A marine engine manufacturing whole-process decision prediction method combined with machine learning

Abstract

This invention discloses a decision-making and prediction method for the entire manufacturing process of marine engines, incorporating machine learning, and relates to the field of marine engine manufacturing and testing technology. It addresses problems such as model prediction bias, confidence imbalance, and response lag caused by differences in the statistical distribution of key operating condition characteristics between bench testing and sea trials. The method selects adversarial domain adaptation or confidence calibration algorithms for model training, constructs a prediction model based on a deep neural network, and improves the model's generalization ability and confidence reliability through adversarial optimization or post-processing calibration of the feature extraction network, the main task prediction sub-network, and the domain discrimination sub-network. The model is deployed and validated in real time in the sea trial environment, and closed-loop decision control is executed based on the model prediction results by calculating the error index between the predicted and measured values. Through data layer alignment, model layer adaptation, validation layer verification, and execution layer control, the prediction accuracy and operational safety of marine engines during the sea trial phase are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of marine engine manufacturing and testing technology, and more specifically, to a decision prediction method for the entire marine engine manufacturing process that incorporates machine learning. Background Technology

[0002] In the manufacturing and commissioning of marine engines, performance verification is typically performed on a land-based test bench before sea trials. Test bench conditions are controlled, with stable ambient temperature, humidity, and fuel quality. The sensor layout is complete, and data acquisition frequency is high, providing accurate samples for performance model establishment. However, the sea trial environment is complex and variable, affected by seawater temperature, air humidity, wave loads, and fuel fluctuations, leading to frequent changes in engine operating conditions. Some measurement points cannot be placed, or there is significant signal noise. Current technologies often use test bench data to train models, then transfer them to sea trials using empirical coefficient correction or static bias methods, but this ignores the distribution offset between the two domains.

[0003] The statistical characteristics of input features vary greatly under different environments, leading to problems such as prediction bias, confidence imbalance, and response lag in the model during sea trials. Especially during sea state fluctuations or equipment aging, the model's power and fuel consumption prediction errors often exceed engineering tolerances, failing to reliably guide energy consumption optimization or load control. Current technologies also lack quantification and gating mechanisms for model uncertainty; model updates rely on manual retraining, hindering adaptive correction and continuous optimization, resulting in insufficient prediction reliability and high maintenance costs.

[0004] To address the above problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a decision prediction method for the entire manufacturing process of marine engines that incorporates machine learning, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A decision-making and prediction method for the entire manufacturing process of marine engines that incorporates machine learning, including the following steps;

[0008] Step S1: Collect the operating conditions, process parameters and performance indicators involved in bench tests and sea trials, construct a data map with operating condition buckets as the basic unit, and select samples to form an anchoring calibration set based on the two criteria of operating condition coverage and time stability. According to the criterion results, identify the trainable and non-trainable areas.

[0009] Step S2: Select a training path. When there is a significant deviation between the working conditions of the test bench and the sea trial, use an adversarial domain adaptation algorithm to map and converge the representation space. When the deviation is within the allowable range, use a confidence calibration method to process the model output and adjust the model structure and weights to generate a fixed representation.

[0010] Step S3: During the sea trial operation, load the trained model, input real-time collected multi-source data, perform continuous prediction, and establish a dual release mechanism by setting the anchor bucket accuracy and residual fluctuation threshold, as well as the consistency and uncertainty limits of the neighborhood bucket.

[0011] In step S4, the model is deployed using a two-stage verification method with shadow channel and canary ratio verification. The predicted and measured residual trajectories are monitored online, and fine-tuning is performed within the allowable range of bias and confidence parameters. All running records are then synchronously written back to the spectral system.

[0012] In a preferred embodiment, step S1 includes the following:

[0013] Collect raw data during the test bench phase of marine engines, including key parameters such as engine speed, torque, fuel consumption rate, exhaust temperature, and vibration acceleration.

[0014] The raw data is preprocessed, including outlier removal and noise filtering, and the time series data is divided into multiple training samples.

[0015] Extract the feature parameters of each sample, including mean, peak value, fluctuation range, and spectral energy, and associate them with the expected output value or state label;

[0016] Data is filtered based on the working condition coverage criterion and the domain difference criterion to obtain the flag parameters, which are used to guide the selection of the model training algorithm path in step S2;

[0017] The working condition coverage criterion determines whether the data volume corresponding to each typical working condition is not lower than the preset minimum sample proportion threshold by statistically analyzing the data volume and distribution of each working condition period.

[0018] Domain difference criteria quantify the degree of domain shift by calculating the distribution differences of key features between the source domain and the target domain, including using statistical comparison methods or kernel methods to calculate the maximum mean difference index (MMD).

[0019] The filtered data are normalized or standardized, and data augmentation is performed, including superimposing simulated noise and random perturbation parameter values ​​on the original time series, or generating virtual samples based on physical models.

[0020] In a preferred embodiment, step S2 includes the following:

[0021] Based on the flag parameters, select either the adversarial domain adaptation algorithm or the confidence calibration algorithm;

[0022] When the flag parameter indicates that the domain difference index between the source domain data and the target domain data is greater than the preset domain difference threshold, the adversarial domain adaptation algorithm is adopted.

[0023] When the flag parameter indicates that the domain difference index is less than or equal to the preset domain difference threshold, a confidence calibration algorithm is used.

[0024] The training process of the adversarial domain adaptation algorithm includes:

[0025] Iterative training is performed using source domain samples and target domain samples;

[0026] Calculate the main task loss for the source domain samples and update the parameters of the main task prediction sub-network and the feature extraction network;

[0027] The domain classification loss is calculated for samples from the source and target domains. After gradient inversion, the parameters of the feature extraction network are updated to increase the domain confusion.

[0028] In a preferred embodiment, step S3 includes the following:

[0029] The obtained model is deployed in the sea trial environment to acquire engine sensor data in real time and perform predictive calculations.

[0030] Record the measured values ​​and calculate the error indices between the predicted and measured values, including the relative error percentage and the mean square error (MSE).

[0031] Based on the error index and preset threshold, determine whether the model prediction is inaccurate; if it is, trigger the model adjustment process.

[0032] Model tuning includes retraining, fine-tuning model parameters, or adjusting calibration parameters;

[0033] In the model adjustment process, for adversarial domain adaptation models, if the prediction error is mainly concentrated in the operating range corresponding to the operating condition bucket that is not marked as covered by the operating condition coverage criterion, then new operating condition data will be included to retrain or fine-tune the model; for confidence calibration models, if the confidence level is inaccurate, then the calibration parameters will be recalculated.

[0034] In a preferred embodiment, step S4 includes the following:

[0035] Based on the model prediction results, control strategies are formulated, such as adjusting the fuel injection quantity, cooling flow rate or boost pressure, to avoid malfunctions or insufficient performance.

[0036] The quality assessment is conducted using model prediction results. If the prediction matches the actual measurement and meets the requirements, the engine passes the test; otherwise, the design is improved or the engine is remanufactured.

[0037] The sea trial data is fed back to the dataset in step S1 to update the training sample library and criterion thresholds for subsequent engine testing and model optimization.

[0038] The technical effects and advantages of the present invention, which combines machine learning with a decision-making and prediction method for the entire manufacturing process of marine engines, are as follows:

[0039] First, by establishing a test bench-sea trial condition map and anchoring calibration set at the data layer, the alignment and distribution mapping of multi-source data were achieved, solving the problem of inconsistent input features in traditional models. Second, cross-domain stable representation learning and physical consistency constraints were introduced at the model layer, enabling the model to maintain the physical interpretability of the input-output relationship under environmental changes, fundamentally reducing the error accumulation caused by distribution shifts. Third, a confidence assessment and dual-criteria verification mechanism were introduced to perform bidirectional verification of accuracy and stability under overlapping test bench and sea trial conditions, and uncertainty gating was established under extrapolation conditions, thereby compensating for the lack of anomaly detection and risk control in traditional models. Finally, an online adaptation and closed-loop write-back mechanism was formed at the execution layer, achieving small-step stable updates of the model through shadow channels and canary ratio verification, avoiding the lag and untraceability problems of manual retraining. Through these four progressive steps, this invention achieves a transformation from static experience correction to dynamic domain alignment, self-calibration, and reliable decision-making, significantly improving the prediction accuracy and operational safety of the engine during the sea trial phase. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a decision prediction method for the entire manufacturing process of marine engines that combines machine learning, according to the present invention.

[0041] Figure 2 This is a schematic diagram of the structure of a decision prediction method for the entire manufacturing process of marine engines that combines machine learning, according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example

[0044] Please see Figures 1-2 As shown, this invention discloses a decision-making and prediction method for the entire manufacturing process of marine engines that combines machine learning, including the following steps:

[0045] Step S1: Collect the operating conditions, process parameters and performance indicators involved in bench tests and sea trials, construct a data map with operating condition buckets as the basic unit, and select samples to form an anchoring calibration set based on the two criteria of operating condition coverage and time stability. According to the criterion results, identify the trainable and non-trainable areas.

[0046] Step S2: Select a training path. When there is a significant deviation between the working conditions of the test bench and the sea trial, use an adversarial domain adaptation algorithm to map and converge the representation space. When the deviation is within the allowable range, use a confidence calibration method to process the model output and adjust the model structure and weights to generate a fixed representation.

[0047] Step S3: During the sea trial operation, load the trained model, input real-time collected multi-source data, perform continuous prediction, and establish a dual release mechanism by setting the anchor bucket accuracy and residual fluctuation threshold, as well as the consistency and uncertainty limits of the neighborhood bucket.

[0048] In step S4, the model is deployed using a two-stage verification method with shadow channel and canary ratio verification. The predicted and measured residual trajectories are monitored online, and fine-tuning is performed within the allowable range of bias and confidence parameters. All running records are then synchronously written back to the spectral system.

[0049] In step S1, the operating conditions, process parameters, and performance indicators involved in bench tests and sea trials are collected to construct a data map with operating condition buckets as the basic unit. Based on the two criteria of operating condition coverage and time stability, samples are selected to form an anchoring calibration set. Based on the criterion results, trainable and non-trainable regions are identified. Specific content includes:

[0050] Collect raw data from various sensors operating under different conditions during the test bench phase of marine engines, including key parameters such as engine speed, torque, fuel consumption rate, exhaust temperature, and vibration acceleration.

[0051] Data acquisition covers all steady-state and dynamic operating conditions of the engine from low load to full load and different speed steps, ensuring that the source domain data has comprehensive operating condition representativeness;

[0052] It should be noted that since bench tests are usually conducted on fixed land test benches with relatively stable environmental conditions, which differ from the complexity of the actual sea operating environment, some measures to simulate sea conditions can be consciously introduced during the data acquisition process. For example, random fluctuations can be superimposed in the bench loading control to simulate changes in ship propulsion load, or the changes in ambient temperature and humidity can be artificially controlled to approximate the marine environment, thereby improving the representativeness of the collected data to the real sea trial conditions.

[0053] The collected raw data is stored in a preset format, such as recording the readings of each sensor in time series form, and undergoes preliminary preprocessing: removing sensor outliers that show distortion and performing necessary noise filtering. Outlier identification can be achieved using statistical methods, such as the 3σ criterion or cross-validation based on sensor redundancy signals, to ensure the reliability of the input data.

[0054] Continuous time series data is divided into multiple training samples, for example, by operating conditions or fixed time windows. Each sample contains all sensor feature sequences or extracted statistical features within that time period.

[0055] To facilitate model learning, a set of feature parameters can be further extracted for each sample, such as the mean, peak value, fluctuation amplitude, and spectral energy of each sensor signal. Alternatively, the time series signal can be directly used as the model input in a deep learning framework. Each sample is also associated with a desired output value or state label, which comes from the corresponding performance indicators or test records measured in bench tests. For example, the actual power output and fuel consumption of the engine in the corresponding time period of the sample, or the working condition health status determined by the test personnel, including normal and abnormal.

[0056] Data that does not meet the requirements is filtered out based on the working condition coverage criterion: This criterion ensures that each typical working condition has a sufficient sample proportion by statistically analyzing the data volume and distribution of each working condition segment. It determines whether the data volume corresponding to each typical working condition is not lower than the preset minimum sample proportion threshold, so as to determine whether the working condition is considered to have effective training coverage. The minimum sample proportion threshold is preset based on historical experimental statistics or model training stability requirements. It is used to determine whether a certain working condition is representative enough to participate in model training. For example, the data volume of each type of working condition is not less than 10% of the total data. If it is found that the data corresponding to a certain working condition is seriously insufficient, it is necessary to supplement the corresponding experiments to obtain more samples, or reduce the dependence weight of this working condition in the analysis to avoid skewed training data distribution.

[0057] After processing using this criterion, sample segments that deviate too much from the main operating conditions or have too little data can be removed. Then, the degree of difference between the bench data and the sea trial environment can be evaluated based on the domain difference criterion: this criterion quantifies the degree of domain offset by calculating the distribution difference of key features between the source domain and the target domain. For example, statistical comparison methods can be used to calculate the relative deviation Δ between the bench data and historical sea trial data in terms of statistical characteristics such as parameter mean and standard deviation; kernel methods can also be introduced to calculate the maximum mean difference index MMD to measure the distance between the data distributions of the two domains.

[0058] In addition, a simple domain discrimination model, such as a binary classifier, can be trained to distinguish between bench data and target domain data. If the classifier's discrimination accuracy on the test data is much higher than the random level of 50%, it indicates that there is a significant difference between the two distributions. To compare the above-mentioned differences in domain distribution, it is necessary to select representative target domain data as a reference.

[0059] The target domain data can be derived from the sea trial test data of similar engine models in the past. If real data is lacking before the first sea trial of a brand-new model, simulated sea operation data can be generated by engine mechanism simulation model to replace it.

[0060] The calculation of the domain difference criterion can integrate the differences of multiple parameters and assign different weights to different parameters to reflect their sensitivity to engine performance, thereby obtaining a weighted comprehensive difference index. In practical implementation, the domain difference threshold... It can be selected based on experience or design requirements. For example, based on the statistical analysis of historical tests of multiple engines, a 5% distribution deviation of a certain key parameter can be determined as the upper limit of the acceptable performance range.

[0061] When the calculated values ​​of indicators such as Δ or MMD are lower than the threshold If the source domain and target domain are similarly distributed, it is considered that the source domain and target domain are similar; otherwise, it is determined that there is a significant domain shift, and special adaptation measures are required.

[0062] For example, after calculation and analysis of the distribution difference of a certain key parameter: the mean value of this parameter in the bench test data is about 300℃, while the mean value of the reference sea trial data is about 320℃, with a deviation of about 6.7%. If the pre-set threshold is 5%, this deviation exceeds the allowable range, and it is determined that there is a significant domain shift, which requires special measures such as domain adaptation in subsequent model training.

[0063] Through the above dual criteria screening, abnormal or insufficient training samples are eliminated on the one hand, and quantitative assessment of the representativeness and domain difference of the source domain data is given on the other hand. The judgment results of the dual criteria will be recorded as flag parameters for subsequent steps, where 0 indicates that the working condition coverage is sufficient and the domain difference is acceptable, and 1 indicates that the working condition coverage is insufficient and / or the domain difference is significant.

[0064] This flag parameter will be used to guide the selection of the model training algorithm path in the next step, that is, to determine whether it is necessary to use an adversarial domain adaptation algorithm or a confidence calibration algorithm to improve the generalization performance of the model in order to solve the domain shift problem from bench testing to sea trial environment.

[0065] The features of the selected dataset are normalized or standardized as needed to eliminate the influence of different parameter units on training, thereby ensuring that the training sample set is of good quality and representative, and preparing it for model training.

[0066] To improve the robustness of the model, the selected training data can be augmented, for example by superimposing simulated noise on the original time series, randomly perturbing certain parameter values, or generating additional virtual samples based on the physical model, so as to enrich the diversity of the dataset and allow the model to see more changes in similar sea trial environments during training, thereby improving its adaptability to noise and operating condition changes.

[0067] It should be noted that the minimum working condition unit is obtained by discretizing the working condition bucket using a three-dimensional grid of speed × load × ambient temperature and humidity.

[0068] Anchor buckets are those that cover both the source and target domains and have a sample size ≥ The working condition bucket;

[0069] The neighborhood bucket is the set of buckets that are adjacent to the current bucket in any dimension;

[0070] The accuracy of the anchor bucket is the moving average of the relative error between the prediction and the actual measurement within the anchor bucket.

[0071] Neighborhood consistency is the maximum difference between the predicted mean of the current bucket and its neighboring buckets.

[0072] In step S2, a training path is selected. When there is a significant deviation between the operating conditions of the test bench and the sea trial, an adversarial domain adaptation algorithm is used to map and converge the representation space. When the deviation is within the allowable range, a confidence calibration method is used to process the model output, and the model structure and weights are adjusted to generate a fixed representation. Specific details include:

[0073] The operating condition offset between bench test and sea trial refers to the difference in statistical distribution between bench test data (source domain) and sea trial operation data (target domain) in at least one key operating condition feature dimension. The key operating condition feature is the operating condition element, process parameter or performance index collected in step S1 and used for model training, which includes one or more of engine speed, load state, temperature parameter or vibration feature.

[0074] The operating condition deviation is quantitatively evaluated by the domain difference criterion, which is based on the statistical distribution difference between the source domain data and the target domain data on the key operating condition features to calculate the difference index. This difference index can be obtained by statistical comparison method or distribution distance measurement method.

[0075] When the difference index is greater than the preset domain difference threshold, it is determined that there is a working condition offset between the source domain data and the target domain data that requires domain adaptation processing; when the difference index is not greater than the domain difference threshold, it is determined that the working condition difference between the two is within an acceptable range.

[0076] Based on the criterion results, perform adversarial domain adaptation or confidence calibration model training. After completing the data screening, establish a machine learning model for predicting the performance or status of marine engines, and select an appropriate training algorithm path based on the criterion results obtained in step S1.

[0077] Specifically, when the flag parameter indicates a significant domain difference, and the domain difference index between the source domain data and the target domain data is greater than the preset domain difference threshold, the adversarial domain adaptation algorithm is used for model training; conversely, when the source domain data is sufficiently representative of the target domain, that is, when the flag parameter indicates that the domain difference index is less than or equal to the preset domain difference threshold, the confidence calibration algorithm is used to train the model to enhance its generalization ability.

[0078] It should be noted that the operating condition offset between bench testing and sea trials refers to the difference in the statistical distribution of source domain data and target domain data in at least one key operating condition feature dimension.

[0079] Differences are quantitatively evaluated using domain difference criteria, which can be obtained by comparing the statistical differences, distribution distances, or domain separability indices of the corresponding working condition characteristics between the source domain and the target domain. These criteria include, but are not limited to, mean difference, variance difference, maximum mean difference indices, or discrimination results obtained by training a domain discrimination model.

[0080] When the difference index corresponding to the domain difference criterion exceeds the preset domain difference threshold, it is determined that there is a working condition deviation between the bench test data and the sea trial operation data that requires domain adaptation processing; wherein, the preset domain difference threshold is preset based on the statistical results of historical test data or the stability requirements of model training.

[0081] The initial structure of the prediction model is constructed. For example, to construct a regression or classification model based on a deep neural network, an appropriate network structure can be selected according to the feature type of the input data. For sensor data in the form of time series, a long short-term memory network or a one-dimensional convolutional neural network that can capture temporal correlation can be used as the feature extraction layer of the model. For the extracted static feature vectors, a multi-layer fully connected feedforward network can be used.

[0082] The input to the model is the multidimensional feature vector filtered in step S1, which contains parameters of various engine sensors. Several auxiliary parameters that characterize external environmental conditions, such as ambient temperature, humidity, and sea state, can also be added to the feature vector so that the model can learn the impact of these factors on engine performance.

[0083] Since these environmental parameters are relatively constant in bench tests, but will change significantly in sea trials, incorporating them into the model input helps improve the model's adaptability to the target domain. The output is the target prediction index, such as power output, fuel consumption, or failure probability in sea trials. For cases with significant domain differences, an adversarial domain adaptation model framework is introduced: the model includes a shared feature extraction network F, and two branch sub-networks, namely the main task prediction sub-network C and the domain discrimination sub-network D.

[0084] During training, the source domain, i.e., bench test data, carries supervision labels, such as actual power or health status, while the target domain, i.e., sea trial environment data, can select unlabeled samples, such as those based on previous similar engine sea trial data or simulation data.

[0085] By employing gradient reversal layers or equivalent adversarial training strategies, the feature extraction network F minimizes the source domain task loss. Simultaneously, it maximizes the obfuscation of the domain discriminant subnetwork D to reduce the difference in feature distribution between the source and target domains. In implementation, this adversarial optimization strategy can be easily achieved by introducing a gradient reversal layer between the feature extraction network and the domain discriminant network.

[0086] During training, a batch of source domain samples and target domain samples are simultaneously extracted in each iteration: the main task loss is calculated on the source domain samples. The parameters of C and F are updated to reduce this loss, and the source and target domain samples are input into the domain discriminant subnetwork D to calculate the domain classification loss. After gradient reversal, the parameters of F are updated to increase the domain confusion.

[0087] Domain discrimination subnetwork D first passes The probability that the output feature comes from the source domain; where, The probability value output by the domain discriminant subnetwork D represents the input feature. The probability of s from the source domain is 1. If the value is close to 1, it is determined to be from the source domain. If it is close to 0, it is determined to be from the target domain.

[0088] Input features The output feature vector of the feature extraction network F is given by the input features of the sample after the screening in step one, such as the extracted features of sensor parameters like speed and exhaust temperature.

[0089] To determine the weight parameters and bias terms of the domain discriminant subnetwork D, the domain classification loss is minimized during training. Iterative optimization yielded the following results;

[0090] The sigmoid activation function is used to map the output of the linear layer to the [0,1] interval, which meets the probability output requirements of the binary classification task.

[0091] Then through

[0092] Calculate the domain classification loss, and update the parameters of F after gradient inversion to increase the domain confusion; where, The domain classification loss is used to guide the parameter updates of the domain discrimination subnetwork D. The smaller the loss value, the stronger the domain discrimination ability of D.

[0093] The number of source domain samples and target domain samples participating in the iteration is specified. For example, in step two, the source domain training set has about 1000 samples and the target domain unlabeled samples have about 500 samples. Each iteration extracts a batch of samples, such as batch size 32.

[0094] These are the source domain sample set and a single source domain sample, respectively. Samples from bench tests after screening in step one;

[0095] These are the target domain sample set and a single target domain sample, respectively. Data derived from historical sea trial measurements or simulation-generated data;

[0096] This training cycle continues until the validation set metrics converge, indicating that the model has achieved the required accuracy and domain indistinguishability in the source domain. When the domain discriminant subnetwork D cannot effectively distinguish between the source and target domains, for example, when its discrimination accuracy drops to about 50%, it means that the feature distribution alignment is successful.

[0097] After the initial training of the model is completed, for example, the source domain training set contains about 1,000 samples and the target domain unlabeled samples contain about 500 samples. After training with the Adam optimization algorithm with a learning rate of 0.001 for about 100 epochs, the various losses of the model converge, and the classification accuracy of the domain discriminator drops to about 50%, which is close to the level of random guessing. The training of the adversarial domain adaptation model is completed.

[0098] If a small number of labeled samples in the target domain are subsequently obtained, a supervised fine-tuning strategy can be used to further optimize the model. That is, the feature extraction layer F is fixed, and only the prediction layer C is adjusted to adapt to the new samples in the target domain, thereby further improving the prediction accuracy of the target domain.

[0099] It should be noted that, in this embodiment, unsupervised adversarial training is sufficient to guarantee the model's generalization performance. On the other hand, for cases where the criterion results indicate acceptable domain differences, a confidence calibration algorithm is used to train and adjust the model.

[0100] At this point, the source domain data is sufficient to train the model's foundation. Therefore, the source domain labeled data is used to perform conventional supervised training on the prediction model, such as minimizing the mean squared error or cross-entropy loss, until the validation set converges. During the training process, techniques such as L2 regularization and Dropout random deactivation can be introduced to prevent the model from overfitting the source domain data, thereby further improving its generalization ability to new environments.

[0101] To improve the reliability of the model's confidence in the target domain, a post-processing confidence calibration strategy is adopted. For example, a temperature scaling method is introduced to calibrate the model's output probability: using the source domain validation set or a small number of target domain validation samples, a temperature parameter T is learned to scale the model's logits, thereby adjusting the shimmy of the output distribution and making the predicted probability more consistent with the actual occurrence rate.

[0102] The calibration effect can be verified by methods such as reliability curves or calculation of expected calibration error. For example, by plotting the reliability curve between the model's predicted probability and the actual frequency of the result, and observing the prediction accuracy of each confidence interval, it can be found that before calibration, the actual frequency of an event that the model gives a 90% confidence level may only be 70%. However, after calibration such as temperature scaling, the meaning of the probability output by the model is corrected, and a 90% prediction confidence level can correspond to approximately 90% of the actual occurrence rate, thereby significantly improving the reliability of the confidence level.

[0103] After calibration, the decision threshold can be reset based on the calibrated model output. For example, the alarm threshold can be adjusted to adapt to the risk level in the target domain environment based on the model output failure probability prediction, thereby ensuring that the alarm strategy achieves an ideal balance between false alarms and false alarms in the target domain.

[0104] For regression prediction, the predicted value can be adjusted based on historical deviations using linear correction methods. For example, the mean of the model prediction deviation can be calculated using known samples in the target domain, and this mean can be used as a calibration offset for subsequent predictions. The confidence calibration algorithm thus ensures that the model will not make overconfident judgments due to source domain deviations in the sea trial environment, thereby improving the overall reliability of the prediction results. After the corresponding training and calibration processes in the above two paths, the obtained machine learning model parameters are finally solidified and ready for inputting sea trial data and making decision predictions in the next step.

[0105] In step S3, during the sea trial, the trained model is loaded, multi-source data collected in real time is input, continuous prediction is performed, and a dual release mechanism is established by setting the anchor bucket accuracy and residual fluctuation threshold, as well as the consistency and uncertainty limits of the neighborhood bucket. Specific details include:

[0106] The model is applied in the sea trial environment and the prediction results are verified in real time. The trained and calibrated model is deployed to the sea trial operation phase of the marine engine, and the actual operation data is predicted and analyzed. The generalization effect of the model is verified by comparing it with the measured results.

[0107] Specifically, the model obtained in step S2 is integrated into the sea trial monitoring. The model can be deployed on the ship's industrial control computer or a shore-based monitoring server. It acquires engine sensor data streams via real-time data bus or wireless transmission and performs predictive calculations at a set period, such as a frequency synchronized with the sensor sampling frequency. During the sea trial after the engine is installed on the ship, sensor data, including parameters of the same type as those in bench tests, such as speed, torque, and temperature, are acquired in real time and input into the model for predictive calculations.

[0108] For example, the model can output predicted values ​​of key engine performance indicators in real time, or provide a health score and probability of failure for the current operating status. As sea trials proceed, it simultaneously records actual measured engine performance values ​​and status events, such as actual power output curves, fuel consumption, and whether fault alarms occur. The model's predictions are compared with the corresponding measured data to calculate error indicators, such as relative error percentage and mean square error (MSE). The mean square error is calculated through... The calculation is used to evaluate the overall prediction bias over multiple periods; where n is the number of sampling periods. For example, the predicted exhaust temperature in the i-th sampling period is given by the model prediction. The measured values ​​of the sea trial in the i-th sampling period, such as the measured exhaust temperature in the i-th second; and the classification prediction performance index, to evaluate the prediction accuracy and confidence of the model in the target domain;

[0109] It is worth noting that the error threshold here can be set according to the engine performance acceptance standards and test requirements: for example, the prediction error is usually required to be no more than 5% for continuous performance parameters such as power and torque, and no missed reports and the false alarm rate should be minimized for key fault events. If the deviation between the predicted value and the measured value is found to be within the allowable range, such as the prediction error of the main performance indicators being less than 5%, it means that the model has successfully generalized and the prediction results are reliable.

[0110] If certain indicators deviate beyond the threshold and persist for multiple sampling periods, or if the model provides a high-confidence prediction that does not match the actual situation, the model verification and adjustment process is triggered. This determination incorporates a duration criterion to avoid false triggers caused by transient fluctuations. For example, under a steady-state cruise condition, the model predicts the engine's output torque to be 2000 N·m, while the actual measured average torque during that period is 1980 N·m. The difference is only 1%, lower than the preset 5% error threshold, indicating accurate prediction. Another example is that the model predicts the exhaust temperature peak will reach 500℃ after a rapid load change, triggering a high-temperature alarm. However, the actual measured peak temperature is 490℃, and no alarm is triggered. In this case, the model overestimates the temperature rise, exhibiting a tendency to false alarms, which falls under the category of deviations exceeding the threshold.

[0111] For the latter, the model prediction will be determined to be inaccurate, and the subsequent verification and adjustment process will be triggered. The predicted value, measured value and error at each moment will be presented in real time through the monitoring interface and stored in the log file at the same time, so that test engineers can view it at any time and conduct in-depth analysis of the model and engine performance afterward.

[0112] Determine the source of the problem based on the type of deviation: For models trained using adversarial domain adaptation, if errors are concentrated in specific uncovered operating conditions, i.e., the prediction errors are mainly concentrated in the operating range corresponding to the operating condition buckets that are not marked as covered by the operating condition coverage criterion, it may indicate that the source domain sample coverage is still insufficient. In this case, the new operating condition data from the sea trial can be included for retraining or fine-tuning the model. Among them, uncovered operating conditions refer to operating condition buckets that do not meet the operating condition coverage criterion in step S1 and are not included in the anchoring calibration set. Their corresponding operating data have not formed a stable representation during the model training phase.

[0113] Specifically, the new working condition data can be added to the training sample set, and the model can be retrained offline or its parameters fine-tuned using the new data during the test intervals. Then, the updated model can be loaded and deployed for continued use. If the amount of new data is small and immediate adaptation is required, an online incremental learning algorithm can be used to make small updates to the model parameters without interrupting the test.

[0114] For models using confidence calibration, if a confidence level inaccuracy occurs, such as when the model predicts a high probability but the corresponding event does not actually occur, the calibration parameters need to be adjusted, such as recalculating the temperature scaling factor or offset. Since the amount of computation required to update such calibration parameters is small, the calculation can be completed in real time after the inaccuracy is detected and the new parameters can be applied to the model to correct the prediction bias immediately.

[0115] After the model is adjusted, the predictions for the working conditions with large deviations can be re-compared with the actual measurements to verify the adjustment effect. For example, after fine-tuning and correction, in the high temperature prediction scenario mentioned above where false alarms occurred, the model predicted a peak temperature of 492℃, which is only about 0.4% different from the actual value of 490℃, showing a significant improvement.

[0116] After adjustment, the key performance prediction errors of the model on the sea trial data are all kept within the predetermined threshold range, for example, the average error is less than 3% and the classification prediction accuracy reaches more than 95%, which meets the design accuracy requirements. If the model prediction accuracy still does not meet the requirements after the above correction, return to step S1 to add or optimize training data, and adjust the model structure or algorithm accordingly. Then repeat the process of steps S2 and S3 until the model performance indicators meet the requirements.

[0117] During model validation, the model's predictions can also serve as supplementary information for test monitoring: when a prediction indicates that a parameter is about to exceed safety limits, test personnel or automatic safety protection systems can intervene in advance, such as proactively reducing engine load or performing an emergency shutdown, thereby avoiding potential equipment damage or accidents. This demonstrates that the model improves safety during the testing phase while providing predictions.

[0118] In step S4, the model deployment involves a two-stage validation of the prediction output using shadow channels and canary ratios. The predicted and measured residual trajectories are monitored online, and fine-tuning is performed within the allowable range of bias and confidence parameters. All runtime records are synchronously written back to the spectral system. Specific details include:

[0119] Based on the model prediction results, closed-loop decision control is executed and feedback optimization is performed. After the model verification and necessary adjustments during the sea trial phase, the final reliable prediction model results are used to assist or automatically execute decision control on the engine, realizing closed-loop management of the entire manufacturing process.

[0120] Based on the model's predictions of engine status and performance, corresponding control strategies or operational decisions are formulated. For example, if the model predicts that the engine exhaust temperature will exceed the safety threshold under a certain operating condition, control measures such as reducing the fuel injection quantity or increasing the cooling flow under that operating condition are taken in advance to avoid overheating failures during actual sea trials. If the model predicts that the engine cannot reach the design power requirements under extreme high loads, the boost pressure and fuel supply are adjusted in a timely manner to increase output. If necessary, the test is suspended for recalibration to avoid performance deficiencies being exposed only after delivery.

[0121] If predictions indicate that emissions at a certain stage will exceed the specified thresholds, combustion parameters can be optimized in advance or backup emission reduction devices can be activated to reduce emissions and ensure that the test process meets environmental protection requirements. These prediction-based measures are automatically executed by the control system in a closed loop: for example, the engine electronic control unit adjusts execution parameters such as fuel injection quantity and boost pressure in real time according to the instructions output by the prediction model to implement the corresponding measures;

[0122] The model continuously monitors engine operating parameters, predicts potential deviations, and provides real-time feedback control commands to the engine to correct performance deviations or eliminate potential problems, thus achieving closed-loop regulation of prediction, decision-making, and control. In addition, the prediction system can also provide the above prediction results and proposed control measures to the operator through the human-machine interface for reference, so that manual confirmation or intervention can be carried out when necessary, thereby ensuring the transparency and safety of the decision-making and execution process.

[0123] Using model prediction results to assist in quality assessment and manufacturing improvement decisions: After the sea trial is completed, the key performance indicators predicted by the model are compared with the actual measured indicators to evaluate whether the engine meets the design requirements and acceptance standards. If the prediction and the actual measurement are consistent and both meet the requirements, the engine is confirmed to have passed the test and can be delivered for use.

[0124] If the model has predicted a problem and corrected it through control measures, the corresponding problem and solution are recorded and included in the manufacturing process knowledge base; otherwise, if both the prediction results and the actual measured values ​​indicate that the engine's key performance indicators continue to fail to meet the predetermined targets, the engine is determined to be non-compliant with delivery standards and requires further design improvements or remanufacturing before testing.

[0125] For unforeseen anomalies, the analysis feeds back to model improvement and manufacturing process optimization. For example, if the model repeatedly predicts that a certain temperature parameter may exceed the limit during the sea trial and actual verification confirms that there is indeed a tendency to exceed the limit, the manufacturing department can improve the cooling system design or adjust the calibration strategy to address this weakness and eliminate the potential problem in subsequent products. Conversely, if a fault occurs during the sea trial that the model failed to predict in advance, the fault analysis will identify the relevant features that the model did not capture and incorporate them into the training data and feature set, thereby improving the model's ability to predict similar faults.

[0126] The actual data, model predictions and decision execution records collected during this sea trial were organized and fed back into the dataset and criterion settings in step S1: the training sample library was expanded to include new operating condition data, and the thresholds or rules of the dual criteria were updated. For example, the threshold of the domain difference criterion was adjusted for the offset pattern discovered this time, so as to further improve the accuracy and adaptability of the model in subsequent engine manufacturing and testing.

[0127] In subsequent manufacturing and testing of the same engine model, the model and calibration parameters obtained from this training can be directly used as the initial model, reducing the time and debugging costs of retraining; by continuously inheriting and optimizing the model, product consistency and development efficiency will be improved.

[0128] A test report containing a comparison analysis of predictions and actual measurements is generated, serving as one of the bases for quality acceptance. The model's parameters and prediction results are also archived and saved, providing data reference and experience support for subsequent similar engine projects. In addition, after the engine is delivered and put into use, the prediction model constructed in this embodiment can continue to be used for online monitoring and predictive maintenance: by analyzing operational data in real time, potential fault signs can be detected in advance. In the above-mentioned service phase application, a closed-loop mechanism similar to steps S3 and S4 can be adopted. By periodically collecting engine data during ship operation, the model parameters are verified and updated, so that the model can continuously adapt to the specific ship's operating conditions, thereby always maintaining a high accuracy in predicting the engine's health status, further extending the application value of this decision prediction method throughout the engine's entire life cycle.

[0129] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0131] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0132] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0133] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0134] In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A decision-making and prediction method for the entire manufacturing process of marine engines, combined with machine learning, characterized in that, Including the following steps: Step S1: Collect the operating conditions, process parameters and performance indicators involved in bench tests and sea trials, construct a data map with operating condition buckets as the basic unit, and select samples to form an anchoring calibration set based on the two criteria of operating condition coverage and time stability. According to the criterion results, identify the trainable and non-trainable areas. Step S2: Select a training path. When the statistical distribution difference between bench test data and sea trial operation data exceeds a preset domain difference threshold in at least one key working condition feature dimension, use an adversarial domain adaptation algorithm to map and converge the representation space. When the offset is within the allowable range, use a confidence calibration method to process the model output and adjust the model structure and weights to generate a fixed representation. Step S3: During the sea trial operation, load the trained model, input real-time collected multi-source data, perform continuous prediction, and establish a dual release mechanism by setting the anchor bucket accuracy and residual fluctuation threshold, as well as the consistency and uncertainty limits of the neighborhood bucket. In step S4, the model is deployed using a two-stage verification method with shadow channel and canary ratio verification. The predicted and measured residual trajectories are monitored online, and fine-tuning is performed within the allowable range of bias and confidence parameters. All running records are then synchronously written back to the spectral system.

2. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 1, characterized in that: Collect raw data during the test bench phase of marine engines, including key parameters such as engine speed, torque, fuel consumption rate, exhaust temperature, and vibration acceleration. The raw data is preprocessed, including outlier removal and noise filtering, and the time series data is divided into multiple training samples. Extract the feature parameters of each sample, including mean, peak value, fluctuation range, and spectral energy, and associate them with the expected output value or state label; Data is filtered based on the working condition coverage criterion and the domain difference criterion to obtain the flag parameters, which are used to guide the selection of the model training algorithm path in step S2.

3. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 2, characterized in that, The working condition coverage criterion determines whether the data volume corresponding to each typical working condition is not lower than the preset minimum sample proportion threshold by statistically analyzing the data volume and distribution of each working condition period. Domain difference criteria quantify the degree of domain shift by calculating the distributional differences of key features between the source and target domains, including using statistical comparison methods or kernel methods to calculate the maximum mean difference index (MMD).

4. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 2, characterized in that, The filtered data are normalized or standardized, and data augmentation is performed, including superimposing simulated noise and random perturbation parameter values ​​on the original time series, or generating virtual samples based on physical models.

5. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 1, characterized in that, Based on the flag parameters, select either the adversarial domain adaptation algorithm or the confidence calibration algorithm; When the flag parameter indicates that the domain difference index between the source domain data and the target domain data is greater than the preset domain difference threshold, the adversarial domain adaptation algorithm is adopted. When the flag parameter indicates that the domain difference index is less than or equal to the preset domain difference threshold, a confidence calibration algorithm is used.

6. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 5, characterized in that, The training process of the adversarial domain adaptation algorithm includes: Iterative training is performed using source domain samples and target domain samples; Calculate the main task loss for the source domain samples and update the parameters of the main task prediction sub-network and the feature extraction network; The domain classification loss is calculated for samples from the source and target domains. After gradient inversion, the parameters of the feature extraction network are updated to increase the domain confusion.

7. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 5, characterized in that, The confidence calibration algorithm involves scaling the model's logits using the source domain validation set or a small number of target domain validation samples' temperature parameter T, making the predicted probability more consistent with the actual occurrence rate. Verify the calibration effect using reliability curves or expected calibration errors.

8. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 1, characterized in that, The obtained model is deployed in the sea trial environment to acquire engine sensor data in real time and perform predictive calculations. Record the measured values ​​and calculate the error indices between the predicted and measured values, including the relative error percentage and the mean square error (MSE). Based on the error index and preset threshold, determine whether the model prediction is inaccurate; if it is, trigger the model adjustment process. Model tuning includes retraining, fine-tuning model parameters, or adjusting calibration parameters.

9. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 8, characterized in that, In the model adjustment process, for adversarial domain adaptation models, if the prediction error is mainly concentrated in the operating range corresponding to the operating condition bucket that is not marked as covered by the operating condition coverage criterion, then new operating condition data will be included to retrain or fine-tune the model; for confidence calibration models, if the confidence level is inaccurate, then the calibration parameters will be recalculated.

10. The decision-making and prediction method for the entire manufacturing process of marine engines combined with machine learning as described in claim 1, characterized in that, Based on the model prediction results, control strategies are formulated, such as adjusting the fuel injection quantity, cooling flow rate or boost pressure, to avoid malfunctions or insufficient performance. The quality assessment is conducted using model prediction results. If the prediction matches the actual measurement and meets the requirements, the engine passes the test; otherwise, the design is improved or the engine is remanufactured. The sea trial data is fed back to the dataset in step S1 to update the training sample library and criterion thresholds for subsequent engine testing and model optimization.

Images