Underwater operation risk assessment method based on generative adversarial network data enhancement

Abstract

The invention discloses an underwater operation risk assessment method based on generative adversarial network data enhancement, and belongs to the field of underwater robot operation risk assessment. The implementation method comprises the following steps: respectively constructing ten risk quantification models such as an ocean current risk factor quantification model, a simplified temperature risk quantification model and a water density risk quantification model according to an underwater robot operation risk factor and an underwater robot operation risk mechanism; collected original data is quantized, quantized risk values form ten-dimensional input of an original data set, original data set output is formed through expert evaluation, small-scale original data is formed by network input and output, and data expansion is performed on the premise of ensuring the quality of generated data by adopting a generative adversarial network principle. A larger-scale risk data set is obtained to train a risk assessment network model, and a more accurate risk prediction result is realized according to the trained risk assessment network model, so that the accuracy and efficiency of underwater robot operation risk assessment are improved.

Description

Technical Field

[0001] The invention relates to an underwater operation risk assessment method based on generative adversarial network data expansion, and belongs to the field of underwater robot operation risk assessment. Background Art

[0002] At present, the risk assessment methods for underwater vehicles are mainly divided into three categories, namely Bayesian networks, Markov chains, and system dynamics modeling and analysis. The three methods can perform offline primary analysis of the risk status of underwater robots to a certain extent, but the risk assessment based on the above methods is complex to implement, has a large amount of calculation, lacks real-time and usability, and cannot meet the actual real-time requirements of the risk assessment system during the operation of underwater robots.

[0003] Due to the high risk and irreversible loss of underwater robots, a real-time risk assessment system must be established for their operation and movement process, and the prediction accuracy should be met at the same time. Therefore, the data-driven risk assessment method can meet the requirements of real-time and accuracy at the same time, that is, the prediction of underwater robot operation risks can be achieved through neural network structure design and neural network training. High-accuracy neural networks require a lot of data support, but there is less data on underwater robot operation risks. Therefore, appropriate data enhancement methods should be considered to increase the amount of data, thereby improving the accuracy of the prediction model. Summary of the invention

[0004] In view of the problem that the existing underwater robot risk prediction methods have poor practicality and cannot form an implementation risk assessment system for underwater robot operation risk assessment, the main purpose of the present invention is to provide an underwater operation risk assessment method based on generative adversarial network data enhancement. The method proposes a risk quantification and assessment model by analyzing the action mechanism of each risk factor, quantifies the collected original data, and the quantified risk values ​​constitute the ten-dimensional input of the original data set. After expert evaluation, the original data set network output is formed. The network input and output constitute small-scale original data. The generative adversarial network principle is used to expand the data while ensuring the quality of the generated data, so as to obtain a larger-scale risk data set for training the risk assessment network model, and achieve more accurate risk prediction results based on the trained risk assessment network model. According to the risk prediction results, the accuracy and efficiency of the underwater robot's risk avoidance actions are improved.

[0005] The purpose of the present invention is achieved through the following technical solutions.

[0006] The underwater operation risk assessment method based on generative adversarial network data enhancement disclosed in the present invention comprises the following steps:

[0007] Step 1: Analyze the risk factors faced by underwater robot operations. The risk factors of underwater robot operations include ten risk factors: ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. According to the risk factors of underwater robot operations and the risk mechanism of underwater robot operations, ten quantitative models are constructed, including the quantitative model of ocean current risk factor, simplified temperature risk quantitative model, water density risk quantitative model, reef risk factor quantitative model, aquatic plant belt risk quantitative model, swimming fish risk quantitative model, power system risk quantitative model, power system quantitative model, operation depth risk factor quantitative model and operation distance risk factor quantitative model. The data collected in the quantitative model of ocean current risk factors include ocean current speed, angle between the robot's forward direction and the ocean current direction, ambient temperature in Celsius, ambient water density, horizontal width of the reef, distance between the underwater robot and the reef, angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, body length of the swimming fish, swimming speed of the swimming fish, distance between the swimming fish and the underwater robot, horizontal width of the waterweed belt, density of the waterweed belt, distance between the center of the waterweed belt and the center of the underwater robot, angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, battery service time, and battery remaining power percentage , the current depth accounts for the percentage of the rated maximum operating depth and the current mission distance accounts for the percentage of the maximum rated navigation distance. The risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operating depth risk factor and operating distance risk factor obtained after quantification processing by the respective quantitative models of the ten risk factors are used as the ten-dimensional input of the original data set for training the underwater robot operation risk factor generation adversarial network.

[0008] The quantitative model of ocean current risk factors is constructed as shown in formula (1):

[0009] r cur =αv cur |cosθ cur |+βv cur sinθ cur (1)

[0010] Among them, r cur represents the risk value of the ocean current to the underwater robot, α represents the longitudinal threat coefficient of the ocean current to the underwater robot, β represents the lateral threat coefficient of the ocean current to the underwater robot, and v cur is the ocean current velocity, θ cur represents the angle between the ocean current and the underwater robot’s forward direction, θ∈[0,π]. The underwater robot’s ocean current speed v cur The range is vcur ∈[0,10], the risk value r caused by ocean current to underwater robot cur The range is r cur ∈[0,1], according to the underwater robot current speed v cur The risk value r of range and ocean current to underwater robots cur The longitudinal threat coefficient α of the ocean current to the underwater robot and the lateral threat coefficient β of the current to the underwater robot satisfy the constraints shown in the following equations (2) and (3):

[0011]

[0012] α<β (3)

[0013] The simplified temperature risk quantification model using formula (4)

[0014]

[0015] Among them, r tmp represents the risk value of the ambient temperature to the underwater robot, r tem ∈[0,1], t is the ambient temperature in Celsius, combined with the actual operation scenario and appropriate expansion, T∈[-30,30].

[0016] The water density risk quantification model shown in formula (5) is used

[0017] r den =493.8(ρ-1.045) 2 (5)

[0018] In the formula, r den is the density risk value, r den ∈[0,1], ρ is the density of seawater, and after limit extension, ρ∈[1.00,1.09].

[0019] The quantitative model of reef risk factors used is:

[0020]

[0021] In the formula, r sto is the reef risk value, θ sto is the angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, θ sto ∈[0,π],k(θ sto ) is the angle risk factor, L sto is the horizontal width of the reef, d sto is the distance between the underwater robot and the reef. Considering the actual operation scenario and the measurable range, L sto The value range is 0~10m, d stoThe detection range is 0.1 to 5 m. When the value is less than 0.6 and reefs exist, r sto Directly take the value as 1, k(θ sto ) is taken as formula (7)

[0022]

[0023] In the formula, r sto ∈[0,1], when r sto >1, set it to 1.

[0024] The risk quantification model of aquatic weed belts is adopted using formula (8)

[0025]

[0026] Among them, r gra Indicates the risk value of aquatic plants to the underwater robot, L gra represents the horizontal width of the waterweed belt, ρ gra represents the area density of the aquatic weed belt, that is, the number of aquatic weeds per unit area, d gra represents the distance between the center of the waterweed belt and the center of the underwater robot, θ gra is the angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, θ gra ∈[0,π], considering the actual operation scenario and the measurable range, when the risk factor of the length of the waterweed belt is not saturated, L gra The value range is 0~5m, d gra The detection range is 0.1 to 5 m. However, in order to ensure the absolute safety of the underwater robot, when its value is less than 0.6 and there are aquatic plants, r gra The value is 1, gra The value range is 0~20 plants / m 2 , coefficient k(θ gra ) takes a value of

[0027]

[0028] To ensure gra ∈[0,1], when r gra >1, set it to 1.

[0029] The risk quantification model of swimming fish shown in formula (10) is adopted

[0030]

[0031] Among them, r fih is the risk value caused by swimming fish to the underwater robot, and its range is r fih ∈[0,1],L fih is the body length of the swimming fish, and its consideration range is L fih∈[0.1,1],v fih is the swimming speed of the fish, v fish ∈[0,10], v is the navigation speed of the underwater robot, v∈[0,10], d fih is the distance between the fish and the underwater robot, d fih ∈[0.1,5], to ensure the safety of the underwater robot, when its value is less than 0.6 and there are swimming fish, let r fih =1,θ fih is the angle between the speed direction of the underwater robot and the speed direction of the swimming fish, θ fih ∈[0,π], take k fih =0.1, r fih >1, set it to 1.

[0032] The power system risk quantification model shown in formula (11) is adopted

[0033] r pow =0.1t (11)

[0034] In the formula, r pow The risk value of the underwater robot power system, that is, the risk coefficient of the power system is proportional to the service life of the battery.

[0035] The quantitative model of the power system caused by the decrease in power percentage is defined as

[0036]

[0037] In the formula, r dc is the power system risk caused by the normalized power percentage decrease, p 1 is the percentage of remaining battery power. According to the data and rules expressed in the research experiment, the parameter k is solved according to the exponential decay law of battery capacity. 1 , k 2 , k 3 .

[0038] The total risk value of the power system adds up the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in power. The power system risk is defined as

[0039] r dyn =γr gra +(1-γ)r dc (13)

[0040] Where: γ is the weight coefficient of the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in electricity, which is specified as a maximum value between 0 and 1.

[0041] Formula (14) is used to establish a quantitative model of operation depth risk factor

[0042]

[0043] In the formula, r dc is the power system risk caused by the normalized power percentage decrease, p 2 is the percentage of the current depth to the rated maximum operating depth. When the depth percentage is 0, r dph =0, when the depth percentage is 0.8, r dph =0.2, when the depth percentage is 1, r dph =1, solve the parameter k based on the above three percentage data 4 , k 5 , k 6 .

[0044] The operation distance risk quantification model used is:

[0045] r dist =p 3 (15)

[0046] In the formula, p 3 is the percentage of the current mission distance to the maximum rated navigation distance, r dist is the risk value of the underwater robot’s operating distance.

[0047] After quantification according to the ten risk factor quantification models, we obtain the risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. The risk value data are used as the ten-dimensional input of the original underwater robot operation risk assessment data set for training the underwater robot operation risk factor generative adversarial network.

[0048] Step 2: Solicit a predetermined number of experts to evaluate the risk level corresponding to the input data of the original data set generated after the quantification in step 1, and divide the evaluation levels into low risk, relatively low risk, relatively high risk and high risk, and encode the risk levels as 1, 2, 3, and 4 respectively. Take the mode of the evaluation results as the risk level corresponding to the corresponding risk vector to form the output of the original data set, and complete the construction of the small-scale original underwater robot operation risk assessment data set. The input dimension of the underwater robot operation risk assessment data set is ten dimensions, and the output dimension is one dimension, and the one-dimensional output is the risk level.

[0049] Step 3: Randomly generate noise input based on the generator of the fully connected network structure. The output of the generator has the same dimension as the real data in the original underwater robot operation risk assessment dataset. The real data in the original underwater robot operation risk assessment dataset and the data generated by the generator are mixed and input into the discriminator composed of the fully connected network. The discriminator measures the authenticity of the generated data, and the average square error is used as the loss function. The network parameters of the generator are adjusted and optimized through continuous training. Through continuous iteration and joint optimization of the generator and the discriminator, the generator and the discriminator are trained against each other, and the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original dataset is smaller and smaller, until the optimal parameters of the generator model that meet the preset accuracy requirements are solved, and the Nash equilibrium state is reached, the quality of the generated data is improved, and the generated data that is more in line with the distribution of the original data is obtained, thereby improving the risk assessment accuracy and efficiency of underwater robots in underwater operation scenarios.

[0050] In order to obtain a risk assessment model with higher accuracy, a data enhancement generative adversarial network for underwater robot operation risk factors is constructed, and the data set is expanded using the generative adversarial network. The generator adopts a fully connected network structure, whose input is a randomly generated noise vector, and the output is the generated data with the same dimension as the input data in the original underwater robot operation risk assessment data set. The loss function uses formula (16)

[0051]

[0052] In the formula, m is the number of data in a noise data batch, z i is the random noise input to the generator G. This formula directly takes the average value of the discrimination results of the discriminator D on the data generated by the generator G as the loss function of the generator G.

[0053] The discriminator D also uses a fully connected network structure. The input data is the real input in the original underwater robot operation risk assessment data set. The output result is a continuous value between 0 and 1. The discriminator output result is 1 when the real data is input, and the discriminator output result is 0 when the generated data is input. The loss function of the discriminator is:

[0054]

[0055] Where n is the number of data in a mixed data set of a single batch of real data and data generated by the generator, and y i For x i The corresponding expected value is when x i To generate data y i =0, when x i When y is real data i =1.

[0056] By continuously iteratively jointly optimizing the generator and the discriminator, and training the generator and the discriminator against each other, the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original data set becomes smaller and smaller, until the optimal parameters of the generator model that meet the preset accuracy requirements are solved, the Nash equilibrium state is reached, and a high-quality data generator is obtained, thereby expanding the underwater robot operation risk factor data set to conform to the original data distribution, thereby improving the accuracy and efficiency of risk identification in underwater robot operation scenarios.

[0057] Step 4: Use the original data in the underwater robot operation risk factor dataset to train a fully connected neural network, use the fully connected neural network as a verification tool to verify the quality of the data generated by the generator, input the data generated by the generator trained in step 3 into the verification tool neural network, and judge the quality of the generated data by the prediction accuracy of the verification tool neural network. After the verification, a predetermined proportion of the original data is used to verify the quality of the data generated by the generator. The original data in the underwater robot operation risk factor dataset and the newly generated data are mixed to train the fully connected network as the final underwater robot operation risk assessor. The underwater robot operation risk assessor is used to assess the risk of the underwater robot in the underwater operation scene, and the risk identification accuracy and efficiency of the underwater robot in the underwater operation scene are improved.

[0058] Beneficial effects:

[0059] 1. The underwater operation risk assessment method based on generative adversarial network data enhancement disclosed in the present invention systematically analyzes the risk factors faced by underwater robot operations, and constructs ten quantitative models according to the underwater robot operation risk factors and underwater robot operation risk mechanisms, including ocean current risk factor quantification model, simplified temperature risk quantification model, water density risk quantification model, reef risk factor quantification model, aquatic plant belt risk quantification model, swimming fish risk quantification model, power system risk quantification model, power system quantification model, operation depth risk factor quantification model and operation distance risk factor quantification model. Risk value data is quantitatively and accurately generated according to the ten quantitative models, and the risk value data is used as the ten-dimensional input of the original data set for training the underwater robot operation risk factor generative adversarial network, and the artificially extracted risk mechanism prior knowledge is integrated into the data, thereby reducing the difficulty of model training.

[0060] 2. The underwater operation risk assessment method based on generative adversarial network data enhancement disclosed in the present invention combines ten risk factor quantification models with expert evaluation to establish a small-scale data set for underwater robot operation risk assessment, providing raw data for establishing a data-driven underwater robot risk assessment model. Integrating expert knowledge into the raw data solves the problem of the difficulty in obtaining the original supervised data set for underwater robot operation risk assessment, and improves the accuracy and generalization ability of the risk assessment model.

[0061] 3. The underwater operation risk assessment method based on generative adversarial network data enhancement disclosed by the present invention inputs randomly generated noise into a generator based on a fully connected network structure, mixes the real data in the original underwater robot operation risk assessment data set and the data generated by the generator into the discriminator composed of the fully connected network, measures the authenticity of the generated data through the discriminator, takes the average square error as the loss function, and continuously trains to adjust and optimize the generator network parameters. Through continuous iteration and joint optimization of the generator and the discriminator, the generator and the discriminator are trained against each other, and the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original data set is smaller and smaller, until the optimal parameters of the model are solved, the Nash equilibrium state is reached, the quality of the generated data is improved, and the generated data that is more in line with the distribution of the original data is obtained, and a high-quality underwater robot operation risk assessment expanded data set is obtained. The data-driven underwater robot operation risk assessment model trained with the enhanced data set can improve the risk assessment accuracy and efficiency of underwater robots in underwater operation scenarios.

[0062] 4. The underwater operation risk assessment method based on generative adversarial network data enhancement disclosed by the present invention uses the original data in the underwater robot operation risk factor data set to train a fully connected neural network as a verification tool, and inputs the data generated by the generator trained in step 3 into the verification tool neural network, and judges the quality of the generated data by the prediction accuracy. The quality of the data generated by the generative adversarial network is tested, so that the evaluation effect of the underwater robot operation risk assessment model trained with the expanded data set is more accurate and stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the underwater operation risk assessment method based on generative adversarial network data enhancement of the present invention.

[0064] Figure 2 Generator fully connected network structure diagram.

[0065] Figure 3 Discriminator fully connected network structure diagram.

[0066] Figure 4 Generate adversarial network structure diagram.

[0067] Figure 5 Nash equilibrium diagram of the game between the generator and the discriminator.

[0068] Figure 6 Diagram of the fully connected network structure of the risk assessor. DETAILED DESCRIPTION

[0069] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. The technical problems solved by the technical solution of the present invention and the beneficial effects are also described. It should be noted that the described embodiments are only intended to facilitate the understanding of the present invention and do not have any limiting effect on it.

[0070] This example establishes a risk assessment model for underwater robot operations. Figure 1 First, create the initial data set, and then follow the steps in Figure 2 The generator shown and Figure 3 The discriminator fully connected network structure shown in Figure 4 The connection method shown builds the generator and the discriminator into a generative adversarial network and designs a loss function for training. Finally, the generated data and the original data are mixed for training. Figure 5 The evaluator shown is used to establish a risk assessment model.

[0071] like Figure 1 As shown, the underwater operation risk assessment method based on generative adversarial network data enhancement disclosed in this embodiment has the following specific implementation steps:

[0072] Step 1: Analyze the risk factors faced by underwater robot operations. The risk factors of underwater robot operations include ten risk factors: ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. According to the risk factors of underwater robot operations and the risk mechanism of underwater robot operations, ten quantitative models are constructed, including the quantitative model of ocean current risk factor, simplified temperature risk quantitative model, water density risk quantitative model, reef risk factor quantitative model, aquatic plant belt risk quantitative model, swimming fish risk quantitative model, power system risk quantitative model, power system quantitative model, operation depth risk factor quantitative model and operation distance risk factor quantitative model. The data collected in the quantitative model of ocean current risk factors include ocean current speed, angle between the robot's forward direction and the ocean current direction, ambient temperature in Celsius, ambient water density, horizontal width of the reef, distance between the underwater robot and the reef, angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, body length of the swimming fish, swimming speed of the swimming fish, distance between the swimming fish and the underwater robot, horizontal width of the waterweed belt, density of the waterweed belt, distance between the center of the waterweed belt and the center of the underwater robot, angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, battery service time, and battery remaining power percentage , the current depth accounts for the percentage of the rated maximum operating depth and the current mission distance accounts for the percentage of the maximum rated navigation distance. The risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operating depth risk factor and operating distance risk factor obtained after quantification processing by the respective quantitative models of the ten risk factors are used as the ten-dimensional input of the original data set for training the underwater robot operation risk factor generation adversarial network.

[0073] The quantitative model of ocean current risk factors is constructed as shown in formula (1):

[0074] r cur =αv cur |cosθ cur |+βv cur sinθ cur (18)

[0075] Among them, r cur represents the risk value of the ocean current to the underwater robot, α represents the longitudinal threat coefficient of the ocean current to the underwater robot, β represents the lateral threat coefficient of the ocean current to the underwater robot, and v cur is the ocean current velocity, θ cur represents the angle between the ocean current and the underwater robot’s forward direction, θ∈[0,π]. The underwater robot’s ocean current speed v curThe range is v cur ∈[0,10], the risk value r caused by ocean current to underwater robot cur The range is r cur ∈[0,1], according to the underwater robot current speed v cur Risks of range and ocean currents to underwater robots rcur The longitudinal threat coefficient α of the ocean current to the underwater robot and the lateral threat coefficient β of the current to the underwater robot satisfy the constraints shown in the following equations (2) and (3):

[0076]

[0077] α<β (20)

[0078] In combination with the mechanism, constraints and expert advice, α=0.05 and β=0.087 are taken in this embodiment.

[0079] The simplified temperature risk quantification model using formula (4)

[0080]

[0081] Among them, r tmp represents the risk value of the ambient temperature to the underwater robot, r tem ∈[0,1], t is the ambient temperature in Celsius, combined with the actual operation scenario and appropriate expansion, T∈[-30,30].

[0082] The water density risk quantification model shown in formula (5) is used

[0083] r den =493.8(ρ-1.045) 2 (twenty two)

[0084] In the formula, r den is the density risk value, r den ∈[0,1], ρ is the density of seawater, and after limit extension, ρ∈[1.00,1.09].

[0085] The quantitative model of reef risk factors used is:

[0086]

[0087] In the formula, r sto is the reef risk value, θ sto is the angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, θ sto ∈[0,π],k(θ sto ) is the angle risk factor, L sto is the horizontal width of the reef, d stois the distance between the underwater robot and the reef. Considering the actual operation scenario and the measurable range, L sto The value range is 0~10m, d sto The detection range is 0.1 to 5 m. When the value is less than 0.6 and reefs exist, r sto Directly take the value as 1, k(θ sto ) is taken as formula (7)

[0088]

[0089] In the formula, r sto ∈[0,1], when r sto >1, set it to 1.

[0090] The risk quantification model of aquatic weed belts is adopted using formula (8)

[0091]

[0092] Among them, r gra Indicates the risk value of aquatic plants to the underwater robot, L gra represents the horizontal width of the waterweed belt, ρ gra represents the area density of the aquatic weed belt, that is, the number of aquatic weeds per unit area, d gra represents the distance between the center of the waterweed belt and the center of the underwater robot, θ gra is the angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, θ gra ∈[0,π], considering the actual operation scenario and the measurable range, when the risk factor of the length of the waterweed belt is not saturated, L gra The value range is 0~5m, d gra The detection range is 0.1 to 5 m. However, in order to ensure the absolute safety of the underwater robot, when its value is less than 0.6 and there are aquatic plants, r gra The value is 1, gra The value range is 0~20 plants / m 2 , coefficient k(θ gra ) takes a value of

[0093]

[0094] To ensure gra ∈[0,1], when r gra >1, set it to 1.

[0095] The risk quantification model of swimming fish shown in formula (10) is adopted

[0096]

[0097] Among them, r fihis the risk value caused by swimming fish to the underwater robot, and its range is r fih ∈[0,1],L fih is the body length of the swimming fish, and its consideration range is L fih ∈[0.1,1],v fih is the swimming speed of the fish, v fish ∈[0,10], v is the navigation speed of the underwater robot, v∈[0,10], d fih is the distance between the fish and the underwater robot, d fih ∈[0.1,5], to ensure the safety of the underwater robot, when its value is less than 0.6 and there are swimming fish, let r fih =1,θ fih is the angle between the speed direction of the underwater robot and the speed direction of the swimming fish, θ fih ∈[0,π], take k fih =0.1, r fih >1, set it to 1.

[0098] The power system risk quantification model shown in formula (11) is adopted

[0099] r pow =0.1t (28)

[0100] In the formula, r pow The risk value of the underwater robot power system, that is, the risk coefficient of the power system is proportional to the service life of the battery.

[0101] The quantitative model of the power system caused by the decrease in power percentage is defined as

[0102]

[0103] In the formula, r dc is the power system risk caused by the normalized power percentage decrease, p 1 is the percentage of remaining battery power. According to the data and rules expressed in the research experiment, the parameter k is solved according to the exponential decay law of battery capacity. 1 =0.232, k 2 =-3.852, k 3 =-0.0127.

[0104] The total risk value of the power system adds up the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in power. The power system risk is defined as

[0105] r dyn =γr gra +(1-γ)r dc (30)

[0106] Wherein: γ is the weight coefficient of the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in electricity, which is specified as a maximum value between 0 and 1. Combined with expert advice, γ=0.3 is taken in this embodiment.

[0107] Formula (14) is used to establish a quantitative model of operation depth risk factor

[0108]

[0109] In the formula, r dc is the power system risk caused by the normalized power percentage decrease, p 2 is the percentage of the current depth to the rated maximum operating depth. When the depth percentage is 0, r dph =0, when the depth percentage is 0.8, r dph =0.2, when the depth percentage is 1, r dph =1, solve the parameters based on the above three percentage data to get k 4 =0.032, k 5 =8.047, k 6 =0.

[0110] The operation distance risk quantification model used is:

[0111] r dist =p 3 (32)

[0112] In the formula, p 3 is the percentage of the current mission distance to the maximum rated navigation distance, r dist is the risk value of the underwater robot’s operating distance.

[0113] After quantification according to the ten risk factor quantification models, we obtain the risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. The risk value data are used as the ten-dimensional input of the original underwater robot operation risk assessment data set for training the underwater robot operation risk factor generative adversarial network.

[0114] Step 2: Solicit a predetermined number of experts to evaluate the risk levels corresponding to the input data of the original data set generated after the quantification in step 1, and divide the evaluation levels into low risk, relatively low risk, relatively high risk and high risk, and encode the risk levels as 1, 2, 3, and 4 respectively. Take the mode of the evaluation results as the risk level corresponding to the corresponding risk vector to form the output of the original data set, and complete the construction of a small-scale original underwater robot operation risk assessment data set. The input dimension of the underwater robot operation risk assessment data set is ten dimensions, and the output dimension is one dimension, and the one-dimensional output is the risk level. The data distribution of the risk assessment data set of the underwater robot formed in this embodiment is shown in the following table

[0115]

[0116] Step 3: Randomly generate noise input based on the generator of the fully connected network structure. The output of the generator has the same dimension as the real data in the original underwater robot operation risk assessment dataset. The real data in the original underwater robot operation risk assessment dataset and the data generated by the generator are mixed and input as follows: Figure 2 The discriminator composed of the fully connected network shown in Figure 3 The discriminator of the fully connected network structure shown in the figure measures the authenticity of the generated data, uses the mean square error as the loss function, and continuously trains to adjust and optimize the generator network parameters. Figure 4 The generative adversarial network structure shown in the figure jointly optimizes the generator and the discriminator through continuous iterations, and trains the generator and the discriminator against each other. The weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original data set becomes smaller and smaller, until the optimal parameters of the generator model that meet the preset accuracy requirements are solved, the Nash equilibrium state is reached, the quality of the generated data is improved, and the generated data that is more in line with the distribution of the original data is obtained, thereby improving the risk assessment accuracy and efficiency of the underwater robot in the underwater operation scenario.

[0117] In order to obtain a risk assessment model with higher accuracy, a data enhancement generative adversarial network for underwater robot operation risk factors is constructed, and the data set is expanded using the generative adversarial network. The generator adopts a fully connected network structure, whose input is a randomly generated noise vector, and the output is the generated data with the same dimension as the input data in the original underwater robot operation risk assessment data set. The loss function uses formula (16)

[0118]

[0119] In the formula, m is the number of data in a noise data batch, z iis the random noise input to the generator G. This formula directly takes the average value of the discrimination results of the discriminator D on the data generated by the generator G as the loss function of the generator G.

[0120] The discriminator D also uses a fully connected network structure. The input data is the real input in the original underwater robot operation risk assessment data set. The output result is a continuous value between 0 and 1. The discriminator output result is 1 when the real data is input, and the discriminator output result is 0 when the generated data is input. The loss function of the discriminator is:

[0121]

[0122] Where n is the number of data in a mixed data set of a single batch of real data and data generated by the generator, and y i For x i The corresponding expected value is when x i To generate data y i =0, when x i When y is real data i =1.

[0123] By continuously iterating and jointly optimizing the generator and the discriminator, and training the generator and the discriminator against each other, the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original data set is made smaller and smaller, until the optimal parameters of the generator model that meet the preset accuracy requirements are solved, the Nash equilibrium state is reached, and a high-quality data generator is obtained, thereby expanding the underwater robot operation risk factor data set in accordance with the original data distribution, and improving the accuracy and efficiency of risk identification in the underwater robot operation scene. The game process of the generator and the adversary in this embodiment is as follows: Figure 5 As shown, a better Nash equilibrium is achieved.

[0124] Step 4: Training with the original data from the underwater robot operation risk factor dataset Figure 6 The fully connected neural network of the structure shown is used as a verification tool for verifying the quality of the data generated by the generator. The data generated by the generator trained in step 3 is input into the verification tool neural network, and the quality of the generated data is judged by the prediction accuracy of the verification tool neural network. In this embodiment, after the verification is passed, a predetermined proportion of the original data is divided to verify the quality of the data generated by the generator. The division proportion is shown in the following table

[0125]

[0126] The original data in the underwater robot operation risk factor dataset and the newly generated data are mixed to train a fully connected network as the final underwater robot operation risk assessor, and the risk level assessment accuracy reaches 89%. In this embodiment, the risk of the underwater robot in the underwater operation scene is assessed according to the underwater robot operation risk assessor, thereby improving the accuracy and efficiency of risk identification in the underwater robot underwater operation scene.

[0127] The present invention provides a risk data expansion method based on a generative adversarial network. The method proposes a risk quantification and evaluation function by analyzing the action mechanisms of ten risk factors, namely, ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. After evaluation by a predetermined number of experts, small-scale initial data for underwater robot operation risk evaluation is formed. Then, the generative adversarial network principle is adopted to expand the data under the premise of ensuring the quality of the generated data to obtain a larger-scale risk data set for training a risk evaluation network model, thereby achieving a more accurate risk prediction effect and improving the risk identification accuracy and efficiency in the underwater operation scene of the underwater robot.

[0128] The specific description above further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An underwater operation risk assessment method based on generative adversarial network data enhancement, characterized by: The following steps are included: Step 1: Analyze the risk factors faced by underwater robot operations. The risk factors of underwater robot operations include ten risk factors: ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor; According to the risk factors and risk mechanisms of underwater robot operations, ten quantitative models are constructed, including ocean current risk factor quantitative model, simplified temperature risk quantitative model, water density risk quantitative model, reef risk factor quantitative model, aquatic plant belt risk quantitative model, fish risk quantitative model, power system risk quantitative model, power system quantitative model, operation depth risk factor quantitative model and operation distance risk factor quantitative model; the data collected in the ocean current risk factor quantitative model include ocean current speed, angle between the robot's forward direction and the ocean current direction, ambient temperature in Celsius, ambient water density, reef width, distance between the underwater robot and the reef, angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, body length of fish, swimming speed of fish, and the distance between fish and underwater robot. The distance between robots, the width of the waterweed belt, the density of the waterweed belt, the distance between the center of the waterweed belt and the center of the underwater robot, the angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, the battery service time, the percentage of remaining battery power, the current depth as a percentage of the rated maximum operating depth, and the current mission distance as a percentage of the maximum rated navigation distance. The risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, waterweed risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operating depth risk factor and operating distance risk factor obtained after the collected data are quantified by ten quantitative models are used as the ten-dimensional input of the original data set for training the underwater robot operation risk factor generation adversarial network; Step 2: Solicit a predetermined number of experts to evaluate the risk levels corresponding to the input data of the original data set generated after the quantification in step 1, and divide the evaluation levels into low risk, relatively low risk, relatively high risk and high risk, and encode the risk levels as 1, 2, 3 and 4 respectively. Take the mode of the evaluation results as the risk level corresponding to the corresponding risk vector to form the output of the original data set, and complete the construction of the small-scale original underwater robot operation risk assessment data set; the input dimension of the underwater robot operation risk assessment data set is ten dimensions, and the output dimension is one dimension, and the one-dimensional output is the risk level; Step 3: Randomly generate noise input based on the generator of the fully connected network structure. The output of the generator has the same dimension as the real data in the original underwater robot operation risk assessment dataset. The real data in the original underwater robot operation risk assessment dataset and the data generated by the generator are mixed and input into the discriminator composed of the fully connected network. The discriminator measures the authenticity of the generated data, and the mean square error is used as the loss function. The network parameters of the generator are adjusted and optimized through continuous training; the generator and the discriminator are jointly optimized through continuous iteration, and the generator and the discriminator are trained against each other. According to the loss function, the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted, so that the difference between the data generated by the generator and the real data in the original dataset is smaller and smaller, until the optimal parameters that meet the preset accuracy requirements of the generator model are solved, and the Nash equilibrium state is reached, the quality of the generated data is improved, and the generated data that is more in line with the distribution of the original data is obtained, thereby improving the risk assessment accuracy and efficiency of the underwater robot in the underwater operation scene; Step 4: Use the original data in the underwater robot operation risk factor dataset to train a fully connected neural network, and use the fully connected neural network as a verification tool to verify the quality of the data generated by the generator. The data generated by the generator trained in step 3 is input into the verification tool neural network, and the quality of the generated data is judged by the prediction accuracy of the verification tool neural network. After the verification, a predetermined proportion of the original data is allocated to verify the quality of the generated data. The original data in the underwater robot operation risk factor dataset and the newly generated data are mixed to train the fully connected network as the final underwater robot operation risk assessor. The risk of the underwater robot in the underwater operation scenario is assessed according to the underwater robot operation risk assessor, so as to improve the accuracy and efficiency of risk identification in the underwater operation scenario of the underwater robot.

2. The underwater operation risk assessment method based on generative adversarial network data enhancement according to claim 1 is characterized in that: In step one, The quantitative model of ocean current risk factors is constructed as shown in formula (1): r cur =αv cur |cosθ cur |+βv cur sinθ cur (1) Among them, r cur represents the risk value of the ocean current to the underwater robot, α represents the longitudinal threat coefficient of the ocean current to the underwater robot, β represents the lateral threat coefficient of the ocean current to the underwater robot, and v cur is the ocean current velocity, θ cur represents the angle between the ocean current and the underwater robot's forward direction, θ∈[0,π]; the underwater robot's ocean current speed v cur The range is v cur ∈[0,10], the risk value r caused by ocean current to underwater robot cur The range is r cur ∈[0,1], according to the underwater robot current speed v cur The risk value r of range and ocean current to underwater robots cur The longitudinal threat coefficient α of the ocean current to the underwater robot and the lateral threat coefficient β of the current to the underwater robot satisfy the constraints shown in the following equations (2) and (3): α＜β (3) The simplified temperature risk quantification model using formula (4) Among them, r tmp represents the risk value of the ambient temperature to the underwater robot, r tem ∈[0,1], t is the ambient temperature in Celsius, combined with the actual operation scenario and appropriately expanded, T∈[-30,30]; The water density risk quantification model shown in formula (5) is used r den =493.8(ρ-1.045) 2 (5) In the formula, r den is the density risk value, r den ∈[0,1], ρ is the density of seawater, and after limit extension, ρ∈[1.00,1.09]; The quantitative model of reef risk factors used is: In the formula, r sto is the reef risk value, θ sto is the angle between the line connecting the reef and the center of the underwater robot and the forward direction of the underwater robot, θ sto ∈[0,π],k(θ sto ) is the angle risk factor, L sto is the horizontal width of the reef, d sto is the distance between the underwater robot and the reef; considering the actual operation scenario and the measurable range, L sto The value range is 0~10m, d sto The detection range is 0.1 to 5 m. When the value is less than 0.6 and reefs exist, r sto Directly take the value as 1, k(θ sto ) is taken as formula (7) In the formula, r sto ∈[0,1], when r sto >1, set it to 1; The risk quantification model of aquatic weed belts is adopted using formula (8) Among them, r gra Indicates the risk value of aquatic plants to the underwater robot, L gra represents the horizontal width of the waterweed belt, ρ gra represents the area density of the aquatic weed belt, that is, the number of aquatic weeds per unit area, d gra represents the distance between the center of the waterweed belt and the center of the underwater robot, θ gra is the angle between the line connecting the center of the waterweed belt and the center of the underwater robot and the forward direction of the underwater robot, θ gra ∈[0,π], considering the actual operation scenario and the measurable range, when the risk factor of the length of the waterweed belt is not saturated, L gra The value range is 0~5m, d gra The detection range is 0.1 to 5 m. However, in order to ensure the absolute safety of the underwater robot, when its value is less than 0.6 and there are aquatic plants, r gra The value is 1, gra The value range is 0~20 plants / m 2 , coefficient k(θ gra ) takes a value of To ensure gra ∈[0,1], when r gra >1, set it to 1; The risk quantification model of swimming fish shown in formula (10) is adopted Among them, r fih is the risk value caused by swimming fish to the underwater robot, and its range is r fih ∈[0,1],L fih is the body length of the swimming fish, and its consideration range is L fih ∈[0.1,1],v fih is the swimming speed of the fish, v fish ∈[0,10], v is the navigation speed of the underwater robot, v∈[0,10], d fih is the distance between the fish and the underwater robot, d fih ∈[0.1,5], to ensure the safety of the underwater robot, when its value is less than 0.6 and there are swimming fish, let r fih =1,θ fih is the angle between the speed direction of the underwater robot and the speed direction of the swimming fish, θ fih ∈[0,π], take k fih =0.1, r fih >1, set it to 1; The power system risk quantification model shown in formula (11) is adopted r pow ＝0.1t (11) In the formula, r pow The risk value of the underwater robot power system, that is, the risk factor of the power system is proportional to the battery service time; The power system quantitative model caused by the decrease in power percentage is defined as In the formula, r dc is the power system risk caused by the decrease in the normalized power percentage, and p1 is the remaining battery power percentage; according to the data and rules described in the research experiment, the parameters k1, k2, and k3 are solved according to the exponential decay law of battery capacity; The total risk value of the power system adds up the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in power. The power system risk is defined as r dyn =γr gra +(1-γ)r dc (13) Where: γ is the weight coefficient of the environmental risk caused by aquatic plants and the risk caused by the percentage decrease in electricity, which is specified as a maximum value between 0 and 1; Formula (14) is used to establish a quantitative model of operation depth risk factor In the formula, r dc is the power system risk caused by the normalized power percentage decrease, p2 is the percentage of the current depth to the rated maximum operating depth; when the depth percentage is 0, r dph =0, when the depth percentage is 0.8, r dph =0.2, when the depth percentage is 1, r dph =1, solve the parameters k4, k5, k6 according to the above three percentage data; The operation distance risk quantification model used is: r dist ＝p3 (15) Where p3 is the percentage of the current mission distance to the maximum rated navigation distance, r dist is the risk value of the underwater robot operation distance; After quantification according to the ten risk factor quantification models, we obtain the risk value data of ocean current risk factor, temperature risk factor, water density risk factor, reef risk factor, aquatic plant risk factor, swimming fish risk factor, power system risk factor, power system risk factor, operation depth risk factor and operation distance risk factor. The risk value data are used as the ten-dimensional input of the original underwater robot operation risk assessment data set for training the underwater robot operation risk factor generative adversarial network.

3. The underwater operation risk assessment method based on generative adversarial network data enhancement according to claim 2 is characterized in that: The implementation method of step three is: In order to obtain a risk assessment model with higher accuracy, a data enhancement generative adversarial network for underwater robot operation risk factors is constructed, and the data set is expanded using the generative adversarial network. The generator adopts a fully connected network structure, whose input is a randomly generated noise vector, and the output is the generated data with the same dimension as the input data in the original underwater robot operation risk assessment data set. The loss function uses formula (16) In the formula, m is the number of data in a noise data batch, z i is the random noise input to the generator G. This formula directly takes the average value of the discriminator D’s discrimination results on the data generated by the generator G as the loss function of the generator G; The discriminator D also uses a fully connected network structure. The input data is the real input in the original underwater robot operation risk assessment data set. The output result is a continuous value between 0 and 1. The discriminator output result is 1 when the real data is input, and the discriminator output result is 0 when the generated data is input. The loss function of the discriminator is: Where n is the number of data in a mixed data set of a single batch of real data and data generated by the generator, and y i For x i The corresponding expected value is when x i To generate data y i =0, when x i When y is real data i =1; By continuously iteratively jointly optimizing the generator and the discriminator, and training the generator and the discriminator against each other, the weight parameters of the adversarial network composed of the generator and the discriminator are continuously optimized and adjusted according to the loss function, so that the difference between the data generated by the generator and the real data in the original data set becomes smaller and smaller, until the optimal parameters that meet the preset accuracy requirements of the generator model are solved, the Nash equilibrium state is reached, and a high-quality data generator is obtained, thereby expanding the underwater robot operation risk factor data set to conform to the original data distribution, thereby improving the accuracy and efficiency of risk identification in underwater robot operation scenarios.

Images