Fault pre-identification method for gearbox of traveling system of corn cob and stalk harvester

Abstract

The invention discloses a method for pre-identifying faults of a gearbox of a traveling system of a corn cob and stalk harvester. The method comprises an information collection module, an information processing module and a fault pre-identification module for use with a certain type of corn cob and stalk harvester, wherein the information collection module is electrically connected to the information processing module via a wire, and the information processing module is electrically connected to the fault pre-identification module via a wire; the information collection module processes the collected vehicle data and stores the data in a database, and manages the database; without adding hardware to the corn cob and stalk harvester, the invention only utilizes existing firmware, processes historical data and real-time data through the method, and completes pre-identification of faults of a gearbox of a traveling system of the harvester, and the fault identification rate can reach 75%. At the current stage, the implementation cost of intelligent agricultural machinery is reduced and the production efficiency of agricultural machinery is improved.

Description

Technical Field

[0001] The invention relates to the technical field of agricultural automation machinery fault prediction, in particular to a method for pre-identifying faults of a gear box of a walking system of a corn cob and stalk harvester. Background Art

[0002] With the changes in my country's population structure and the further growth of urban population, the level of agricultural mechanization and intelligence must be continuously improved to meet the needs of agricultural development in the new era. As an important agricultural tool, the corn cob and stalk harvester has a relatively high gearbox failure rate in its travel system. If a failure occurs during the busy farming season, it will cause unnecessary losses to agricultural production.

[0003] With the development of artificial intelligence and the Internet of Things, intelligent diagnosis of mechanical failures has become increasingly popular. As an important production machinery, the intelligence of agricultural machinery has received more and more attention.

[0004] To realize the intelligentization of agricultural machinery, on the one hand, it is necessary to rely on a large number of high-precision sensors to accurately perceive the state of the machinery, thereby laying the foundation for subsequent processing strategies. On the other hand, it is necessary to process the data, find out the internal correlation, and realize the differentiated processing of the machinery. The gear box of the walking system of the corn cob and stalk harvester is an important part of agricultural machinery. The main causes of its failure are generally long-term overload operation, frequent power switching in a short period of time, resulting in gear impact, etc.

[0005] However, due to cost considerations, harvesters are currently equipped with only a small number of sensors to monitor their basic status, and the monitoring effect is difficult to guarantee. Summary of the invention

[0006] The purpose of the present invention is to utilize the engine power sensor and vehicle speed sensor that the harvester is equipped with, and identify the health status of the gearbox of the corn harvester's travel system through an intelligent algorithm arranged in the cloud, so as to promptly inform the driver of reasonable operation and remind maintenance personnel to inspect and maintain, thereby avoiding the occurrence of gearbox failures in the harvester's travel system, and providing a low-cost solution for the intelligence of corn harvesters.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A gearbox fault pre-identification system for a corn cob and stalk harvester travel system, comprising an information collection module, an information processing module and a fault pre-identification module for use with a certain type of corn cob and stalk harvester, wherein the information collection module is electrically connected to the information processing module via a wire, and the information processing module is electrically connected to the fault pre-identification module via a wire;

[0009] The information collection module processes the collected vehicle data, stores it in a database, and manages the database;

[0010] The information processing module extracts feature data using the information in the database and organizes it into a data set and input data for the prediction module;

[0011] The fault pre-identification module uses the data set to determine the parameters of the prediction network and then evaluates the vehicle status;

[0012] The information collection module collects data in a manner of sending vehicle data to a central database via a vehicle terminal. The types of data collected are instantaneous engine power, instantaneous vehicle speed, total vehicle mileage, and vehicle operating status.

[0013] As a further solution of the present invention: the method for data processing by the information collection module is: taking 5 consecutive minutes as an interval (there is monitoring data for at least 3 minutes in the interval, otherwise no record is made), averaging the power and speed data collected by the sensor within 5 minutes (Formula 1 or 2) and the engine work (Formula 5), ​​and then comparing the average with the data of the previous interval to obtain the change (Formulas 7 and 8).

[0014]

[0015] Δp=|p 前 -p 当前 | (6)

[0016] Δυ=|υ 前 -υ 当前 | (7)

[0017] In the formula, is the overall mean; d max-i is the i-th maximum data in the interval; is the mean of the larger values ​​in the interval; is the mean of the smaller values ​​in the interval; d i is the data collected in the interval; np is the number of data collected in the interval; d min-i is the i-th minimum data in the interval; d sum is the accumulated work done up to the current time; d sum-b is the accumulated work done in the previous recording time; Δp is the power change compared with the previous interval; p 前 is the power value of the previous interval; p 当前 is the power value of the current interval; Δυ is the speed change compared with the previous interval; ν 前 is the power value of the previous interval; v 当前 is the power value of the current interval;

[0018] The information collection module processes the data and stores it in the database. The stored contents are time, instantaneous power average within the interval, instantaneous speed average within the interval, power change, speed change, accumulated total work, current total mileage, and vehicle operating status.

[0019] As a further solution of the present invention: the information collection module manages the data in the database, and each vehicle saves all 2160 hours of data. The data exceeding the limit is updated every hour, and the earliest hour of data in the database is deleted according to the first-in-first-out principle; for vehicles with data retention time less than 480 hours, it is marked that subsequent data processing will not be carried out until the data volume meets the requirement.

[0020] The information processing module provides network parameter training data and input data for "fault prediction at the exact moment" and "fault prediction in the next 6 hours".

[0021] As a further solution of the present invention: the method of providing a data set for the cluster center of "punctual fault pre-identification" and the SVM classifier parameter determination of "fault pre-identification in the next 6 hours" by the information processing module is: extracting the data of the latest 480 hours from the database, and extracting the mean square value (Formula 8), kurtosis (Formula 9) and permutation entropy feature data from the four types of data: instantaneous power mean, instantaneous speed mean, power change, and speed change:

[0022]

[0023] Where, d msv is the mean square value; d j Data collected for information purposes; is the mean of each type of monitoring information; n is the number of each type of monitoring information; d kur is the kurtosis; σ is the standard deviation of each type of monitoring information;

[0024] Permutation entropy does not consider the numerical value of the time series, but compares and analyzes adjacent sample points to obtain corresponding feature information. As a sequence feature extraction method, it can better capture the slight changes of the sequence. For a time series u(1),u(2),....,u(N) (u(i) is the discrete value of the signal data), its specific implementation steps are:

[0025] Step 1) Set the parameter embedding dimension m=6 and the delay parameter τ=3 in the permutation entropy calculation formula; the data length is N=5760 and arranged in time;

[0026] Step 2) Reconstruct the four types of data, namely, the instantaneous power mean, the instantaneous speed mean within the interval, the power change, and the speed change, respectively, to obtain each subsequence X(i), where X(i) = u(i), u(i+τ), ..., u(i+(m-1)τ);

[0027] Where: X(i) is the i-th reconstruction vector;

[0028] Step 3) Sort Xi in ascending order, i.e. u(i+(j1-1)τ)≤...≤u(i+(j m -1)τ), obtain the index sequence Y(i){j1,j2,...,j m};

[0029] Where: Y(i) is the sorted vector of the i-th X(i); m is the sequence number after sorting;

[0030] Step 4) Calculate the probability P(i) of the index sequence appearing (Formula 10).

[0031]

[0032] Where: Num(·) represents the number;

[0033] Step 5) Calculate its Shannon entropy H PE (Formula 11).

[0034]

[0035] Step 6) Finally calculate the permutation entropy H NPE (Formula 12).

[0036]

[0037] After the feature data is extracted, a label is assigned according to the repair status reported by the after-sales service and the vehicle owner. If the vehicle is in normal condition, the label value is 1; if the vehicle is faulty, the label value is 2;

[0038] Each vehicle has 14 characteristic values ​​(mean square value, kurtosis, and permutation entropy of speed mean; mean square value, kurtosis, and permutation entropy of power mean; mean square value, kurtosis, and permutation entropy of power variation; mean square value, kurtosis, and permutation entropy of speed variation; accumulated power; and accumulated history) and 1 state label. The characteristic data of all vehicles are aggregated and normalized (Formula 13):

[0039]

[0040] Where: x′ is the normalized value; x is the original value; min(x) is the minimum value in the feature data; max(x) is the maximum value in the feature data.

[0041] As a further solution of the present invention: the method of providing a data set for the LSTM network of "fault pre-identification in the next 6 hours" by the information processing module is: extracting the data of the most recent 54 hours (instantaneous power and change, instantaneous speed and change) from the database, taking hours as the unit, selecting 3 medians of 12 data in the hour and calculating their average values ​​as the data of the hour, and arranging to obtain time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change; taking the first 48 hours as the input data of LSTM, and the last 6 hours as the output data of LSTM; randomly shuffling the order of the data set and passing it to the fault pre-identification module;

[0042] The method by which the information processing module provides input data for the SVM classification of “fault prediction in the next 6 hours” is: extracting the 14-dimensional feature data of a single vehicle in the specified time period;

[0043] The method by which the information processing module provides input data for the LSTM network of "fault prediction in the next 6 hours" is: extract the data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change of a single vehicle in the past 48 hours, take hours as the unit, select the 3 medians of the 12 data in the hour and calculate their average value as the data for the hour, organize the time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change, and pass the 4 time series to the fault prediction module.

[0044] As a further solution of the present invention: the fault pre-identification module is divided into "punctual fault pre-identification" and "fault pre-identification in the next 6 hours". The method of the fault pre-identification module "punctual fault pre-identification" is: determine the cluster center point k of the vehicle with good state at 23:40 every day -1 and the faulty vehicle cluster center k -2 , from 1 o'clock, the cluster center is used to predict vehicle failures at the right time;

[0045] The method for determining the cluster center of the "punctual fault pre-identification" of the fault pre-identification module is:

[0046] Step 1) Let the number of vehicles be na, then the 14-dimensional feature data of all vehicles will form an na*14 matrix d 14 :

[0047]

[0048] Where: d i-j is the jth data of the i-th vehicle; d i is the data vector of the i-th vehicle;

[0049] Step 2) De-standardize the data to obtain the matrix d i-new ;

[0050]

[0051] In the formula, s i is the standard deviation of each row of original data; d i Represents each row of original data; represents the mean of each row of original data; Step 3) calculate the covariance matrix R;

[0052]

[0053] Step 4) Calculate the eigenvector set of the covariance matrix R by the formula And select the first 3 rows to form the feature vector

[0054] Where: λ is the eigenvalue of the covariance matrix R; is the eigenvector set of the covariance matrix R;

[0055]

[0056] Step 5) Multiplying the 14-dimensional feature data of all vehicles by the feature vector to form a na*14 matrix can obtain a new n*3 matrix d3, where each row represents the features of a vehicle after the feature dimension is compressed;

[0057]

[0058] Step 6) Use the formula to calculate the good vehicle cluster center point k -1 and the faulty vehicle cluster center k -2 ;

[0059]

[0060] Where: k x1 ,k y1 ,k z1 is the xyz coordinate value of the good vehicle cluster center; d i-x1 ,d i-y1 ,d i-z1 is the feature after the good vehicle feature dimension is compressed; k x2 ,k y2 ,k z2 is the xyz coordinate value of the faulty vehicle cluster center; d i-x2 ,d i-y2 ,d i-z1 It is the feature after the feature dimension of the faulty vehicle is compressed.

[0061] As a further solution of the present invention: the method of "punctual fault pre-identification" of the fault pre-identification module is: starting from 1 o'clock every day, the 14-dimensional feature data of nearly 480 hours provided by the data processing module is used to and the determined eigenvectors Calculate the vehicle's current 3D feature data FV x , calculate the Euclidean distance dist(FV x ,k -1 ) and the Euclidean distance between the feature data and the cluster center of the faulty vehicle (FV x ,k -2 ); finally the probability of failure p is given fault-x , the formula is:

[0062]

[0063] Where: dist(·) is the spatial distance between two data.

[0064] When the probability value is between 0% and 50%, it indicates "good operating status"; when the probability value is between 50% and 70%, it indicates "low risk of failure, please drive according to the guidance requirements"; when the probability value is between 70% and 90%, it indicates "medium risk of failure, please stop the vehicle for maintenance at an appropriate time"; when the probability value is above 90%, it indicates "high risk of failure, please stop the vehicle for maintenance immediately";

[0065] The method for determining the LSTM network parameters of the "fault pre-identification in the next 6 hours" of the fault pre-identification module is as follows: at 23:00 every day, LSTM uses the data set provided by the data processing module to set the initial parameters of LSTM: the input layer dimension is 1 dimension; the LSTM parameter sets the hidden unit to 200, the output mode uses 'last', the state activation function uses 'tanh', and the gate activation function uses 'sigmoid'; a 3-layer LSTM network is used; a drop layer is connected after the 3 layers, and the ratio is set to 30%; then 2 fully connected layers are connected, and finally the output layer is 1 dimension; the LSTM is trained for 450 rounds, and the accuracy of the verification set is 85%, which means that the network training is completed, and then the parameters of the LSTM prediction network of instantaneous power, instantaneous power change, instantaneous speed, and instantaneous speed change are stored;

[0066] The LSTM network prediction method of the "fault prediction in the next 6 hours" of the fault prediction module is: every day (1 o'clock, 5 o'clock, 9 o'clock, 13 o'clock, 17 o'clock, 21 o'clock), for vehicles with 6 hours of monitoring data in the past 24 hours, use the data provided by the data processing module, and use the LSTM prediction network to predict the data for the next 6 hours, and output the instantaneous power, instantaneous power change, instantaneous speed, and instantaneous speed change in units of hours.

[0067] As a further solution of the present invention: the method for determining the SVM classifier of the "fault pre-identification in the next 6 hours" of the fault pre-identification module is: at 23:00 every day, using the training set provided by the data processing module, the type is selected as C-support vector classifier; the kernel function is selected as RBF radial basis kernel function; the SVM classifier is trained using the optimized bat algorithm, and the network training is considered to be completed when the accuracy rate reaches 90%.

[0068] As a further solution of the present invention: the optimized bat algorithm is: in the formula of the standard bat algorithm individual for local search, an adaptive speed weight factor ω is introduced i , improve the optimization speed.

[0069]

[0070] In the formula, rand(·) generates a (0,1) random number; x i is the position of bat i; x mean is the mean of the global position;

[0071] Calculate ω i Finally, update the speed formula of the bat algorithm:

[0072]

[0073] In the formula, is the speed of bat i at time t+1; is the speed of bat i at time t; is the position of bat i at time t; x * is the optimal position of the bat population at time t; f i is the pulse emission frequency of bat i at time t.

[0074] The SVM fault pre-identification method of the "fault pre-identification in the next 6 hours" of the fault pre-identification module is: every day (1 o'clock, 5 o'clock, 9 o'clock, 13 o'clock, 17 o'clock, 21 o'clock), use the 14-dimensional feature data extracted from the data of the past 480 hours and the 6 hours predicted by LSTM provided by the data processing module, input the SVM classifier for fault pre-identification, if the prediction result is that a fault may occur in the next 6 hours, then send "it is recommended to stop for maintenance during a break" to the vehicle equipment terminal; if the prediction result is that a fault is unlikely to occur in the next 6 hours, then send "it is recommended to maintain the current operating status" to the vehicle equipment terminal.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] Without adding any hardware to the corn cob and stalk harvester, only the existing firmware is used to process historical data and real-time data through this method to complete the pre-identification of the gearbox fault in the harvester's travel system. The fault identification rate can reach 75%. At the current stage, the implementation cost of agricultural machinery intelligence is reduced and the production efficiency of agricultural machinery is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is the overall flow chart of the gearbox fault pre-identification method for the traveling system of a corn cob and stalk harvester.

[0078] Figure 2 Flowchart for training an SVM classifier to optimize the Bat Algorithm. DETAILED DESCRIPTION

[0079] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0080] See also Figures 1-2 In an embodiment of the present invention, a gearbox fault pre-identification system for a corn cob and stalk harvester travel system includes an information collection module, an information processing module and a fault pre-identification module for use with a certain type of corn cob and stalk harvester, wherein the information collection module is electrically connected to the information processing module through a wire, and the information processing module is electrically connected to the fault pre-identification module through a wire;

[0081] The information collection module processes the collected vehicle data, stores it in a database, and manages the database;

[0082] The information processing module extracts feature data using the information in the database and organizes it into a data set and input data for the prediction module;

[0083] The fault pre-identification module uses the data set to determine the parameters of the prediction network and predict the fault;

[0084] The method in which the information processing module provides a data set for the fault pre-identification module is:

[0085] Step 1: Extract the latest 480 hours of data from the database, and extract the mean square value, kurtosis and permutation entropy feature data from the four types of data: instantaneous power mean, instantaneous speed mean, power change, and speed change;

[0086] Step 2: After the feature data is extracted, a label is assigned according to the repair status reported by the after-sales service and the vehicle owner. If the vehicle is in normal condition, the label value is 1; if the vehicle is faulty, the label value is 2; each vehicle has 14 feature values ​​(mean square value, kurtosis, and permutation entropy of the mean speed; mean square value, kurtosis, and permutation entropy of the mean power; mean square value, kurtosis, and permutation entropy of the power change; mean square value, kurtosis, and permutation entropy of the speed change; cumulative power; cumulative history) and 1 state label. The feature data of all vehicles are aggregated and normalized;

[0087] The parameters for extracting the permutation entropy feature data are set to embedding dimension m=6, delay parameter τ=3; data length is N=5760;

[0088] The method in which the information processing module provides a data set for the LSTM network of the fault pre-identification module is:

[0089] Step 1: Extract the data of the most recent 54 hours (instantaneous power and change, instantaneous speed and change) from the database, take the hour as the unit, select the 3 medians of the 12 data in the hour and calculate their average value as the data of the hour, and sort out the time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change;

[0090] Step 2: Take the first 48 hours as the input data of LSTM, and the last 6 hours as the output data of LSTM;

[0091] Step 3: Randomly shuffle the order of the data set and pass it to the fault pre-identification module;

[0092] The steps of determining the cluster center of the fault pre-identification module are as follows:

[0093] Step 1) Let the number of vehicles be na, then the 14-dimensional feature data of all vehicles will form an na*14 matrix d 14 :

[0094]

[0095] Where: d i-j is the jth data of the i-th vehicle; d i is the data vector of the i-th vehicle;

[0096] Step 2) De-standardize the data to obtain the matrix d i-new ;

[0097]

[0098] In the formula, s i is the standard deviation of each row of original data; d i Represents each row of original data; represents the mean of each row of original data; Step 3) calculate the covariance matrix R;

[0099]

[0100] Step 4) Calculate the eigenvector set of the covariance matrix R by the formula And select the first 3 rows to form the feature vector

[0101] Where: λ is the eigenvalue of the covariance matrix R; is the eigenvector set of the covariance matrix R;

[0102]

[0103] Step 5) Multiplying the 14-dimensional feature data of all vehicles by the feature vector to form a na*14 matrix can obtain a new n*3 matrix d3, where each row represents the features of a vehicle after the feature dimension is compressed;

[0104]

[0105] Step 6) Use the formula to calculate the good vehicle cluster center point k -1 and the faulty vehicle cluster center k -2 ;

[0106]

[0107]

[0108] Where: k x1 ,k y1 ,k z1 is the xyz coordinate value of the good vehicle cluster center; d i-x1 ,d i-y1 ,d i-z1 is the feature after the good vehicle feature dimension is compressed; k x2 ,k y2 ,k z2 is the xyz coordinate value of the faulty vehicle cluster center; d i-x2 ,d i-y2 ,d i-z1 It is the feature after the feature dimension of the faulty vehicle is compressed;

[0109] The method for determining the LSTM network parameters of the fault pre-identification module is:

[0110] Step 1: LSTM uses the data set provided by the data processing module and sets the initial parameters of LSTM: the input layer dimension is 1D; the LSTM parameter sets the hidden unit to 200, the output mode uses 'last', the state activation function uses 'tanh', and the gate activation function uses 'sigmoid'; a 3-layer LSTM network is used; a drop layer is connected after the 3 layers, and the ratio is set to 30%; then two fully connected layers are connected, and the final output layer is 1D;

[0111] Step 2: Train LSTM for 450 rounds. When the accuracy of the validation set is 85%, the network training is considered complete.

[0112] The method for determining the SVM classifier of the fault pre-identification module is:

[0113] Step 1: Use the training set provided by the data processing module, select the C-support vector classifier as the type, and select the RBF radial basis kernel function as the kernel function;

[0114] Step 2: Use the optimized bat algorithm to train the SVM classifier. When the accuracy reaches 90%, the network training is considered complete.

[0115] The optimized bat algorithm is:

[0116] Step 1) Introduce the adaptive speed weight factor ω into the formula for local search of the standard bat algorithm individual i , improve the optimization speed.

[0117]

[0118] In the formula, rand(·) generates a (0,1) random number; x i is the current position; x mean is the mean of the global position;

[0119] Step 2) Calculate ω i Finally, update the speed formula of the bat algorithm:

[0120]

[0121] In the formula, is the speed of bat i at time t+1; is the speed of bat i at time t; is the position of bat i at time t; x * is the optimal position of the bat population at time t; f i is the pulse emission frequency of bat i at time t;

[0122] Example

[0123] Take 5 consecutive minutes as an interval (at least 3 minutes of monitoring data in the interval, otherwise no record), and calculate the power average of the power and speed data collected by the sensor within 5 minutes Speed ​​mean Engine work Then the power change Δp is obtained i and the velocity change Δv i .

[0124] After the data is initially processed, the updated data is stored in the database. The stored contents are time i, the instantaneous power mean value in the interval, The average instantaneous speed in the interval Power change Δp i , speed change Δv i , cumulative work P sum-i 、Current total mileage L sum-i .

[0125] At the exact time, the latest 480 hours of data is used to extract 14-dimensional feature data, and then the 14-dimensional special data is reduced to 3 dimensions using the determined feature vector to obtain the current 3-dimensional feature data FV of the vehicle. x , calculate the Euclidean distance dist(FV x ,k -1 ) and the Euclidean distance between the feature data and the cluster center of vehicles in good condition dist(FV x ,k -2 ), the probability of failure p fault-x .

[0126] When the probability value is between 0% and 50%, it indicates "good operating status"; when the probability value is between 50% and 70%, it indicates "low risk of failure, please drive according to the guidance requirements"; when the probability value is between 70% and 90%, it indicates "medium risk of failure, please stop for maintenance at an appropriate time"; when the probability value is above 90%, it indicates "high risk of failure, please stop for maintenance immediately"; every day (1:00, 5:00, 9:00, 13:00, 17:00, 21:00), for vehicles with 6 hours of monitoring data in the past 24 hours, the latest 48 hours of data in the database are used to calculate the mean and change of instantaneous power and instantaneous speed in one hour in units of 1 hour as the monitoring data for this hour, and the 48 groups (4 in each group) of data obtained are used as the input of LSTM network prediction, and 6 groups (4 in each group) of prediction values ​​for the next 6 hours are output. 14-dimensional feature data are extracted from the nearly 480 hours of data in the database and the 6 hours of estimated data, and the SVM classifier is used to estimate whether the vehicle may fail in the next 6 hours.

[0127] At 23:00 every day, for vehicles with at least 54 hours of monitoring data in the database within the past 120 hours, the instantaneous power and instantaneous speed of the vehicles in the past 54 hours are sorted out and estimated. The instantaneous power average, instantaneous power change, instantaneous speed average, and instantaneous speed change within one hour are calculated in units of 1 hour. Each vehicle has 4 groups of data (54 in each group), and the first 48 are used as input data and the last 6 as output data.

[0128] The vehicles are randomly arranged and divided into training set and validation set according to 8:2. Then, the input data and output data of instantaneous power mean, instantaneous power change, instantaneous speed mean, and instantaneous speed change are input into the LSTM network in turn. The network sets the input layer dimension to 1D. The LSTM parameter sets the hidden unit to 200, the output mode to use 'last', the state activation function to use 'tanh', and the gated activation function to use 'sigmoid'. A 3-layer LSTM network is used. A drop layer is connected after the 3 layers, and the ratio is set to 30%. Then, 2 fully connected layers are connected, and the output layer is 1D. After the settings are completed, training is carried out. If the accuracy of the validation set is 85%, the network training is considered complete, and the LSTM network parameters are saved for the next day's prediction.

[0129] At 23:00 every day, for all vehicles in the database with nearly 480 hours of monitoring data, 14-dimensional characteristic data are extracted from the instantaneous power mean, instantaneous power change, instantaneous speed mean, and instantaneous speed change, and labels are assigned according to the vehicle status.

[0130] The vehicles are randomly arranged and divided into training set and validation set in the ratio of 8:2. The training set is then input into the SVM classifier. The parameters for constructing the SVM classifier are as follows: the type uses C-support vector classifier; the kernel function uses RBF radial basis kernel function; the improved bat algorithm is used to quickly determine the hyperparameters of the SVM. When the accuracy reaches 90%, the network training is considered complete, and the SVM classifier parameters are saved for the next day's prediction.

[0131] At 23:40 every day, for all vehicles in the database with nearly 480 hours of monitoring data, the principal component analysis method is used to reduce the dimension of the feature data, and the cluster center point k of vehicles in good condition is calculated using formula 16 -1 and the faulty vehicle cluster center k -2 , the two cluster centers are used as the pre-identification of vehicle failures for the next day.

[0132] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0133] The original technical solution of the present invention

[0134] The present invention utilizes the engine power sensor and vehicle speed sensor that the harvester is equipped with, and identifies the health status of the gearbox of the corn harvester's travel system through an intelligent algorithm arranged in the cloud, so as to promptly inform the driver of reasonable operation and remind maintenance personnel to inspect and maintain, thereby avoiding the occurrence of gearbox failures in the harvester's travel system, and providing a low-cost solution for the intelligence of corn harvesters.

[0135] The technical solution adopted by the present invention to achieve its technical purpose includes three modules, namely, an information collection module, an information processing module, and a fault pre-identification module.

[0136] The information collection module processes the collected vehicle data and stores it in the database, and manages the database; the information processing module uses the information in the database to extract feature data and organizes it into a data set and input data for the recognition module; the fault pre-recognition module uses the data set to determine the parameters of the prediction network and then evaluates the vehicle status.

[0137] The information collection module collects data by sending vehicle data to the central database through the vehicle terminal. The types of data collected are engine instantaneous power, vehicle instantaneous speed, vehicle total mileage, and vehicle operating status.

[0138] The information collection module processes data in the following way: taking 5 consecutive minutes as an interval (there must be monitoring data for at least 3 minutes in the interval, otherwise no record will be made), averaging the power and speed data collected by the sensor within 5 minutes (Formula 1 or 2) and the engine work (Formula 5), ​​and then comparing the average with the data of the previous interval to obtain the change (Formulas 7 and 8).

[0139]

[0140]

[0141] Δp=|p 前 -p 当前 | (29)

[0142] Δυ=|υ 前 -υ 当前 | (30)

[0143] In the formula, is the overall mean; d max-i is the i-th maximum data in the interval; is the mean of the larger values ​​in the interval; is the mean of the smaller values ​​in the interval; d iis the data collected in the interval; np is the number of data collected in the interval; d min-i is the i-th minimum data in the interval; d sum is the accumulated work done up to the current time; d sum-b is the accumulated work done in the previous recording time; Δp is the power change compared with the previous interval; p 前 is the power value of the previous interval; p 当前 is the power value of the current interval; Δυ is the speed change compared with the previous interval; ν 前 is the power value of the previous interval; ν 当前 It is the power value of the current interval.

[0144] The information collection module processes the data and stores it in the database. The stored contents are time, instantaneous power average within the interval, instantaneous speed average within the interval, power change, speed change, accumulated total work, current total mileage, and vehicle operating status.

[0145] The information collection module manages the data in the database. Each vehicle saves all 2160 hours of data. For data exceeding this limit, it is updated every hour and the earliest hour of data in the database is deleted according to the first-in-first-out principle. For vehicles with data retention time of less than 480 hours, the module will mark the vehicle and will not perform subsequent data processing until the data volume meets the requirement.

[0146] The information processing module provides network parameter training data and input data for "fault prediction at the exact moment" and "fault prediction in the next 6 hours".

[0147] The information processing module provides a data set for the cluster center of "on-time fault pre-identification" and the SVM classifier parameter determination of "fault pre-identification in the next 6 hours" by extracting the latest 480 hours of data from the database, and extracting the mean square value (Formula 8), kurtosis (Formula 9) and permutation entropy feature data from the four types of data: instantaneous power mean, instantaneous speed mean, power change, and speed change:

[0148]

[0149] Where, d msv is the mean square value; d j Data collected for information purposes; is the mean of each type of monitoring information; n is the number of each type of monitoring information; d kur is the kurtosis; σ is the standard deviation of each type of monitoring information.

[0150] Permutation entropy does not consider the numerical value of the time series, but compares and analyzes adjacent sample points to obtain corresponding feature information. As a sequence feature extraction method, it can better capture the slight changes of the sequence. For a time series u(1),u(2),....,u(N) (u(i) is the discrete value of the signal data), its specific implementation steps are:

[0151] Step 1) Set the parameter embedding dimension m=6 and the delay parameter τ=3 in the permutation entropy calculation formula; the data length is N=5760 and arranged in time;

[0152] Step 2) Reconstruct the four types of data, namely, the instantaneous power mean, the instantaneous speed mean within the interval, the power change, and the speed change, respectively, to obtain each subsequence X(i), where X(i) = u(i), u(i+τ), ..., u(i+(m-1)τ);

[0153] Where: X(i) is the i-th reconstruction vector;

[0154] Step 3) Sort Xi in ascending order, i.e. u(i+(j1-1)τ)≤...≤u(i+(j m -1)τ), obtain the index sequence Y(i){j1,j2,...,j m};

[0155] Where: Y(i) is the sorted vector of the i-th X(i); m is the sequence number after sorting;

[0156] Step 4) Calculate the probability P(i) of the index sequence appearing (Formula 10).

[0157]

[0158] Where: Num(·) represents the number;

[0159] Step 5) Calculate its Shannon entropy H PE (Formula 11).

[0160]

[0161] Step 6) Finally calculate the permutation entropy H NPE (Formula 12).

[0162]

[0163] After the feature data is extracted, labels are assigned according to the repair status reported by after-sales and vehicle owners. If the vehicle is in normal condition, the label value is 1; if the vehicle is faulty, the label value is 2.

[0164] Each vehicle has 14 characteristic values ​​(mean square value, kurtosis, and permutation entropy of speed mean; mean square value, kurtosis, and permutation entropy of power mean; mean square value, kurtosis, and permutation entropy of power variation; mean square value, kurtosis, and permutation entropy of speed variation; accumulated power; and accumulated history) and 1 state label. The characteristic data of all vehicles are aggregated and normalized (Formula 13):

[0165]

[0166] Where: x′ is the normalized value; x is the original value; min(x) is the minimum value in the feature data; max(x) is the maximum value in the feature data.

[0167] The method by which the information processing module provides a data set for the LSTM network of "fault prediction in the next 6 hours" is as follows: extract the data of the most recent 54 hours (instantaneous power and change, instantaneous speed and change) from the database, take hours as the unit, select the 3 medians of the 12 data within the hour and calculate their average value as the data for the hour, and sort out the time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change; take the first 48 hours as the input data of the LSTM, and the next 6 hours as the output data of the LSTM; and randomly shuffle the order of the data set and pass it to the fault prediction module.

[0168] The method by which the information processing module provides input data for the SVM classification of "fault prediction within the next 6 hours" is to extract the 14-dimensional feature data of a single vehicle within a specified time period.

[0169] The method by which the information processing module provides input data for the LSTM network of "fault prediction in the next 6 hours" is: extract the data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change of a single vehicle in the past 48 hours, take hours as the unit, select the 3 medians of the 12 data in the hour and calculate their average value as the data for the hour, organize the time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change, and pass the 4 time series to the fault prediction module.

[0170] The fault prediction module is divided into "fault prediction at the exact moment" and "fault prediction in the next 6 hours".

[0171] The method of the fault pre-identification module "punctual time fault pre-identification" is: determine the cluster center point k of the good vehicles at 23:40 every day -1 and the faulty vehicle cluster center k -2 , starting from 1 o'clock, the cluster center is used to predict vehicle failures at the right time.

[0172] The method for determining the cluster center of the "punctual fault pre-identification" of the fault pre-identification module is:

[0173] Step 1) Let the number of vehicles be na, then the 14-dimensional feature data of all vehicles will form an na*14 matrix d 14 :

[0174]

[0175] Where: d i-j is the jth data of the i-th vehicle; d i is the data vector of the i-th vehicle;

[0176] Step 2) De-standardize the data to obtain the matrix d i-new ;

[0177]

[0178] In the formula, s i is the standard deviation of each row of original data; d i Represents each row of original data; represents the mean of each row of original data; Step 3) calculate the covariance matrix R;

[0179]

[0180] Step 4) Calculate the eigenvector set of the covariance matrix R by the formula And select the first 3 rows to form the feature vector

[0181] Where: λ is the eigenvalue of the covariance matrix R; is the eigenvector set of the covariance matrix R;

[0182]

[0183] Step 5) Multiplying the 14-dimensional feature data of all vehicles by the feature vector to form a na*14 matrix can obtain a new n*3 matrix d3, where each row represents the features of a vehicle after the feature dimension is compressed;

[0184]

[0185] Step 6) Use the formula to calculate the good vehicle cluster center point k -1 and the faulty vehicle cluster center k -2 ;

[0186]

[0187] Where: k x1 ,k y1 ,k z1is the xyz coordinate value of the good vehicle cluster center; d i-x1 ,d i-y1 ,d i-z1 is the feature after the good vehicle feature dimension is compressed; k x2 ,k y2 ,k z2 is the xyz coordinate value of the faulty vehicle cluster center; d i-x2 ,d i-y2 ,d i-z1 It is the feature after the feature dimension of the faulty vehicle is compressed.

[0188] The fault pre-identification module's "punctual fault pre-identification" method is: starting at 1:00 every day, the 14-dimensional feature data of nearly 480 hours provided by the data processing module is used to and the determined eigenvectors Calculate the vehicle's current 3D feature data FV x , calculate the Euclidean distance dist(FV x ,k -1 ) and the Euclidean distance between the feature data and the cluster center of the faulty vehicle (FV x ,k -2 ); finally gives the probability of failure p fault-x , the formula is:

[0189]

[0190] Where: dist(·) is the spatial distance between two data.

[0191] When the probability value is between 0% and 50%, it prompts "good operating status"; when the probability value is between 50% and 70%, it prompts "low risk of failure, please drive according to the instructions"; when the probability value is between 70% and 90%, it prompts "medium risk of failure, please stop and repair at an appropriate time"; when the probability value is above 90%, it prompts "high risk of failure, please stop and repair immediately".

[0192] The method for determining the LSTM network parameters of the "fault pre-identification in the next 6 hours" of the fault pre-identification module is as follows: at 23:00 every day, LSTM uses the data set provided by the data processing module to set the initial parameters of LSTM: the input layer dimension is 1 dimension; the LSTM parameter sets the hidden unit to 200, the output mode uses 'last', the state activation function uses 'tanh', and the gated activation function uses 'sigmoid'; a 3-layer LSTM network is used; a drop layer is connected after the 3 layers, and the ratio is set to 30%; then 2 fully connected layers are connected, and finally the output layer is 1 dimension; the LSTM is trained for 450 rounds, and the accuracy of the verification set is 85%, which means that the network training is completed, and then the parameters of the LSTM prediction network of instantaneous power, instantaneous power change, instantaneous speed, and instantaneous speed change are stored.

[0193] The LSTM network prediction method of the "fault prediction in the next 6 hours" of the fault prediction module is: every day (1 o'clock, 5 o'clock, 9 o'clock, 13 o'clock, 17 o'clock, 21 o'clock), for vehicles with 6 hours of monitoring data in the past 24 hours, use the data provided by the data processing module, and use the LSTM prediction network to predict the data for the next 6 hours, and output the instantaneous power, instantaneous power change, instantaneous speed, and instantaneous speed change in units of hours.

[0194] The method for determining the SVM classifier of the "fault pre-identification within the next 6 hours" of the fault pre-identification module is: at 23:00 every day, use the training set provided by the data processing module, select the C-support vector classifier as the type; select the RBF radial basis kernel function as the kernel function; use the optimized bat algorithm to train the SVM classifier, and the network training is considered completed when the accuracy reaches 90%.

[0195] The optimized bat algorithm is: In the formula for local search of the standard bat algorithm individual, an adaptive speed weight factor ω is introduced i , improve the optimization speed.

[0196]

[0197] In the formula, rand(·) generates a (0,1) random number; x i is the position of bat i; x mean is the mean of the global position;

[0198] Calculate ω i Finally, update the speed formula of the bat algorithm:

[0199]

[0200] In the formula, is the speed of bat i at time t+1; is the speed of bat i at time t; is the position of bat i at time t; x * is the optimal position of the bat population at time t; f i is the pulse emission frequency of bat i at time t.

[0201] The SVM fault pre-identification method of the "fault pre-identification in the next 6 hours" of the fault pre-identification module is: every day (1 o'clock, 5 o'clock, 9 o'clock, 13 o'clock, 17 o'clock, 21 o'clock), use the 14-dimensional feature data extracted from the data of the past 480 hours and the 6 hours predicted by LSTM provided by the data processing module, input the SVM classifier for fault pre-identification, if the prediction result is that a fault may occur in the next 6 hours, then send "it is recommended to stop for maintenance during a break" to the vehicle equipment terminal; if the prediction result is that a fault is unlikely to occur in the next 6 hours, then send "it is recommended to maintain the current operating status" to the vehicle equipment terminal.

[0202] The present invention proposes a method for pre-identifying faults of the gear box of the traveling system of a corn cob and stalk harvester, which has the beneficial effect of: without adding hardware to the corn cob and stalk harvester, only using existing firmware, through this method, historical data and real-time data are processed to complete the pre-identification of the gear box faults of the harvester's traveling system, and the fault identification rate can reach 75%. At the current stage, the implementation cost of agricultural machinery intelligence is reduced and the production efficiency of agricultural machinery is improved.

[0203] Take 5 consecutive minutes as an interval (at least 3 minutes of monitoring data in the interval, otherwise no record), and calculate the power average of the power and speed data collected by the sensor within 5 minutes Speed ​​mean Engine work Then the power change Δp is obtained i and the velocity change Δv i .

[0204] After the data is initially processed, the updated data is stored in the database. The stored contents are time i, the instantaneous power mean value in the interval, The average instantaneous speed in the interval Power change Δp i , speed change Δv i , cumulative work P sum-i 、Current total mileage L sum-i .

[0205] At the exact time, the latest 480 hours of data is used to extract 14-dimensional feature data, and then the 14-dimensional special data is reduced to 3 dimensions using the determined feature vector to obtain the current 3-dimensional feature data FV of the vehicle. x , calculate the Euclidean distance dist(FV x,k -1 ) and the Euclidean distance between the feature data and the cluster center of vehicles in good condition dist(FV x ,k -2 ), the probability of failure p fault-x .

[0206] When the probability value is between 0% and 50%, it indicates "good operating status"; when the probability value is between 50% and 70%, it indicates "low risk of failure, please drive according to the guidance requirements"; when the probability value is between 70% and 90%, it indicates "medium risk of failure, please stop for maintenance at an appropriate time"; when the probability value is above 90%, it indicates "high risk of failure, please stop for maintenance immediately"; every day (1:00, 5:00, 9:00, 13:00, 17:00, 21:00), for vehicles with 6 hours of monitoring data in the past 24 hours, the latest 48 hours of data in the database are used to calculate the mean and change of instantaneous power and instantaneous speed in one hour in units of 1 hour as the monitoring data for this hour, and the 48 groups (4 in each group) of data obtained are used as the input of LSTM network prediction, and 6 groups (4 in each group) of prediction values ​​for the next 6 hours are output. 14-dimensional feature data are extracted from the nearly 480 hours of data in the database and the 6 hours of estimated data, and the SVM classifier is used to estimate whether the vehicle may fail in the next 6 hours.

[0207] At 23:00 every day, for vehicles with at least 54 hours of monitoring data in the database within the past 120 hours, the instantaneous power and instantaneous speed of the vehicles in the past 54 hours are sorted out and estimated. The instantaneous power average, instantaneous power change, instantaneous speed average, and instantaneous speed change within one hour are calculated in units of 1 hour. Each vehicle has 4 groups of data (54 in each group), and the first 48 are used as input data and the last 6 as output data.

[0208] The vehicles are randomly arranged and divided into training set and validation set according to 8:2. Then, the input data and output data of instantaneous power mean, instantaneous power change, instantaneous speed mean, and instantaneous speed change are input into the LSTM network in turn. The network sets the input layer dimension to 1D. The LSTM parameter sets the hidden unit to 200, the output mode to use 'last', the state activation function to use 'tanh', and the gated activation function to use 'sigmoid'. A 3-layer LSTM network is used. A drop layer is connected after the 3 layers, and the ratio is set to 30%. Then, 2 fully connected layers are connected, and the output layer is 1D. After the settings are completed, training is carried out. If the accuracy of the validation set is 85%, the network training is considered complete, and the LSTM network parameters are saved for the next day's prediction.

[0209] At 23:00 every day, for all vehicles in the database with nearly 480 hours of monitoring data, 14-dimensional characteristic data are extracted from the instantaneous power mean, instantaneous power change, instantaneous speed mean, and instantaneous speed change, and labels are assigned according to the vehicle status.

[0210] The vehicles are randomly arranged and divided into training set and validation set in the ratio of 8:2. The training set is then input into the SVM classifier. The parameters for constructing the SVM classifier are as follows: the type uses C-support vector classifier; the kernel function uses RBF radial basis kernel function; the improved bat algorithm is used to quickly determine the hyperparameters of the SVM. When the accuracy reaches 90%, the network training is considered complete, and the SVM classifier parameters are saved for the next day's prediction.

[0211] At 23:40 every day, for all vehicles in the database with nearly 480 hours of monitoring data, the principal component analysis method is used to reduce the dimension of the feature data, and the cluster center point k of vehicles in good condition is calculated using formula 16 -1 and the faulty vehicle cluster center k -2 , the two cluster centers are used as the pre-identification of vehicle failures for the next day.

Claims

1. A method for pre-identifying faults in a gearbox of a corn cob and stalk harvester travel system, characterized in that: It includes an information collection module, an information processing module and a fault pre-identification module used for a certain type of corn cob and stalk harvester, wherein the information collection module is electrically connected to the information processing module through a wire, and the information processing module is electrically connected to the fault pre-identification module through a wire; The information collection module processes the collected vehicle data, stores it in a database, and manages the database; The information processing module extracts feature data using the information in the database and organizes it into a data set and input data for the prediction module; The fault pre-identification module uses the data set to determine the parameters of the prediction network and predict the fault; The method in which the information processing module provides a data set for the fault pre-identification module is: Step 1: Extract the latest 480 hours of data from the database, and extract the mean square value, kurtosis and permutation entropy feature data from the four types of data: instantaneous power mean, instantaneous speed mean, power change, and speed change; Step 2: After the feature data is extracted, a label is assigned according to the repair status reported by the after-sales service and the vehicle owner. If the vehicle is in normal condition, the label value is 1; if the vehicle is faulty, the label value is 2; each vehicle has 14 feature values: mean square value, kurtosis, and permutation entropy of the mean speed; mean square value, kurtosis, and permutation entropy of the mean power; mean square value, kurtosis, and permutation entropy of the power variation; mean square value, kurtosis, and permutation entropy of the speed variation; accumulated power; accumulated history and 1 state label. The feature data of all vehicles are aggregated and normalized; The steps of determining the cluster center of the fault pre-identification module are as follows: Step 1) Let the number of vehicles be na, then the 14-dimensional feature data of all vehicles will form an na*14 matrix d 14 : Where: d i-j is the jth data of the i-th vehicle; d i is the data vector of the i-th vehicle; Step 2) De-standardize the data to obtain the matrix d i-new ; In the formula, s i is the standard deviation of each row of original data; d i Represents each row of original data; represents the mean of each row of original data; Step 3) calculate the covariance matrix R; Step 4) Calculate the eigenvector set of the covariance matrix R by the formula And select the first 3 rows to form the feature vector Where: λ is the eigenvalue of the covariance matrix R; is the eigenvector set of the covariance matrix R; Step 5) Multiplying the 14-dimensional feature data of all vehicles by the feature vector to form a na*14 matrix can obtain a new n*3 matrix d3, where each row represents the features of a vehicle after the feature dimension is compressed; Step 6) Use the formula to calculate the good vehicle cluster center point k -1 and the faulty vehicle cluster center k -2 ; Where: k x1 ,k y1 ,k z1 is the xyz coordinate value of the good vehicle cluster center; d i-x1 ,d i-y1 ,d i-z1 is the feature after the good vehicle feature dimension is compressed; k x2 ,k y2 ,k z2 is the xyz coordinate value of the faulty vehicle cluster center; d i-x2 ,d i-y2 ,d i-z1 It is the feature after the feature dimension of the faulty vehicle is compressed; The method for determining the LSTM network parameters of the fault pre-identification module is: Step 1: LSTM uses the data set provided by the data processing module and sets the initial parameters of LSTM: the input layer dimension is 1D; the LSTM parameter sets the hidden unit to 200, the output mode uses 'last', the state activation function uses 'tanh', and the gate activation function uses 'sigmoid'; a 3-layer LSTM network is used; a drop layer is connected after the 3 layers, and the ratio is set to 30%; then two fully connected layers are connected, and the final output layer is 1D; Step 2: Train LSTM for 450 rounds. When the accuracy of the validation set is 85%, the network training is considered complete. The method for determining the SVM classifier of the fault pre-identification module is: Step 1: Use the training set provided by the data processing module, select the C-support vector classifier as the type, and select the RBF radial basis kernel function as the kernel function; Step 2: Use the optimized bat algorithm to train the SVM classifier. When the accuracy reaches 90%, the network training is considered complete.

2. The method for pre-identifying faults of the gear box of the traveling system of the corn cob and stalk harvester according to claim 1 is characterized in that: The parameters for extracting the permutation entropy feature data are set as embedding dimension m=6, delay parameter τ=3; and data length N=5760.

3. The method for pre-identifying faults of the gear box of the traveling system of the corn cob and stalk harvester according to claim 1 is characterized in that: The method in which the information processing module provides a data set for the LSTM network of the fault pre-identification module is: Step 1: Extract the data of the most recent 54 hours from the database: instantaneous power and change, instantaneous speed and change, in units of hours, select the 3 medians of the 12 data within the hour and calculate their average value as the data for the hour, and sort out the time series data of instantaneous power, instantaneous speed, instantaneous power change, and instantaneous speed change; Step 2: Take the first 48 hours as the input data of LSTM, and the last 6 hours as the output data of LSTM; Step 3: Randomly shuffle the order of the data set and pass it to the fault pre-identification module.

4. The method for pre-identifying faults of the gear box of the traveling system of the corn cob and stalk harvester according to claim 1 is characterized in that: The optimized bat algorithm is: Step 1) Introduce the adaptive speed weight factor ω into the formula for local search of the standard bat algorithm individual i , improve the optimization speed; In the formula, rand(·) generates a (0,1) random number; x i is the current position; x mean is the mean of the global position; Step 2) Calculate ω i Finally, update the speed formula of the bat algorithm: In the formula, is the speed of bat i at time t+1; is the speed of bat i at time t; is the position of bat i at time t; x * is the optimal position of the bat population at time t; f i is the pulse emission frequency of bat i at time t.

Images