A deep neural network-based underground fruit damage testing system and testing method
By combining deep neural networks with soil physical properties and underground fruit damage prediction models, the stress state of underground fruits can be monitored in real time, solving the problem of insufficient field trial data and realizing the development of accurate prediction of underground fruit damage and low-damage harvesting technology.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING AGRICULTURAL UNIVERSITY
- Filing Date
- 2025-09-01
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies make it difficult to monitor the stress state of underground fruits in real time in a field environment, resulting in insufficient reliability of simulation analysis for research on the harvest damage mechanism of underground fruits. Furthermore, the limited field test data makes it impossible to accurately assess the damage to underground fruits.
A deep neural network-based underground fruit damage testing system is adopted, which combines a soil physical property prediction model and an underground fruit damage prediction model. The system uses a servo motor to drive the top plate to make linear reciprocating motion in the soil box. Data is collected in real time using strain gauge sensors and soil moisture sensors. The system combines deep neural networks to predict the mechanical parameters of the soil and fruit, thereby achieving accurate prediction of the probability of underground fruit damage.
This study enabled the prediction of damage probability of underground fruits under cyclic loading in a laboratory environment, providing a theoretical basis for low-loss harvesting of underground fruits and improving harvest quality and the reliability of simulation experiments.
Smart Images
Figure CN121347727B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of crop damage testing equipment design and artificial intelligence, and particularly relates to a test system and method for testing underground fruit damage based on deep neural networks. Background Technology
[0002] In agricultural fields, there are already tests on different tissues such as stems and fruits of various crops, including tensile, compression, and bending tests, to obtain data such as elastic modulus and breaking stress. The testing equipment used is similar in principle to the universal testing machine and is called a texture analyzer.
[0003] However, whether using a universal testing machine or a dedicated texture analyzer, the determination of the mechanical properties of various parts of a crop requires the test sample to have a specific, regular shape to facilitate subsequent data processing based on relevant mechanical theories. Furthermore, these testing methods all fall under the category of measuring the basic mechanical properties of the test sample, obtaining its mechanical parameters under specified conditions. However, obtaining these basic mechanical parameters is only the first step in scientific research. The stress states of actual objects are highly variable, and basic mechanical parameters alone cannot accurately determine the stress state of the test sample under complex stress conditions. Therefore, specific mechanical tests are needed based on the specific stress forms experienced by the object under test.
[0004] In agricultural engineering, especially during the harvesting of underground fruits, the complex interactions between the crop-soil-machine system and the working components in contact with the soil can easily damage the fruits, affecting harvest quality. Currently, the investigation of damage mechanisms during underground fruit harvesting typically relies on simulation and experimental methods. Simulation methods primarily depend on measured basic mechanical data, followed by simulations of actual harvesting operations. However, due to the complex real-world environment of field trials and the lack of effective data comparison, the reliability of simulation analysis results cannot be guaranteed. Experimental methods mainly depend on actual field operations, but measured data is very limited. In particular, field trials cannot monitor the real stress state of the underground fruits in real time to determine their damage.
[0005] Therefore, this invention provides a novel underground fruit damage testing system and method based on deep neural networks, which combines the advantages of simulation analysis and field experiments. In a laboratory environment, it can detect the stress changes of actuators, soil, and underground fruits under different cyclic vibration conditions in real time, explore the influence law of underground fruit damage under cyclic load, provide experimental basis for low-damage harvesting of underground fruits, and provide experimental evidence for simulation experiments. Summary of the Invention
[0006] To address the shortcomings of existing technologies, this invention provides a deep neural network-based underground fruit damage testing system and method, which enables the prediction of damage probability of underground fruits under cyclic loading, providing theoretical and experimental basis for the study of vibration damage mechanisms of underground fruits and the development of mechanized low-damage harvesting technology.
[0007] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0008] A deep neural network-based underground fruit damage testing system includes a steel frame profile and a soil box and a fixing plate fixedly mounted on the steel frame profile. A linear guide rail and a reducer are fixedly mounted on the fixing plate. The output shaft of the reducer is keyed to an amplitude adjustment disc, and the input end of the reducer is keyed to the output shaft of a servo motor. The amplitude adjustment disc is hinged to a connecting rod, the other end of which is hinged to a top rod, and the other end of the top rod is hinged to the top plate of the soil box. Driven by the servo motor, the top plate performs linear reciprocating motion inside the soil box. Strain gauge sensors are attached to the outer surface of the top rod, and multiple soil moisture sensors are evenly distributed inside the soil box. Both the strain gauge sensors and the soil moisture sensors are connected to a computer via a signal acquisition device.
[0009] Based on the constructed soil physical property prediction model and underground fruit damage prediction model, the computer-based system predicts the distribution of parameters such as elastic modulus, moisture, and density of the soil in the target test box, as well as data such as stress distribution, centroid location, and damage probability of the underground fruit, and displays them visually.
[0010] The method for testing underground fruit damage using the aforementioned deep neural network-based underground fruit damage testing system includes the following steps:
[0011] Step 1: Correlate the relationship between soil elastic modulus and moisture content and density, as well as the relationship between stress state and moisture content and density, as physical constraints for subsequent predictions;
[0012] Soil compression samples were prepared and compression tests were conducted using a universal testing machine. The initial elastic modulus of the soil was obtained based on the stress-strain curve. This elastic modulus corresponds to different moisture contents and densities, thus establishing a pre-established relationship between the soil elastic modulus and moisture content and density. The soil sample was placed in a sealed compression container with an internal moisture sensor, and a compression load was applied. The load-displacement and moisture content-displacement curves were recorded and converted into density-moisture content curves. The stress-density-moisture content surface was then fitted.
[0013] Step 2: Collect initial soil physical data to provide a baseline for subsequent tests;
[0014] In the initial state, the underground fruit damage testing system is stationary. The initial moisture content is obtained from the data collected by the soil moisture sensor, the initial density is calculated by measuring the mass and volume of the soil in the soil box, and the initial stress is zero. The initial elastic modulus is obtained from step 1.
[0015] Step 3: The underground fruit damage testing system starts working. The servo motor starts and drives the top plate to make linear reciprocating motion. During this process, the computer receives the sensor detection data and converts the strain data on the top rod into pressure data by combining the elastic modulus and cross-sectional area of the top rod, thereby obtaining cyclic load data. This data, along with the initial soil density and moisture content data, is input into the pre-constructed soil physical property prediction model and underground fruit damage prediction model to obtain the real-time distribution changes of soil density, elastic modulus, stress, and moisture content in the soil box with the load, as well as the real-time location, stress distribution, and damage probability of the underground fruit.
[0016] Step 4: Visualize the results obtained in Step 3 on the computer.
[0017] Furthermore, in step 3, the method for constructing the soil physical property prediction model is as follows:
[0018] S1: Collect bench test data of soil density, elastic modulus, stress, and moisture content under different loads to form a soil physical property dataset;
[0019] S2: Construct a soil physical property prediction generator, which includes an input layer, a hidden layer, and an output layer;
[0020] Input layer: Receives parameters from six dimensions, including real-time load data from the test bench and initial soil physical data. The number of neurons is matched one-to-one with the parameter dimensions. Normalizes the input data, removes unreasonable data, and reconstructs the load data into a two-dimensional matrix.
[0021] Hidden Layers: A hybrid architecture of convolutional feature extraction and fully connected feature fusion is adopted. The convolutional feature extraction module, targeting the load time-series data, sets one convolutional block with 32 5*1 convolutional kernels to extract global load change time-series features. ReLU activation function is used to achieve non-linear mapping, and each convolutional block is followed by a max pooling layer. The fully connected feature fusion module sets a flattening layer to flatten the time-series features output by the convolutional module and concatenates them with the initial soil physical data to form a fused feature vector. A two-layer fully connected network is set. The first layer has 128 neurons, which further fuse multi-source features through ReLU activation function. The second layer has 64 neurons, which uses LeakReLU activation to enhance the non-linear expression capability of features. Dropout is added between the fully connected layers to randomly deactivate some neurons and enhance the model's generalization ability.
[0022] Output layer: The spatiotemporal coupling matrix is adopted, and the output dimension is set as: time step × number of spatial grids × 4. The time step is matched with the temporal length of the input load, and the number of spatial grids corresponds to the number of preset observation points in the soil box. The four physical quantity channels are density, water content, elastic modulus and stress. A linear activation function is used to ensure physical meaning: the output values of density and water content are non-negative and within the range of natural soil properties. The output values of elastic modulus and stress are constrained by the pre-obtained values under different water contents and densities, and maintain mechanical correlation with load amplitude, frequency and initial soil state. The spatial distribution shows physical continuity.
[0023] S3: Construct a soil physical performance evaluator based on vibration test data of a test bench without underground fruit;
[0024] S4: Construct a soil physical performance evaluator based on vibration test data of underground fruit test benches;
[0025] S5: Using the two evaluators S3 and S4 and the soil physical property dataset, the soil physical property prediction generator is trained and tested in two stages. The qualified soil physical property prediction generator is used as the soil physical property prediction model. This model is a neural network architecture that takes load data and initial soil physical data as input and predicts soil moisture content, density, elastic modulus and stress data as output.
[0026] Furthermore, the construction process of the soil physical performance evaluator based on vibration test data of the platform without underground fruit is as follows: based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established, and the calculated losses are weighted and summed as the evaluation rules of the soil physical performance evaluator based on vibration test data of the platform without underground fruit, which is used for the evaluation of soil physical properties.
[0027] The process of constructing a soil physical performance evaluator based on vibration test data of a test bench containing underground fruits is as follows: Based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established, and the calculated losses are weighted and summed to serve as the evaluation rules for the soil physical performance evaluator based on vibration test data of a test bench containing underground fruits, which is used to evaluate soil physical properties.
[0028] Furthermore, the two-stage collaborative training of the soil physical property prediction generator includes:
[0029] The soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench without underground fruit are compared with the data in the soil physical performance dataset, and the loss between the two is taken as soil loss; the soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench with underground fruit are compared with the data in the soil physical performance dataset, and the loss between the two is taken as underground fruit loss.
[0030] Furthermore, in step 3, the method for constructing the underground fruit damage prediction model is as follows:
[0031] S1: Construct a load data and soil physics dataset using relevant data collected during bench tests and simulation tests;
[0032] S2: Construct a generator to predict underground fruit damage;
[0033] S3: Construct an assessment tool for the damage performance of underground fruits;
[0034] S4: Construct a real-time location evaluator for underground fruits;
[0035] S5: Construct an underground fruit damage probability evaluator;
[0036] S6: Based on the underground fruit damage performance evaluator, underground fruit real-time location evaluator, underground fruit damage probability evaluator, and load data and soil physics dataset, the underground fruit damage prediction generator undergoes three-stage collaborative training and testing using underground fruit damage performance data, real-time location data, and damage probability data. The underground fruit damage prediction generator that passes the test is used as the damage prediction model. The three-stage collaborative training and testing includes: comparing the maximum stress of the inner and outer layers of the fruit output by the underground fruit damage performance evaluator with simulation data under the same conditions, and using the loss between the two as the stress assessment loss; comparing the real-time location output by the underground fruit real-time location evaluator with simulation data, and using the loss between the two as the location assessment loss; and comparing the damage probability of the inner and outer layers of the underground fruit output by the underground fruit damage probability evaluator with simulation test data and bench test data, and using the loss between the two as the probability assessment loss.
[0037] Furthermore, the stress assessment loss is determined based on the maximum stress in each layer of the underground fruit structure, as shown in the following formula:
[0038]
[0039] In the formula, L S For stress assessment loss value, α j is the weighting coefficient; m is the number of layers in the underground fruit, j is the layer number; S javeThe maximum stress simulation value for each layer of the underground fruit under the same conditions is s. jave These are the maximum stresses of each layer of the underground fruit output by the underground fruit damage performance assessment device.
[0040] Furthermore, the location loss assessment is determined based on the real-time location of the underground fruit during the simulation test, as shown in the following formula:
[0041]
[0042] In the formula, L L The location is used to evaluate the loss value; α, β, and γ are the weighting coefficients in the range [0,1], and x Ct y Ct z Ct X represents the triaxial coordinates of the centroid of the underground fruit at time t, output by the real-time position estimator of the underground fruit. Ct Y Ct Z Ct The values of the centroid of the underground fruit at time t are obtained from simulation calculations; T is the final time.
[0043] Furthermore, the probability assessment loss is determined based on the maximum stress and compressive shear strength of each layer of the underground fruit, as shown in the following formula:
[0044] L P =(pP ave ) 2
[0045] In the formula, L P To assess the loss value using probability; P ave Let be the probability of damage to underground fruits under the same conditions; p is the damage probability output by the underground fruit damage probability estimator, and its expression is shown in the following formula:
[0046]
[0047] In the formula, β j σ is the weighting coefficient of the j-th layer of underground fruit; j ] represents the actual allowable strength of the underground fruit layer j; σ is the equivalent allowable strength. sj σ represents the maximum equivalent strength of the j-th layer of the actual underground fruit; j The maximum equivalent stress of the j-th layer of the underground fruit is predicted.
[0048] Furthermore, the underground fruit damage prediction generator is a neural network architecture that takes real-time load data from bench testing, initial position of underground fruit, and initial soil physical data as input, and outputs current stress distribution of underground fruit, real-time position of underground fruit, and damage probability.
[0049] The construction process of the underground fruit damage performance evaluator is as follows: Based on load data and soil physics dataset, the stress distribution and real-time centroid position of each layer inside and outside the underground fruit are predicted to form an expanded dataset; a calculation model for underground fruit damage performance evaluation is constructed, and the maximum stress loss of each layer of the underground fruit obtained by the calculation model is weighted and summed as the evaluation rule for underground fruit damage performance, which is used for the evaluation of underground fruit damage performance.
[0050] The construction process of the underground fruit real-time location evaluator is as follows: extract the data output by the underground fruit damage performance evaluator, combine it with simulation data, establish a loss calculation model, and perform enhanced summation on the calculated triaxial loss as the evaluation rule for the real-time location of the underground fruit.
[0051] The process of constructing the underground fruit damage probability evaluator is as follows: extract the data output by the underground fruit damage performance evaluator, combine it with bench test and simulation data, establish a loss calculation model, and use this model as the damage probability evaluation rule for underground fruits to evaluate the damage probability of underground fruits.
[0052] The present invention has the following beneficial effects:
[0053] This invention achieves accurate prediction of underground fruit damage through the construction of a soil physical property prediction model and an underground fruit damage prediction model. The soil physical property prediction model was obtained through two stages of collaborative training and testing: one based on soil physical property evaluators containing underground fruit vibration test data, and the other based on soil physical property evaluators without underground fruit vibration test data. This resulted in a model capable of predicting the physical and mechanical properties of the soil within the test chamber. The underground fruit damage prediction model was obtained through three stages of coordinated training: an underground fruit damage performance evaluator, an underground fruit real-time location evaluator, and an underground fruit damage probability evaluator. This invention solves the problem of inaccurate assessment of underground fruit damage in the current research and development of low-damage harvesting technologies, providing a prerequisite for reducing mechanized harvesting damage and improving harvest quality. It also has significant guiding significance for the research of intelligent harvesting technologies for underground fruits. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the underground fruit damage testing system based on deep neural networks described in this invention;
[0055] Figure 2 This is a schematic diagram showing the arrangement of the strain gauge sensor and soil moisture sensor described in this invention;
[0056] Figure 3 This is a flowchart of the underground fruit damage testing method described in this invention;
[0057] Figure 4 This is a schematic diagram of the two-stage collaborative training process for predicting soil physical properties as described in this invention;
[0058] Figure 5 This is a schematic diagram of the three-stage collaborative training process for predicting underground fruit damage as described in this invention.
[0059] In the diagram: 1-Servo motor; 2-Reducer; 3-Fixing plate; 4-Amplitude adjustment disc; 5-Soil box; 6-Connecting rod; 7-Steel frame profile; 8-Top rod; 9-Guide rail; 5-Soil box; 51-Cover plate; 52-Back plate; 53-Left side plate; 54-Right side plate; 55-Top plate; 56-Bottom plate; 10-Strain gauge sensor; 11-Soil moisture sensor. Detailed Implementation
[0060] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.
[0061] like Figure 1 As shown, the underground fruit damage testing system based on deep neural networks of the present invention includes a steel frame profile 7 and a soil box 5 and a fixing plate 3 fixedly installed on the steel frame profile 7. The soil box 5 consists of a cover plate 51, a back plate 52, a left side plate 53, a right side plate 54, a top plate 55, and a bottom plate 56. The top plate 55 can reciprocate linearly within the soil box 5. The other plates all have mutually matching slots and are mutually fixed by L-shaped brackets. A linear guide rail 9 and a reducer 2 are fixedly installed on the fixing plate 3. The output shaft of the reducer 2 is connected to the amplitude adjustment disk 4 by a key, and the input end of the reducer 2 is connected to the output shaft of the servo motor 1 by a key. The amplitude adjustment disk 4 is hinged to the connecting rod 6. One end of the top rod 8 is hinged to the connecting rod 6, and the other end is hinged to the top plate 55 of the soil box 5. The middle part is fixed on the slider of the linear guide rail 9.
[0062] like Figure 1 As shown, the amplitude adjustment disc 4 has circular holes at different positions with the same size as the circular hole at the head of the connecting rod 6. The disc is hinged by bolts and, together with the top rod 8, the top plate 55, and the guide rail 9, can convert the rotation of the servo motor 1 into the linear reciprocating motion of the top plate 55.
[0063] like Figure 2 As shown, the underground fruit damage testing system based on deep neural networks also includes strain gauge sensors 10 attached to the outer surface of the top rod 8 and several soil moisture sensors 11 evenly distributed inside the soil box 5. Each sensor is connected to a signal acquisition device, which is connected to a computer to transmit data to the computer for subsequent underground fruit damage testing and analysis.
[0064] like Figure 1As shown, after the test begins, the servo motor 1 starts to rotate, driving the top plate 55 to reciprocate linearly via the reducer 2, amplitude adjustment disc 4, connecting rod 6, and top rod 8. The soil moisture sensor 11 detects and transmits real-time moisture data of the soil in the soil box 5, and the strain gauge sensor 10 detects and transmits real-time strain data of the top rod 8. Based on the constructed soil physical property prediction model and underground fruit damage prediction model, the computer predicts the distribution of parameters such as elastic modulus, moisture, and density of the soil in the target test box, as well as the stress distribution, centroid position, and damage probability of the underground fruit. The computer then visualizes the predicted distribution of various physical quantities of the soil and underground fruit, the predicted centroid position of the underground fruit, and the predicted damage probability value.
[0065] The method for testing underground fruit damage based on deep neural networks using the above-mentioned testing system is as follows: Figure 3 As shown, the specific process includes the following:
[0066] Step 1: Correlate the relationship between soil elastic modulus, stress state, and moisture content and density;
[0067] Soil compression samples are prepared and compression tests are conducted using a universal testing machine. The initial elastic modulus of the soil is obtained based on the stress-strain curve. This elastic modulus corresponds to different water contents and densities, thus establishing a pre-established relationship between the soil elastic modulus and water content and density: E = f(x,y), where E is the elastic modulus, x is the water content, and y is the density. This elastic modulus serves as a physical constraint for subsequent predicted values.
[0068] For a given soil, the density increases after compression, water is squeezed out, and the measured moisture content increases. A soil sample is placed in a sealed compression container containing a moisture sensor. A compression load is applied, and the load-displacement and moisture content-displacement curves are recorded, which are then converted into a density-moisture content curve y = f(x). The stress-density-moisture content surface σ = f(x,y) is fitted. The moisture content and stress are marked beforehand, with σ representing stress. This density and stress serve as physical constraints for subsequent predictions.
[0069] Step 2: Initial soil physical data collection, including moisture content, density, elastic modulus, and stress;
[0070] In the initial state, the underground fruit damage testing system is stationary, and the soil troughs in the soil box 5 are evenly distributed. The initial moisture content is obtained by collecting data from several soil moisture sensors 11 evenly distributed in the soil box 5. The initial density is calculated by measuring the mass and volume of the soil in the soil box 5. The initial stress is set to zero in the stationary state. The initial elastic modulus can be given by step 1. The above data are the initial soil condition parameters, which provide a benchmark for subsequent tests.
[0071] Step 3: The underground fruit damage testing system starts working. The servo motor 1 starts, and the power is transmitted to the top plate 55 through the reducer 2, amplitude adjustment disk 4, connecting rod 6, and top rod 8, causing the top plate 55 to perform linear reciprocating motion. During this process, the strain data on the top rod 8 is combined with the elastic modulus and cross-sectional area of the top rod 8 to convert it into pressure data, thereby obtaining cyclic load data. This data, along with the initial soil density and moisture content data, is input into the pre-constructed soil physical property prediction model and underground fruit damage prediction model to obtain the real-time distribution changes of soil density, elastic modulus, stress, and moisture content in the soil box 5 with the load, as well as the real-time location, stress distribution, and damage probability of the underground fruit.
[0072] The soil physical performance prediction model is a model obtained after two stages of collaborative training and testing, including vibration test data of the underground fruit platform and vibration test data of the underground fruit platform, and is used to predict the soil physical performance in the soil box.
[0073] like Figure 4 As shown, the method for constructing the soil physical property prediction model is as follows:
[0074] S1: Construct a soil physical property dataset, which involves collecting test data on soil density, elastic modulus, stress, moisture content, etc. under different loads to form a soil physical property dataset. The test data comes from bench tests.
[0075] S2: Construct a soil physical property prediction generator, which includes an input layer, a hidden layer, and an output layer, as detailed below:
[0076] Input layer: Receives real-time load data (amplitude, frequency) and initial soil physical data (density, moisture content, elastic modulus, stress (initially set to zero stress)) from six dimensions. The number of neurons is set to 6, matching the parameter dimensions one by one. The input data is normalized to eliminate dimensional differences. The Tongtian IQR method is used to remove unreasonable data, ensuring the validity of the input data. The load data is reconstructed into a two-dimensional matrix (time, load) to prepare for the subsequent convolutional layer to extract temporal features.
[0077] Hidden Layers: A hybrid architecture of "convolutional feature extraction + fully connected feature fusion" is adopted. The convolutional feature extraction module sets one convolutional block and 32 5*1 convolutional kernels for the load time series data to extract global load change time series features (such as period). The ReLU activation function is used to achieve non-linear mapping and avoid gradient vanishing. Each convolutional block is followed by a max pooling layer to reduce the parameter scale and suppress overfitting. The fully connected feature fusion module sets a flattening layer to flatten the time series features output by the convolutional module and concatenate them with the initial soil physical data (density, elastic modulus, water content) to form a fused feature vector. Two fully connected network layers are set. The first layer has 128 neurons and uses ReLU activation to further fuse multi-source features. The second layer has 64 neurons and uses LeakReLU activation to enhance the non-linear expression of features. Dropout is added between the fully connected layers to randomly deactivate some neurons and enhance the model's generalization ability.
[0078] Output layer: The output adopts a "spatiotemporal coupling matrix" format, with the output dimension set as: time step × number of spatial grids × 4. The time step matches the temporal length of the input load, and the number of spatial grids corresponds to the number of preset observation points within soil box 5. The four physical quantity channels are density, water content, elastic modulus, and stress. A linear activation function is used to ensure physical meaning: the output values of density and water content are strictly non-negative and within the range of natural soil properties; the output values of elastic modulus and stress are constrained by the pre-obtained values under different water contents and densities, and maintain mechanical correlation with load amplitude, frequency, and initial soil state; the spatial distribution exhibits physical continuity.
[0079] S3: Construct a soil physical performance evaluator based on vibration test data of a test bench without underground fruit;
[0080] A soil physical performance evaluator based on vibration test data from a test bench without underground fruit is used to analyze soil density, moisture content, and other data output by a soil physical performance prediction generator for soils without underground fruit, and outputs corresponding soil elastic modulus, stress, and other mechanical data; the specific construction process is as follows:
[0081] Based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established. The calculated losses are weighted and summed to serve as the evaluation rule for the soil physical performance evaluator based on vibration test data of the platform without underground fruit, and is used for the evaluation of soil physical properties.
[0082] S4: Construct a soil physical performance evaluator based on vibration test data of underground fruit test benches;
[0083] A soil physical property evaluator based on vibration test data of a test bench containing underground fruit is used to analyze soil density, moisture content, and other data output by the soil physical property prediction generator under underground fruit conditions, and output corresponding soil elastic modulus, stress, and other mechanical data; the specific construction process is as follows:
[0084] Based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established. The calculated losses are weighted and summed to serve as the evaluation rule for the soil physical performance evaluator based on vibration test data of the platform containing underground fruits, and is used for the evaluation of soil physical properties.
[0085] S5: Two-stage collaborative training and testing of the soil physical property prediction generator;
[0086] The soil physical performance prediction generator was trained and tested in two stages using a soil physical performance evaluator based on vibration test data of a platform without underground fruit, a soil physical performance evaluator based on vibration test data of a platform with underground fruit, and the soil physical performance dataset in S1. The qualified soil physical performance prediction generator was used as the soil physical performance prediction model. This model is a neural network architecture that takes load data and initial soil physical data as input and predicts soil moisture content, density, elastic modulus, and stress data as output.
[0087] The two-stage collaborative training of the soil physical property prediction generator includes:
[0088] The soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench without underground fruit are compared with the data in the S1 soil physical performance dataset, and the loss between the two is taken as soil loss; the soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench with underground fruit are compared with the data in the S1 soil physical performance dataset, and the loss between the two is taken as underground fruit loss.
[0089] Soil loss is determined based on moisture content, density, elastic modulus, and stress, as shown in equation (1):
[0090]
[0091] In the formula, L Soil Let λ be the soil loss value. i The weighting coefficients for the i-th parameter sum to 1; n = 4, representing four physical quantities; X represents the characteristic parameters of the actual working state; x aveThe soil physical performance evaluator outputs characteristic parameters based on vibration test data of a test bench without underground fruit, including soil moisture content, density, elastic modulus, and stress data.
[0092] The loss of underground fruit is also determined based on moisture content, density, elastic modulus, and stress, as shown in equation (2) below:
[0093]
[0094] In the formula, L Crop w represents the loss value of underground fruit. i Y represents the weight coefficient of the i-th parameter, and their sum is 1; Y represents the characteristic parameters of the actual working state; y ave The soil physical performance evaluator outputs characteristic parameters based on vibration test data of underground fruit test benches. These characteristic parameters include soil density, elastic modulus, stress, and moisture content data.
[0095] After the soil physical property prediction generator is trained, it is tested using sample data that was not used in the training. After evaluation by two evaluators, the soil physical property prediction generator that meets the performance index requirements is qualified and is used as the soil physical property prediction model.
[0096] The underground fruit damage prediction model is a model obtained after two stages of collaborative training and testing using underground fruit damage performance data and damage probability data, used to predict the probability of damage to underground fruits inside the box.
[0097] like Figure 5 As shown, the method for constructing the underground fruit damage prediction model is as follows:
[0098] S1: Construct a dataset using test samples, namely, construct load data and soil physics dataset. The test samples include relevant data collected during bench tests and simulation tests.
[0099] S2: Construct a generator for predicting damage to underground fruits; the specific construction process is as follows:
[0100] The underground fruit damage prediction generator is a neural network architecture that takes real-time load data from bench testing, initial position of underground fruit, and initial soil physical data as input, and outputs current stress distribution of underground fruit, real-time position of underground fruit, and damage probability.
[0101] S3: Construct an assessment tool for the damage performance of underground fruits;
[0102] The underground fruit damage performance evaluator is used to assess the stress state of underground fruits during bench testing and outputs the evaluation results, namely the stress distribution of each layer inside and outside the underground fruit; the specific construction process is as follows:
[0103] Based on the load data and soil physics dataset constructed using S1, the stress distribution and real-time centroid position of each layer inside and outside the underground fruit are predicted to form an expanded dataset. A calculation model for assessing the damage performance of the underground fruit is constructed. The maximum stress loss of each layer of the underground fruit obtained by the calculation model is weighted and summed as the assessment rule for the damage performance of the underground fruit.
[0104] S4: Construct a real-time location evaluator for underground fruits;
[0105] The underground fruit real-time position evaluator is used to evaluate the real-time position status of underground fruits during bench testing and outputs the evaluation result, i.e., the real-time position of the underground fruits; the specific construction process is as follows:
[0106] Data from the underground fruit damage performance evaluator was extracted and combined with simulation data to establish a loss calculation model. The calculated triaxial losses were then summed to serve as the evaluation rule for the real-time location of the underground fruit.
[0107] S5: Construct an underground fruit damage probability evaluator;
[0108] The underground fruit damage probability estimator is used to assess the probability of damage to underground fruits based on their stress state and output the assessment result, i.e., the damage probability of the underground fruit. The specific construction process is as follows:
[0109] Data from the underground fruit damage performance assessment device was extracted and combined with bench test and simulation data to establish a loss calculation model. This model was then used as a damage probability assessment rule for underground fruits to evaluate the probability of damage.
[0110] S6: Three-stage collaborative training and testing of the underground fruit damage prediction generator;
[0111] Based on the underground fruit damage performance evaluator, underground fruit real-time location evaluator, underground fruit damage probability evaluator, and the load data and soil physics dataset constructed by S1, the underground fruit damage prediction generator undergoes three-stage collaborative training and testing using underground fruit damage performance data, real-time location data, and damage probability data. The underground fruit damage prediction generator that passes the test is used as the damage prediction model. The three-stage collaborative training and testing of the underground fruit damage prediction generator using underground fruit damage performance data, real-time location data, and damage probability data includes:
[0112] The maximum stress of the inner and outer layers of the fruit output by the underground fruit damage performance evaluator is compared with the simulation data under the same conditions, and the loss between the two is taken as the stress assessment loss; the real-time position output by the underground fruit real-time position evaluator is compared with the simulation data, and the loss between the two is taken as the position assessment loss; the damage probability of the inner and outer layers of the underground fruit output by the underground fruit damage probability evaluator is compared with the simulation test data and the bench test data, and the loss between the two is taken as the probability assessment loss.
[0113] The stress assessment loss is determined based on the maximum stress in each layer of the underground fruit structure, as shown in equation (3) below:
[0114]
[0115] In the formula, L S For stress assessment loss value, α j is the weighting coefficient; m is the number of layers in the underground fruit, j is the layer number; S jave The maximum stress simulation value for each layer of the underground fruit under the same conditions is s. jave These are the maximum stresses of each layer of the underground fruit output by the underground fruit damage performance assessment device;
[0116] The location loss assessment is determined based on the real-time location of the underground fruit during the bench test, as shown in equation (4) below:
[0117]
[0118] In the formula, L L The location is used to evaluate the loss value; α, β, and γ are the weighting coefficients in the range [0,1], and x Ct y Ct z Ct X represents the triaxial coordinates of the centroid of the underground fruit at time t, output by the real-time position estimator of the underground fruit. Ct Y Ct Z Ct The values of the centroid of the underground fruit at time t are obtained from simulation calculations; T is the final time.
[0119] The probability assessment of loss is determined based on the maximum stress and compressive shear strength of each layer of the underground fruit, as shown in equation (5) below:
[0120] L P =(pP ave ) 2 ( 5)
[0121] In the formula, L P To assess the loss value using probability; P aveLet be the probability of damage to underground fruits under the same conditions; p is the damage probability output by the underground fruit damage probability estimator, and its expression is shown in the following formula (6):
[0122]
[0123] In the formula, β j The weight coefficients of the j-th layer of underground fruit are 1; [σ j ] represents the equivalent allowable strength of the actual underground fruit layer j, given by the power law criterion based on shear and compressive strength; σ sj σ represents the maximum equivalent strength of the j-th layer of the actual underground fruit, given by simulation values under identical conditions; j The maximum equivalent stress of the j-th layer of the underground fruit is predicted.
[0124] Step 4: The computer will visualize the real-time distribution changes of soil moisture content, density, elastic modulus, and stress with load in soil box 5 obtained in Step 3, as well as the real-time location, stress distribution, and damage probability of underground fruit.
[0125] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for testing damage to underground fruits based on deep neural networks, characterized in that, The process includes the following: Step 1: Correlate the relationship between soil elastic modulus and moisture content and density, as well as the relationship between stress state and moisture content and density, as physical constraints for subsequent predictions; Soil compression tests were conducted using a universal testing machine. The initial elastic modulus of the soil was obtained based on the stress-strain curve. This elastic modulus corresponds to different moisture contents and densities, thus establishing the relationship between the soil elastic modulus and moisture content and density. Soil samples were placed in a sealed compression container with an internal moisture sensor, and a compression load was applied. The load-displacement and moisture content-displacement curves were recorded and converted into density-moisture content curves. The stress-density-moisture content surface was then fitted. Step 2: Collect initial soil physical data to provide a baseline for subsequent tests; In the initial state, the underground fruit damage test system is stationary. The initial moisture content is obtained by the data collected by the soil moisture sensor (11). The initial density is obtained by measuring the mass and volume of the soil in the soil box (5). The initial stress is zero. The initial elastic modulus is obtained from step 1. Step 3: The underground fruit damage test system is working. The servo motor (1) is started, driving the top plate (55) to make linear reciprocating motion. During this process, after receiving the sensor detection data, the computer terminal converts the strain data on the top rod (8) into pressure data by combining the elastic modulus and cross-sectional area of the top rod (8), thereby obtaining cyclic load data. This data, along with the initial soil density and moisture content data, is input into the pre-constructed soil physical performance prediction model and underground fruit damage prediction model to obtain the real-time distribution changes of soil density, elastic modulus, stress, and moisture content in the soil box (5) with the load, as well as the real-time position, stress distribution, and damage probability of the underground fruit. Step 4: Visualize the results obtained in Step 3 on the computer. In step 3, the method for constructing the underground fruit damage prediction model is as follows: Using data collected from bench tests and simulation tests, a load data and soil physics dataset was constructed. Construct a generator to predict underground fruit damage; Construct a damage performance assessment tool for underground fruits; Construct a real-time location estimator for underground fruits; Construct a probability estimator for damage to underground fruits; Based on the underground fruit damage performance evaluator, underground fruit real-time location evaluator, underground fruit damage probability evaluator, and load data and soil physics dataset, the underground fruit damage prediction generator is trained and tested in three stages using underground fruit damage performance data, real-time location data, and damage probability data. The underground fruit damage prediction generator that passes the test is used as the damage prediction model. The three-stage collaborative training and testing includes: comparing the maximum stress of the inner and outer layers of the fruit output by the underground fruit damage performance evaluator with the simulation data under the same conditions, and using the loss between the two as the stress assessment loss. The real-time position output by the underground fruit real-time position evaluator is compared with the simulation data, and the loss between the two is taken as the position evaluation loss; the damage probability of the inner and outer layers of the underground fruit output by the underground fruit damage probability evaluator is compared with the simulation test data and bench test data, and the loss between the two is taken as the probability evaluation loss.
2. The method for testing underground fruit damage based on deep neural networks according to claim 1, characterized in that, In step 3, the method for constructing the soil physical property prediction model is as follows: Data on soil density, elastic modulus, stress, and moisture content were collected from bench tests under different loads to form a soil physical properties dataset. A soil physical property prediction generator is constructed, which includes an input layer, a hidden layer, and an output layer; Input layer: Receives parameters from six dimensions, including real-time load data from the test bench and initial soil physical data. The number of neurons is matched one-to-one with the parameter dimensions. Normalizes the input data, removes unreasonable data, and reconstructs the load data into a two-dimensional matrix. Hidden layer: Adopts a hybrid architecture of convolutional feature extraction + fully connected feature fusion; The convolutional feature extraction module sets one convolutional block and 32 5*1 convolutional kernels for the load time series data to extract global load change time series features. The ReLU activation function is used to achieve non-linear mapping, and each convolutional block is connected to a max pooling layer. The fully connected feature fusion module uses a flattening layer to flatten the temporal features output from the convolutional module and concatenate them with the initial soil physical data to form a fused feature vector. It employs a two-layer fully connected network: the first layer has 128 neurons, which further fuse multi-source features using the ReLU activation function; the second layer has 64 neurons, which use LeakReLU activation to enhance the non-linear expression of features. Dropout is added between the fully connected layers to randomly deactivate some neurons, enhancing the model's generalization ability. Output layer: The spatiotemporal coupling matrix is adopted, and the output dimension is set as: time step × number of spatial grids × 4; the time step is matched with the temporal length of the input load, and the number of spatial grids corresponds to the number of observation points preset in the soil box (5). The four physical quantity channels are density, water content, elastic modulus and stress, respectively. The linear activation function is used to ensure the physical meaning: the output values of density and water content are non-negative and within the range of natural soil properties. The output values of elastic modulus and stress are constrained by the values obtained in advance under different water content and density, and maintain mechanical correlation with load amplitude, frequency and initial soil state. The spatial distribution shows physical continuity. Construct a soil physical performance evaluator based on vibration test data of a test bench excluding underground fruits; Construct a soil physical performance evaluator based on vibration test data of a test bench containing underground fruits; A soil physical performance prediction generator was trained and tested in two stages using a soil physical performance evaluator based on vibration test data of a test platform without underground fruit, a soil physical performance evaluator based on vibration test data of a test platform with underground fruit, and a soil physical performance dataset. The qualified soil physical performance prediction generator was used as the soil physical performance prediction model. This model is a neural network architecture that takes load data and initial soil physical data as input and outputs predicted soil moisture content, density, elastic modulus, and stress data.
3. The method for testing underground fruit damage based on deep neural networks according to claim 2, characterized in that, The construction process of the soil physical performance evaluator based on vibration test data of the platform without underground fruit is as follows: Based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established, and the calculated losses are weighted and summed as the evaluation rules of the soil physical performance evaluator based on vibration test data of the platform without underground fruit, which is used for the evaluation of soil physical properties. The process of constructing a soil physical performance evaluator based on vibration test data of a test bench containing underground fruits is as follows: Based on the distribution data of soil physical and mechanical parameters output by the soil physical performance prediction generator, a loss calculation model is established, and the calculated losses are weighted and summed to serve as the evaluation rules for the soil physical performance evaluator based on vibration test data of a test bench containing underground fruits, which is used to evaluate soil physical properties.
4. The method for testing underground fruit damage based on deep neural networks according to claim 2, characterized in that, The two-stage collaborative training of the soil physical property prediction generator includes: The soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench without underground fruit are compared with the data in the soil physical performance dataset, and the loss between the two is taken as soil loss; the soil moisture content, density, elastic modulus, and stress data output by the soil physical performance evaluator based on vibration test data of the test bench with underground fruit are compared with the data in the soil physical performance dataset, and the loss between the two is taken as underground fruit loss.
5. The method for testing underground fruit damage based on deep neural networks according to claim 1, characterized in that, The stress assessment loss is determined based on the maximum stress in each layer of the underground fruit structure, as shown in the following formula: ； In the formula, For stress assessment loss values, These are the weighting coefficients; The number of layers in an underground fruit. This refers to the sequence number of the layer; The values represent the maximum stress simulation values for each layer of the underground fruit under the same conditions. These are the maximum stresses of each layer of the underground fruit output by the underground fruit damage performance assessment device.
6. The method for testing underground fruit damage based on deep neural networks according to claim 1, characterized in that, The location assessment loss is determined based on the real-time location of the underground fruit during the simulation test, as shown in the following formula: ； In the formula, To assess the loss value for the location; , , The weighting coefficients are [0,1]. , , The output of the real-time position estimator for underground fruits is the triaxial coordinate value of the centroid of the underground fruit at time t. , , These are the triaxial coordinates of the centroid of the underground fruit at time t, obtained from simulation calculations. This is the final time.
7. The method for testing underground fruit damage based on deep neural networks according to claim 1, characterized in that, The probability assessment loss is determined based on the maximum stress and compressive shear strength of each layer of the underground fruit, as shown in the following formula: ； In the formula, To assess the loss value based on probability; This represents the probability of damage to underground fruits under the same conditions. The damage probability output by the underground fruit damage probability estimator is expressed as follows: ； In the formula, For underground fruit Layer weight coefficients; For actual underground fruit The equivalent allowable strength of the layer; For the actual underground fruit The maximum equivalent strength of the layer; For the predicted underground fruit The maximum equivalent stress of the layer; The number of layers in an underground fruit. This is the sequence number for the layer.
8. The method for testing underground fruit damage based on deep neural networks according to claim 1, characterized in that, The underground fruit damage prediction generator is a neural network architecture that takes real-time load data from bench testing, initial position of underground fruit, and initial soil physical data as input, and outputs current stress distribution of underground fruit, real-time position of underground fruit, and damage probability. The process of constructing the underground fruit damage performance evaluator is as follows: based on load data and soil physics dataset, predict the stress distribution and real-time centroid position of each layer inside and outside the underground fruit to form an expanded dataset. A computational model for assessing the damage performance of underground fruits is constructed. The maximum stress loss of each layer of underground fruits is weighted and summed using this model as the assessment rule for the damage performance of underground fruits. The process of constructing the real-time location evaluator for underground fruits is as follows: extract the data output by the damage performance evaluator for underground fruits, combine it with simulation data, establish a loss calculation model, and perform a weighted summation of the calculated triaxial losses as the evaluation rule for the real-time location of underground fruits. The process of constructing the underground fruit damage probability evaluator is as follows: extract the data output by the underground fruit damage performance evaluator, combine it with bench test and simulation data, establish a loss calculation model, and use this model as the damage probability evaluation rule for underground fruits to evaluate the damage probability of underground fruits.
9. A deep neural network-based underground fruit damage testing system for implementing the deep neural network-based underground fruit damage testing method of claim 1, characterized in that, Includes a steel frame profile (7) and a soil box (5) and a fixing plate (3) fixedly installed on the steel frame profile (7); a linear guide rail (9) and a reducer (2) are fixedly installed on the fixing plate (3). The output shaft of the reducer (2) is keyed to the amplitude adjustment disc (4), and the input end of the reducer (2) is keyed to the output shaft of the servo motor (1); the amplitude adjustment disc (4) is hinged to the connecting rod (6), the other end of the connecting rod (6) is hinged to the top rod (8), and the other end of the top rod (8) is hinged to the top plate (55) of the soil box (5). Under the drive of the servo motor (1), the top plate (55) moves linearly back and forth in the soil box (5); strain gauge sensors (10) are pasted on the outer surface of the top rod (8), and multiple soil moisture sensors (11) are evenly distributed in the soil box (5). The strain gauge sensors (10) and the soil moisture sensors (11) are all connected to the computer through a signal acquisition device; Based on the constructed soil physical property prediction model and underground fruit damage prediction model, the computer predicts the parameter distribution of the soil in the target test box, including the distribution of elastic modulus, moisture and density parameters. At the same time, it predicts the stress distribution, centroid location and damage probability data of the underground fruit and displays them visually.