation trend of the physical damping feature parameter relative to the physical reference damping parameter, so as to obtain the corrected target fault classification probability. The target health assessment result includes the target failure classification probability.
6. The method according to claim 1, characterized in that, The step of storing the initial fault classification probability into a time-series buffer includes: The fault classification model is continuously run at a preset inference period and an initial fault classification probability with a timestamp is generated. The inference period is from 10 milliseconds to 100 milliseconds. Construct a time-series buffer with a preset buffer duration. The time-series buffer is a sliding buffer queue. The buffer duration is 2 to 5 seconds. The buffer duration is greater than or equal to the time window length required to calculate the physical damping characteristic parameters. When generating the physical damping characteristic parameters, an initial fault classification probability is selected from the time-series buffer based on the timestamp of the generation time, with a timestamp no later than the generation time and the smallest time difference between the timestamp and the generation time.
7. The method according to claim 1, characterized in that, It also includes the still water condition assessment steps: When the coefficient of variation of the transient torque signal is detected to be lower than the preset still water determination threshold for a preset duration, it is determined that the transient excitation event has not occurred, and the system switches to still water disturbance detection mode. The perturbation signal of the blade passing frequency torque generated by the propeller rotation is obtained, wherein the blade passing frequency is the product of the number of propeller blades and the shaft rotation frequency; The time-domain averaging of the blade passage frequency torque perturbation signal is performed using a multi-cycle synchronous averaging method to amplify the perturbation characteristic amplitude. The average number of cycles of the multi-cycle synchronous averaging method is determined based on the signal-to-noise ratio of the blade passage frequency torque perturbation signal. The average number of cycles is determined by the larger of a preset minimum number of cycles and a ratio k / SNR, where k is a preset constant and SNR is the signal-to-noise ratio. The shaft system micro-deformation dynamic characteristic parameters are calculated based on the amplified perturbation characteristic amplitude. These parameters are the equivalent damping coefficients under still water conditions. The shaft system micro-deformation dynamic characteristic parameters are used as alternative physical constraint parameters to correct the initial fault classification probability read from the time-series buffer, thereby obtaining the target health assessment result.
8. The method according to claim 1, characterized in that, Before inputting the multimodal operating state data (excluding the transient torque signal) into a preset fault classification model for inference, the method further includes: The multimodal operating status data also includes vibration signals; An adaptive Kalman filter is constructed based on a preset shaft inertia matrix and stiffness matrix. The state variables of the adaptive Kalman filter include shaft torsional angular displacement and angular velocity. The gain matrix of the adaptive Kalman filter is dynamically adjusted according to the real-time rotational speed of the host machine. The vibration signal is decoupled using the adaptive Kalman filter to separate the coupling interference between the main machine torsional vibration component and the auxiliary machine torsional vibration component, and the decoupled vibration signal is then input into the fault classification model.
9. The method according to claim 5, characterized in that, After obtaining the target health assessment results, the following is also included: When the probability of a preset fault category in the target fault classification probability exceeds the corresponding safety threshold, a load limiting instruction corresponding to the fault category is generated and sent to the ship's main engine control system to dynamically adjust the maximum allowable load of the main engine. The ship's energy efficiency operation index parameters are updated based on the target fault classification probability. The energy efficiency operation index parameters include the energy efficiency operation index and the carbon intensity index.
10. An intelligent ship health assessment system, characterized in that, include: The data acquisition module is used to acquire multimodal operating status data of the ship propulsion system, including transient torque signals. The AI inference and caching module is used to input the multimodal operating state data other than the transient torque signal into a preset fault classification model for inference, obtain the initial fault classification probability, and store the initial fault classification probability in the time-series cache area. The physical feature calculation module is used to determine the transient excitation event based on the fluctuation characteristics of the transient torque signal, extract the torque attenuation envelope corresponding to the transient excitation event, and obtain physical damping feature parameters based on the torque attenuation envelope. The fusion adjudication module is used to read the corresponding initial fault classification probability from the time-series buffer at the time of generation of the physical damping characteristic parameters, and to correct the initial fault classification probability using the physical damping characteristic parameters to obtain the target health assessment result.
