Method, device and equipment for safety assessment of deep-sea manned submersible

Abstract

The application provides a deep-sea manned submersible safety evaluation method, device and equipment, the method comprises the following steps: dividing historical sensor measurement data according to whether a safety state label is marked to obtain a labeled data set and an unlabeled data set; identifying distribution mismatched unlabeled data from the unlabeled data set based on trained deep support vector networks; selecting a preset number of unlabeled data based on a weighted robust structure sparse representation method according to an active labeling budget to perform active labeling; performing semi-supervised learning on a safety evaluation model based on a new labeled data set and a new unlabeled data set to obtain a trained safety evaluation model; collecting real-time sensor measurement data of the deep-sea manned submersible, inputting the trained safety evaluation model, and obtaining the current safety state of the deep-sea manned submersible. Based on the identification of distribution mismatched samples, active labeling is performed, and the accuracy of safety evaluation is improved.

Description

Technical Field

[0001] This invention relates to the field of safety assessment technology, specifically to a method, apparatus, and equipment for assessing the safety of deep-sea manned submersibles based on active semi-supervised learning. Background Technology

[0002] Deep-sea manned submersibles typically operate in high-salt, high-pressure, and highly uncertain deep-sea environments, fraught with unknown risks. If a malfunction occurs and cannot be addressed promptly, it can not only prevent the completion of underwater operations but also directly threaten the lives of the submersible's crew. Therefore, safety assessments of deep-sea manned submersibles are crucial. These assessments must combine data collected by sensors to evaluate the overall safety of the submersible system in real time, provide early warnings for potential malfunctions, and ensure the safety of the crew through self-rescue systems. However, current safety assessments of deep-sea manned submersibles rely primarily on the subjective judgment of the crew. Due to the limited visibility and attention of the crew underwater, misjudgments are easily made, potentially leading to irreversible consequences. Therefore, there is an urgent need to research data-driven safety assessment methods to assist crew members in making self-rescue decisions.

[0003] Deep-sea manned submersibles collect abundant data through sensors during dives, but this data lacks safety status labels. Labeling this data is costly and often difficult to obtain sufficient labeled data. Semi-supervised learning can extract information from both labeled and unlabeled data, making it suitable for assessing the safety of deep-sea manned submersibles. For deep-sea manned submersibles, the same safety state may correspond to multiple failure scenarios. For example, a moderate safety state could be caused by leaks in the adjustable ballast tank, the auxiliary manipulator's oil level compensation system, or the main manipulator's oil level compensation system. Some of these failure scenarios may only have unlabeled data, lacking corresponding system safety information; these are called unknown scenarios. While semi-supervised learning can extract information from both labeled and unlabeled data, traditional semi-supervised learning models do not consider the existence of unknown scenarios in unlabeled data. Unknown scenarios can impair the performance of traditional semi-supervised learning models, preventing them from accurately assessing system safety when unknown scenarios occur, thus hindering correct self-rescue decisions and creating safety hazards. Summary of the Invention

[0004] This invention provides a method, apparatus, and equipment for safety assessment of deep-sea manned submersibles based on active semi-supervised learning, in order to solve the problem of low assessment accuracy of existing methods when unknown scenarios exist.

[0005] In a first aspect, embodiments of the present invention provide a method for safety assessment of deep-sea manned submersibles based on active semi-supervised learning, comprising:

[0006] Historical sensor measurement data of deep-sea manned submersibles were acquired, and the historical sensor measurement data were divided into labeled datasets and unlabeled datasets based on whether safety status labels were attached.

[0007] Using a labeled dataset, deep support vector networks are trained for each security state label in the security state label space. Based on the trained deep support vector networks, mismatched unlabeled data are identified from the unlabeled dataset.

[0008] According to the active annotation budget, a preset number of unlabeled data are selected based on the weighted robust structural sparse representation method, and active annotation is performed to obtain a new labeled dataset and a new unlabeled dataset.

[0009] Semi-supervised learning is performed on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain a trained security assessment model.

[0010] Real-time sensor measurement data from the deep-sea manned submersible is collected and input into a trained safety assessment model to obtain the current safety status of the deep-sea manned submersible.

[0011] In one embodiment, the loss function of the deep support vector network is as follows:

[0012]

[0013] in, It is a deep support vector network φ k (·;W k The set of weight parameters, L k It is the number of hidden layers, n k It is the number of data samples in the labeled dataset that belong to the safe state label k. Let be the i-th data sample in the labeled dataset that belongs to the safe state label k, where λ is the regularization hyperparameter, and c k It is the center point of the hypersphere in the feature space, which is generally fixed as the average value of the network output generated by the initial forward propagation of the training samples.

[0014] In one embodiment, identifying mismatched unlabeled data from an unlabeled dataset based on pre-trained deep support vector networks includes:

[0015] For each data sample in the unlabeled dataset, a pre-trained deep support vector network with various security state labels is used. Determine the distribution mismatch score s of the data sample with respect to each security status label. k (x) is as follows:

[0016]

[0017] If the distribution mismatch score of a data sample for each security status label is greater than the preset threshold θ, k If k = 1, 2, ..., M, then the data sample is determined to be a distribution mismatch.

[0018] In one embodiment, according to the active annotation budget, a preset number of unlabeled data points are selected based on a weighted robust structural sparse representation method, including:

[0019] Construct the following optimization objective function J:

[0020]

[0021]

[0022] Where X is the selected representative unlabeled sample, D is the weight matrix, γ is the hyperparameter, n is the number of selected representative unlabeled samples, and A is the optimization objective matrix.

[0023] Solving the objective function yields the optimal matrix A corresponding to the objective matrix. * ;

[0024] Calculate the optimal matrix A * Sum the absolute values ​​of all elements of the row vectors, sort them in descending order, and select the unlabeled data corresponding to the first preset number of row vectors.

[0025] In one embodiment, when X is the entire unlabeled dataset, the elements in the weight matrix D are determined according to the following expression:

[0026]

[0027] Where ρ is a constant greater than 1, ID represents distribution matching, M is the number of security state labels in the security state label space, and s k (x i ) is the i-th unlabeled data point x in the unlabeled dataset. i The distribution of the safety status label k does not match the score.

[0028] In one embodiment, when X represents identified unlabeled data with a mismatched distribution, the weight matrix can be defined as D = D O :

[0029]

[0030]

[0031] Where M is the number of security state labels in the security state label space, and s k (x i X is unlabeled data with a mismatched distribution. o The i-th unlabeled data x i Regarding the distribution mismatch score for the security status label k, It represents the number of unlabeled data points with mismatched distributions.

[0032] In one embodiment, when X represents identified unlabeled data with a mismatched distribution, the weight matrix can be defined as follows: Where I represents the identity matrix, It represents the number of unlabeled data points with mismatched distributions.

[0033] Secondly, embodiments of the present invention provide a safety assessment device for deep-sea manned submersibles, comprising:

[0034] The acquisition module is used to acquire historical sensor measurement data of deep-sea manned submersibles and divide the historical sensor measurement data into labeled datasets and unlabeled datasets based on whether they are labeled with safety status tags.

[0035] The recognition module is used to train deep support vector networks for each security state label in the security state label space using a labeled dataset, and to identify unlabeled data with mismatched distribution from the unlabeled dataset based on the trained deep support vector networks.

[0036] The annotation module is used to select a preset number of unlabeled data points according to the active annotation budget and based on the weighted robust structural sparse representation method, and actively annotate them to obtain a new labeled dataset and a new unlabeled dataset.

[0037] The training module is used to perform semi-supervised learning on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain the trained security assessment model.

[0038] The evaluation module is used to collect real-time sensor measurement data of the deep-sea manned submersible, input it into the trained safety evaluation model, and obtain the current safety status of the deep-sea manned submersible.

[0039] Thirdly, embodiments of the present invention provide an electronic device, comprising:

[0040] At least one processor and memory;

[0041] The memory stores instructions that the computer executes;

[0042] At least one processor executes computer execution instructions stored in memory, causing the at least one processor to perform a safety assessment method for a deep-sea manned submersible as described in any of the first aspects.

[0043] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the safety assessment method for deep-sea manned submersibles as described in any of the first aspects.

[0044] The present invention provides a method, apparatus, and device for safety assessment of deep-sea manned submersibles based on active semi-supervised learning. This method acquires historical sensor measurement data from the deep-sea manned submersible and divides this data into labeled and unlabeled datasets based on whether safety status labels are attached. Using the labeled datasets, deep support vector networks are trained for each safety status label in the safety status label space. Based on these trained networks, mismatched unlabeled data are identified from the unlabeled datasets. According to an active labeling budget, a predetermined number of unlabeled data points are selected and actively labeled using a weighted robust sparse representation method, resulting in new labeled and unlabeled datasets. The safety assessment model is then semi-supervised based on these datasets to obtain a trained model. Real-time sensor measurement data from the deep-sea manned submersible is collected and input into the trained model to obtain the current safety status of the submersible, allowing for appropriate self-rescue measures to be taken. Considering the existence of unknown scenarios in unlabeled datasets, a deep support vector description network is constructed to identify samples with mismatched distributions. Representative samples are selected for active labeling based on a weighted robust sparse representation method, which improves the accuracy and reliability of security assessment. Attached Figure Description

[0045] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0046] Figure 1 A flowchart illustrating a method for assessing the safety of a deep-sea manned submersible based on active semi-supervised learning, as provided in an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of a safety assessment device for deep-sea manned submersibles based on active semi-supervised learning, provided in an embodiment of the present invention.

[0048] Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0049] The accompanying drawings have illustrated specific embodiments of the invention, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the invention in any way, but rather to illustrate the concept of the invention to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.

[0051] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.

[0052] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).

[0053] Traditional semi-supervised learning models do not consider the presence of unknown scenarios in unlabeled data. The existence of these unknown scenarios prevents traditional semi-supervised learning models from accurately assessing system safety when they occur, thus hindering correct self-rescue decisions and creating safety hazards. To address this issue, this application proposes an active semi-supervised safety assessment method that considers unknown scenarios for the safety assessment of deep-sea manned submersible self-rescue systems. This method fully considers the presence of unknown scenarios in unlabeled datasets, constructs a deep support vector description network to identify distributionally mismatched samples, and actively labels representative samples based on a weighted robust sparse representation method. This method can be combined with any semi-supervised learning method to improve the accuracy of the model's safety assessment, especially its ability to handle unknown scenarios. Specific embodiments will be used below to illustrate the method provided in this application in detail.

[0054] Figure 1 This is a flowchart illustrating a method for safety assessment of deep-sea manned submersibles based on active semi-supervised learning, provided as an embodiment of the present invention. Figure 1 As shown, the deep-sea manned submersible safety assessment method based on active semi-supervised learning provided in this embodiment may include:

[0055] S101. Obtain historical sensor measurement data of the deep-sea manned submersible, and divide the historical sensor measurement data into labeled datasets and unlabeled datasets according to whether they are labeled with safety status tags.

[0056] The measurement data in this embodiment is obtained through sensors carried by the deep-sea manned submersible. These sensors can collect data such as cabin pressure, humidity, temperature, 24V power supply voltage, 24V power supply current, oxygen concentration, carbon dioxide concentration, main manipulator oil level compensation, auxiliary manipulator oil level compensation, bow trim level, stern trim level, adjustable water tank level, and computer tank temperature. Taking a submersible carrying V sensors as an example, the historical sensor measurement data obtained is as follows: Because labeling these data with security status labels is costly, usually only a portion of the data is labeled with security status labels y∈Y={1,2,...,M}, where Y is the security status label space and M is the number of system security states, i.e. the number of security status labels in the security status label space.

[0057] Historical sensor measurement data can be divided into labeled datasets based on whether or not they are labeled with safety status tags. and unlabeled datasets Where N1 represents the number of labeled data points and N2 represents the number of unlabeled data points.

[0058] S102. Using labeled datasets, train deep support vector networks for each security state label in the security state label space. Based on the trained deep support vector networks, identify unlabeled data with mismatched distributions from the unlabeled dataset.

[0059] Using a labeled dataset as training data, a deep support vector network φ is trained for each category k∈Y in the safety state label space Y={1,2,...,M}. k Then, using a pre-trained deep support vector network, the distribution mismatch score of each unlabeled data point in the unlabeled dataset with respect to each security state label is calculated, and based on this, the unlabeled dataset D is identified. u Mismatched distribution data in This represents the number of data samples with mismatched distributions. It's understandable that a higher mismatch score indicates a higher probability that the data corresponds to an unknown scenario. Learning from unlabeled data with mismatched distributions helps improve the ability to assess the safety of unknown scenarios.

[0060] S103. According to the active annotation budget, select a preset number of unlabeled data based on the weighted robust structural sparse representation method, perform active annotation, and obtain a new labeled dataset and a new unlabeled dataset.

[0061] In this embodiment, the active annotation budget can be a quantity budget, a time budget, or a monetary budget. Regardless of how the active annotation budget is limited, the preset quantity is positively correlated with the active annotation budget. In an optional implementation, unlabeled data with mismatched distributions can be prioritized for active annotation. Specifically, a preset number of unlabeled data with mismatched distributions can be selected as the objects of active annotation, ranked from highest to lowest based on the sum of the mismatch scores of the unlabeled data with respect to each security state label. It is understood that after actively annotating the unlabeled data, the amount of data in the unlabeled dataset will decrease, while the amount of data in the labeled dataset will increase. Taking the selection of q unlabeled data for active annotation as an example, a new labeled dataset will be obtained. and new unlabeled datasets

[0062] S104. Perform semi-supervised learning on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain a trained security assessment model.

[0063] Based on the new labeled dataset after active annotation and new unlabeled datasets This allows for semi-supervised learning, and the method provided in this embodiment can be combined with any semi-supervised learning method. In one optional implementation, a semi-supervised learning method based on course learning pseudo-labels can be used. Specifically, this may include the following steps:

[0064] S1041: Select and initialize a suitable classification network structure and parameters, denoted as C. θ Its input is sensor measurement data. The output is a security label space Y = {1,2,...,M}, where θ represents the network parameters;

[0065] S1042, Based on labeled datasets Train the classification network and obtain The loss function is as follows

[0066]

[0067] Where CE represents the cross-entropy loss function.

[0068] S1043. Iteratively train the model using the concept of course learning. The loss function for the t-th iteration is as follows:

[0069]

[0070] in, This represents the model after t-1 training rounds. For sample x j Prediction results for The conditions are met Data sample set, Representation Model The maximum value output. return middle The r-th result of the calculation t Percentile values. During iteration, set r... t The percentage increases from 20% to 100% in increments of 20%. Represents a set of data samples The number of samples.

[0071] S105. Collect real-time sensor measurement data of the deep-sea manned submersible, input it into the trained safety assessment model, and obtain the current safety status of the deep-sea manned submersible.

[0072] After obtaining a well-trained safety assessment model based on semi-supervised learning, online evaluation can be performed by inputting real-time sensor measurement data into the model. The safety assessment model will output the current safety status of the deep-sea manned submersible, allowing for appropriate self-rescue measures to be taken. For example, if the current safety status is dangerous, the submersible can be stopped immediately; if the current safety status is moderately safe, relevant warnings can be given to the submersible crew or they can be guided to eliminate potential safety hazards.

[0073] The deep-sea manned submersible safety assessment method based on active semi-supervised learning provided in this embodiment fully considers the existence of unknown scenarios in unlabeled datasets. It identifies mismatched samples in the unlabeled dataset by constructing a deep support vector description network, and selects representative samples for active labeling based on a weighted robust sparse representation method. Semi-supervised learning is then performed based on the actively labeled data, which improves the accuracy and reliability of safety assessment.

[0074] Based on the above embodiments, the following will further explain in detail how to extract data from the unlabeled dataset D. u Identify data with mismatched distributions Specifically, the following steps are included:

[0075] S1021. For each category k∈Y in the security state label space Y={1,2,...,M}, train a deep support vector network φ. k Its loss function is as follows:

[0076]

[0077] in, It is a deep support vector network φ k (·;W k The set of weight parameters, L k It is the number of hidden layers, n k There is a labeled dataset D l The number of samples belonging to category k. Let be the i-th data sample in the labeled dataset that belongs to the safe state label k, where λ is the regularization hyperparameter, and c k It is the center point of the hypersphere in the feature space, which is generally fixed as the average value of the network output generated by the initial forward propagation of the training samples.

[0078] S1022. Given any unlabeled data sample x∈D u Calculate the distribution mismatch score s for each category k. k (x) is as follows:

[0079]

[0080] Among them, W k * This represents the parameters of the trained deep support vector network.

[0081] S1023, Given a threshold θ k For k = 1, 2, ..., M, the following formula is used to determine whether an unlabeled sample x belongs to a distribution mismatch:

[0082]

[0083] Where OOD indicates a mismatched distribution, and ID indicates a matched distribution. Threshold θ k This can be determined by assuming that a certain proportion of the training data belonging to class k is mismatched in distribution. Generally, we can assume that the proportion of mismatched distribution in the training data is 1%.

[0084] S1024. Based on steps S1021 to S1023, from the unlabeled dataset D... u Identify data with mismatched distributions

[0085] In other words, the semi-supervised learning-based deep-sea manned submersible safety assessment method provided in this embodiment identifies mismatched unlabeled data from the unlabeled dataset based on the trained depth support vector networks. Specifically, this may include: for each data sample in the unlabeled dataset, using the trained depth support vector networks for each safety state label to determine the mismatch score of the data sample with respect to each safety state label; if the mismatch score of a data sample with respect to each safety state label is greater than a preset threshold, then the data sample is determined to be mismatched data.

[0086] Based on the above embodiments, the following will further explain in detail how to select q representative samples for active annotation based on the weighted robust sparse representation method, given an active annotation budget q, i.e., the preset number is q. Specifically, it includes the following steps:

[0087] S1031. Construct the following optimization objective function J:

[0088]

[0089]

[0090] Where X is the selected representative unlabeled sample, D is the weight matrix to be designed, γ is the hyperparameter, n is the number of selected representative unlabeled samples, and A is the optimization objective matrix. X and D can be selected and designed according to any of the strategies provided in step S1032.

[0091] S1032. Three strategies are provided for selecting and designing matrices X and D:

[0092] Strategy S1: Select all unlabeled data samples, i.e., X = D u The elements of the weight matrix D are calculated as follows:

[0093]

[0094] Where ρ is a constant greater than 1, ID represents distribution matching, M is the number of security state labels in the security state label space, and s k (x i ) is the i-th unlabeled data point x in the unlabeled dataset. i The score for the distribution mismatch of the security status label k.

[0095] Strategy S2: Select unlabeled data samples with mismatched distributions, i.e., X = X o The weight matrix D = D O The calculation is as follows:

[0096]

[0097]

[0098] Where M is the number of security state labels in the security state label space, and s k (x i X is unlabeled data with a mismatched distribution. o The i-th unlabeled data x i Regarding the distribution mismatch score for the security status label k, It represents the number of unlabeled data points with mismatched distributions.

[0099] Strategy S3: Select unlabeled data samples with mismatched distributions, i.e., X = X o Weight matrix Where I represents the identity matrix, It represents the number of unlabeled data points with mismatched distributions.

[0100] S1033: Select the corresponding strategy from step S1032, solve the objective function J, and obtain the optimal matrix A after solving. * The specific solution process includes the following steps:

[0101] S10331. Calculate diagonal matrices U and V:

[0102]

[0103]

[0104] Where, xi and a i Let a represent the i-th column vector of matrices X and A respectively. i Let i represent the i-th row vector of matrix A. It is a reasonably chosen constant greater than 0, which can generally be set to the relative precision of a floating-point number.

[0105] S10332. Calculate the column vectors a of matrix A according to the following formula. i ,1≤i≤n:

[0106] a i =u i (u i X T X+γVD) -1 X T x i

[0107] S10333, repeat iterative steps S10331 and S10332 until the objective function J converges or the maximum number of iterations is reached.

[0108] S1034. Calculate the optimal matrix A * Sum the absolute values ​​of all elements of the row vectors, sort them in descending order, select the first q row vectors and their corresponding unlabeled samples, label them with their corresponding safety state labels, and obtain a new labeled dataset after active labeling. and new unlabeled datasets

[0109] Combining the above embodiments, this embodiment provides a method for safety assessment of deep-sea manned submersibles based on active semi-supervised learning, including the following steps:

[0110] Step 1: Collect sensor measurement data from the deep-sea manned submersible Part of the data is labeled with safety state labels y∈Y={1,2,...,M}. Here, V is the number of sensors, and M is the number of system safety states. The training data is divided into labeled datasets. and unlabeled datasets

[0111] Step 2: Identify the unlabeled dataset D u Mismatched distribution data in This refers to the number of data samples with a mismatched distribution. The specific steps include:

[0112] Step 2.1: For each category k∈Y in the security label space Y={1,2,...,M}, train a deep support vector network φ. k Its loss function is as follows:

[0113]

[0114] in, It is a deep support vector network φ k (·;W k The set of weight parameters, L k It is the number of hidden layers, n k There is a labeled dataset D l The number of samples belonging to category k. Let be the i-th data sample in the labeled dataset that belongs to the safe state label k, and let c be the regularization hyperparameter. k It is the center point of the hypersphere in the feature space, which is generally fixed as the average value of the network output generated by the initial forward propagation of the training samples;

[0115] Step 2.2: Given any unlabeled data sample x∈D u Calculate the distribution mismatch score s for each category k. k (x) is as follows:

[0116]

[0117] Among them W k * This represents the parameters of the trained deep support vector network.

[0118] Step 2.3: Given a threshold θ k For k = 1, 2, ..., M, the following formula is used to determine whether an unlabeled sample x belongs to a distribution mismatch:

[0119]

[0120] Where OOD indicates a mismatched distribution, and ID indicates a matched distribution. Threshold θ k This can be determined by assuming that a certain proportion of the training data belonging to class k is mismatched in distribution. Generally, we can assume that the proportion of mismatched distribution in the training data is 1%.

[0121] Step 2.4: Based on steps 2.1 to 2.3, extract data from the unlabeled dataset D. u Identify data with mismatched distributions

[0122] Step 3: Based on the weighted robust sparse representation method, given an active annotation budget q, select q representative samples for active annotation, specifically including the following steps:

[0123] Step 3.1: Construct the following optimization objective function:

[0124]

[0125] Where X is the selected representative unlabeled sample, D is the weight matrix to be designed, and γ is the hyperparameter. X and D can be selected and designed according to any of the strategies provided in step 3.2;

[0126] Step 3.2: Provide three strategies for selecting and designing matrices X and D:

[0127] Strategy S1: Select all unlabeled data samples, i.e., X = D u The elements of the weight matrix D are calculated as follows:

[0128]

[0129] Where ρ is a constant greater than 1.

[0130] Strategy S2: Select unlabeled data samples with mismatched distributions, i.e., X = X o The weight matrix D = D O Calculation as follows

[0131]

[0132]

[0133] Strategy S3: Select unlabeled data samples with mismatched distributions, i.e., X = X o Weight matrix Where I represents the identity matrix;

[0134] Step 3.3: Select the corresponding strategy from Step 3.2 and solve the objective function (4) to obtain the solved matrix A. * The specific solution process includes the following steps:

[0135] Step 3.3.1: Calculate the diagonal matrices U and V:

[0136]

[0137]

[0138] Where, x i and a i Let a represent the i-th column vector of matrices X and A respectively. i Let i represent the i-th row vector of matrix A. It is a reasonably chosen constant greater than 0, which can generally be set to the relative precision of a floating-point number;

[0139] Step 3.3.2: Calculate the column vectors a of matrix A sequentially according to the following formula. i ,1≤i≤n

[0140] ai =u i (u i X T X+γVD) -1 X T x i (10),

[0141] Step 3.3.3: Repeat steps 3.3.1 and 3.3.2 until the objective function (4) converges or the maximum number of iterations is reached;

[0142] Step 3.4: Calculate matrix A * Sum the absolute values ​​of all elements of the row vectors, sort them in descending order, select the first q row vectors and their corresponding unlabeled samples, and label them with their corresponding safety labels. This yields the actively labeled dataset. and unlabeled datasets

[0143] Step 4: Based on the labeled dataset after active annotation and unlabeled datasets Semi-supervised learning is performed. This invention can be combined with any semi-supervised learning method; here, a semi-supervised learning method based on course learning pseudo-labels is adopted, specifically including the following steps:

[0144] Step 4.1: Select and initialize a suitable classification network structure and parameters, denoted as C. θ Its input is sensor measurement data. The output is a security label space Y = {1,2,...,M}, where θ represents the network parameters;

[0145] Step 4.2: Based on labeled datasets Train the classification network and obtain The loss function is as follows

[0146]

[0147] Where CE represents the cross-entropy loss function.

[0148] Step 4.3: Iteratively train the model using the concepts learned in the course. The loss function for the t-th iteration is as follows:

[0149]

[0150] in, This represents the model after t-1 training rounds. For sample x j Prediction results for The conditions are met Data sample set, Representation Model The maximum value output. return middle The r-th result of the calculation t Percentile values. During iteration, set r... t The percentage increases from 20% to 100% in increments of 20%. Represents a set of data samples The number of samples.

[0151] Step 5: Input the system's online measurement data x test to the trained network The security assessment result y was obtained test Based on the safety assessment results, take corresponding self-rescue measures.

[0152] This application addresses the issue of unknown scenarios in unlabeled data for safety assessment tasks of deep-sea manned submersible self-rescue systems. Specifically, it proposes a safety assessment method based on active semi-supervised learning. Considering the presence of unknown scenarios in unlabeled datasets, a deep support vector description network is constructed to identify distributionally mismatched samples. Based on the identification of these mismatched samples, representative samples are actively labeled using a weighted robust structural sparse representation method. This method can be combined with any semi-supervised learning method. By introducing active labeling, it improves the accuracy and reliability of the model's safety assessment, especially its ability to handle unknown scenarios.

[0153] To further illustrate the improvement in safety assessment accuracy of the active semi-supervised learning-based deep-sea manned submersible safety assessment method provided in this application, a dataset from a dive of the "Jiaolong" deep-sea manned submersible is used as an example for verification:

[0154] The measurement data collected by the sensors carried by the "Jiaolong" deep-sea manned submersible were used as the model input. In this example, 13 sensor measurement variables of relevant subsystems were considered as model input. The specific variable symbols and meanings are shown in Table 1.

[0155] Table 1

[0156] <![CDATA[x1]]> cabin pressure <![CDATA[x8]]> Main robotic arm oil level compensation <![CDATA[x2]]> cabin humidity <![CDATA[x9]]> Sub-manipulator oil level compensation <![CDATA[x3]]> cabin temperature <![CDATA[x 10 ]]> Bow tilt level <![CDATA[x4]]> 24V power supply voltage <![CDATA[x 11 ]]> Stern tilt level <![CDATA[x5]]> 24V power supply current <![CDATA[x 12 ]]> Adjustable water tank level <![CDATA[x6]]> oxygen concentration <![CDATA[x 13 ]]> Computer tank temperature <![CDATA[x7]]> carbon dioxide concentration

[0157] In addition to data from the normal system operation scenario, data from four fault scenarios were also simulated. The system safety status corresponding to different scenarios is shown in Table 2:

[0158] Table 2

[0159] Normal scenario 1 (Safety) Fault 1: Leakage in the main robotic arm's oil level compensation system. 2 (Medium security) Fault 2: Oil level compensation leak in the auxiliary robotic arm 2 (Medium security) Fault 3: Leakage in the adjustable water tank 2 (Medium security) Fault 4: Clogged oxygen supply device 3 (Danger)

[0160] Consider fault scenarios 1 and 2 as unknown scenarios. For each unknown scenario, collect 100 unlabeled data samples and add them to the training set. For all other known scenarios, collect 100 labeled data samples and 100 unlabeled data samples and add them to the training set. The test set is divided into two cases: all scenarios and unknown scenarios. When the test set covers all scenarios, collect 100 data samples for each scenario. When the test set covers unknown scenarios, collect 100 data samples for each unknown scenario.

[0161] To verify the effectiveness of the method of this invention, the active annotation budget q = 50, hyperparameters λ = 1e-6, γ = 5, ρ = 10, and the deep support vector network φ were set. k The number of neurons in each layer is set to [13, 9, 5], and the classification network C θ Each hidden layer contains a fully connected layer and a BatchNorm layer, with the number of neurons in each layer set to [13, 13, 6, 6, 6, 3]. The safety assessment accuracy of the proposed method for all scenarios and unknown scenarios was calculated under the three proposed active annotation strategies (S1, S2, S3). Furthermore, to verify the superiority of the method, it was compared with the original semi-supervised learning method (CL) based on course-learning pseudo-labels, and methods combining Random, IE, BvSB, RRSS, and KRRSS active annotation strategies respectively. All methods were run 5 times, and the results were averaged. The safety assessment accuracy results on the test set are shown in Table 3.

[0162] Table 3

[0163]

[0164] Analysis of Table 3 shows that the method proposed in this invention significantly improves the accuracy of semi-supervised security assessment tasks and outperforms other active annotation strategies compared to this invention. In particular, it significantly improves the accuracy of security assessments in unknown scenarios. The test results demonstrate the effectiveness and superiority of the method proposed in this invention in semi-supervised security assessment tasks considering unknown scenarios.

[0165] Figure 2 This is a schematic diagram of a deep-sea manned submersible safety assessment device based on active semi-supervised learning, provided in an embodiment of the present invention. Figure 2 As shown, this embodiment provides a deep-sea manned submersible safety assessment device 20 based on active semi-supervised learning, which may include: an acquisition module 201, an identification module 202, an annotation module 203, a training module 204, and an assessment module 205.

[0166] The acquisition module 201 is used to acquire historical sensor measurement data of the deep-sea manned submersible and divide the historical sensor measurement data into labeled datasets and unlabeled datasets based on whether they are labeled with safety status tags.

[0167] The recognition module 202 is used to train a deep support vector network for each security state label in the security state label space using a labeled dataset, and to identify unlabeled data with mismatched distribution from the unlabeled dataset based on the trained deep support vector networks.

[0168] The annotation module 203 is used to select a preset number of unlabeled data according to the active annotation budget and based on the weighted robust structural sparse representation method, and perform active annotation to obtain a new labeled dataset and a new unlabeled dataset.

[0169] Training module 204 is used to perform semi-supervised learning on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain the trained security assessment model.

[0170] The evaluation module 205 is used to collect real-time sensor measurement data of the deep-sea manned submersible, input it into the trained safety evaluation model, and obtain the current safety status of the deep-sea manned submersible.

[0171] The apparatus of this embodiment can be used to perform Figure 1 The technical solutions of the method embodiments shown are similar in principle and in effect, and will not be described again here.

[0172] This invention also provides an electronic device, please refer to [link to relevant documentation]. Figure 3 As shown, the embodiments of the present invention are only used as examples. Figure 3 The examples are provided for illustration only and do not imply that the invention is limited to these examples. Figure 3 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of the present invention. Figure 3 As shown, the electronic device 30 provided in this embodiment may include: a memory 301, a processor 302, and a bus 303. The bus 303 is used to connect the various components.

[0173] The memory 301 stores a computer program, which, when executed by the processor 302, can implement the technical solutions of any of the above method embodiments.

[0174] The memory 301 and processor 302 are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines, such as bus 303. The memory 301 stores a computer program that implements a safety assessment method for deep-sea manned submersibles, including at least one software functional module that can be stored in the memory 301 in the form of software or firmware. The processor 302 executes various functional applications and data processing by running the software program and modules stored in the memory 301.

[0175] The memory 301 may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 301 stores programs, and the processor 302 executes the programs after receiving execution instructions. Furthermore, the software programs and modules within the memory 301 may also include an operating system, which may include various software components and / or drivers for managing system tasks (such as memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components.

[0176] Processor 302 can be an integrated circuit chip with signal processing capabilities. The aforementioned processor 302 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. It is understood that... Figure 3 The structure shown is for illustrative purposes only and may include more... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown. Figure 3 The components shown can be implemented in hardware and / or software.

[0177] This invention also provides a computer-readable storage medium storing a computer program thereon, which is executed by a processor to implement the technical solutions of any of the above method embodiments.

[0178] The various embodiments in this disclosure are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0179] The scope of protection of this disclosure is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its scope and spirit. If such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A method for safety assessment of deep-sea manned submersibles based on active semi-supervised learning, characterized in that, include: Historical sensor measurement data of a deep-sea manned submersible are acquired, and the historical sensor measurement data are divided into labeled datasets and unlabeled datasets based on whether they are labeled with safety status tags. Using the labeled dataset, deep support vector networks are trained for each security state label in the security state label space. Based on the trained deep support vector networks, mismatched unlabeled data are identified from the unlabeled dataset. According to the active annotation budget, a preset number of unlabeled data are selected based on the weighted robust structural sparse representation method, and active annotation is performed to obtain a new labeled dataset and a new unlabeled dataset. Semi-supervised learning is performed on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain a trained security assessment model. Real-time sensor measurement data of the deep-sea manned submersible is collected and input into the trained safety assessment model to obtain the current safety status of the deep-sea manned submersible.

2. The method according to claim 1, characterized in that, The loss function of the deep support vector network is as follows: in, It is a deep support vector network φ k (·;W k The set of weight parameters, L k It is the number of hidden layers, n k It is the number of data samples in the labeled dataset that belong to the safe state label k. Let be the i-th data sample in the labeled dataset that belongs to the safe state label k, where λ is the regularization hyperparameter, and c k It is the center point of the hypersphere in the characteristic space.

3. The method according to claim 1, characterized in that, The process of identifying mismatched unlabeled data from the unlabeled dataset based on the trained deep support vector networks includes: For each data sample in the unlabeled dataset, a pre-trained deep support vector network with various security state labels is used. Determine the distribution mismatch score s of the data sample with respect to each security status label. k (x) is as follows: If the distribution mismatch score of a data sample for each security status label is greater than the preset threshold θ, k If k = 1, 2, ..., M, then the data sample is determined to be a distribution mismatch.

4. The method according to any one of claims 1-3, characterized in that, The step of selecting a preset number of unlabeled data points based on the weighted robust structural sparse representation method according to the active labeling budget includes: Construct the following optimization objective function J: Where X is the selected representative unlabeled sample, D is the weight matrix, γ is the hyperparameter, n is the number of selected representative unlabeled samples, and A is the optimization objective matrix. Solving the objective function yields the optimal matrix A corresponding to the objective matrix. * ; Calculate the optimal matrix A * Sum the absolute values ​​of all elements of the row vectors, sort them in descending order, and select the unlabeled data corresponding to the first preset number of row vectors.

5. The method according to claim 4, characterized in that, When X is the unlabeled dataset, the elements in the weight matrix D are determined according to the following expression: Where ρ is a constant greater than 1, ID represents distribution matching, M is the number of security state labels in the security state label space, and s k (x i ) is the i-th unlabeled data point x in the unlabeled dataset. i The score for the distribution mismatch of the security status label k.

6. The method according to claim 4, characterized in that, When X consists of unlabeled data with a mismatched distribution, the weight matrix is ​​defined as D = D O : Where M is the number of security state labels in the security state label space, and s k (x i X is unlabeled data with a mismatched distribution. o The i-th unlabeled data x i Regarding the distribution mismatch score for the security status label k, It represents the number of unlabeled data points with mismatched distributions.

7. The method according to claim 4, characterized in that, When X consists of unlabeled data with a mismatched distribution, the weight matrix is ​​defined as follows: Where I represents the identity matrix, It represents the number of unlabeled data points with mismatched distributions.

8. A safety assessment device for deep-sea manned submersibles, characterized in that, include: The acquisition module is used to acquire historical sensor measurement data of the deep-sea manned submersible, and divide the historical sensor measurement data into labeled datasets and unlabeled datasets according to whether they are labeled with safety status tags. The identification module is used to train a deep support vector network for each security state label in the security state label space using the labeled dataset, and to identify unlabeled data with mismatched distribution from the unlabeled dataset based on the trained deep support vector networks. The annotation module is used to select a preset number of unlabeled data points according to the active annotation budget and based on the weighted robust structural sparse representation method, and actively annotate them to obtain a new labeled dataset and a new unlabeled dataset. The training module is used to perform semi-supervised learning on the security assessment model based on the new labeled dataset and the new unlabeled dataset to obtain a trained security assessment model. The evaluation module is used to collect real-time sensor measurement data of the deep-sea manned submersible, input it into the trained safety evaluation model, and obtain the current safety status of the deep-sea manned submersible.

9. An electronic device, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the deep-sea manned submersible safety assessment method based on active semi-supervised learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the deep-sea manned submersible safety assessment method based on active semi-supervised learning as described in any one of claims 1-7.

Images