A method for identifying and fault-tolerant control of power damage state of a twin-propeller driven unmanned surface vehicle

Abstract

The application provides a kind of double propeller driven unmanned ship power damage state identification and fault-tolerant control method.The method is first based on the motion effect of measured data to obtain the navigation characteristics of unmanned ship, which consists of normal state navigation and damage state navigation data to form training and test set;Then record the steering behavior and motion response during navigation, compare the training characteristics, and establish a power damage state recognition algorithm through a deep neural network;Finally, based on the damage degree recognition algorithm and the control end adaptive software and hardware system, the propeller command setting method is determined to control the goal of reducing speed and keeping direction, so that the unmanned ship can still maintain a certain autonomous navigation ability under the damage state.This scheme is convenient to solve the problem that the double propeller driven unmanned ship has low response ability to power system failure, so that it can still maintain a certain degree of autonomous navigation ability under abnormal power state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned ship navigation control, in particular, and particularly relates to a dual-propeller-driven unmanned ship power damage state identification and fault-tolerant control method. BACKGROUND

[0002] The unmanned ship has the characteristics of high cost performance, strong seaworthiness, wide combat use and high intelligence, and is an important new combat equipment in future intelligent war. Compared with manned ships, surface unmanned ships have significant advantages in tracking and striking, anti-submarine warfare, environmental monitoring, maritime transportation and other scenarios. They do not need to consider the design limitations of manned design, so they can promote the upgrading of ship functions, enhance the task execution capability and improve the task execution efficiency. In order to fully exert the advantages, in recent years, new type of power propulsion system suitable for unmanned concept and adaptive control method have gradually become a new discussion hotspot. For unmanned ships, the design process does not need to consider the influence of main engine power equipment layout on temperature and noise of the crew, so the ship has a more free layout range and design upper limit, and it is more likely to break through the performance limit of the traditional power propulsion system.

[0003] However, the reliability and stability of the power system are always the key technical challenges in the development process of surface ship unmanned. The power system failure is not a special problem in the development process of unmanned ship, but also exists in manned ship, but for unmanned ship, the importance of power system failure is higher and the challenge is greater. First of all, autonomy and strong survivability are important indicators of unmanned ship. Once the power failure occurs, it will seriously affect the autonomous navigation performance and task execution capability of the unmanned ship, and may also threaten the navigation safety and cause unnecessary loss. Secondly, the unmanned ship is far away from human intervention for a long time, and cannot rely on direct intervention of personnel to solve the problem. The maintenance response is not timely, and the characteristic signal of power anomaly cannot be fed back in time, so the self-sustaining ability and reliability of the power system are required to be higher. In addition, one of the advantages of unmanned ship is that it can perform tasks in special working conditions and extreme environments. In this case, the ship is more likely to have various failures or abnormal states due to environmental factors.

[0004] In summary, power damage identification and fault-tolerant control are the key challenges of unmanned ship. The strong autonomous navigation capability of unmanned ship requires the ability of power system anomaly identification and response as the foundation. Therefore, how to accurately identify the power abnormal state of unmanned ship and take effective fault-tolerant control measures has become a key technology to ensure the reliable operation of unmanned ship. SUMMARY

[0005] The purpose of the present application is to provide a dual-propeller driven unmanned ship power damage state recognition and fault-tolerant control method, to construct a steering motion mapping for an unmanned ship driven by a stern differential dual propeller, to perform power damage degree recognition based on a deep neural network, to design a control rate in accordance with the damage degree recognition result, to perform fault-tolerant control with the goal of reducing speed and preserving direction, and to solve the problem of abnormal power system.

[0006] The specific technical solution of the present application is to provide a dual-propeller driven unmanned ship power damage state recognition and fault-tolerant control method, which comprises the following steps:

[0007] Step 1: Based on the measured data, a mapping of unmanned ship steering and motion characteristics is constructed, and based on the outlier detection algorithm of navigation data, the data validity is ensured.

[0008] Step 2: A power damage state recognition algorithm is established by a deep neural network to qualitatively analyze the power damage degree of the unmanned ship.

[0009] Step 3: Based on the damage degree recognition algorithm, the propeller command setting method is determined.

[0010] Further, in step 1, the following steps are further included:

[0011] Step 11: A mapping of unmanned ship steering and motion characteristics is constructed.

[0012] Step 12: Based on the outlier detection algorithm of navigation data, the data validity is ensured.

[0013] In step 12, the following is further included:

[0014] Step 121: Screening based on Z-score.

[0015] Step 122: Further screening based on local anomaly factor.

[0016] Further, in step 11, the steering motion mapping relationship is:

[0017] a→P

[0018]

[0019] a is the current acceleration, P is the propeller power, F is the three-degree-of-freedom force and torque matrix, m is the mass of the unmanned ship, and v is the speed of the unmanned ship.

[0020] Based on the dual-propeller differential drive unmanned ship, the decomposition mapping of acceleration in the body inertia system can be written as:

[0021] P2:(a x ,a y )→(Tl T r )

[0022] wherein a x ,a y are the acceleration of the unmanned surface vehicle perpendicular to the heading direction and parallel to the heading direction respectively, T l ,T r are the left and right propeller throttle respectively.

[0023] Further, in step 121, the Z-score is expressed as:

[0024]

[0025] The expected value of any variable X is E(X), and the standard deviation is σ(X);

[0026] The sample size is selected as n, and the upper limit of the absolute value of the Z-score of a single data point is selected as σ, that is, if a data point does not constitute a significant difference with other values within t seconds before and after it, it is recorded as valid data.

[0027] Further, in step 122, based on the local anomaly factor, the screening is performed again as:

[0028] Given the parameter p, the p-reachable distance d k (p,o) of the data point p to the data point o is expressed as:

[0029] d k (p,o) = max{d k (o),d(p,o)}

[0030] o is the nearest point to p, d k (o) is the distance between o and the kth nearest point to o, and d(p,o) is the distance between p and o;

[0031] The local reachable density of the data point p is defined as:

[0032]

[0033] N k (p) is the k-nearest neighbor of p composed of all points with a distance less than d k (p) from p;

[0034] The local anomaly factor LOF of the data point p is the ratio of the average local reachable density of the points in the neighborhood of the point p to the local reachable density of the data point p:

[0035]

[0036] LOF k (p) represents the local anomaly factor of the point p neighborhood N k(p) the average of the ratio of the local reachable density of other points to the local reachable density of point p;

[0037] If the ratio is closer to 1, it means that the density of the neighborhood points of point p is similar, and point p may belong to the same cluster as the neighborhood; if the ratio is less than 1, it means that the density of point p is higher than that of its neighborhood points, and point p is a dense point; if the ratio is greater than 1, it means that the density of point p is less than that of its neighborhood points, and point p may be an abnormal point.

[0038] Further, in step 2, the following steps are further included:

[0039] Step 21, define the damage degree;

[0040] The absolute damage degree of the propeller is defined as:

[0041]

[0042] Where, Δ X is the absolute damage degree of the propeller with high power loss, Δ Y is the absolute damage degree of the propeller with low power loss, Δ X > Δ Y ; the control end throttle of the propeller with high power loss is T X , the control end throttle of the other propeller is T Y , and the equivalent propulsion throttle of the actual movement of the two propellers is

[0043] The damage degree is defined as:

[0044] Δ = Δ X - Δ Y

[0045] Δ is the damage degree;

[0046] Step 22, establish a power damage state recognition through a deep neural network.

[0047] Further, in step 22, the following steps are further included:

[0048] Step 221: filter the speed, acceleration, left and right throttle, and time and effective navigation information corresponding to each group of data to form input feature values;

[0049] Step 222: perform activation function transformation on the hidden variables as the input of the next fully connected layer; the expression is:

[0050] ReLU(x) = max(x, 0)

[0051] ReLU(x) is a ReLU function, and x is an input feature value of the activation function, which is effective navigation information after filtering;

[0052] Step 223: The fully connected layer is composed of multiple hidden layers with extended widths, and the hidden layer output is transformed by an activation function;

[0053] Step 224: The output layer uses the Huber function as the loss function for the regression problem. The Huber error is defined as follows:

[0054]

[0055] Among them, y is the true value, f(x) is the predicted value, and δ is a hyperparameter used to control the turning point of the error.

[0056] Furthermore, in step 3, the following steps are also included:

[0057] Step 31: Calculate the initial power reduction rate;

[0058] The difference between the calculated throttle value and the true value is calculated. When the damage Δ is 0, that is, when the ship's propulsion system is normal, the difference should be a curve close to 0. The damage time interval is [t0,t n ], n+1 groups of data are recorded in the interval, namely:

[0059]

[0060] Indicates t n Acceleration information recorded at all times, {a} represents the acceleration time series of the unmanned boat;

[0061] A power damage state recognition algorithm is established based on a deep neural network. The heading and vertical acceleration are input and the output is the equivalent propulsion throttle:

[0062]

[0063] Indicates the equivalent throttle of the left and right propellers calculated from the nth set of recorded acceleration information. Represents the time series of the equivalent propulsion throttle of the unmanned boat propeller;

[0064] In the time interval [t0,t n ], calculate the cumulative deviation of the boost throttle:

[0065]

[0066] T δX (n) represents the cumulative throttle deviation of the X propeller with the higher power loss in the nth group of recorded information, T δY (n) represents the cumulative throttle deviation of the Y propeller with the lower power loss in the nth group of recorded information. represents the X propeller equivalent throttle calculated by the i-th set of recorded acceleration information, T iX represents the X propeller control end throttle recorded at the time of the i-th set of information, represents the Y propeller equivalent throttle calculated by the i-th set of recorded acceleration information, T iY represents the Y propeller control end throttle recorded at the time of the i-th set of information;

[0067] and the absolute damage degree of the left and right propellers at t n is calculated:

[0068]

[0069] Δ X (n) represents the X propeller absolute damage degree calculated by the n-th set of recorded information, Δ Y (n) represents the Y propeller absolute damage degree calculated by the n-th set of recorded information;

[0070] then the warship damage degree at t n is:

[0071] Δ(n) = Δ X (n) - Δ Y (n)

[0072] Step 32: multiple iterations are performed until the relative damage degree of the motion effect power after adjustment is less than a preset value as the termination condition.

[0073] The beneficial effects achieved by the present application are:

[0074] The unmanned warship power damage state recognition and fault-tolerant control method provided by the present application can identify problems and take corresponding measures to maximize the stability and safety of the warship in the case of power damage or failure of the unmanned warship, without the need for additional installation of device systems, and can be migrated to most unmanned warships driven by double propellers;

[0075] The unmanned warship damage state recognition provided by the present application is realized based on the measured data during navigation, which not only saves the complex fault diagnosis process, but also avoids the problem of difficult attribution, and the motion effect is used to identify the operating behavior, which is simple, reliable and low in cost. The identification is based on the training of the neural network in the early test stage, and the output of the navigation data is more immediate, and the error correction adjustment is fast;

[0076] After the method provided by the present application identifies that the power of the propeller end is abnormal, a fault-tolerant control method suitable for power damage can be designed to adjust the operating instructions of other propellers, so as to realize the navigation target of reducing speed and keeping direction, so that the unmanned warship can still save certain autonomous navigation and task execution ability in the case of power abnormality, which is highly consistent with the navigation demand. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 is a flowchart of a dual-propeller driven unmanned ship power damage state recognition and fault-tolerant control method according to an embodiment of the present application;

[0078] Figure 2 is a calculation chart for verifying the recognition accuracy of known damage degrees through motion effects according to an embodiment of the present application;

[0079] Figure 3 is a schematic diagram of tracking control effects after recognizing unknown damage degrees according to an embodiment of the present application. DETAILED DESCRIPTION

[0080] The technical solutions of the present application will be further described below with reference to the accompanying drawings.

[0081] Embodiment One

[0082] As shown in the drawings, the present embodiment provides a dual-propeller driven unmanned ship power damage state recognition and fault-tolerant control method. Figure 1

[0083] The dual-propeller driven unmanned ship power damage state recognition and fault-tolerant control method comprises the following steps:

[0084] Step 1, based on the measured data, construct the unmanned ship control and motion feature mapping, determine the data related to the mapping model and its acquisition and setting method. Develop an outlier detection algorithm based on navigation data to ensure data validity;

[0085] In step 1, the following steps are further included:

[0086] Step 11, construct the unmanned ship control and motion feature mapping;

[0087] The target control motion mapping relationship is set as:

[0088] According to Newton's second law, the dynamic process of the three-degree-of-freedom ship control motion model can be expressed as F = Ma, where F is the three-degree-of-freedom force and moment matrix, M is the three-degree-of-freedom mass and inertia matrix, and a is the current acceleration.

[0089] The narrow-sense propulsion force can be represented by the propulsion power and the ship speed as:

[0090]

[0091] Where the propulsion power P can be regarded as a physical quantity linearly related to the throttle, so that the relationship between the control behavior and the motion response can be constructed as:

[0092] ​

[0093] Thus, the mapping to be constructed is:

[0094] P1: a→P

[0095] For a twin-propeller differential drive unmanned surface vehicle, the decomposition of acceleration in the body inertial frame is investigated, and the above mapping can be written as:

[0096] P2: (a x ,a y )→(T l ,T r )

[0097] where T l ,T r are the left and right propeller throttle, respectively. In practical applications, the motion data obtained during navigation are discrete points.

[0098] Step 12, outlier detection algorithm based on navigation data, to ensure data validity;

[0099] The significance of the outlier detection algorithm is set as:

[0100] GPS and inertial navigation system may be subject to various disturbances in practical applications, resulting in unavailability of perception data, including obstructions, multipath effects, weather conditions, electromagnetic interference, reading drift, etc. For the navigation perception system of an unmanned surface vehicle, these disturbances can cause some data errors or unavailability, and short-term unavailability does not affect the navigation and task execution of the vehicle. However, for the data set of power damage state recognition, the errors in perception data will cause the training set to be contaminated, leading to degraded training results. Due to the large amount of data, manual screening is not realistic, so it is necessary to develop an effective navigation information screening algorithm. The goal of screening is to filter, remove outliers, and correct errors in perception data using appropriate algorithms and techniques to improve data quality and availability. Based on the experience of real ship testing, different types of errors caused by various disturbances are different, for example, weather conditions may cause the GPS update rate to decrease, and consecutive data values (latitude, longitude, speed, heading angle, etc.) may be the same. Electromagnetic interference may cause data to jump, and a number of discrete data points are significantly different from their neighbors.

[0101] To solve the above problems, an outlier detection algorithm based on navigation data can be used, which can be divided into the following steps:

[0102] Step 121: Z-score screening. For many types of failed navigation information, the Z-score of a certain value in the sample composed of its nearby data can be used to perform coarse-precision failed value screening. Assuming that the expected value of any variable X is E(X) and the standard deviation is σ(X), the Z-score is represented as

[0103]

[0104] The sample capacity is selected as n, and the upper limit of the absolute value of the score of a single data point Z is selected as σ, that is, if each data point does not form a significant difference with other values within t seconds before and after it, it is recorded as valid data.

[0105] Step 122: Local anomaly factor screening. The Z-score method can only achieve sample point screening under coarse accuracy, and has certain limitations in the establishment of a dynamic damage state recognition data set. Since the unmanned ship propulsion signal is a step signal, the propeller speed can be quickly changed, resulting in a step change in the dispersion of adjacent samples. When a smaller upper limit σ of the absolute value of the Z-score is selected, such a step change may be considered as an abnormal point and be removed, and when a larger upper limit σ of the absolute value of the Z-score is selected, a jump value close to the neighborhood value may be considered as a normal step and be retained, thereby polluting the data set. Therefore, the data set screened by the Z-score coarse accuracy is further screened based on the local anomaly factor in this step.

[0106] The local anomaly factor (Local Outlier Factor, LOF) is a method for describing abnormal values, which is commonly used to identify outliers that do not conform to the trend in a large data set. The LOF algorithm measures the abnormality of a point by investigating the relative density of the point and its neighboring data points, allowing for uneven data distribution and different densities. Therefore, when facing the navigation data anomaly problem in this part, the screening error caused by the step change of the data can be avoided. The method flow is as follows:

[0107] In the same sample set, the point p under study is denoted as o, and the distance between p and the kth nearest point to p is denoted as d k (p).

[0108] All points with a distance p less than d k (p) are defined as the k-nearest neighbors of p, denoted as N k (p).

[0109] When a parameter p is given, the p-reachable distance from data point p to data point o is represented as

[0110] d k (p,o) = max{d k (o),d(p,o)}

[0111] The limitation of the reachable distance significantly reduces the statistical fluctuations of all data points p close to o. The strength of this smoothing effect can be controlled by the parameter k. The higher the value of k, the more similar the reachable distances of objects in the same neighborhood, and the smoother the result.

[0112] Then, the local reachable density of data point p is defined as

[0113]

[0114] The local reachable density is based on the inverse of the average reachable distance of the k-nearest neighbors of p, the greater the distance, the smaller the density, that is, the average distance of the k-nearest neighbors of the observation point to the observation point, and the sample size of the dynamic sample is set to n. According to the definition of the local reachable density, if the observation data point is relatively discrete relative to other data points, the local reachable density is smaller, and the local relative density (local outlier factor LOF) of the data point p is the ratio of the average local reachable density of the points in the neighborhood of the point p to the local reachable density of the data point p, that is:

[0115]

[0116] The local outlier factor represents the average of the ratio of the local reachable density of other points in the neighborhood N k (p) of point p to the local reachable density of point p. If the ratio is closer to 1, it means that the density of the neighborhood points of point p is similar, and point p may belong to the same cluster as the neighborhood; if the ratio is less than 1, it means that the density of point p is higher than that of its neighborhood points, and point p is a dense point; if the ratio is greater than 1, it means that the density of point p is less than that of its neighborhood points, and point p may be an abnormal point. Considering that the speed and acceleration information may fluctuate when the working condition changes, and combining the test effect, γ is selected as the upper limit of the local outlier factor LOF, and when it exceeds the upper limit, it is regarded as receiving an anomaly, and the interpolation of the neighborhood data in the same group is used as the information at that time.

[0117] In addition, the LOF algorithm can also solve the problem of multiple data being the same or close due to receiving delay, signal interruption, etc. If there are more than or equal to k repeated or extremely close points in the sample, the average reachable distance of these points is close to zero, and the local reachable density tends to infinity. When the local reachable density p k of the sample set is greater than the set threshold β, it can be judged that a data error caused by signal interference has occurred, and an error signal should be transmitted to the control end.

[0118] Step 2, the concept of damage degree is proposed, and a dynamic power damage state recognition algorithm is established through a deep neural network to qualitatively analyze the damage degree of the unmanned ship;

[0119] In step 2, the following steps are further included:

[0120] Step 21, define the damage degree;

[0121] The definition and calculation method of the damage degree are set as:

[0122] Without loss of generality, assume that in a double-screw differential propulsion unmanned ship, let the control end throttle of the propeller with higher power loss be TX , the other side propeller control end throttle is T Y , the actual movement of the two propellers is equivalent to the propeller The absolute damage degree of the propeller is defined as

[0123]

[0124] Where, Δ X >Δ Y .

[0125] The damage degree Δ is defined as

[0126] Δ=Δ X -Δ Y

[0127] The damage degree here is a relative value, that is, it is possible that both sides of the propeller are damaged. This setting is because the control effect is the design goal of the recognition algorithm, and the output power of the two propellers is not quantitatively analyzed, but the relationship between the relative propelling power is analyzed, in order to realize the control effect of speed reduction and direction preservation. Therefore, only the calculation method of the equivalent propelling power of the actual movement of the two propellers can be used to calculate the damage degree of the propeller. The calculation method is essentially equivalent to the mapping P2, but due to the floating of micro data points, the overall trend quantitative calculation method in a period of time needs to be constructed, that is, the mapping is established:

[0128]

[0129] Step 22, the power damage state recognition is established by a deep neural network;

[0130] Further, in step 2, the neural network studying the mapping relationship between the control behavior and the movement effect is composed of feature input, activation function, batch normalization, full connection layer and regression function, and is specifically set as:

[0131] Step 221: feature input. The feature input layer is the data collection part of the neural network, and the input content here includes speed, acceleration, left and right throttle and time corresponding to each group of data, which constitutes the input feature value after the data setting and effective navigation information filtering described in the foregoing. Among them, the acceleration is in units of gravity acceleration g, the propelling throttle is a positive integer in [20, 100], indicating the current throttle ratio to the limit throttle.

[0132] Step 222: activation function. The hidden variable is transformed by the activation function as the input of the next full connection layer. In the training of the deep neural network, the ReLU function is used as the activation function, and its expression is

[0133] ReLU(x)=max(x,0)

[0134] Step 223: full connection layer. The full connection layer is composed of multiple expanded width hidden layers, and the hidden layer output is transformed by an activation function. The hidden layer extracts more meaningful and abstract features from the original input data by learning weight and bias parameters. By increasing the number of hidden layers and the number of neurons, the expression ability of the neural network can be increased, while preserving important information and removing redundancy and noise. The Adam algorithm is used to update the weight parameters of the neural network. The Adam algorithm combines the ideas of momentum method and adaptive learning rate, and considers the first and second moment estimates of the gradient to calculate the update step.

[0135] Step 224: output layer. The output layer uses the Huber function as the loss function of the regression problem, which is a compromise between squared loss and absolute value loss within a certain range. The definition of Huber error is as follows:

[0136]

[0137] Where y is the true value, f(x) is the predicted value, and δ is a hyperparameter that controls the turning point of the error.

[0138] The input data of the neural network is the motion response, and the output is the steering behavior. The training set contains input and output information, and the validation set gives the input data to detect whether the output effect is consistent with the true steering throttle. For any i, input the motion effect data in a period of time, such as acceleration information {(a ix ,a iy )}, and obtain the steering behavior data, which can achieve the goal of damage degree recognition algorithm. The actual construction of the neural network proposed in the present application is

[0139]

[0140] Through the above steps, the scheme proposed in the present application can calculate the relative damage degree of the propeller of the double-paddle driven unmanned ship through the acceleration record value in the navigation motion process, and further predict the actual propelling effect of the propeller.

[0141] Step 3, based on the damage degree recognition algorithm and the control end adaptive software and hardware system, determine the propeller command setting method to control the target of reducing speed and keeping direction, so that the unmanned ship can still maintain a certain autonomous navigation ability under the damaged state.

[0142] In step 3, the following steps are further included:

[0143] The propeller command setting method is set as:

[0144] The power damage degree of one side propeller X relative to the other side propeller Y can be determined by the identification algorithm in step 2, and then the propulsion instruction of propeller Y is actively reduced. The difference of the initial reduction is the damage degree corresponding to the throttle, so that the two side propellers rotate at the same power state under the initial detected damage degree. Then, considering the complex nonlinear mapping relationship between the control throttle and the power response, the control effect of the initial reduction throttle may still not achieve the heading target, at this time the navigation data is obtained again through the perception system, and is added to the identification algorithm and the control algorithm iteration as feedback, and finally the goal of heading control is achieved.

[0145] Step 31: Initial power reduction degree calculation: In a period of time, the difference between the calculated value and the true value of the throttle is calculated, and when the damage degree Δ is 0, that is, the ship power propulsion system is normal, the difference should be a curve close to 0. Record the time interval [t0, t n ] for studying the damage degree, and record n+1 groups of data in the interval, respectively

[0146]

[0147] According to the deep neural network algorithm designed in step 2, the heading and vertical acceleration are input, and the output is the equivalent propulsion throttle

[0148]

[0149] In the time interval [t0, t n ], the cumulative deviation of the propulsion throttle is calculated

[0150]

[0151] Thus, the absolute damage degree of the left and right propellers at t n can be calculated

[0152]

[0153] Then the damage degree of the ship at t n is

[0154] Δ(n)=Δ X (n)-Δ Y (n)

[0155] The fluctuation of the damage degree decreases as the ship sailing time increases, and when the fluctuation of the damage degree Δ(n) is less than the set threshold η for a continuous time τ, the damage degree identification is considered to be completed, and the initial calculated compensation reduction coefficient can be obtained.

[0156] Step 32: Iterative control law setting. Although in some cases, the initial calculation can obtain relatively accurate results, there may still be throttle ranges with fewer training data sets, lower recognition accuracy, and sparser characteristic values ​​in the sample space composed of all control behaviors. At this time, the instructions calculated initially may have low accuracy. Therefore, multiple iterative calculations are required to achieve sufficient accuracy. After the power compensation reduction, repeat the contents of steps 2 and 3, re-identify the damage degree, and update the power reduction coefficient. The present invention uses the adjusted relative damage degree of the motion effect power to be less than α as the termination condition, and α is initially set to 5% according to the test ship type.

[0157] Example 2

[0158] In this embodiment, two experiments were carried out to verify the feasibility of the scheme.

[0159] The hull parameters selected for the test are shown in Table 1.

[0160] Table 1 Test ship parameters

[0161]

[0162]

[0163] In the test, the parameter values ​​of the Z-score method selected according to the test ship type are shown in Table 2.

[0164] Table 2 Z score method parameters

[0165]

[0166] In the test, the parameter values ​​of the local anomaly factor method selected according to the test ship type are shown in Table 3.

[0167] Table 3 Parameters of local anomaly factor method

[0168]

[0169] In the test, the neural network training parameters selected according to the data conditions and training results are shown in Table 4.

[0170] Table 4 Neural network training parameters based on measured data

[0171]

[0172]

[0173] In Test 1, during actual navigation, power attenuation was applied to one propeller output, and the accuracy of power damage identification was verified through navigational testing. This test artificially set a 20% loss of propeller thrust on one propeller, with the goal of identifying a 20% damage level based on navigational results.

[0174] According to the actual construction of the neural network mapping Input the heading and vertical acceleration, and get the output as the equivalent propulsion throttle

[0175]

[0176] In the time interval [t0, t n ], the cumulative deviation of the propulsion throttle is calculated, as shown in Figure 1 -(a), Figure 1 -(b).

[0177]

[0178] Thus, the absolute damage degree of the right propeller at t n time can be calculated

[0179]

[0180] Then, the damage degree of the ship at t n time Δ(n) = Δ X (n) - Δ Y (n) is shown in Figure 1 -(c). The volatility of the damage degree decreases as the ship's sailing time increases. When the damage degree Δ(n) has a volatility less than the set threshold η for a continuous time τ, it is considered that the damage degree recognition is completed, and the compensation reduction coefficient is obtained. In this test, the relative damage degree is about -20.8256%, that is, the power of the right propeller is reduced by about 20%, which is consistent with the set condition. This test accurately reflects the degree of power damage.

[0181] Test 2: When the propeller is wrapped with water grass, fishing nets and other foreign matters to cause abnormal power output, the fault-tolerant control reliability is verified through power damage degree recognition and control effect verification in the tracking navigation state.

[0182] According to the actual construction of the neural network mapping Input the heading and vertical acceleration, and get the output as the equivalent propulsion throttle

[0183]

[0184] In the time interval [t0, t n ], the cumulative deviation of the propulsion throttle is calculated

[0185]

[0186] Thus, the absolute damage degree of the right propeller at t n time can be calculated

[0187]

[0188] then t n Time Ship Damage Degree Δ(n) = Δ X (n) - Δ Y (n) as shown in Figure 3 -(a). The fluctuation of the damage degree decreases as the ship sailing time increases, and when the fluctuation of the damage degree Δ(n) is less than the set threshold η for a continuous time τ, it is considered that the damage degree recognition is completed, and the compensation reduction coefficient is obtained by the initial calculation. In this test, the relative damage degree is about 45.5232%, that is, the left propeller power is reduced by about 45%, and the accuracy of the number is verified by the fault-tolerant tracking control effect.

[0189] The right throttle is actively reduced to the original 54.4768%, which is the initial accuracy instruction of the directional control. The PID control method is used for verification, and the target point is set for the path tracking navigation. The effect of fault-tolerant control can be verified by the tracking effect of the actual ship, and whether the target of directional navigation is achieved is determined. The navigation trajectory and the throttle time sequence curve are shown in Figure 3 -(b), Figure 3 -(c). In the test, the damage degree can achieve the ideal effect after one calculation, and iteration correction is not needed.

[0190] As can be seen from Figure 3 -(b), the target track (orange broken line) and the actual track (blue curve) have high adhesion, and the control end left and right throttle values are not the same when the actual track is approximately a straight line. The left propeller throttle value is about 22-28, and the right propeller throttle value is about 10-20. The actual throttle value of the left propeller at the output end needs to be reduced to 54.4768% of the original, and at this time the actual throttle values of the left and right propellers are almost the same. The actual ship should be approximately straight, which is consistent with the actual motion effect. Thus, the power damage degree calculation and fault-tolerant control effect of Test 2 can be preliminarily verified.

[0191] The technical solutions of the present application are described in detail above with reference to the drawings. The present application proposes a power damage state recognition and fault-tolerant control method for a double-propeller driven unmanned ship. The control motion mapping of the unmanned ship driven by the rear double propellers is constructed, the power damage degree is identified based on a deep neural network, the control rate is designed according to the damage degree identification result, the speed reduction and directional control are performed as the target, and the problem of abnormal power system is solved.

[0192] The steps in the present application can be adjusted, combined and deleted in sequence according to actual needs.

[0193] Although the present application has been disclosed with reference to the attached drawings, it will be understood that the above examples are merely illustrative of the principles of the application, and are not intended to limit the scope of the application as defined in the appended claims. The scope of the application is defined by the appended claims, and encompass numerous variants, modifications and equivalents that fall within the scope of the application as defined in the appended claims.

[0194] The present application is not limited to the above specific embodiments, and can be implemented in other various specific embodiments by those skilled in the art based on the disclosure of the embodiments and the drawings. Therefore, any design that uses the design structure and the idea of the present application, and makes some simple changes or modifications, falls within the scope of the present application.

Claims

1. A method for identifying and fault-tolerantly controlling the power damage state of a double-propeller driven unmanned boat, characterized in that: The method for identifying and fault-tolerantly controlling the power damage state of a dual-propeller driven unmanned boat includes the following steps: Step 1: Construct the UAV maneuvering and motion feature map based on the measured data, and use the outlier detection algorithm based on the navigation data to ensure the data validity; Step 2: Establish a power damage state recognition algorithm through a deep neural network to qualitatively analyze the degree of power damage of the unmanned boat; Step 3: Determine the thruster command setting method based on the damage degree identification algorithm; In step 2, the following steps are also included: Step 21, defining the damage degree; The absolute propeller damage is defined as: Among them, Δ X is the absolute damage degree of the propeller with higher power loss, Δ Y is the absolute damage degree of the propeller with lower power loss, Δ X >Δ Y The throttle at the propeller control end with higher power loss is T X , the throttle at the propeller control end on the other side is T Y The equivalent propulsion throttles of the actual motion of the two propellers are Damage is defined as: D=D X -D Y Δ is the degree of damage; Step 22, establishing power damage state recognition through deep neural network; In step 22, the following steps are also included: Step 221: Filter the speed, acceleration, left and right throttles, and the time corresponding to each set of data and valid navigation information to form input feature values; Step 222: Perform activation function transformation on the hidden variable and use it as the input of the next fully connected layer; its expression is: ReLU(x)=max(x,0) ReLU(x) is the ReLU function, x is the input feature value of the activation function, which is the filtered valid navigation information; Step 223: The fully connected layer is composed of multiple hidden layers with extended widths, and the hidden layer output is transformed by an activation function; Step 224: The output layer uses the Huber function as the loss function for the regression problem. The Huber error is defined as follows: Where y is the true value, f(x) is the predicted value, and δ is a hyperparameter used to control the turning point of the error; In step 3, the following steps are also included: Step 31: Calculate the initial power reduction rate; The difference between the calculated throttle value and the true value is calculated. When the damage Δ is 0, that is, when the ship's propulsion system is normal, the difference should be a curve close to 0. The damage time interval is [t0,t n ], n+1 groups of data are recorded in the interval, namely: Indicates t n Acceleration information recorded at all times, {a} represents the acceleration time series of the unmanned boat; A power damage state recognition algorithm is established based on a deep neural network. The heading and vertical acceleration are input and the output is the equivalent propulsion throttle: Indicates the equivalent throttle of the left and right propellers calculated from the nth set of recorded acceleration information. Represents the time series of the equivalent propulsion throttle of the unmanned boat propeller; In the time interval [t0,t n ], calculate the cumulative deviation of the boost throttle: T δX (n) represents the cumulative throttle deviation of the X propeller with the higher power loss in the nth group of recorded information, T δY (n) represents the cumulative throttle deviation of the Y propeller with the lower power loss in the nth group of recorded information. represents the equivalent throttle of the X propeller calculated from the i-th group of recorded acceleration information, T iX Indicates the throttle of the X propeller control end recorded at the moment of the i-th group of information, represents the equivalent throttle of the Y propeller calculated from the i-th group of recorded acceleration information, T iY represents the Y propeller control end throttle recorded at the moment of the i-th group of information; And calculate to t n The absolute damage degree of the propeller at the time: Δ X (n) represents the absolute damage degree of X propeller calculated from the nth set of recorded information, Δ Y (n) represents the absolute damage degree of the Y propeller calculated from the nth set of recorded information; Then t n The ship's damage at this moment is: Δ(n)=Δ X (n)-D Y (n) Step 32: Iterate multiple times until the adjusted relative damage degree of the motion effect is less than a preset value, which serves as a termination condition.

2. The method for identifying and fault-tolerantly controlling the power damage state of a double-propeller driven unmanned boat according to claim 1 is characterized in that: In step 1, the following steps are also included: Step 11, constructing the UAV manipulation and motion feature map; Step 12: An outlier detection algorithm based on navigation data is used to ensure data validity. Step 12 also includes: Step 121: Screening based on Z score; Step 122: Screen again based on the local abnormality factor.

3. The method for identifying and fault-tolerantly controlling the power damage state of a double-propeller driven unmanned boat according to claim 2 is characterized in that: In step 11, the manipulation motion mapping relationship is: a→P a is the acceleration at the current moment, P is the propulsion power, F is the three-degree-of-freedom force and torque matrix, m is the mass of the unmanned boat, and v is the speed of the unmanned boat; Based on the unmanned ship with dual propeller differential drive, the decomposition mapping of acceleration in the inertial system can be written as: P2:(a x ,a y )→(T l ,T r ) Among them, a x ,a y are the accelerations of the unmanned boat perpendicular to the bow direction and parallel to the bow direction, T l ,T r These are the left and right propeller thrust throttles respectively.

4. The method for identifying and fault-tolerantly controlling the power damage state of a double-propeller driven unmanned boat according to claim 3 is characterized in that: In step 121, the Z score is expressed as: The expected value of any variable X is E(X) and the standard deviation is σ(X); The sample size is selected as n, and the upper limit of the absolute value of the Z score of a single data point is taken as σ. That is, if each data point does not constitute a significant difference with other values ​​within t seconds before and after it, it is recorded as valid data.

5. The method for identifying and fault-tolerantly controlling the power damage state of a double-propeller driven unmanned boat according to claim 3 is characterized in that: In step 122, the selection is performed again based on the local abnormality factor: Given a parameter p, the distance d from data point p to data point o is p k (p,o) is expressed as: d k (p,o)=max{d k (o),d(p,o)} o is the point closest to p, d k (o) is the distance between o and the kth closest point to o, and d(p,o) is the distance between p and o; The local reachable density of data point p is defined as: N k (p) is all distances p less than d k The points of (p) constitute the k nearest neighbors of p; The local outlier factor LOF of a data point p is the ratio of the average local reachability density of points in the neighborhood of point p to the local reachability density of data point p: LOF k (p) Local outlier factor represents the neighborhood N of point p k The average of the ratios of the local reachability density of other points in (p) to the local reachability density of point p; If the ratio is closer to 1, it means that the density of point p's neighborhood points is similar, and point p may belong to the same cluster as its neighborhood points; if the ratio is less than 1, it means that the density of point p is higher than the density of its neighborhood points, and point p is a dense point; if the ratio is greater than 1, it means that the density of point p is lower than the density of its neighborhood points, and point p may be an outlier.

Images