A soil compaction detection method based on multi-sensor and neural network

Abstract

The present invention discloses a soil compaction detection method based on multiple sensors and neural networks, which can be applied to various terrains, including: collecting detection data of soil samples with various textures; performing unscented Kalman filtering and data cleaning to output noise-removed three-dimensional information of force, time and ground distance and soil multi-parameter information data; training a physical information neural network model based on the three-dimensional information of force, time and ground distance and the soil multi-parameter information data; calculating the force distribution, power distribution, dynamic distribution, frequency domain characteristics and time-frequency characteristics of each soil insertion process, and training a deep neural network model in combination with the three-dimensional information of force, time and ground distance; fusing the trained models and introducing a harmonic parameter α to train a hybrid model; using the three-dimensional information of force, time and ground distance and temperature and humidity data of the sample soil to be tested as inputs of the hybrid model, and outputting the predicted soil compaction.

Description

Technical Field

[0001] The present invention relates to the technical field of soil detection, and more particularly to a soil compaction detection method based on multiple sensors and a neural network. Background Art

[0002] Soil compaction is a key factor affecting agricultural production and environmental protection. Appropriate soil compaction is beneficial to crop growth, reduces water evaporation, and minimizes water and fertilizer waste.

[0003] Existing soil compaction measurement methods are mostly based on a single data source, making it difficult to fully reflect the soil's physical properties and environmental changes. In recent years, the rapid development of drones, sensor technology, and big data analytics has made real-time, high-precision soil compaction measurement possible. However, current research and production processes still face challenges that require further improvement and optimization. A soil compaction measurement technology that is applicable to various terrains, performs in real time, and offers excellent visualization is needed. Summary of the Invention

[0004] In view of this, the present invention provides a soil compaction detection method based on multiple sensors and neural networks. Under various terrain conditions, soil compaction is detected by analyzing multi-source data and the results are visualized. This can improve digital solutions and data support for geographical research and planting planning. Starting from the positive correlation between soil compaction and the force required during soil insertion, a portable, fast and effective soil compaction detection method is proposed.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A soil compaction detection method based on multiple sensors and neural networks, comprising:

[0007] Collect test data of soil samples with various textures;

[0008] Perform unscented Kalman filtering and data cleaning on the detection data, and output the noise-removed three-dimensional force-time-ground distance information and soil multi-parameter information data based on the timestamp corresponding to the acquisition process;

[0009] The physical information neural network model is trained based on the three-dimensional information of force, time and distance to the ground and the multi-parameter information of soil.

[0010] Based on the three-dimensional information of force, time, and distance to the ground, the force distribution, power distribution, dynamic distribution, frequency domain characteristics, and time-frequency characteristics of each soil insertion process are calculated, and a deep neural network model is trained based on the relevant physical and chemical properties of the soil.

[0011] The trained physical information neural network model and deep neural network model are integrated, and the harmonic parameter α is introduced to train the hybrid model;

[0012] The three-dimensional information of force, time and distance to the ground and the temperature and humidity data of the soil sample to be tested are used as the input of the hybrid model, and the predicted soil compaction is output.

[0013] Preferably, the test data collected from soil samples of various textures specifically include:

[0014] Multiple groups of soil samples of three textures, sandy soil, clay soil, and loam, were selected. For each soil sample of each texture, soil temperature and humidity, soil friction coefficient, soil elastic modulus, the magnitude of the force applied when inserting into the soil, the true value of soil compaction, the distance from the ground during the insertion process, and the corresponding position information and posture information were collected.

[0015] Preferably, the specific process of unscented Kalman filtering is:

[0016] Define the state vector: x t =[F,t,d,temp,H,μ,E] T

[0017] Where F is the true value of the force applied when inserting into the soil, t is the time, d is the true value of the distance from the ground during the insertion process, temp is the true value of the soil temperature, H is the true value of the soil moisture, μ is the true value of the soil friction coefficient, and E is the true value of the soil elastic modulus;

[0018] Define the observation vector: z t =[F m ,d m ,temp m ,H m ,μ m ,] T

[0019] Among them F m is the force applied when inserting into the soil, d m is the distance from the ground during the insertion process, temp m is the measured soil temperature, H m is the measured soil moisture;

[0020] Define dynamic model: x t+1 =f(x t ,w t ), z t =h(x t ,v t )

[0021] Where f and h are the state transfer function and observation function respectively, w t and vt are process noise and observation noise.

[0022] Preferably, it also includes:

[0023] After normalizing the predicted soil compaction, a three-dimensional coordinate is generated for each prediction result according to the latitude, longitude and altitude, and the predicted soil compaction results are visualized in three dimensions.

[0024] Preferably, it also includes:

[0025] Based on the predicted soil compaction, data fusion is performed in combination with location information, and the prediction results are presented on the topographic map to obtain the soil compaction distribution of the measured soil area.

[0026] Preferably, the calculation process of the harmonic parameter α is:

[0027] Initialize the value of the harmonic parameter α, α=∈[0,1];

[0028] According to the α value, calculate the mixed model y pred The predicted value is calculated as follows:

[0029] y pred =α*y1+(1-α)*y2

[0030] Where y1 is the predicted value of the physical information neural network model, and y2 is the predicted value of the deep neural network model;

[0031] Calculate the predicted value y pred and the true value of soil compaction y true The difference between error:

[0032] error=y pred -y true ;

[0033] Calculate the square of the prediction error for each sample: squared error =error 2 ;

[0034] Compute the mean of the squared prediction errors over all samples:

[0035]

[0036] Calculate the root mean square error RMSE:

[0037]

[0038] A Gaussian process regression model is constructed based on the root mean square error (RMSE) and used to calculate the acquisition function (EI). The Gaussian process regression model is used to predict the objective function value for each possible value of α, including the predicted mean μ(α) and the predicted standard deviation σ(α). This can be used to calculate the expected improvement (EI) and guide the optimization of the parameter α.

[0039] Based on the acquisition function EI, the iterative update of the α value is guided and the value of EI(α) is calculated. That is, given a value of α, the EI(α) measures the extent to which the objective function can be improved under the α. The formula is:

[0040]

[0041]

[0042] Where μ(α) is the predicted mean; σ(α) is the predicted standard deviation; ξ is the exploration parameter, set to 0.01; Φ(Z) and represent the cumulative distribution function and probability density function of the standard normal distribution respectively;

[0043] Find the maximum EI(α) in the search space of α, and the corresponding α value is the optimal harmonic parameter α best , that is, the α value that makes the objective function value optimal (maximum or minimum, depending on the nature of the problem) in the existing samples. In each iteration, the new α best Update, find the maximum EI(α) in the search space of α, and the corresponding α value will be set as the new α best , used for the next iteration.

[0044] Preferably, during the data cleaning process, missing values ​​are deleted; outliers are detected; and the data are standardized and normalized so that the data are within the same magnitude range.

[0045] Preferably, the motion state of the probe is obtained through the IMU module, and the probe posture is represented based on the quaternion. Combined with the angle between the probe and the ground edge, the magnitude of the force applied when inserting into the soil and the distance from the ground during the insertion process are corrected.

[0046] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a soil compaction detection method based on multiple sensors and neural networks, which can be applied to various terrains and has a good degree of visualization. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0048] Figure 1 The accompanying drawing is a flow chart of the soil compaction detection method based on multiple sensors and neural networks provided by the present invention.

[0049] Figure 2 The accompanying drawing is a design diagram of the sensor position distribution provided by the present invention.

[0050] Among them, 1. Piezoresistive force sensor, 2. Temperature and humidity sensor, 3. Laser ranging sensor, 4. Temperature and humidity sensor probe, 5. Spring. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, a soil compaction detection method based on multiple sensors and neural networks includes:

[0053] Collect test data of soil samples with various textures;

[0054] Perform unscented Kalman filtering and data cleaning on the detection data, and output the noise-removed three-dimensional force-time-ground distance information and soil multi-parameter information data based on the timestamp corresponding to the acquisition process;

[0055] The physical information neural network model is trained based on the three-dimensional information of force, time and distance to the ground and the multi-parameter information of soil.

[0056] The force distribution, power distribution, dynamic distribution, frequency domain characteristics, and time-frequency characteristics of each soil insertion process are calculated based on the three-dimensional information of force, time, and distance to the ground. The deep neural network model is trained using this three-dimensional information.

[0057] The trained physical information neural network model and deep neural network model are integrated, and the optimal harmonic parameter α is introduced to train the hybrid model;

[0058] The three-dimensional information of force, time and distance to the ground and the temperature and humidity data of the soil sample to be tested are used as the input of the hybrid model, and the predicted soil compaction is output.

[0059] In this embodiment, the detection data collected from soil samples of various textures specifically includes:

[0060] Multiple groups of soil samples of three textures, namely sandy soil, clay soil, and loam, were selected. For each group of soil samples of each texture, soil temperature and humidity, soil friction coefficient, soil elastic modulus, the magnitude of the force applied when inserting into the soil, the true value of soil compaction, the distance from the ground during the insertion process, and the corresponding position information and posture information were collected.

[0061] More specifically, Figure 2 As shown, multiple temperature and humidity sensor probes 4 are placed at the lower end of the soil compaction detector, and the probe ends are vertically facing downward by default. The space above the temperature and humidity sensor 2 is used to place the piezoresistive force sensor 1, leaving space for the probe to be fixed with a spring 5. After fixation, the piezoresistive force sensor data is adjusted to zero. At the same time, a laser ranging sensor 3 is installed to measure the depth of insertion into the soil, a core board is placed for control, and a built-in GPS module and IMU module are used to collect data on the geographic location and movement posture of the detector.

[0062] For each sample, the force required to penetrate the soil is measured using a piezoresistive force sensor (1), soil temperature and humidity are measured using a soil temperature and humidity sensor (2), and the distance to the ground during insertion is measured using a laser rangefinder (3). Data is output based on the corresponding moments, indicating changes in force and distance to the ground over time. The GPS and IMU modules measure the current position and posture. Calibration is used to measure the soil position and adjust the posture of the soil during insertion. A soil compaction meter measures the true compactness of the soil sample. A friction coefficient meter measures the soil friction coefficient at the corresponding point, and an appropriate elastic modulus is uniformly selected based on the reference elastic modulus of each soil type.

[0063] Data were collected at least three times for each sample to obtain stable measurement results.

[0064] Specifically, this embodiment is applied to soil compaction testing in mountainous terrain. Based on the mountain characteristics determined in the preliminary survey, it is recommended to select multiple test points every 3 or 5 meters on the mountainous terrain, and collect data three times at each test point. It should be noted that the method of the present invention can also be applied to other terrain types; this embodiment is merely a specific example.

[0065] Furthermore, when the laser ranging sensor inserts the temperature and humidity sensor probe vertically into the soil, the temperature and humidity sensor probe length is defined as L1, the distance to the ground is L2, and the allowable error is L0. The following situation will occur:

[0066] (1) If L2-L1 is greater than L0, it means that sampling preparation is not complete.

[0067] (2) If L2-L1 is less than or equal to L0 and L2-L1 is greater than 0, then prepare for collection.

[0068] (3) If L2-L1 is less than 0, it is in the data collection process.

[0069] like Figure 2 As shown, after collecting data, this embodiment uses the IMU module to perform attitude correction. When used in mountains, the terrain will have a direct impact on the accuracy of the sensor module data of the detector. First, it is necessary to overcome the problem that the direction of insertion into the soil is not perpendicular to the ground. According to the real-time measurement of the acceleration, angular velocity and geomagnetic field information of the probe by the IMU module, the attitude of the temperature and humidity sensor probe in space is analyzed and represented by the quaternion method. The specific steps are as follows:

[0070] Taking piezoresistive force sensor as an example

[0071] When the temperature and humidity sensor probe is placed vertically on the ground with the detector, the quaternion q IMU Expressed as:

[0072] q IMU =[w,x,y,z];

[0073] Where w is the scalar part, [w,x,y,z] is the vector part, the rotation of the probe relative to the ground coordinate system (or reference coordinate system),

[0074] Define a vector v that represents the direction of the force sensor in the probe coordinate system probe , for example [0,0,1] means the probe measures force along the Z axis of its own coordinate system.

[0075] In order to transform the vector v in the probe coordinate system probe Convert to vector v in the ground coordinate system ground , we need to use the attitude quaternion q IMU v probe Rotate.

[0076] The quaternion rotation formula is as follows:

[0077]

[0078] in, is the inverse (or conjugate) of the quaternion. For the unit quaternion (i.e. ||q IMU ||=1), the calculation formula of the inverse quaternion is Note that * here indicates quaternion multiplication.

[0079] The quaternion multiplication formula is as follows:

[0080] q1*q2=[w1*w2-x1*x2-y1*y2-z1*z2,

[0081] w1*x2-x1*w2-y1*z2-z1*y2,

[0082] w1*y2-x1*z2-y1*w2-z1*x2,

[0083] w1*z2-x1*y2-y1*x2-z1*w2,]

[0084] The quaternion rotation and multiplication formula can calculate v probe , which is the vector representing the direction of the force sensor in the ground coordinate system. The measured force value can then be projected onto this direction, thereby correcting the force through the IMU module. Specifically, the force value and v can be calculated probe The dot product of is used to obtain the corrected force value.

[0085] The processing of the ground distance during the insertion process by IMU attitude correction is the same as the above part.

[0086] Define the vector of the laser ranging sensor pointing to the ground in the probe coordinate system: v laser-probe , for example [0,0,-1], which means that the laser ranging sensor measures the distance along the negative Z axis of the probe's own coordinate system.

[0087] V laser-probe Convert to vector v in the ground coordinate system laser-ground Specifically, the attitude quaternion q measured by the IMU is required IMU v laser-probe To perform a rotation:

[0088]

[0089] Calculate the distance to the ground during the measurement insertion process (assuming d laser ) and v laser-ground The dot product of the distance measured by the laser rangefinder is the vertical component in the ground coordinate system. Let the unit vector of the ground coordinate system be v up =[0,0,-1], then the distance d from the ground after the correction ground It can be calculated by the following formula:

[0090] d ground =d laser *(v laser-ground ·v up ).

[0091] In this embodiment, data preprocessing is first to process the initial data using the unscented Kalman filter algorithm. This algorithm can effectively improve the accuracy and stability of nonlinear data. The main steps of the unscented Kalman filter algorithm are as follows:

[0092] a. Initialization preparation:

[0093] Initialize the state vector x0 and covariance matrix p0

[0094] Define the state vector: x t =[F,t,d,temp,H,μ,E] T

[0095] Where F is the true value of the force applied when inserting into the soil, t is the time, d is the true value of the distance from the ground during the insertion process, temp is the true value of the soil temperature, H is the true value of the soil moisture, μ is the true value of the soil friction coefficient, and E is the true value of the soil elastic modulus.

[0096] Define the observation vector: z t =[F m ,d m ,temp m ,H m ,] T

[0097] Among them F m is the force applied when inserting into the soil, d m is the distance from the ground during the insertion process, temp m is the measured soil temperature, H m is the measured soil moisture;

[0098] Define dynamic model: x t+1 =f(x t ,w t ), z t =h(x t ,v t )

[0099] Where f and h are the state transfer function and observation function respectively, w t and v t For x t process noise and observation noise in the state.

[0100] According to the current state x t and process noise w t , calculate the predicted state transfer matrix F t and the predicted state covariance matrix P t :

[0101]

[0102]

[0103] Among them, 0 means that the process noise is a zero vector, and Q is the process noise covariance matrix;

[0104] b. Calculate sigma point:

[0105]

[0106]

[0107]

[0108] in, is the first sigma point, which is the estimated value of the current state. That is, a vector is added to the current state at the next n sigma points. The vector is the P of the state covariance matrix. t Appropriate columns for the corresponding dimensions; It is based on the current estimated state minus a vector, which is the P of the state covariance matrix t The appropriate column of the corresponding dimension; λ is a tuning parameter of the unscented Kalman filter algorithm, which is used to control the parameters of the sigma point distribution. It is usually taken as 3-n, where n is the dimension of the state vector. represents the square root of the i-th eigenvalue of the matrix, p t To predict the state x t The covariance matrix of , i represents the i-th dimension corresponding to the state variable.

[0109] c. Predict sigma point:

[0110]

[0111] in, is the i-th sigma point of the predicted value at time t obtained at time t-1, and f is the state transfer function.

[0112] d. Calculate the predicted state vector and the predicted covariance matrix P t∣t-1 :

[0113]

[0114]

[0115]

[0116]

[0117]

[0118] in, is the state mean after the prediction step, P t∣t-1 is the covariance matrix after the prediction step, and is the weight coefficient, α and β are the parameters of the unscented Kalman filter algorithm, which are generally set to 0.1, and Q is the process noise covariance matrix;

[0119] e. Calculate the observed sigma point:

[0120] That is, the sigma point after the state transfer, v t is the observation noise, v t The size depends mainly on the sensor accuracy;

[0121] i. Calculate the predicted observation mean and the observation covariance matrix S t :

[0122]

[0123] Calculate the mean of the predicted measurements. The predicted measurements at each sigma point The corresponding weight After weighted summation, the predicted measured value mean is obtained

[0124]

[0125] Calculate the predicted measurement covariance matrix S t It represents the uncertainty of the predicted measurement value. The deviation of the predicted measurement value at each sigma point (i.e. ) is represented by its corresponding weight After weighting, multiply the matrix formed by its transpose and then sum it. R is the covariance matrix of measurement noise, which is used to consider the impact of measurement noise.

[0126] f. Calculate the Kalman gain K t :

[0127]

[0128]

[0129] Among them C t is the cross covariance between the predicted state and the predicted observation value, which is used to represent the correlation between the state value and the measured value.

[0130] g. Update the state vector and covariance matrix:

[0131]

[0132] x t|t Represents the state vector updated after obtaining the observation value at time t. t represents the actual observation vector at time t, and represents the mean of the observations at time t predicted at time t-1.

[0133]

[0134] P t|t P represents the updated covariance matrix after obtaining the observation vector at time t. t|t-1 is the covariance matrix at time t predicted at time t-1, S t is the covariance matrix of the observation at time t predicted at time t-1, K t is the Kalman gain matrix at time t, K t It is the key factor in updating state estimation. Kalman gain K t Determines the ratio between the predicted state and the measured value when the state is updated. If the Kalman gain is large, the influence of the measured value will be greater; if the Kalman gain is small, the influence of the predicted state will be greater.

[0135] In this embodiment, data cleaning mainly includes: deleting missing values, detecting outliers using box plots or the 3σ principle, and using the zscore function to achieve data standardization, thereby obtaining reliable force-time-ground distance three-dimensional information and providing model training.

[0136] In this embodiment, a physical information neural network (PINN) model is trained based on the three-dimensional information of force-time-ground distance and soil multi-parameter information data (soil temperature and humidity temp, soil friction coefficient μ, soil elastic modulus E).

[0137] In this embodiment, the training of the deep neural network model. The statistical model requires a large amount of feature data as the input layer:

[0138] (1) Dynamic characteristics: rate of change of force (derivative of force), rate of change of power (derivative of power);

[0139] (2) Statistical characteristics: median force, median power, maximum value, minimum value, standard deviation, peak-to-average ratio (PAR) of power. The sensor data acquisition process is time-series, and the force-time-ground distance coordinate is displayed as a continuous surface. For example, the peak value in the axial direction of the force can be used to obtain the maximum force. Integrating force and time can obtain the power, and integrating force and ground distance can obtain the work.

[0140] (3) Frequency domain characteristics: spectral characteristics of force signals (such as energy spectrum density, spectrum center frequency, etc.), spectral characteristics of power signals (such as energy spectrum density, spectrum center frequency, etc.);

[0141] (4) Time-frequency characteristics: Use time-frequency analysis methods such as short-time Fourier transform (STFT) and wavelet transform to extract the time-frequency characteristics of the signal;

[0142] Furthermore, while enriching the input layer, the minimum redundancy maximum relevance (mRMR) feature selection method is used to select the above-mentioned groups of feature data, maintaining the maximum relevance while maintaining the minimum redundancy, avoiding the occurrence of overfitting, and avoiding the introduction of highly correlated features.

[0143] Furthermore, the deep neural network model is configured with 3 to 5 hidden layers, 100 to 200 neurons per hidden layer, a ReLU activation function, and an L1 or L2 regularization term. The output layer predicts soil density, and a linear activation function can be used for direct output.

[0144] In this embodiment, the hybrid model y pred The harmonic parameter α is introduced to balance the contribution of the physical information neural network model and the deep neural network model to the prediction results. The value range of α is [0,1].

[0145] Furthermore, Bayesian optimization is used to find the optimal harmonic parameter α best , so that y pred While maintaining physical rationality, improve y pred The accuracy and robustness of the prediction. Initialize the α value, and construct a Gaussian process regression model based on the prediction error corresponding to the α value. The Gaussian process regression is used to describe the mixed model y pred Instability, select the α value corresponding to the acquisition function ExpectedImprovement (EI) for iterative update, and finally find the maximum EI (α), which is the optimal harmonic parameter α best The specific process is as follows:

[0146] Initialize the value of the harmonic parameter α, α=∈[0,1];

[0147] According to the α value, calculate the mixed model y pred The predicted value of . The calculation formula is:

[0148] y pred =α*y1+(1-α)*y2,

[0149] Where y1 is the predicted value of the physical information neural network model, and y2 is the predicted value of the deep neural network model;

[0150] Calculate the predicted value y pred and the true value of soil compaction y true The difference between error:

[0151] error=y pred -y true ;

[0152] Calculate the square of the prediction error for each sample: squared error =error 2 ;

[0153] Compute the mean of the squared prediction errors over all samples:

[0154]

[0155] Calculate the root mean square error RMSE:

[0156] Construct a Gaussian process regression model based on the root mean square error to describe the mixed model y pred The instability of the proposed method is investigated, thereby providing guidance for optimizing the harmonic parameter α. The Gaussian process regression model is used to predict the objective function value corresponding to each possible value of α, including the predicted mean μ(α) and the predicted standard deviation σ(α), which can be used to calculate the expected improvement (EI) value and guide the optimization of the parameter α.

[0157] Based on the acquisition function EI, the iterative update of the α value is guided and the value of EI(α) is calculated. That is, given a value of α, the EI(α) measures the extent to which the objective function can be improved under the α. The formula is:

[0158]

[0159]

[0160] Where μ(α) is the predicted mean; σ(α) is the predicted standard deviation; ξ is the exploration parameter, set to 0.01; Φ(Z) and represent the cumulative distribution function and probability density function of the standard normal distribution respectively;

[0161] Find the maximum EI(α) in the search space of α, and the corresponding α value is the optimal harmonic parameter α best , that is, the α value that makes the objective function value optimal (maximum or minimum, depending on the nature of the problem) in the existing samples. In each iteration, the new α best Update, find the maximum EI(α) in the search space of α, and the corresponding α value will be set as the new α best , used for the next iteration.

[0162] In this embodiment, after the predicted soil compaction data is normalized, the latitude, longitude and altitude data are integrated into a matrix using coordinates, and a three-dimensional coordinate is generated for each prediction result. The predicted soil compaction results can be visualized in three dimensions using a scatter plot with color mapping in MATLAB, that is, a three-dimensional scatter plot is drawn using scatter3(), and the color mapping is specified using colormap() to reflect the size of the compaction.

[0163] In this embodiment, based on the terrain point cloud data of the mountain and the vegetation height data measured by the drone, combined with the predicted soil compaction results, a three-dimensional model of the mountain terrain with vegetation cover and soil compaction is created. Each sampling point is colored according to the soil compaction data to display the soil compaction distribution in the mountainous area, so that users can intuitively understand the distribution of soil compaction in geographic space.

[0164] Effectively detecting and predicting soil compaction in mountainous areas will provide valuable information for agricultural production, soil conservation, and ecological restoration, contributing to sustainable development and ecological conservation. Furthermore, this method has broad application prospects in other terrain types.

[0165] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0166] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A soil compaction detection method based on multiple sensors and neural networks, characterized in that: include: Collect test data of soil samples with various textures; Perform unscented Kalman filtering and data cleaning on the detection data, and output the noise-removed three-dimensional force-time-ground distance information and soil multi-parameter information data based on the timestamp corresponding to the acquisition process; The physical information neural network model is trained based on the three-dimensional information of force, time and distance to the ground and soil multi-parameter information data, including soil temperature and humidity, soil friction coefficient and soil elastic modulus. The force distribution, power distribution, dynamic distribution, frequency domain characteristics, and time-frequency characteristics of each soil insertion process are calculated based on the three-dimensional information of force, time, and distance to the ground. The deep neural network model is trained using this three-dimensional information. Fuse the trained physical information neural network model and deep neural network model and introduce the harmonization parameter Train the hybrid model; adjust parameters The calculation process is: a. Initialize the reconciliation parameters The value of ; b.According to Value, calculate the predicted value of the mixed model , the calculation formula is: ； in is the predicted value of the physical information neural network model, is the predicted value of the deep neural network model; c. Calculate the predicted value True value of soil compaction The difference between : ； d. Calculate the square of the prediction error for each sample: ; e. Calculate the mean of the squared prediction errors of all samples: ； f. Calculate the root mean square error : ; g. Construct a Gaussian process regression model based on the root mean square error and calculate the acquisition function through the Gaussian process regression model ; h. Based on acquisition function guide Iterative update of values, calculation The value of is: ； in, is the predicted mean; is the forecast standard deviation; is the exploration parameter, set to 0.01; and represent the cumulative distribution function and probability density function of the standard normal distribution respectively; exist Find the largest , corresponding to The value is the optimal harmonic parameter ; The three-dimensional information of force, time and distance to the ground and the temperature and humidity data of the soil sample to be tested are used as the input of the hybrid model, and the predicted soil compaction is output.

2. The soil compaction detection method based on multiple sensors and neural networks according to claim 1, characterized in that: The test data collected from soil samples of various textures include: Multiple groups of soil samples of three textures, sandy soil, clay soil, and loam, were selected. For each soil sample of each texture, soil temperature and humidity, soil friction coefficient, soil elastic modulus, the magnitude of the force applied when inserting into the soil, the true value of soil compaction, the distance from the ground during the insertion process, and the corresponding position information and posture information were collected.

3. The soil compaction detection method based on multiple sensors and neural networks according to claim 2, characterized in that: The initialization preparation process of the unscented Kalman filter is: Define the state vector at time t: ; in is the true value of the force applied when inserting into the soil, For time, is the true value of the distance to the ground during the insertion process, is the true value of soil temperature, is the true value of soil moisture, is the true value of soil friction coefficient, is the true value of soil elastic modulus; Define the observation vector: ; in is the force applied when inserting into the soil, is the distance from the ground during the insertion process, is the soil temperature, is soil moisture; Define the dynamic model: ; in and are the state transfer function and observation function respectively, and are process noise and observation noise.

4. The soil compaction detection method based on multiple sensors and neural networks according to claim 1, characterized in that: Also includes: After normalizing the predicted soil compaction, a three-dimensional coordinate is generated for each prediction result according to the latitude, longitude and altitude, and the predicted soil compaction results are visualized in three dimensions.

5. The soil compaction detection method based on multiple sensors and neural networks according to claim 1, characterized in that: Also includes: Based on the predicted soil compaction, data fusion is performed in combination with location information, and the prediction results are presented on the topographic map to obtain the soil compaction distribution of the measured soil area.

6. The soil compaction detection method based on multiple sensors and neural networks according to claim 1, characterized in that: During the data cleaning process, missing values ​​are deleted; outliers are detected; and the data is standardized and normalized so that the data are within the same magnitude range.

7. The soil compaction detection method based on multiple sensors and neural networks according to claim 2, characterized in that: The motion state of the probe is obtained through the IMU module, and the probe posture is expressed based on the quaternion. Combined with the angle between the probe and the ground edge, the force applied when inserting into the soil and the distance from the ground during the insertion process are corrected.

Images