A method for predicting the concentration of atmospheric environment gas of a submarine based on SP-ARIMA

Abstract

The application provides a submarine atmospheric environment gas concentration prediction method based on SP-ARIMA, and relates to the technical field of submarine atmospheric environment control and safety guarantee. The method first converts submarine gas concentration sensor data into a supervised learning sample set through a sliding window method, constructs an online ARIMA prediction model containing differential processing and defines a loss function; then through algorithm initialization, joint optimization target construction and alternative optimization strategy iteration, model training from easy to difficult is realized. The submarine atmospheric environment management is upgraded from passive monitoring to active early warning, the prediction accuracy and robustness are significantly improved, excellent prediction performance is shown under smooth and disturbed working conditions, and it is a general prediction framework, belongs to the domestic advanced level, is suitable for concentration prediction of multiple key submarine gases such as CO2, CO, O2 and the like, provides a valuable time window for submarine atmospheric environment control system regulation and control and personnel emergency response, and greatly improves the safety guarantee level of submarine cabin environment.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment control and safety assurance technology for submersibles, specifically to an intelligent prediction method applicable to the long-term closed cabin environment of submersibles, capable of high-precision and interference-resistant prediction of key gas concentrations. Background Technology

[0002] As specialized equipment operating in a sealed environment underwater for extended periods, the stability and safety of the cabin's atmospheric environment are core issues for ensuring the health and safety of operators and sustaining continuous missions. To achieve this, modern submersibles are equipped with sophisticated atmospheric environment control systems. These systems not only possess crucial functions such as oxygen regeneration and carbon dioxide absorption, but also integrate catalytic purification capabilities for carbon monoxide, hydrogen, hydrocarbons, and other trace pollutants. Furthermore, they utilize air filters to adsorb and treat acidic and odorous gases, collectively maintaining clean and suitable air within the cabin. Atmospheric environment monitoring, as one of the most critical components of the life support system, provides the basis for decision-making regarding atmospheric treatment and control.

[0003] The use of centralized atmospheric environment monitoring systems on submersibles has been around for nearly forty years. Current mainstream technologies primarily rely on mass spectrometry to monitor critical gas components that affect human safety in real time. However, these traditional monitoring systems are inherently reactive, triggering alarms or activating purification equipment only after a gas concentration exceeds a preset threshold, lacking necessary predictability. For some highly toxic gas components, even a brief exceedance of the concentration limit can cause irreversible damage to the health of operators. Therefore, this reactive approach presents significant safety hazards and limitations.

[0004] A more significant challenge lies in the fact that the interior of a submersible is a dynamic and complex environment. The daily activities of operators, the start-up and shutdown of various equipment, and the periodic fluctuations of the ventilation system all inevitably introduce significant noise and outliers into the data collected by gas concentration sensors. This highly noisy environment constitutes a major obstacle to data-driven prediction. When data is severely contaminated, changes in data points across time series can be deceptive, greatly impacting the accuracy of the model.

[0005] Against this backdrop, traditional time-series forecasting models, such as the classic Autoregressive Integrated Moving Average (ARIMA) model, have revealed their inherent vulnerabilities. These models experience a sharp decline in performance and severe instability when dealing with noisy data containing numerous outliers. The root cause lies in the model's difficulty in distinguishing between real data patterns and random noise during training. It is easily misled by outliers, leading to deviations in model parameter estimates and an inability to accurately capture the true underlying patterns of gas concentration changes. This noise-induced model instability makes it difficult for traditional forecasting methods to provide reliable early warning information in the practical application scenarios of underwater vehicles.

[0006] Therefore, existing atmospheric environment protection technologies for submersibles face an urgent need to transition from passive monitoring to proactive early warning. Developing a robust prediction method that can effectively resist strong noise interference and accurately predict the changing trends of key gas concentrations inside the submersible cabin is of vital practical significance for achieving early warning of atmospheric conditions, comprehensively improving the survival protection level of operators, and enhancing the overall operational effectiveness of submersibles. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for predicting atmospheric gas concentrations for underwater vehicles based on SP-ARIMA. This method aims to overcome the deficiencies of existing underwater vehicle atmospheric monitoring technologies, such as lack of predictive ability, susceptibility of prediction models to noise interference, and poor robustness. It primarily solves the following technical problems: 1) Instability of the prediction model in a strong noise environment.

[0008] 2) The problems of premature convergence and insufficient accuracy of traditional prediction algorithms, and the problem that standard time series models are easily misled by abnormal data points in the early stage of training, which leads to the model getting stuck in local optima and failing to accurately model.

[0009] 3) Shift from passive monitoring to proactive early warning: This aims to upgrade atmospheric environmental management from passive response to proactive early warning, by predicting the trend of gas concentration changes several minutes or even hours in advance, providing a valuable time window for the regulation of atmospheric environmental control systems and emergency response by personnel.

[0010] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting atmospheric gas concentrations for underwater vehicles based on SP-ARIMA includes the following steps: Step 1: Construct a time series prediction model for gas concentration in submersibles Concentration sensor data of target gas in a specific compartment of an underwater vehicle are collected and arranged in chronological order to form an original time series. The sliding window method is used to convert the original time series into a supervised learning training sample set. An online ARIMA model is constructed as the basic prediction framework. The non-stationary time series is processed by the difference technique, and the linear combination of historical data points is used to characterize the gas concentration prediction relationship. The mean squared error loss function is defined to quantify the error between the model prediction value and the true value. Step 2: Design and execute the self-stepping SP-ARIMA algorithm to train the prediction model The algorithm initializes the ARIMA model with weights and initial loss threshold; a joint optimization objective function including model weights and sample selection weights is constructed, and a self-step regularization term is introduced to control the sample selection process; the model weights, sample selection weights, and loss threshold parameters are iteratively updated using an alternative optimization strategy until the preset termination condition is met. Step 3: Gas Concentration Prediction and Active Early Warning The trained SP-ARIMA model is applied to predict the future time step of the target gas concentration of the submersible, and the concentration prediction results are output. If the prediction results show that the gas concentration will exceed the preset safety threshold, the active atmospheric environment warning of the submersible is immediately triggered, providing a basis for decision-making for environmental regulation and emergency response.

[0011] Furthermore, the specific method for generating the training sample set using the sliding window method described in step 1 is as follows: The gas concentration data collected from sensors in a specific compartment of the submersible are arranged in chronological order to form a time series. ,in At any moment Observed values ​​of gas concentration; To utilize historical information from time series, a one-dimensional original gas concentration sequence is generated. Convert the sample set required for supervised learning; Construct the original sequence as training samples, It is the total length of the original time series, the first... training samples The generation method is as follows: in, Indicates the first A generated training sample pair, ; It is the first The input feature vector of each sample is derived from... It consists of a series of continuous historical concentration observations; It is the first The target output label for each sample is the true concentration value at the time immediately following the input feature vector. The length of the sliding window determines the number of historical data points used for a single prediction.

[0012] Furthermore, the online ARIMA model described in step 1 at time... concentration value The concentration prediction expression is: in, At any moment The vector of model parameters that needs to be learned; yes The The component represents the first component. The contribution weights of each difference term to the prediction result; These are the hyperparameters of the ARIMA model, which together determine the range of historical information referenced by the model and the order of difference. Indicates historical observation values conduct The order difference is used to extract trend information from the sequence.

[0013] Furthermore, the expression for the mean squared error loss function mentioned in step 1 is: in, It is the model at the current weight parameters Next time The loss value generated by predicting data points.

[0014] Furthermore, the specific operations for algorithm initialization in step 2 include: 1) Weight initialization: Initialize the weight vector of the ARIMA model. Perform initialization, for example, using random small values ​​or zero values, denoted as . ; 2) Initial loss calculation: using initial weights For all generated in step 1.1 Calculate the initial prediction loss for each training sample. ; 3) Initialize the loss threshold parameter: Initialize the loss threshold. for To ensure that only a small subset of the simplest samples are selected during the initial training phase, It is typically set as a specific proportion of the average initial loss of all samples to initiate the self-step learning process.

[0015] Furthermore, the expression for the joint optimization objective function in step 2 is: in, It is the weight parameter vector of the ARIMA model; It is a potential selection weight vector. Representing the Whether a training sample is selected. This indicates that the sample is considered a simple sample and participates in this round of training. This indicates that the sample was considered a difficult sample and was not included. It is the first Prediction loss for each sample; This is a self-stepping regularization term used to control the sample selection process. It employs a hard regularizer and takes the form of: ,parameter A loss threshold is defined to distinguish between easy and difficult samples, thus controlling the pace of learning.

[0016] Furthermore, the iterative update of the alternative optimization strategy described in step 2 includes: 1) Update model weights In sample selection weights With the weights fixed, the gradient descent method is used to update the model weights. For the defined mean squared error loss, its weights The partial derivatives are: The update formula for batch gradient descent is: in, The learning rate determines the step size for updating the model parameters; The core of this update formula is that only the gradients of the selected samples are accumulated, i.e. The model automatically ignores samples that it considers difficult or noisy when calculating the overall update direction, thus ensuring the stability of the model in the early stages of training. 2) Update sample selection weights In model weights Under fixed conditions, for Recalculate the predicted loss for each sample. Then, based on the loss value and the current parameters The comparison updates the selection weights for each sample. : The updated sample selection weight The mechanism acts like a dynamic filter, adapting to the model. With continuous optimization, its ability to fit data has improved, and the loss value of some samples that previously had large losses may be reduced to [a lower value]. Therefore, they are re-identified as simple samples and included in subsequent training. 3) Update the loss threshold parameter After a complete round of parameter and sample selection updates, the learning criteria need to be relaxed so that the model can learn more complex patterns by increasing the parameters. To achieve this goal: in, It is greater than The update factor controls the rate of learning pace; The gradual increase of the value means that the model's definition of simple samples becomes more and more relaxed, which allows difficult samples that were previously excluded due to excessive loss to be gradually included in the training set. This helps the model avoid overfitting to simple samples and improves its final generalization ability and robustness.

[0017] Furthermore, the preset termination condition in step 2 is that the number of iterations reaches a preset maximum number of iterations, or the model weights of two adjacent iterations reach a certain threshold. The change is less than the preset minimum threshold.

[0018] Furthermore, the target gas includes one or more of CO2, CO, O2, hydrogen, and hydrocarbons; The active early warning is triggered by local audible and visual alarms in the submersible compartment and by control commands from the environmental control system. This method is suitable for atmospheric environmental monitoring in long-term sealed compartments of submersibles and can predict gas concentration trends for the next few minutes to hours.

[0019] This invention provides a method for predicting atmospheric gas concentrations in underwater vehicles based on SP-ARIMA. It offers the following advantages: 1. This invention provides a method for predicting atmospheric gas concentration in underwater vehicles based on SP-ARIMA. Through a training mechanism that progresses from easy to difficult, it effectively avoids the interference of noise and outliers in underwater vehicle sensor data, enabling the model to learn the true laws of gas concentration changes. Compared with traditional ARIMA models, it has higher prediction accuracy and stronger robustness in noisy environments.

[0020] 2. This invention provides a method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA, which can predict the trend of gas concentration changes over a period of time in advance, and in particular, can provide early warning of potential risks of concentration exceeding the standard. It transforms the traditional passive monitoring-alarm mode into an active prediction-early warning mode, thus buying valuable time for taking preventive control measures or protecting personnel.

[0021] 3. This invention provides a method for predicting atmospheric gas concentrations in submersibles based on SP-ARIMA. It is a general framework that only requires inputting time series data of different types of gases (such as CO2, CO, O2, etc.) to train a corresponding robust prediction model. It is applicable to the prediction of various key atmospheric components inside submersibles. Attached Figure Description

[0022] Figure 1 This is a flowchart of the SP-ARIMA-based method for predicting atmospheric gas concentrations in underwater vehicles according to the present invention. Figure 2 This is a flowchart of the SP-ARIMA algorithm of the present invention; Figure 3 This is a training data diagram for an implementation case of the present invention; Figure 4 This is a training data diagram for implementation case two of the present invention; Figure 5 This is a comparison chart of the prediction effects under one working condition in the implementation case of the present invention; Figure 6 This is a comparison chart of the prediction effects under the second implementation case of the present invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0024] In the description of this invention, unless otherwise explicitly specified and limited, the terms "connected," "linked," and "fixed" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0025] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can include direct contact between the first and second features, or contact between the first and second features through another feature between them. Furthermore, "above," "over," and "on top" of the second feature includes the first feature directly above or diagonally above the second feature, or simply indicates that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature includes the first feature directly below or diagonally below the second feature, or simply indicates that the first feature is at a lower horizontal level than the second feature.

[0026] In the description of this embodiment, the terms "upper," "lower," "right," etc., refer to the orientation or positional relationship shown in the accompanying drawings. They are used only for ease of description and simplification of operation, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. In addition, the terms "first" and "second" are used only for distinction in description and have no special meaning.

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0028] like Figures 1-2 As shown, this embodiment of the invention provides a method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA, including the following steps: Step 1: Construct a time series prediction model for gas concentration in submersibles Concentration sensor data of target gas in a specific compartment of an underwater vehicle are collected and arranged in chronological order to form an original time series. The sliding window method is used to convert the original time series into a supervised learning training sample set. An online ARIMA model is constructed as the basic prediction framework. The non-stationary time series is processed by the difference technique, and the linear combination of historical data points is used to characterize the gas concentration prediction relationship. The mean squared error loss function is defined to quantify the error between the model prediction value and the true value. The specific method for generating the training sample set using the sliding window method is as follows: The gas concentration data collected from sensors in a specific compartment of the submersible are arranged in chronological order to form a time series. ,in At any moment Observed values ​​of gas concentration; To utilize historical information from time series, a one-dimensional original gas concentration sequence is generated. Convert the sample set required for supervised learning; Construct the original sequence as training samples, It is the total length of the original time series, the first... training samples The generation method is as follows: in, Indicates the first A generated training sample pair, ; It is the first The input feature vector of each sample is derived from... It consists of a series of continuous historical concentration observations; It is the first The target output label for each sample is the true concentration value at the time immediately following the input feature vector. The length of the sliding window determines the number of historical data points used for a single prediction. Online ARIMA model at time concentration value The concentration prediction expression is: in, At any moment The vector of model parameters that needs to be learned; yes The The component represents the first component. The contribution weights of each difference term to the prediction result; These are the hyperparameters of the ARIMA model, which together determine the range of historical information referenced by the model and the order of difference. Indicates historical observation values conduct The order difference is used to extract trend information from the sequence; The quality of a model's predictions is assessed through a loss function. To measure this, it quantifies the predicted value. Compared with the true value The error between them is usually expressed as mean squared error (MSE) or other error measures; The expression for the mean squared error loss function is: in, It is the model at the current weight parameters Next time The loss value generated by predicting data points.

[0029] Step 2: Design and execute the self-stepping SP-ARIMA algorithm to train the prediction model The algorithm initializes the ARIMA model with weights and initial loss threshold; a joint optimization objective function including model weights and sample selection weights is constructed, and a self-step regularization term is introduced to control the sample selection process; the model weights, sample selection weights, and loss threshold parameters are iteratively updated using an alternative optimization strategy until the preset termination condition is met. The specific operations for algorithm initialization include: 1) Weight initialization: Initialize the weight vector of the ARIMA model. Perform initialization, for example, using random small values ​​or zero values, denoted as . ; 2) Initial loss calculation: using initial weights For all generated in step 1.1 Calculate the initial prediction loss for each training sample. ; 3) Initialize the loss threshold parameter: Initialize the loss threshold. for To ensure that only a small subset of the simplest samples are selected during the initial training phase, It is typically set to a specific proportion of the average initial loss of all samples to initiate the self-step learning process; The expression for the joint optimization objective function is: in, It is the weight parameter vector of the ARIMA model; It is a potential selection weight vector. Representing the Whether a training sample is selected. This indicates that the sample is considered a simple sample and participates in this round of training. This indicates that the sample was considered a difficult sample and was not included. It is the first Prediction loss for each sample; This is a self-stepping regularization term used to control the sample selection process. It employs a hard regularizer and takes the form of: ,parameter A loss threshold was defined to distinguish between easy and difficult samples, thus controlling the pace of learning; Iterative updates to alternative optimization strategies include: 1) Update model weights In sample selection weights With the weights fixed, the gradient descent method is used to update the model weights. For the defined mean squared error loss, its weights The partial derivatives are: The update formula for batch gradient descent is: in, The learning rate determines the step size for updating the model parameters; The core of this update formula is that only the gradients of the selected samples are accumulated, i.e. The model automatically ignores samples that it considers difficult or noisy when calculating the overall update direction, thus ensuring the stability of the model in the early stages of training. 2) Update sample selection weights In model weights Under fixed conditions, for Recalculate the predicted loss for each sample. Then, based on the loss value and the current parameters The comparison updates the selection weights for each sample. : Update sample selection weights The mechanism acts like a dynamic filter, adapting to the model. With continuous optimization, its ability to fit data has improved, and the loss value of some samples that previously had large losses may be reduced to [a lower value]. Therefore, they are re-identified as simple samples and included in subsequent training. 3) Update the loss threshold parameter After a complete round of parameter and sample selection updates, the learning criteria need to be relaxed so that the model can learn more complex patterns by increasing the parameters. To achieve this goal: in, It is greater than The update factor controls the rate of learning pace; The gradual increase of the value means that the model's definition of simple samples becomes more and more relaxed, which allows difficult samples that were previously excluded due to excessive loss to be gradually included in the training set. This helps the model avoid overfitting to simple samples and improves its final generalization ability and robustness.

[0030] The preset termination condition is that the number of iterations reaches the preset maximum number of iterations, or the model weights of two consecutive iterations reach a certain threshold. The change is less than the preset minimum threshold.

[0031] Step 3: Gas Concentration Prediction and Active Early Warning The trained SP-ARIMA model is applied to predict the future time step of the target gas concentration of the submersible, and the concentration prediction results are output. If the prediction results show that the gas concentration will exceed the preset safety threshold, the active atmospheric environment warning of the submersible is immediately triggered, providing a basis for decision-making for environmental regulation and emergency response.

[0032] The target gas includes one or more of CO2, CO, O2, hydrogen, and hydrocarbons; The active early warning is triggered by local audible and visual alarms in the submersible compartment and by control commands from the environmental control system. This method is suitable for atmospheric environmental monitoring in long-term sealed compartments of submersibles and can predict gas concentration trends for the next few minutes to hours.

[0033] Implementation Case: The effectiveness and superiority of the method described in this invention are verified by using a specific scenario of predicting carbon dioxide (CO2) concentration in a submarine compartment.

[0034] 1) Scene and Data Settings This case study simulates time-series data of CO2 concentration inside a cabin, continuously collected by sensors, to verify the effectiveness of the algorithm of this invention. To comprehensively evaluate the robustness of the model, the dataset includes two typical operating conditions: ① Operating Condition 1 (Stable Period): This stage corresponds to the stable operation of the operators' work and rest schedules and the atmospheric circulation and purification systems. Data characteristics include periodic, small-range fluctuations with an overall stable trend, representing a relatively simple data environment. For example... Figure 3 The image shown is a training data graph for working condition 1.

[0035] ②Scenario Two (Disturbance Period): This stage simulates several non-periodic, severe, and brief spikes in CO2 concentration caused by sudden events such as temporary intensive activities of personnel in specific compartments or adjustments to the start and stop of ventilation and purification equipment. For example... Figure 4The image shown is a training data graph for working condition 2.

[0036] The goal of this case study is to train a prediction model using historical data containing the two operating conditions mentioned above, and to test its accuracy in predicting CO2 concentrations in the future.

[0037] 2) Algorithm parameter settings To conduct comparative verification, this case study will implement two algorithms simultaneously: 1) a control group using a standard ARIMA model; and 2) the SP-ARIMA model proposed in this invention. To ensure fairness in the comparison, the common parameter settings for both models are as follows: ① ARIMA model hyperparameters All set to This indicates that the model uses the observations after the past 14 differences for prediction, and performs first-order differencing on the original sequence to handle non-stationarity; ② Sliding window length Set to 15; ③ Weight parameters within the model Initialize using a standard normal distribution. ; ④ The maximum number of iterations is set to 500; ⑤ Learning rate Set as ; ⑥ The historical dataset used for training is exactly the same.

[0038] The core parameters of the SP-ARIMA model of this invention are set as follows: ① Initial threshold parameter Set to 30% of the initial average loss for all training samples; ②Update factor Set as This allows for a smooth and gradual relaxation of the learning pace.

[0039] 3) Implementation Results Both models were trained on a training dataset containing mixed operating conditions, and CO2 concentration predictions were performed on test set data that were not used for training after the stationary and disturbance periods. The root mean square error (RMSE) was used as the core performance evaluation metric; the smaller the RMSE value, the higher the prediction accuracy of the model.

[0040] like Figure 5 The image shown is a comparison chart of the prediction results for working condition 1. Figure 6 The image shown is a comparison chart of the prediction results for working condition 2.

[0041] The prediction performance comparison is shown in the table below: The data in the table above reveals the significant advantages of the method of the present invention: 1) Under stable operating conditions, the prediction error RMSE of the SP-ARIMA model of the present invention and the standard ARIMA model are both at a low level, at 5.28ppm and 15.10ppm respectively, but the method of the present invention is better.

[0042] 2) Under perturbation conditions, the two models exhibited significant performance differences. The prediction error of the standard ARIMA model increased dramatically, with the RMSE soaring from 15.10 ppm to 41.12 ppm. This indicates that the standard ARIMA model was severely affected by strong noise spikes in the data, misinterpreting them as true data patterns, causing a serious deviation in its learning direction and resulting in highly unreliable predictions.

[0043] 3) In comparison, the model of this invention exhibits superior robustness under perturbation conditions. Its RMSE is only 8.80 ppm, far lower than the standard ARIMA model. The fundamental reason for this is that the self-step learning mechanism intelligently identifies data points caused by noise spikes that result in significant losses as difficult samples in the early stages of training and temporarily excludes them from model training. The model first learns the basic laws of CO2 concentration changes from a large amount of simple, stationary data. As the model gradually stabilizes and the parameters increase, some previously ignored noise data is gradually and controlledly incorporated into the training set. At this point, the model has a strong discriminative ability, thereby minimizing the damage of noise to the final model performance and ensuring the stability and accuracy of the prediction results.

[0044] This case fully verifies the robust prediction method based on self-step learning proposed in this invention. It can effectively overcome the interference of strong noise data in the complex, variable, and noisy environment of the submarine cabin, and achieve high-precision and high-stability prediction of the concentration of key gases. Its performance is significantly better than traditional methods and has extremely high engineering application value.

[0045] The following points should be noted in this article: 1. The accompanying drawings of the embodiments disclosed herein only relate to the structures involved in the embodiments disclosed herein; other structures can be referred to in a general design.

[0046] 2. Where there is no conflict, the embodiments of this disclosure and the features in the embodiments can be combined with each other to obtain new embodiments.

[0047] Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA, characterized in that, Includes the following steps: Step 1: Construct a time series prediction model for gas concentration in submersibles Concentration sensor data of target gas in a specific compartment of an underwater vehicle are collected and arranged in chronological order to form an original time series. The sliding window method is used to convert the original time series into a supervised learning training sample set. An online ARIMA model is constructed as the basic prediction framework. The non-stationary time series is processed by the difference technique, and the linear combination of historical data points is used to characterize the gas concentration prediction relationship. The mean squared error loss function is defined to quantify the error between the model prediction value and the true value. Step 2: Design and execute the self-stepping SP-ARIMA algorithm to train the prediction model The algorithm initializes the ARIMA model with weights and initial loss threshold; a joint optimization objective function including model weights and sample selection weights is constructed, and a self-step regularization term is introduced to control the sample selection process; the model weights, sample selection weights, and loss threshold parameters are iteratively updated using an alternative optimization strategy until the preset termination condition is met. Step 3: Gas Concentration Prediction and Active Early Warning The trained SP-ARIMA model is applied to predict the future time step of the target gas concentration of the submersible, and the concentration prediction results are output. If the prediction results show that the gas concentration will exceed the preset safety threshold, the active atmospheric environment warning of the submersible is immediately triggered, providing a basis for decision-making for environmental regulation and emergency response.

2. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The specific method for generating the training sample set using the sliding window method described in step 1 is as follows: The gas concentration data collected from sensors in a specific compartment of the submersible are arranged in chronological order to form a time series. ,in At any moment Observed values ​​of gas concentration; To utilize historical information from time series, a one-dimensional original gas concentration sequence is generated. Convert the sample set required for supervised learning; Construct the original sequence as training samples, It is the total length of the original time series, the first... training samples The generation method is as follows: in, Indicates the first A generated training sample pair, ; It is the first The input feature vector of each sample is derived from... It consists of a series of continuous historical concentration observations; It is the first The target output label for each sample is the true concentration value at the time immediately following the input feature vector. The length of the sliding window determines the number of historical data points used for a single prediction.

3. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The online ARIMA model described in step 1 at time... concentration value The concentration prediction expression is: in, At any moment The vector of model parameters that needs to be learned; yes The The component represents the first component. The contribution weights of each difference term to the prediction result; These are the hyperparameters of the ARIMA model, which together determine the range of historical information referenced by the model and the order of difference. Indicates historical observation values conduct The order difference is used to extract trend information from the sequence.

4. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The expression for the mean squared error loss function mentioned in step 1 is: in, It is the model at the current weight parameters Next time The loss value generated by predicting data points.

5. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The specific operations for algorithm initialization in step 2 include: 1) Weight initialization: Initialize the weight vector of the ARIMA model. Perform initialization, for example, using random small values ​​or zero values, denoted as . ; 2) Initial loss calculation: using initial weights For all generated in step 1.1 Calculate the initial prediction loss for each training sample. ; 3) Initialize the loss threshold parameter: Initialize the loss threshold. for To ensure that only a small subset of the simplest samples are selected during the initial training phase, It is typically set as a specific proportion of the average initial loss of all samples to initiate the self-step learning process.

6. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The expression for the joint optimization objective function mentioned in step 2 is: in, It is the weight parameter vector of the ARIMA model; It is a potential selection weight vector. Representing the Whether a training sample is selected. This indicates that the sample is considered a simple sample and participates in this round of training. This indicates that the sample was considered a difficult sample and was not included. It is the first Prediction loss for each sample; This is a self-stepping regularization term used to control the sample selection process. It employs a hard regularizer and takes the form of: ,parameter A loss threshold is defined to distinguish between easy and difficult samples, thus controlling the pace of learning.

7. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The iterative update of the alternative optimization strategy in step 2 includes: 1) Update model weights In sample selection weights With the weights fixed, the gradient descent method is used to update the model weights. For the defined mean squared error loss, its weights The partial derivatives are: The update formula for batch gradient descent is: in, The learning rate determines the step size for updating the model parameters; The core of this update formula is that only the gradients of the selected samples are accumulated, i.e. The model automatically ignores samples that it considers difficult or noisy when calculating the overall update direction, thus ensuring the stability of the model in the early stages of training. 2) Update sample selection weights In model weights Under fixed conditions, for Recalculate the predicted loss for each sample. Then, based on the loss value and the current parameters The comparison updates the selection weights for each sample. : The updated sample selection weight The mechanism acts like a dynamic filter, adapting to the model. With continuous optimization, its ability to fit data has improved, and the loss value of some samples that previously had large losses may be reduced to [a lower value]. Therefore, they are re-identified as simple samples and included in subsequent training. 3) Update the loss threshold parameter After a complete round of parameter and sample selection updates, the learning criteria need to be relaxed so that the model can learn more complex patterns by increasing the parameters. To achieve this goal: in, It is greater than The update factor controls the rate of learning pace; The gradual increase of the value means that the model's definition of simple samples becomes more and more relaxed, which allows difficult samples that were previously excluded due to excessive loss to be gradually included in the training set. This helps the model avoid overfitting to simple samples and improves its final generalization ability and robustness.

8. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 7, characterized in that, The preset termination condition mentioned in step 2 is that the number of iterations reaches the preset maximum number of iterations, or the model weights of two adjacent iterations reach a certain threshold. The change is less than the preset minimum threshold.

9. The method for predicting atmospheric gas concentrations of underwater vehicles based on SP-ARIMA according to claim 1, characterized in that, The target gas includes one or more of CO2, CO, O2, hydrogen, and hydrocarbons; The active early warning is triggered by local audible and visual alarms in the submersible compartment and by control commands from the environmental control system. This method is suitable for atmospheric environmental monitoring in long-term sealed compartments of submersibles and can predict gas concentration trends for the next few minutes to hours.

Images