Intelligent agricultural machine state monitoring method and system

Abstract

The invention relates to the field of intelligent monitoring, in particular to an intelligent agricultural machine state monitoring method and system, and the method comprises the following steps: agricultural machine operation data collection, signal preprocessing, feature extraction, deep feature learning, state monitoring and fault diagnosis. According to the method, during feature extraction, a GhostNet lightweight convolutional network is adopted to extract high-frequency vibration signal features, a statistical feature extraction method is adopted for low-frequency operation signals, and high-frequency features are combined for fusion; during state monitoring, a dynamic graph convolutional neural network and bidirectional LSTM are adopted to model a state monitoring process, spatial dependency and time dynamic evolution of an agricultural machine state are captured, a hyper-parameter search mechanism based on differential evolution optimization is designed, high-fitness individuals are stored through a PCM memory pool, manual intervention is reduced, and generalization ability is improved. The system comprises a data acquisition module, a signal preprocessing module, a feature extraction module, a deep feature learning module, a state monitoring module and a fault diagnosis module.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring, and in particular to an intelligent agricultural machinery status monitoring method and system. Background Art

[0002] With the continuous improvement of the level of modern agricultural mechanization, the widespread application of agricultural machinery and equipment has greatly improved agricultural production efficiency. However, the operating status of agricultural machinery and equipment directly affects the efficiency and economic benefits of agricultural operations. Most of the existing agricultural machinery status monitoring methods rely on traditional signal processing methods, such as Fourier transform or wavelet transform, which can only extract limited frequency domain or time-frequency features and cannot fully explore the deep-level features in high-frequency vibration signals and low-frequency operation signals; traditional status monitoring methods are based on simple machine learning algorithms, which are difficult to capture complex time dependencies and cannot effectively model the dynamic evolution process of agricultural machinery operating status. In addition, the monitoring model lacks an intelligent optimization mechanism, the training process converges slowly, and the monitoring accuracy is limited. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides an intelligent agricultural machinery state monitoring method and system. In feature extraction, the GhostNet lightweight convolutional network is used to extract high-frequency vibration signal features, and the low-frequency operation signal is extracted by a statistical feature extraction method, which is combined with high-frequency features for fusion to ensure the effective use of multi-source signals and improve the comprehensive performance of the monitoring model. In state monitoring, a dynamic graph convolutional neural network + bidirectional LSTM modeling state monitoring process is adopted to capture the spatial dependence and temporal dynamic evolution of the agricultural machinery state, improve the accuracy of state monitoring, design a hyperparameter search mechanism based on differential evolution optimization, store high fitness individuals through the PCM memory pool, adaptively adjust hyperparameters such as learning rate, number of network layers, and number of neurons, reduce manual intervention, improve generalization ability, introduce an archive disturbance mechanism, and when the PCM memory pool is not significantly optimized within multiple iteration cycles, trigger a global search to improve the efficiency of hyperparameter search and make the state monitoring model more stable.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent agricultural machinery status monitoring method, the method comprising the following steps:

[0005] Step S1: collecting operation data of agricultural machinery, installing high-frequency vibration sensors and low-frequency signal sensors at key parts of agricultural machinery to collect high-frequency vibration signals and low-frequency operation signals of agricultural machinery respectively, wherein the key parts include bearings, gears and engines;

[0006] Step S2: signal preprocessing, using wavelet transform to perform multi-scale decomposition on the high-frequency vibration signal, and using smoothing filtering and denoising on the low-frequency operation signal to obtain the preprocessed high-frequency vibration signal and low-frequency operation signal;

[0007] Step S3: Feature extraction: GhostNet lightweight convolution is used to extract high-frequency features from the preprocessed high-frequency vibration signal, and the features are enhanced by combining ensemble learning and coordinate attention mechanism. The statistical feature extraction method is used to extract low-frequency features from the preprocessed low-frequency operation signal, and the agricultural machinery operation features are obtained by fusion.

[0008] Step S4: deep feature learning, using deep learning and self-supervised learning framework, combined with autoencoder to conduct deep learning on agricultural machinery operation features to obtain deep features;

[0009] Step S5: state monitoring, building a state monitoring model based on deep features, analyzing the operation state of the agricultural machinery in real time, and obtaining state monitoring results;

[0010] Step S6: Fault diagnosis, constructing a fault diagnosis model, analyzing the status detection results, and obtaining analysis results and risk prediction.

[0011] Further, step S4 specifically includes the following steps:

[0012] Step S41: data preparation, the operation characteristics of agricultural machinery are sorted into a feature matrix in chronological order, the rows represent time steps, the columns represent different features, and the operation characteristics of agricultural machinery at each time step are regarded as a node;

[0013] Step S42: self-supervised target construction, generating two sets of positive samples through data enhancement, and using contrast loss function to learn the node discrimination ability;

[0014] Step S43: constructing a spatiotemporal dependency structure, using a sliding window method to divide the agricultural machinery operation characteristics, constructing a time series data set, calculating a similarity matrix between time steps, and constructing a dynamic adjacency matrix;

[0015] Step S44: autoencoder reconstruction, constructing an autoencoder including an encoder and a decoder, encoder dimension reduction, decoder reconstruction, and setting the optimization goal to minimize the reconstruction loss;

[0016] Step S45: deep feature extraction, using a dynamic graph convolutional neural network combined with LSTM to obtain deep features;

[0017] Step S46: Output depth features.

[0018] Further, step S45 specifically includes the following contents:

[0019] Dynamic graph convolutional neural network part: input the dynamic adjacency matrix and feature matrix, update the node features through dynamic graph convolution operation, extract the spatial dependency, and calculate the characteristic representation of agricultural machinery operation at each time step;

[0020] LSTM module: It uses a two-layer bidirectional LSTM to input the time series dataset extracted from the dynamic graph convolutional neural network, model the time dependency, calculate the long-term and short-term features, and extract the deep features;

[0021] Joint training: An end-to-end training strategy is adopted to partially extract spatial dependencies through a dynamic graph convolutional neural network, and then the temporal dependencies are modeled through LSTM.

[0022] Further, step S5, constructing a condition monitoring model based on the deep features, specifically includes the following steps:

[0023] Step S51: constructing state features, combining the deep features of step S4 to construct a time series feature vector, and processing it by normalization;

[0024] Step S52: Initialize the condition monitoring model, input the time series feature vector processed in step S51, and use LSTM to build the condition monitoring model;

[0025] Step S53: Population initialization: randomly initialize an initial population, where each individual represents a set of hyperparameter configurations, and each individual is represented as follows: ;

[0026] in, For the Individuals, is the learning rate, is the number of network layers, is the number of neurons in each layer, and is the control parameter of differential evolution;

[0027] Calculate the fitness of each individual in the initial population and obtain the fitness value;

[0028] Step S54: Establish a PCM memory pool, set the storage capacity to M, store the first M individuals with the highest current fitness values ​​into the PCM memory pool, set the maximum number of iterations and the global iteration variable, as shown below: ; when When , execute step S55 to step S57;

[0029] in, is a global iteration variable, is the maximum number of iterations, set to 200;

[0030] Step S55: Adaptive search, performing global search and local search, specifically including the following contents:

[0031] Global search: Perform standard differential evolution on each individual in the initial population. The formula used is as follows: ;

[0032] in, , and is an individual randomly selected from the initial population, New individuals generated by global search;

[0033] Local search: Guide mutation of individuals in the PCM memory pool with a probability of 50%. The formula used is as follows: ;

[0034] in, is an individual randomly selected from the PCM memory pool, New individuals generated by local search;

[0035] Step S56: Population update, calculation and The fitness value of Fitness value higher than , then replace Entering the initial population, if If the fitness of is higher than any individual in the PCM memory pool, it is replaced to obtain the replaced initial population and the replaced PCM memory pool;

[0036] Step S57: archive disturbance mechanism, monitoring the changes of the PCM memory pool, and triggering the archive disturbance mechanism when there is no optimization after 50 iterations;

[0037] Step S58: Output the result. When , the iteration ends and the individual with the highest fitness in the final PCM pool is output as the optimal hyperparameter combination.

[0038] The present invention provides an intelligent agricultural machinery state monitoring system, comprising a data acquisition module, a signal preprocessing module, a feature extraction module, a deep feature learning module, a state monitoring module and a fault diagnosis module, which specifically includes the following contents:

[0039] The data acquisition module collects high-frequency vibration signals and low-frequency operation signals of the agricultural machinery;

[0040] The signal preprocessing module preprocesses the high-frequency vibration signal and the low-frequency operation signal;

[0041] The feature extraction module uses GhostNet lightweight convolution to extract high-frequency features from the preprocessed high-frequency vibration signal, and combines ensemble learning and coordinate attention mechanism to enhance features. It uses statistical feature extraction method to extract low-frequency features from the preprocessed low-frequency operation signal, and fuses them to obtain agricultural machinery operation features.

[0042] The deep feature learning module uses a deep learning and self-supervised learning framework, combined with an autoencoder, to perform deep learning on the operation features of agricultural machinery to obtain deep features;

[0043] The state monitoring module constructs a state monitoring model based on deep features, analyzes the operation state of the agricultural machinery in real time, and obtains state monitoring results;

[0044] The fault diagnosis module constructs a fault diagnosis model, analyzes the state detection results, and obtains analysis results and risk predictions.

[0045] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0046] (1) When extracting features, GhostNet lightweight convolutional network is used to extract high-frequency vibration signal features. Statistical feature extraction method is used for low-frequency operation signals, which are combined with high-frequency features for fusion to ensure the effective use of multi-source signals and improve the comprehensive performance of the monitoring model.

[0047] (2) In condition monitoring, a dynamic graph convolutional neural network + bidirectional LSTM model is used to model the condition monitoring process, capture the spatial dependence and temporal dynamic evolution of the agricultural machinery status, improve the accuracy of condition monitoring, and design a hyperparameter search mechanism based on differential evolution optimization. The PCM memory pool is used to store high-fitness individuals, and hyperparameters such as the learning rate, number of network layers, and number of neurons are adaptively adjusted to reduce manual intervention and improve generalization ability. An archive disturbance mechanism is introduced. When the PCM memory pool is not significantly optimized within multiple iteration cycles, a global search is triggered to improve the efficiency of the hyperparameter search and make the condition monitoring model more stable. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a flow chart of an intelligent agricultural machinery status monitoring method provided by the present invention;

[0049] Figure 2 is a schematic flow chart of step S4;

[0050] Figure 3 A module schematic diagram of an intelligent agricultural machinery status monitoring system provided by the present invention.

[0051] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0053] In the description of the present invention, it should be understood that terms such as “upper”, “lower”, “front”, “back”, “left”, “right”, “top”, “bottom”, “inside” and “outside” indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0054] Example 1, see Figure 1 The present invention provides an intelligent agricultural machinery status monitoring method, which comprises the following steps:

[0055] Step S1: collecting operation data of agricultural machinery, installing high-frequency vibration sensors and low-frequency signal sensors at key parts of agricultural machinery to collect high-frequency vibration signals and low-frequency operation signals of agricultural machinery respectively, wherein the key parts include bearings, gears and engines;

[0056] Step S2: signal preprocessing, using wavelet transform to perform multi-scale decomposition on the high-frequency vibration signal, and using smoothing filtering and denoising on the low-frequency operation signal to obtain the preprocessed high-frequency vibration signal and low-frequency operation signal;

[0057] Step S3: Feature extraction: GhostNet lightweight convolution is used to extract high-frequency features from the preprocessed high-frequency vibration signal, and the features are enhanced by combining ensemble learning and coordinate attention mechanism. The statistical feature extraction method is used to extract low-frequency features from the preprocessed low-frequency operation signal, and the agricultural machinery operation features are obtained by fusion.

[0058] Step S4: deep feature learning, using deep learning and self-supervised learning framework, combined with autoencoder to conduct deep learning on agricultural machinery operation features to obtain deep features;

[0059] Step S5: state monitoring, building a state monitoring model based on deep features, analyzing the operation state of the agricultural machinery in real time, and obtaining state monitoring results;

[0060] Step S6: Fault diagnosis, constructing a fault diagnosis model, analyzing the status detection results, and obtaining analysis results and risk prediction.

[0061] Example 2, see Figure 2 This embodiment is based on the above embodiment, and step S4 specifically includes the following steps:

[0062] Step S41: data preparation, the operation characteristics of agricultural machinery are sorted into a feature matrix in chronological order, the rows represent time steps, the columns represent different features, and the operation characteristics of agricultural machinery at each time step are regarded as a node;

[0063] Step S42: self-supervised target construction, generating two sets of positive samples through data enhancement, and using contrast loss function to learn the node discrimination ability;

[0064] Step S43: constructing a spatiotemporal dependency structure, using a sliding window method to divide the agricultural machinery operation characteristics, constructing a time series data set, calculating a similarity matrix between time steps, and constructing a dynamic adjacency matrix;

[0065] Step S44: autoencoder reconstruction, constructing an autoencoder including an encoder and a decoder, encoder dimension reduction, decoder reconstruction, and setting the optimization goal to minimize the reconstruction loss;

[0066] Step S45: deep feature extraction, using a dynamic graph convolutional neural network combined with LSTM to obtain deep features;

[0067] Step S46: Output depth features.

[0068] In this embodiment, the "John Deere 6M Series" tractor is used as the experimental object, and the engine speed, fuel consumption rate, hydraulic system pressure, transmission temperature, vehicle speed and ambient temperature are collected;

[0069] Data was collected through sensors on the tractor for 2 hours, with a total of 72,000 time steps;

[0070] Data storage is represented as follows: Time step Engine speed Fuel consumption rate Hydraulic system pressure Transmission temperature Speed Ambient temperature 1 1500 7.5 8.0 50 6.2 18 2 1510 7.6 8.2 50.2 6.3 18.1 3 1498 7.4 8.1 50.1 6.1 18 4 1520 7.8 8.3 50.3 6.4 18.2 5 1480 7.3 8.0 50.0 6.0 18

[0071] Set the window size to 5 seconds (50 time steps) and the sliding step size to 1 second;

[0072] We get 72000-50+1 = 71951 time segments;

[0073] Gaussian noise is added to the engine speed, and some data are time-shifted to simulate clock drift;

[0074] Using NT-Xent;

[0075] Calculate the Euclidean distance between different time steps and set the threshold: establish a connection when the Euclidean distance is < 0.5, and construct a dynamic adjacency matrix;

[0076] Autoencoder Reconstruction:

[0077] Encoder: 2 fully connected layers: FC1: 6×128, FC2: 128×64;

[0078] Decoder: 2 fully connected layers: FC1: 64×128, FC2: 128×6;

[0079] The optimization objective adopts the mean square error function;

[0080] A 3-layer dynamic graph convolutional neural network is used: the hidden layer sizes are 64, 32, and 16;

[0081] 2-layer LSTM: 64 hidden units, 50 time steps, and 128 output dimensions;

[0082] LSTM output is 128-dimensional;

[0083] The dynamic graph convolutional neural network outputs 16 dimensions;

[0084] The final deep feature vector is 144 dimensions;

[0085] Clustering using K-Means:

[0086] K=3

[0087] Category: normal state, slight abnormality (such as 5% increase in fuel consumption), severe abnormality (such as 20% decrease in hydraulic system pressure);

[0088] Calculate the Mahalanobis distance and give an alarm when the Mahalanobis distance > 3.

[0089] Embodiment 3: This embodiment is based on the above embodiment, and step S45 specifically includes the following contents:

[0090] Dynamic graph convolutional neural network part: input the dynamic adjacency matrix and feature matrix, update the node features through dynamic graph convolution operation, extract the spatial dependency, and calculate the characteristic representation of agricultural machinery operation at each time step;

[0091] LSTM module: It uses a two-layer bidirectional LSTM to input the time series dataset extracted from the dynamic graph convolutional neural network, model the time dependency, calculate the long-term and short-term features, and extract the deep features;

[0092] Joint training: An end-to-end training strategy is adopted to partially extract spatial dependencies through a dynamic graph convolutional neural network, and then the temporal dependencies are modeled through LSTM.

[0093] Embodiment 4: This embodiment is based on the above embodiment. Step S5 constructs a state monitoring model based on deep features, specifically including the following steps:

[0094] Step S51: constructing state features, combining the deep features of step S4 to construct a time series feature vector, and processing it by normalization;

[0095] Step S52: Initialize the condition monitoring model, input the time series feature vector processed in step S51, and use LSTM to build the condition monitoring model;

[0096] Step S53: Population initialization: randomly initialize an initial population, where each individual represents a set of hyperparameter configurations, and each individual is represented as follows: ;

[0097] in, For the Individuals, is the learning rate, is the number of network layers, is the number of neurons in each layer, and is the control parameter of differential evolution;

[0098] Calculate the fitness of each individual in the initial population and obtain the fitness value;

[0099] Step S54: Establish a PCM memory pool, set the storage capacity to M, store the first M individuals with the highest current fitness values ​​into the PCM memory pool, set the maximum number of iterations and the global iteration variable, as shown below: ; when When , execute step S55 to step S57;

[0100] in, is a global iteration variable, is the maximum number of iterations, set to 200;

[0101] Step S55: Adaptive search, performing global search and local search, specifically including the following contents:

[0102] Global search: Perform standard differential evolution on each individual in the initial population. The formula used is as follows: ;

[0103] in, , and is an individual randomly selected from the initial population, New individuals generated by global search;

[0104] Local search: Guide mutation of individuals in the PCM memory pool with a probability of 50%. The formula used is as follows: ;

[0105] in, is an individual randomly selected from the PCM memory pool, New individuals generated by local search;

[0106] Step S56: Population update, calculation and The fitness value of Fitness value higher than , then replace Entering the initial population, if If the fitness of is higher than any individual in the PCM memory pool, it is replaced to obtain the replaced initial population and the replaced PCM memory pool;

[0107] Step S57: archive disturbance mechanism, monitoring the changes of the PCM memory pool, and triggering the archive disturbance mechanism when there is no optimization after 50 iterations;

[0108] Step S58: Output the result. When , the iteration ends and the individual with the highest fitness in the final PCM pool is output as the optimal hyperparameter combination.

[0109] In this embodiment, the codes used are as follows:

[0110] import numpy as np

[0111] import tensorflow as tf

[0112] from tensorflow.keras.models import Sequential

[0113] from tensorflow.keras.layers import LSTM, Dense

[0114] from sklearn.preprocessing import MinMaxScaler

[0115] # Step S51: State feature construction, time series feature vector normalization

[0116] def normalize_data(data):

[0117] scaler = MinMaxScaler(feature_range=(0, 1))

[0118] return scaler.fit_transform(data)

[0119] # Step S52: Build LSTM model

[0120] def create_lstm_model(input_shape, layers, neurons):

[0121] model = Sequential()

[0122] model.add(LSTM(neurons, input_shape=input_shape, return_sequences=True))

[0123] for _ in range(layers - 1):

[0124] model.add(LSTM(neurons, return_sequences=True))

[0125] model.add(Dense(1))

[0126] model.compile(optimizer='adam', loss='mean_squared_error')

[0127] return model

[0128] # Step S53: Population initialization, generating initial population

[0129] def initialize_population(pop_size, param_bounds):

[0130] population = []

[0131] for _ in range(pop_size):

[0132] # Randomly initialize each hyperparameter individual

[0133] l_r = np.random.uniform(param_bounds['l_r'][0], param_bounds['l_r'][1]) # learning rate

[0134] layers = np.random.randint(param_bounds['layers'][0], param_bounds['layers'][1]) # Number of network layers

[0135] neurons = np.random.randint(param_bounds['neurons'][0],param_bounds['neurons'][1]) # Number of neurons in each layer

[0136] F = np.random.uniform(param_bounds['F'][0], param_bounds['F'][1]) # Differential evolution control parameter F

[0137] CR = np.random.uniform(param_bounds['CR'][0], param_bounds['CR'][1]) # Differential evolution control parameter CR

[0138] individual = (l_r, layers, neurons, F, CR)

[0139] population.append(individual)

[0140] return np.array(population)

[0141] # Calculate fitness function: based on model training error

[0142] def fitness_function(model, data, labels, l_r, layers, neurons):

[0143] model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=l_r), loss='mean_squared_error')

[0144] model.fit(data, labels, epochs=10, batch_size=32, verbose=0)

[0145] loss = model.evaluate(data, labels, verbose=0)

[0146] return loss

[0147] # Step S54: Create a PCM memory pool

[0148] def initialize_pcm(population, data, labels):

[0149] fitness_values ​​= []

[0150] for individual in population:

[0151] l_r, layers, neurons, F, CR = individual

[0152] model = create_lstm_model((data.shape[1], data.shape[2]),layers, neurons)

[0153] fitness = fitness_function(model, data, labels, l_r, layers,neurons)

[0154] fitness_values.append(fitness)

[0155] pcm = sorted(zip(population, fitness_values), key=lambda x: x[1])[:10] # Select the first 10 individuals with the best fitness

[0156] return pcm

[0157] # Differential evolution global search

[0158] def global_search(population, F):

[0159] r1, r2, r3 = np.random.choice(len(population), 3, replace=False)

[0160] v_global = population[r1][:-1] + F * (population[r2][:-1] -population[r3][:-1]) # Global mutation

[0161] return v_global

[0162] # Differential evolution local search

[0163] def local_search(population, F, pcm):

[0164] r1, r2, r3 = np.random.choice(len(pcm), 3, replace=False)

[0165] v_local = pcm[r1][:-1] + F * (pcm[r2][:-1] - pcm[r3][:-1]) # Local mutation

[0166] return v_local

[0167] # Step S55: Adaptive search

[0168] def adaptive_search(population, F, pcm):

[0169] new_population = []

[0170] for individual in population:

[0171] v_global = global_search(population, F)

[0172] v_local = local_search(population, F, pcm)

[0173] new_population.append(v_global if np.random.rand() < 0.5 elsev_local)

[0174] return np.array(new_population)

[0175] # Step S56: Population update

[0176] def update_population(population, pcm, new_population):

[0177] updated_population = []

[0178] for i in range(len(population)):

[0179] if new_population[i][-1] < population[i][-1]: # If the new individual has a higher fitness, replace it

[0180] updated_population.append(new_population[i])

[0181] else:

[0182] updated_population.append(population[i])

[0183] return updated_population

[0184] # Step S57: Archive disturbance mechanism

[0185] def archive_disturbance(pcm, t, max_no_improvement=50):

[0186] if t % max_no_improvement == 0:

[0187] # Trigger the disturbance mechanism

[0188] pcm = np.random.shuffle(pcm)

[0189] return pcm

[0190] # Step S58: Output the optimal hyperparameter combination

[0191] def output_best_parameters(pcm):

[0192] best_individual = min(pcm, key=lambda x: x[1]) # Select the individual with the best fitness

[0193] return best_individual

[0194] # Main execution process

[0195] def main():

[0196] # Assume data and labels have been loaded

[0197] data = np.random.rand(100, 10, 5) # Sample data, 100 samples, 10 time steps, 5 features

[0198] labels = np.random.rand(100, 1) # Example labels

[0199] param_bounds = {

[0200] 'l_r': (0.0001, 0.01), # learning rate range

[0201] 'layers': (1, 5), # Network layer range

[0202] 'neurons': (50, 200), # The number of neurons in each layer

[0203] 'F': (0.5, 1.0), # Differential evolution F parameter range

[0204] 'CR': (0.5, 1.0) # Differential evolution CR parameter range

[0205] }

[0206] population_size = 30

[0207] population = initialize_population(population_size, param_bounds)

[0208] pcm = initialize_pcm(population, data, labels)

[0209] max_iterations = 200

[0210] for t in range(1, max_iterations + 1):

[0211] new_population = adaptive_search(population, 0.8, pcm) # Set F to 0.8

[0212] population = update_population(population, pcm, new_population)

[0213] pcm = archive_disturbance(pcm, t)

[0214] if t == max_iterations:

[0215] best_parameters = output_best_parameters(pcm)

[0216] print(f"The best hyperparameter combination: {best_parameters}")

[0217] if __name__ == "__main__":

[0218] main().

[0219] Example 5, see Figure 3 The present invention provides an intelligent agricultural machinery status monitoring system, including a data acquisition module, a signal preprocessing module, a feature extraction module, a deep feature learning module, a status monitoring module and a fault diagnosis module, which specifically includes the following contents:

[0220] The data acquisition module collects high-frequency vibration signals and low-frequency operation signals of the agricultural machinery;

[0221] The signal preprocessing module preprocesses the high-frequency vibration signal and the low-frequency operation signal;

[0222] The feature extraction module uses GhostNet lightweight convolution to extract high-frequency features from the preprocessed high-frequency vibration signal, and combines ensemble learning and coordinate attention mechanism to enhance features. It uses statistical feature extraction method to extract low-frequency features from the preprocessed low-frequency operation signal, and fuses them to obtain agricultural machinery operation features.

[0223] The deep feature learning module uses a deep learning and self-supervised learning framework, combined with an autoencoder, to perform deep learning on the operation features of agricultural machinery to obtain deep features;

[0224] The state monitoring module constructs a state monitoring model based on deep features, analyzes the operation state of the agricultural machinery in real time, and obtains state monitoring results;

[0225] The fault diagnosis module constructs a fault diagnosis model, analyzes the state detection results, and obtains analysis results and risk predictions.

[0226] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0227] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0228] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. An intelligent agricultural machinery status monitoring method, characterized in that: The method comprises the following steps: Step S1: collecting operation data of agricultural machinery, installing high-frequency vibration sensors and low-frequency signal sensors at key parts of agricultural machinery to collect high-frequency vibration signals and low-frequency operation signals of agricultural machinery respectively, wherein the key parts include bearings, gears and engines; Step S2: signal preprocessing, using wavelet transform to perform multi-scale decomposition on the high-frequency vibration signal, and using smoothing filtering and denoising on the low-frequency operation signal to obtain the preprocessed high-frequency vibration signal and low-frequency operation signal; Step S3: Feature extraction: GhostNet lightweight convolution is used to extract high-frequency features from the preprocessed high-frequency vibration signal, and the features are enhanced by combining ensemble learning and coordinate attention mechanism. The statistical feature extraction method is used to extract low-frequency features from the preprocessed low-frequency operation signal, and the agricultural machinery operation features are obtained by fusion. Step S4: deep feature learning, using deep learning and self-supervised learning framework, combined with autoencoder to conduct deep learning on agricultural machinery operation features to obtain deep features; Step S5: state monitoring, building a state monitoring model based on deep features, analyzing the operation state of the agricultural machinery in real time, and obtaining state monitoring results; Step S6: Fault diagnosis, constructing a fault diagnosis model, analyzing the status detection results, and obtaining analysis results and risk prediction.

2. The intelligent agricultural machinery status monitoring method according to claim 1, characterized in that: Building a condition monitoring model based on deep features includes the following steps: Step S51: constructing state features, combining the deep features of step S4 to construct a time series feature vector, and processing it by normalization; Step S52: Initialize the condition monitoring model, input the time series feature vector processed in step S51, and use LSTM to build the condition monitoring model; Step S53: Population initialization: randomly initialize an initial population, where each individual represents a set of hyperparameter configurations, and each individual is represented as follows: ； in, For the Individuals, is the learning rate, is the number of network layers, is the number of neurons in each layer, and is the control parameter of differential evolution; Calculate the fitness of each individual in the initial population and obtain the fitness value; Step S54: Establish a PCM memory pool, set the storage capacity to M, store the first M individuals with the highest current fitness values ​​into the PCM memory pool, set the maximum number of iterations and the global iteration variable, as shown below: ； when When , execute step S55 to step S57; in, is a global iteration variable, is the maximum number of iterations, set to 200; Step S55: Adaptive search, performing global search and local search, specifically including the following contents: Global search: Perform standard differential evolution on each individual in the initial population. The formula used is as follows: ； in, , and is an individual randomly selected from the initial population, New individuals generated by global search; Local search: Guide mutation of individuals in the PCM memory pool with a probability of 50%. The formula used is as follows: ； in, is an individual randomly selected from the PCM memory pool, New individuals generated by local search; Step S56: Population update, calculation and The fitness value of Fitness value higher than , then replace Entering the initial population, if If the fitness of is higher than any individual in the PCM memory pool, it is replaced to obtain the replaced initial population and the replaced PCM memory pool; Step S57: archive disturbance mechanism, monitoring the changes of the PCM memory pool, and triggering the archive disturbance mechanism when there is no optimization after 50 iterations; Step S58: Output the result. When , the iteration ends and the individual with the highest fitness in the final PCM pool is output as the optimal hyperparameter combination.

3. An intelligent agricultural machinery status monitoring system, used to implement an intelligent agricultural machinery status monitoring method as described in any one of claims 1-2, characterized in that: It includes data acquisition module, signal preprocessing module, feature extraction module, deep feature learning module, state monitoring module and fault diagnosis module.

4. The intelligent agricultural machinery status monitoring system according to claim 3 is characterized in that: The data acquisition module collects high-frequency vibration signals and low-frequency operation signals of the agricultural machinery; The signal preprocessing module preprocesses the high-frequency vibration signal and the low-frequency operation signal; The feature extraction module uses GhostNet lightweight convolution to extract high-frequency features from the preprocessed high-frequency vibration signal, and combines ensemble learning and coordinate attention mechanism to enhance features. It uses statistical feature extraction method to extract low-frequency features from the preprocessed low-frequency operation signal, and fuses them to obtain agricultural machinery operation features. The deep feature learning module uses a deep learning and self-supervised learning framework, combined with an autoencoder, to perform deep learning on the operation features of agricultural machinery to obtain deep features; The state monitoring module constructs a state monitoring model based on deep features, analyzes the operation state of the agricultural machinery in real time, and obtains state monitoring results; The fault diagnosis module constructs a fault diagnosis model, analyzes the state detection results, and obtains analysis results and risk predictions.

Images