A threshing low load and blocking load state diagnosis system and method and harvester

Abstract

The present invention provides a threshing low-load and blocking load state diagnosis system and method and a harvester, comprising a signal acquisition module, a control module and an execution module; the signal acquisition module comprises a damping vibration isolation unit and a signal acquisition conversion unit, the damping vibration isolation unit is used to perform preliminary filtering on the original vibration signal generated by the threshing separation device, and the signal acquisition conversion unit is used to collect the vibration signal after preliminary filtering and amplify and convert it into a frequency signal for secondary filtering; the control module comprises a signal storage unit and a signal analysis and processing unit, the signal storage unit is used to segment and store the vibration signal after secondary filtering to form a short-time feature set, the signal analysis and processing unit is used to analyze and process the short-time feature set, and compare it with a preset working condition, so as to identify the real-time load state, and when it is a low-load state or a blocking state, the alarm unit sends an alarm signal. The present invention improves the diagnostic accuracy of the load state, improves work efficiency and reduces the failure rate.

Description

Technical Field

[0001] The invention belongs to the technical field of agricultural machinery, and in particular relates to a threshing low-load and blocking-load state diagnosis system and method, and a harvester. Background Art

[0002] With the continuous increase in the feeding volume of combine harvesters in recent years, the blockage problem of various working parts has become increasingly serious, greatly reducing the working efficiency and delaying the farming time. According to relevant surveys, as the core of the combine harvester, the threshing and separation device has a large proportion of blockage failures. Existing studies have shown that the fundamental reason for the blockage of the threshing and separation device is that the threshing load changes greatly beyond the normal range, and due to the different experience of the operators, the abnormality cannot be discovered in time, which eventually leads to blockage. In addition, there are also cases where the operator's lack of experience causes the combine harvester to work at a low load for a long time, and the machine's operating performance cannot be fully utilized, affecting the operating efficiency. Therefore, there is an urgent need for a method that can effectively monitor the threshing load status in real time, so that the threshing load of the combine harvester can be as stable as possible within the normal range, improve work efficiency and reduce the failure rate, and change the previous situation of relying on the operator's experience to subjectively judge the threshing load status.

[0003] At present, the diagnosis of the threshing load state of the combine harvester often adopts the method of directly measuring the torque and speed of the threshing drum, which can produce certain expected results. However, the torque sensor is large in size, and its application in the combine harvester has certain limitations. In addition, the threshing load state determined by the torque or speed has a certain lag, which often cannot achieve the purpose of early warning.

[0004] Prior art A threshing drum monitoring device for a combine harvester can detect the feed amount, grain moisture content, threshing drum speed, threshing concave plate pressure and transmission chain tension in real time by setting a detection module, so as to timely discover and prevent the occurrence of faults and improve work efficiency and reliability. At the same time, a drum speed change module is also provided, which can timely and effectively adjust the drum speed according to the data information of the detection module to reduce threshing losses. It has the advantages of simple structure and reliable performance. However, the communication between multiple sensors will cause a certain lag in the detection of the blocking condition, which is not conducive to the early warning of the blocking fault.

[0005] Prior art A combine harvester anti-blocking system, based on the speed and torque of the combine harvester's header feeding auger, bridge driving drive shaft and threshing drum detected by the speed sensor and torque sensor, determines whether a blockage occurs by the size of the speed and torque, and automatically decelerates the combine harvester in the event of a blockage until the blockage is eliminated. Automatic anti-blocking control of the combine harvester is achieved, and the operating efficiency of the combine harvester is improved. However, the torque sensor is large in size, and the models that can be applied have certain limitations. Once a blockage is about to occur, the speed will not be significantly reduced immediately, so there is still a certain hysteresis in judging whether a blockage exists through the speed sensor.

[0006] Prior art A threshing drum fault simulation monitoring system for a combine harvester can realize the presetting of threshing drum dynamic imbalance fault, shaft end bearing fault and drum blockage fault on the whole machine by pre-installing faulty sub-disc and brake loading device on the threshing drum, collecting threshing related parameters by using data monitoring components, and loading dynamic load on the threshing drum by controlling the brake, which solves the problem of difficulty in obtaining field fault data and poor repeatability to a certain extent. However, the working conditions of harvesters in the field are complex, and there are many factors that cause faults. Relying solely on pre-installed faulty sub-disc and brake loading device cannot fully simulate the actual fault conditions. Summary of the invention

[0007] In view of the above technical problems, one of the purposes of one embodiment of the present invention is to provide a threshing low load and blocking load state diagnosis system, one of the purposes of one embodiment of the present invention is to provide a control method for the threshing low load and blocking load state diagnosis system, and one of the purposes of one embodiment of the present invention is to provide a harvester including the threshing low load and blocking load state diagnosis system. The present invention can make the harvester threshing load as stable as possible within the normal range, improve the diagnostic accuracy of the load state of the threshing and separation device of the combine harvester, improve work efficiency and reduce the failure rate, and change the previous status quo of relying on the subjective judgment of the threshing load state by the operator's experience.

[0008] Note that the description of these objectives does not prevent the existence of other objectives. One embodiment of the present invention does not need to achieve all of the above objectives. Objectives other than the above objectives can be extracted from the description of the specification, drawings, and claims.

[0009] The technical solution of the present invention is:

[0010] A threshing low load and blocking load state diagnosis system, comprising a signal acquisition module, a control module and an execution module;

[0011] The signal acquisition module includes a damping vibration isolation unit and a signal acquisition conversion unit, wherein the damping vibration isolation unit is used to perform preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit is used to collect the preliminary filtered vibration signal, amplify it and convert it into a frequency signal for secondary filtering;

[0012] The control module includes a signal storage unit and a signal analysis and processing unit. The signal storage unit is used to segment and store the vibration signal after secondary filtering to form a short-time feature set. The signal analysis and processing unit is used to analyze and process the short-time feature set and compare it with the preset working condition to identify the real-time load state.

[0013] The execution module includes a display unit and an alarm unit. The display unit is used to display the current load status. The alarm unit is used to send an alarm signal according to the load status of the control module. When the load status is a low load state or a congestion state, the control unit controls the alarm unit to send an alarm signal.

[0014] In the above solution, the vibration isolation material of the damping vibration isolation unit is a combination of one or more of rubber, nylon and bakelite.

[0015] In the above scheme, the signal acquisition conversion unit in the signal acquisition module includes an acceleration sensor, an amplifier circuit, a signal conversion circuit and a filter circuit;

[0016] The acceleration sensor is used to collect the vibration signal after preliminary filtering by the damping vibration isolation unit, the amplifier circuit is used to amplify the vibration signal after preliminary filtering, the signal conversion circuit is used to convert the amplified vibration signal into frequency, and the filter circuit sets the cutoff frequency for secondary filtering.

[0017] In the above scheme, the length of each time segment is no longer than 2 seconds, and the overlap rate between segments is no longer than 50%.

[0018] A method according to the threshing low load and blocking load state diagnosis system comprises the following steps:

[0019] The damping vibration isolation unit of the signal acquisition module performs preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency before performing secondary filtering;

[0020] The signal storage unit of the control module divides and stores the vibration signal after secondary filtering into segments, constructs a short-time feature set, and the signal analysis and processing unit analyzes and processes the short-time feature set and compares it with the preset working conditions to identify the real-time load status; the display unit of the execution module displays the current load status. When the load status is a low-load state or a congestion state, the control unit controls the alarm unit to send an alarm signal.

[0021] In the above scheme, the construction of the short-term feature set includes the following steps:

[0022] Step S1: Assume that the first group of time-series original vibration signals collected from the outer surface of the threshing and separation device is X 11 , X 12 …X 1n The second set of original vibration signals is X 21 , X 22 …X 2n , ... the original vibration signal of the Mth group of time series is X M1 , X M2 …X Mn , and so on, M groups of time series original vibration signals are collected cumulatively, and the high-frequency random noise part of the signal is filtered out by the damping vibration isolation unit. The initial filtered vibration signal is Y 11 , Y 12 …Y 1n , Y 21 , Y 22 …Y 2n , ...Y M1 , Y M2 …Y Mn By analogy, there are M groups of preliminary filtered vibration signals;

[0023] Step S2: The signal acquisition and conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency, and then performs secondary filtering to output a vibration signal Z. 11 , Z 12 …Z 1n , Z 21 , Z 22 …Z 2n , …Z M1 , Z M2 …Z Mn , and so on, there are M groups of secondary filtered vibration signals;

[0024] Step S3: The filtered vibration signals of the M groups of original vibration signals enter the signal storage unit and are shifted to the short time window W1 to W m Divide into m time sequence segments; Step S4, the signal analysis processing unit analyzes the short time windows W1 to W m Extract time domain feature indicators, and use S1~Sn Represented by , it constitutes m short-term feature sets.

[0025] In the above scheme, the samples in the short-term feature set include five load states of different working conditions: no-load state, low-load state, high-load state, congestion state, and congestion state.

[0026] In the above scheme, the analysis and processing of the short-term feature set includes the following steps:

[0027] Step S1), building a convolutional neural network framework, for the signal characteristics of the original load state of different working conditions in the short-term feature set formed by the signal storage unit, through the convolutional neural network, realize feature extraction, feature processing and load state recognition;

[0028] Step S2), construct a feature attention module of multi-core information attention, adaptively pay attention to the time domain features that affect decision performance, and adaptively weaken the redundant features that interfere with decision performance;

[0029] Step S3), based on the convolutional neural network framework, the feature attention module of multi-core information attention is combined to perform threshing load state diagnosis.

[0030] Furthermore, the step S2) constructs a feature attention module of multi-core information attention and specifically includes the following steps:

[0031] Step S21), group channel-by-channel convolution:

[0032] According to different convolution kernel branches F, channel-by-channel convolution is used to calculate the feature matrix X∈R H×W×C , the channel-by-channel convolution formula is:

[0033] F:X→U∈R H×W×C

[0034] Among them, R represents the feature matrix, H represents the matrix height, W represents the matrix width, C represents the number of channels, and U represents the elements in the branch convolution kernel. After the channel-by-channel convolution, batch normalization is applied to maintain the same distribution of convolution parameters and the nonlinear mapping of the ReLU activation function. The channel-by-channel convolution is fused by element summation to obtain U. 和 , U 和 =U1+U2+U3.............+U n ;

[0035] Step S22), channel attention

[0036] In order to obtain global feature information, global average pooling is introduced ac To obtain channel attention, global average pooling s ac The calculation method is:

[0037]

[0038] Where s∈R C , Uc is the element in the branch convolution kernel of the Cth channel, (i, j) refers to the number of rows and columns,

[0039] After average pooling, ECA is combined to realize information exchange between channels. The attention w of ECA ac The calculation method is as follows:

[0040] w ac =δ(conv1d(s ac ,k))

[0041] δ is the Sigmoid function, k is the number of convolution kernels;

[0042] Step S23), select feature kernel

[0043] In order to adaptively select different feature kernels, cross-channel soft attention is applied, U nc The soft attention vector is:

[0044]

[0045] Among them, TR is a matrix, TR nc is the cth row of the nth branch representing the convolution kernel, TR nc ∈R C×d Finally, the feature matrix X calculated by the feature attention module of the multi-core information attention (n) It is expressed as:

[0046]

[0047] Among them, X (n) represents the nth element in the feature matrix, U nc ∈U n .

[0048] A harvester comprises the threshing low-load and blocking load state diagnosis system.

[0049] A harvester comprises the threshing low load and blocking load state diagnosis system, wherein the threshing low load and blocking load state diagnosis system is controlled according to the method of the threshing low load and blocking load state diagnosis system.

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] According to one embodiment of the present invention, a threshing low load and blocking load state diagnosis system is provided. According to one embodiment of the present invention, a control method of the threshing low load and blocking load state diagnosis system is provided. According to one embodiment of the present invention, a harvester including the threshing low load and blocking load state diagnosis system is provided. According to one embodiment of the present invention, a harvester controlled according to the method of the threshing low load and blocking load state diagnosis system is provided. The present invention can make the threshing load of the harvester as stable as possible within the normal range, improve the diagnostic accuracy of the load state of the threshing and separation device of the combine harvester, improve work efficiency and reduce the failure rate, and change the previous situation of relying on the subjective judgment of the threshing load state by the operator's experience.

[0052] Note that the description of these effects does not prevent the existence of other effects. One embodiment of the present invention does not necessarily have all of the above effects. Effects other than the above can be clearly seen and extracted from the description of the specification, drawings, claims, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of the connection of various modules of the threshing low load and blocking load state diagnosis system signal according to one embodiment of the present invention.

[0054] Figure 2 The figure is a schematic diagram of the installation position of the acceleration sensor of the threshing low load and blocking load state diagnosis system according to one embodiment of the present invention. Figure 3 The figure is a schematic diagram of the working principle of the signal acquisition module of the threshing low load and blocking load state diagnosis system according to one embodiment of the present invention.

[0055] Figure 4 The invention discloses a filter setting for a threshing low load and blocking load state diagnosis system according to an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram of the short-term feature set construction principle of the threshing low-load and blocking load state diagnosis system according to one embodiment of the present invention.

[0057] Figure 6 The present invention is a flowchart of a threshing load state diagnosis method according to an embodiment of the present invention.

[0058] Figure 7 A threshing load state diagnosis method according to one embodiment of the present invention adaptively combines a feature attention module of multi-core information attention with a CNN decision framework diagram.

[0059] Figure 8 This is a schematic diagram of a feature attention module of a multi-core information attention method for threshing load state diagnosis according to an embodiment of the present invention.

[0060] Fig. 9This is a visualization output diagram of the time domain characteristics of sensor No. 4 under five load conditions according to an embodiment of the present invention, where Fig. 9 a is an unloaded state, 9b is a low-load state, 9c is a high-load state, 9d is a blocking state, and 9e is a blocked state.

[0061] Fig.10 This is a curve diagram showing the accuracy change of the threshing load state diagnosis method according to one embodiment of the present invention.

[0062] Fig.11 The figure is a schematic diagram of the working principle of the collection module of the threshing low load and blocking load state diagnosis system according to one embodiment of the present invention. DETAILED DESCRIPTION

[0063] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0064] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "front", "back", "left", "right", "up", "down", "axial", "radial", "vertical", "horizontal", "inside", "outside" and the like indicate positions or positional relationships based on the positions 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 orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0065] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0066] Example 1

[0067] Figure 1 The figure shows a preferred embodiment of the threshing low load and blocking load state diagnosis system, which includes a signal acquisition module, a control module and an execution module;

[0068] The signal acquisition module includes a damping vibration isolation unit and a signal acquisition conversion unit, wherein the damping vibration isolation unit is used to perform preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit is used to collect the preliminary filtered vibration signal, amplify it and convert it into a frequency signal for secondary filtering;

[0069] The control module includes a signal storage unit and a signal analysis and processing unit. The signal storage unit is used to segment and store the vibration signal after secondary filtering to form a short-time feature set. The signal analysis and processing unit is used to analyze and process the short-time feature set and compare it with the preset working condition to identify the real-time load state.

[0070] The execution module includes a display unit and an alarm unit. The display unit is used to display the current load status. The alarm unit is used to send an alarm signal according to the load status of the control module. When the load status is a low load state or a congestion state, the control unit controls the alarm unit to send an alarm signal.

[0071] According to this embodiment, preferably, the vibration isolation material of the damping vibration isolation unit is a combination of one or more of rubber, nylon, and bakelite.

[0072] Combination Figure 3 As shown, according to this embodiment, the signal acquisition conversion unit in the signal acquisition module preferably includes an acceleration sensor, an amplifier circuit, a signal conversion circuit and a filter circuit; the acceleration sensor is used to collect the vibration signal after preliminary filtering by the damping vibration isolation unit, the amplifier circuit is used to amplify the vibration signal after preliminary filtering, the signal conversion circuit is used to convert the amplified vibration signal into a frequency, and the filter circuit sets a cutoff frequency for secondary filtering. The present invention achieves the suppression and attenuation of irrelevant components in the preliminary filtered signal through the signal acquisition conversion unit, and the filtered signal formed after two filterings can improve the accuracy of subsequent analysis and processing.

[0073] The present invention achieves preliminary filtering of high-frequency random noise in the original vibration signal measured by the acceleration sensor on the outer surface of the threshing and separation device when the harvester is working in the field, through the combination of a damping vibration isolation unit and an acceleration sensor. According to the existing related research on diagnosing the grain threshing load state using the vibration signal of the outer surface of the threshing device, since the vibration sources of the combine harvester in the field include the superposition of vibration sources such as the engine, the header, the threshing mechanism, the cleaning mechanism, the transportation mechanism, etc., as well as the impact signal from the ground when walking, the signal directly received by only using a traditional acceleration sensor is relatively complex, which is not conducive to subsequent processing and diagnosis of the load state. The present invention achieves preliminary filtering through a damping vibration isolation unit, and the acceleration sensor is used to collect the vibration signal after preliminary filtering by the damping vibration isolation unit, and the acceleration sensor is small in size and easy to arrange in the limited space of the combine harvester, and has the advantages of high sensitivity.

[0074] According to this implementation, the length of each time segment is preferably no longer than 2s, and the overlap rate between segments is no more than 50%. A short-term feature set with a time of no more than 2s is used to classify and identify the load state. The feature signal is closely related to the current working condition and has a fast response speed, avoiding the shortcomings of the previous technical center using a speed sensor and other untimely responses.

[0075] Example 2

[0076] A method for diagnosing the threshing low load and blocking load state according to the embodiment 1 comprises the following steps:

[0077] The damping vibration isolation unit of the signal acquisition module performs preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency before performing secondary filtering;

[0078] The signal storage unit of the control module divides and stores the vibration signal after secondary filtering into segments, constructs a short-time feature set, and the signal analysis and processing unit analyzes and processes the short-time feature set and compares it with the preset working conditions to identify the real-time load status; the display unit of the execution module displays the current load status. When the load status is a low-load state or a congestion state, the control unit controls the alarm unit to send an alarm signal.

[0079] According to this embodiment, preferably, the construction of the short-term feature set includes the following steps:

[0080] Step S1: Assume that the first group of time-series original vibration signals collected from the outer surface of the threshing and separation device is X 11 , X 12 …X 1n The second set of original vibration signals is X 21 , X 22 …X2n , ... the original vibration signal of the Mth group of time series is X M1 , X M2 …X Mn , and so on, M groups of time series original vibration signals are collected cumulatively, and the high-frequency random noise part of the signal is filtered out by the damping vibration isolation unit. The initial filtered vibration signal is Y 11 , Y 12 …Y 1n , Y 21 , Y 22 …Y 2n , ...Y M1 , Y M2 …Y Mn By analogy, there are M groups of preliminary filtered vibration signals;

[0081] Step S2: The signal acquisition and conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency, and then performs secondary filtering to output a vibration signal Z. 11 , Z 12 …Z 1n , Z 21 , Z 22 …Z 2n , …Z M1 , Z M2 …Z Mn , and so on, there are M groups of secondary filtered vibration signals;

[0082] Step S3: The filtered vibration signals of the M groups of original vibration signals enter the signal storage unit and are shifted to the short time window W1 to W m Divide into m time sequence segments; Step S4, the signal analysis processing unit analyzes the short time windows W1 to W m Extract time domain feature indicators, and use S1~S n Represented by , it constitutes m short-term feature sets.

[0083] According to this embodiment, preferably, the samples in the short-time feature set include load states of 5 different working conditions: no-load state, low-load state, high-load state, congestion state, and blocked state; the signal storage unit preliminarily classifies the samples according to the preset working condition values, with the no-load state as label 1, the low-load state as label 2, the high-load state as label 3, the congestion state as label 4, and the blocked state as label 5.

[0084] According to this embodiment, preferably, the analysis and processing of the short-term feature set includes the following steps:

[0085] Step S1), building a convolutional neural network framework, for the signal characteristics of the original load state of different working conditions in the short-term feature set formed by the signal storage unit, through the convolutional neural network, realize feature extraction, feature processing and load state recognition;

[0086] Step S2), construct a feature attention module of multi-core information attention, adaptively pay attention to the key time domain features that affect decision-making performance, and adaptively weaken the redundant features that interfere with decision-making performance;

[0087] Step S3), based on the convolutional neural network framework, the feature attention module of multi-core information attention is combined to perform threshing load state diagnosis.

[0088] According to this embodiment, preferably, the step S2) constructing a feature attention module of multi-core information attention specifically includes the following steps:

[0089] Step S21), group channel-by-channel convolution:

[0090] According to different convolution kernel branches F, channel-by-channel convolution is used to calculate the feature matrix X∈R H×W×C , the channel-by-channel convolution formula is:

[0091] F:X→U∈R H×W×C

[0092] Among them, R represents the feature matrix, H represents the matrix height, W represents the matrix width, C represents the number of channels, and U represents the elements in the branch convolution kernel. After the channel-by-channel convolution, batch normalization is applied to maintain the same distribution of convolution parameters and the nonlinear mapping of the ReLU activation function. The channel-by-channel convolution is fused by element summation to obtain U. 和 , U 和 =U1+U2+U3.............+U n ;

[0093] Step S22), channel attention

[0094] In order to obtain global feature information, global average pooling is introduced ac To obtain channel attention, global average pooling s ac The calculation method is:

[0095]

[0096] Where s∈R C , Uc is the element in the branch convolution kernel of the Cth channel, (i, j) refers to the number of rows and columns,

[0097] After average pooling, ECA is combined to realize information exchange between channels. The attention w of ECA acThe calculation method is as follows:

[0098] w ac =δ(conv1d(s ac ,k))

[0099] δ is the Sigmoid function, k is the number of convolution kernels;

[0100] Step S23), select feature kernel

[0101] In order to adaptively select different feature kernels, cross-channel soft attention is applied, U nc The soft attention vector is:

[0102]

[0103] Among them, TR is a matrix, TR nc is the cth row of the nth branch representing the convolution kernel, TR nc ∈R C×d Finally, the feature matrix X calculated by the feature attention module of the multi-core information attention (n) It is expressed as:

[0104]

[0105] Among them, X (n) represents the nth element in the feature matrix, U nc ∈U n .

[0106] The present invention relates to the system flow of feature extraction, feature dimension reduction and state diagnosis of the load state signal of the threshing and separation device of the integrated harvester, proposes an intelligent integrated diagnosis technology of deep learning, and provides a new theoretical method for accelerating the process of agricultural intelligence.

[0107] Example 3

[0108] A harvester includes a threshing low-load and blocking load state diagnosis system as described in Example 1, and thus has the beneficial effects of Example 1, which will not be described in detail here.

[0109] Example 4

[0110] A harvester includes the threshing low load and blocking load state diagnosis system described in Example 1, and is controlled by the method of the threshing low load and blocking load state diagnosis system described in Example 2, so it has the beneficial effects of Examples 1 and 2, which will not be repeated here.

[0111] Example 5

[0112] This embodiment is a specific embodiment combining embodiments 1, 2, 3 and 4:

[0113] Combination Figure 1 and Figure 2 A threshing low-load and blocking-load state diagnosis system includes a signal acquisition module, a control module and an execution module. Preferably, the signal acquisition conversion unit of this embodiment is provided with an amplification-conversion-filtering integrated circuit formed by four acceleration sensors and an amplifier circuit, a signal conversion circuit and a filter circuit. The acceleration sensor is connected to the amplification-conversion-filtering integrated circuit by a data transmission line. The control module is connected to the amplification-conversion-filtering integrated circuit through a data transmission line and is arranged inside the harvester body. The execution module includes a display and alarm unit, which is connected to the control module through a data transmission line and is installed in the harvester cab for easy observation by the operator.

[0114] Combination Figure 2 Four acceleration sensors are installed at four measuring points respectively. A damping vibration isolation unit is provided between each acceleration sensor and the measuring point. Among them, measuring point 1 is the bottom of the frame at the front end of the drum, measuring point 2 is the arc from the bottom of the frame to the side, measuring point 3 is the center of the top cover of the separation device, and measuring point 4 is the side of the top cover. The four measuring points are on the same section.

[0115] The vibration isolation material of the damping vibration isolation unit is a combination of one or more of rubber, nylon and bakelite.

[0116] According to this embodiment, preferably, the elastic modulus of the rubber is 0.0078 GPa, and the Poisson's ratio is 0.47; the elastic modulus of the nylon is 2.83 GPa, and the Poisson's ratio is 0.4; the elastic modulus of the bakelite is 1.96-2.94 GPa, and the Poisson's ratio is 0.35-0.38.

[0117] The selection of the vibration isolation material is based on the source of vibration under typical working conditions of the harvester, specifically: 1) Idle condition, the engine is idling after starting, not moving, and the source of vibration is the engine; 2) High-speed standby condition, the engine runs at the maximum speed, and the other parts are not working, and the source of vibration is the engine; 3) Threshing idling condition, the clutch is closed, the working parts are running but not moving, and the source of vibration is the superposition of the engine, threshing separation device, cleaning mechanism and transportation system; 4) Traveling condition, the engine is at high speed, all working parts are running and walking on a crop-free open space, and the harvesting platform is not fed. The source of vibration is the superposition of the engine, threshing separation device, cleaning mechanism, transportation system and ground impact signals; 5) Operating condition, the engine is at high speed, harvesting operation is carried out at normal speed, and the harvesting platform is fed. The source of vibration is the superposition of the ground impact signals of the engine, threshing separation device, cleaning mechanism and transportation system, among which different feed amounts change the load of the threshing system, thereby affecting the vibration signal.

[0118] like Figure 3 and Figure 4The main vibration sources of the harvester are the lateral vibration of the threshing drum, the back-and-forth reciprocating motion of the vibrating screen, the up-and-down swing of the connection between the conveying trough and the threshing frame, and the up-and-down vibration of the engine. In addition, there is also vibration excitation of the whole machine by the road surface. However, when the combine harvester is in normal operation, the frequency of each working part is stable. Among them, the frequency range of the engine (low speed and high speed) vibration signal is 50-90Hz, the frequency range of the threshing separation device is 10-30Hz, the frequency range of the cleaning mechanism is 25-30Hz, and the frequency range of the transportation system and ground impact is within 10Hz. Therefore, the high-frequency components above 50Hz can be filtered out through relevant damping and vibration isolation materials to achieve preliminary filtering. At the same time, the cutoff frequency f1 of the bandpass filter is set to 10Hz and f2 to 30Hz for secondary filtering.

[0119] In this embodiment, the acceleration sensor is preferably a piezoelectric acceleration sensor. The piezoelectric acceleration sensor inputs the preliminary filtered signal through the damping vibration isolation unit into the amplification-conversion-filtering integration circuit. After the combined action of the transformer-coupled isolation amplifier, the voltage / frequency interchange circuit (V / F conversion), and the band-pass filter, the components closely related to the threshing load state have been highlighted, while the irrelevant components have been greatly attenuated and filtered out, forming a vibration signal after secondary filtering.

[0120] Combination Figure 5 As shown, the filtered time series data 1-M enter the signal storage unit in the control module and are moved by short time windows W1-W m The signal is divided into m time segments, and considering the particularity of harvester operation, the occurrence of blockage is only closely related to the current 1-2s working conditions. Therefore, the length of the time segment is not more than 2s, and the overlap rate between segments is not more than 50%. The signal analysis and processing unit processes the short time window W1~W m The extracted feature indexes are S1~S n Represents, thus forming m short-term feature sets. .

[0121] The load state diagnosis method of the threshing and separation device of the present invention comprises the following steps:

[0122] The damping vibration isolation unit of the signal acquisition module performs preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency before performing secondary filtering;

[0123] The signal storage unit of the control module divides and stores the vibration signal after secondary filtering into segments, constructs a short-time feature set, and the signal analysis and processing unit analyzes and processes the short-time feature set and compares it with the preset working conditions to identify the real-time load status; the display unit of the execution module displays the current load status, and the alarm unit of the execution module sends an alarm signal according to the load status of the control module.

[0124] The short-term feature set includes five different load states: no-load state, low-load state, high-load state, congestion state, and jam state. 100 sets of data are extracted for each state, so there are 500 sets of data for the five states. The signal storage unit preliminarily classifies the samples according to the preset value, and the no-load state is labeled 1, the low-load state is labeled 2, the high-load state is labeled 3, the congestion state is labeled 4, and the jam state is labeled 5.

[0125] The control module is provided with a deep learning framework of a convolutional neural network (CNN), and a feature attention module combined with multi-core information attention pays attention to important features that affect decision-making performance and interferes with interference information that affects decision-making performance.

[0126] The convolutional neural network (CNN) performs feature extraction, feature processing and pattern recognition on the status data of five different working conditions, highly integrating the three processes, reducing human intervention and making the decision-making method more intelligent.

[0127] Preferably, the steps of building the convolutional neural network framework are:

[0128] During the training of the convolutional neural network (CNN), its hyperparameters need to be set in advance. After the hyperparameter selection, the Adam optimizer with a learning rate of 0.01 is used to update the weights, biases, and convolution parameters. The batch training size is set to 100, the number of iterations is 100, and the batch normalization method (BN) is applied to keep the convolution kernels with the same parameter distribution. After the convolution calculation, the ReLU activation function is selected for nonlinear mapping, and the Softmax function is applied to the fully connected layer for the final classification recognition. In order to reduce the number of calculation parameters, a single fully connected layer containing 5 neurons is selected. The pooling method in the network involvement process selects the convolution calculation method with a pooling kernel of 2×2 and a step size of 2.

[0129] During the training process, the L2 regularization loss function is used, the regularization coefficient λ is 0.01, and the loss value is calculated as shown in formula (1):

[0130]

[0131] Among them, y i The actual situation category for is the model prediction status category, ω j is the weight of the fully connected layer, b j is the threshold of the fully connected layer.

[0132] Figure 6 The figure shows the convolution calculation process in the convolutional neural network CNN. In order to realize the operation load status diagnosis, the specific network design idea is as follows:

[0133] (1) The sampling points of the five different working condition data are unified. Each acceleration sensor takes 25,000 sampling points, and the four sensors have a total of 100,000 sampling points. They are converted into a 1000×10 form for convolution calculation. The convolution calculation method is used to expand the information channel and then realize feature extraction. Since the load state condition has the characteristic of a small sample size, after parameter pre-adjustment, the single channel is expanded to 20 channels for convolution calculation;

[0134] (2) Design multiple CNN frameworks as shown in Table 1 to determine the optimal diagnostic combination, where C represents convolution calculation, P represents pooling calculation, and FC represents fully connected layer;

[0135] (3) In all CNN frameworks, the first convolution calculation uses the 1×1 convolution calculation form to achieve channel expansion, while the remaining calculations use the 3×3 and 5×5 calculation forms to perform multi-core feature extraction. At the same time, the 5×5 convolution kernel is implemented in the form of 3×3 plus dilated convolution. By adding the Padding term, the matrix size before and after the convolution remains unchanged.

[0136] (4) The output result after convolution is compressed using average pooling and finally input into the fully connected layer for classification.

[0137] 400 samples were randomly selected as training sets to build the deep learning model, and the remaining 100 samples were used as test sets to test the classification performance of the model. Table 1 discusses the recognition results of different numbers of convolutional layers for load condition diagnosis. It can be seen that with the increase in the number of convolutional layers, the accuracy of the deep learning network gradually improves. When the number of convolutional layers reaches 3, the best diagnostic accuracy and the lowest loss value are obtained, with a classification accuracy of 97% and a loss value of 0.0111. Further increasing the number of convolutional layers did not increase the accuracy. Finally, C1-C 3，5 ×3-P-FC network results. Figure 7 This is a diagram of the CNN data processing process under this network framework.

[0138] Table 1 Identification results of load condition diagnosis with different numbers of convolutional layers

[0139]

[0140] Construction of the feature attention module of the multi-core information attention:

[0141] In the convolution process, the size of the convolution kernel directly affects the effectiveness of local feature extraction. A convolution kernel that is too small cannot extract larger receptive field spatial information, and a convolution kernel that is too large is prone to lose local detail features. Therefore, for different features of the data to be convolved, a convolution kernel of a fixed size is prone to underfitting and overfitting of the model to a certain extent. Therefore, a convolution method that adaptively selects the size of the convolution kernel is proposed. At the same time, the features of the convolution area are adaptively focused to improve the CNN decision performance.

[0142] Selective Kernel Networks (SK) achieves adaptive selection of convolution kernels by adding convolution branches and combining channel attention mechanisms. In SK, the attention mechanism is similar to Squeeze-and-Excitation (SE), and global average pooling is performed on the fusion matrix after grouped convolution to obtain channel attention. SE first uses global average pooling for each channel separately, and then uses FC to capture nonlinear cross-channel interactions, including dimensionality reduction to control the complexity of the model. FC dimensionality reduction is inefficient for interactions between all channels, and the weights of the channels need to correspond directly to SE. Therefore, this embodiment utilizes an efficient attention mechanism ECA to use the form of one-dimensional (1D) convolution to exchange cross-channel information without reducing the channel dimensionality.

[0143] In order to increase the convolution branches to realize multi-core feature extraction and realize cross-channel information interaction in the attention calculation process, this embodiment takes the branches of two convolution kernels as an example. Figure 8 The calculation process of attention is shown. The specific process is as follows:

[0144] ①Group channel-by-channel convolution

[0145] According to different convolution kernel branches, channel-by-channel convolution is used to calculate the feature matrix X∈R H×W×C , which can effectively reduce the number of parameters in the convolution process.

[0146]

[0147]

[0148] in, is a 3×3 convolution kernel branch, is a 5×5 convolution kernel branch, R represents the feature matrix, H represents the matrix height, W represents the matrix width, C represents the number of channels, U1 represents the elements in the 3×3 branch convolution kernel, and U2 represents the elements in the 5×5 branch convolution kernel. For higher efficiency, the 5×5 convolution kernel is replaced with a 3×3 kernel with an expansion size of 2. After the channel-by-channel convolution, batch normalization is applied to maintain the same distribution of convolution parameters and the nonlinear mapping of the ReLU activation function. In order to adaptively adjust the convolution kernel and control the flow of multi-scale information, the element-wise summation U 和 The results of channel-by-channel convolution are fused:

[0149] U 和 =U1+U2(4)

[0150] ②Channel Attention

[0151] In order to obtain global feature information, global average pooling is introduced ac To obtain channel attention. Global average pooling s ac The calculation method is:

[0152]

[0153] Where s∈R C , Uc is the element in the branch convolution kernel of the Cth channel, (i, j) refers to the number of rows and columns. After average pooling, ECA is combined to realize information exchange between channels. ECA can effectively reduce the computational complexity of the attention mechanism. It is a local cross-channel interaction strategy without dimensionality reduction, which is effectively implemented through one-dimensional convolution. The attention w of ECA ac The calculation method is as follows:

[0154] w ac =δ(conv1d(s ac ,k))(6)

[0155] δ is the Sigmoid function, k is the number of convolution kernels; in this embodiment, k=3.

[0156] ③Select feature kernel

[0157] In order to adaptively select different feature kernels, cross-channel soft attention is applied, which can be achieved through the softmax operation:

[0158]

[0159]

[0160] Where A,B∈R C×d , a and b are the soft attention vectors of U1 and U2 respectively, A c is the cth row of matrix A, ac is the cth element of a, B c is the cth row of matrix B, b c is the cth element of b. Finally, the feature matrix X calculated by the feature attention module of the multi-core information attention (n) It can be expressed as:

[0161] X (n) =a c U1+b c ·U2a c +b c =1,X (n) ∈R H×W×C (9)

[0162] Among them, X (n) Represents the nth element in the feature matrix. This embodiment only provides a case of two branches, and more convolution kernel branch cases can be expanded through formulas (2), (3), (4), (7), (8) and (9).

[0163] X(n) represents the nth element in the feature matrix. The feature attention module of the multi-core information attention can be inserted after the multi-core convolution calculation to realize the function of adaptively identifying and focusing on the feature indicators closely related to the five load states in the above features. Fig. 9 , Taking the time domain features of sensor No. 4 as an example, the feature numbers 1-11 represent the indicators in Table 2 respectively. It can be seen that after adding the feature attention module of multi-core information attention, there are obvious differences in the feature output values ​​in each state, which can improve the accuracy of subsequent pattern recognition.

[0164] from Fig. 9 It can be seen from a that the output values ​​of features 5, 7, 8, 9, 10, and 11 are significantly greater than 0, which are the key time-domain features that the feature attention module of the multi-core information attention module adaptively pays attention to. The feature importance is 9>11>8>7>10>5; feature 1 has no output value, and feature 6 has a negative output value, which is a redundant feature that is adaptively ignored; features 2, 3, and 4 are slightly greater than 0 and are non-key features that are adaptively paid attention to. Fig. 9 b and Fig. 9 As can be seen from c, the output values ​​of features 5, 7, 8, 9, 10, and 11 are all significantly greater than 0, which are the key features that the feature attention module of the multi-core information attention module adaptively pays attention to. The feature importance is 9>11>7>8>5>10. Fig. 9 It can be seen from b that features 1 and 6 have no output values. Fig. 9 It can be seen from c that feature 1 has no output value and is a redundant feature that is adaptively ignored; Fig. 9 Features 2, 3, 4 and Fig. 9Features 2, 3, 4, and 6 in c are slightly greater than 0 and are non-key features for adaptive attention. Fig. 9 It can be seen from d that the output values ​​of features 5, 7, 8, 9, 10, and 11 are significantly greater than 0, and they are the key features that the feature attention module of the multi-core information attention module adaptively pays attention to. The feature importance is 9>11>8>7>5>10; feature 1 has no output value and is a redundant feature that is adaptively ignored; features 2, 3, 4, and 6 are slightly greater than 0 and are non-key features that are adaptively paid attention to. Fig. 9 It can be seen from e that the output values ​​of features 5, 6, 7, 8, 9, 10, and 11 are significantly greater than 0, and they are the key features that the feature attention module of the multi-core information attention adaptively focuses on. The feature importance is 7>9>11>8>5>10>6; features 1, 2, 3, and 4 have no output values ​​and are redundant features that are adaptively ignored.

[0165] Table 2 Calculation of time domain characteristic indicators

[0166]

[0167] Load state diagnosis combining feature attention module of multi-core information attention and CNN decision:

[0168] Based on the optimal CNN decision structure, the present invention adds a feature attention module with multi-core information attention to realize feature attention. Fig.10 As shown in the accuracy change curve, the decision accuracy reached 99%, and the loss value was 0.0095. The accuracy was increased by 2% and the loss value was reduced by 0.0016. This shows that the introduction of the attention mechanism effectively improved the diagnostic accuracy of the load state of the threshing and separation device of the combine harvester.

[0169] like Fig.11 After the above steps, preferably, when the current load state is identified as low load, the alarm unit flashes a yellow light and uses a buzzer to issue a prompt sound to remind the operator to increase the forward speed, increase the feeding amount, and improve the working efficiency; when the current load state is identified as being blocked, the alarm device flashes a red light and uses a buzzer to issue an alarm sound to remind the operator to slow down, or adjust the threshing gap, reduce the threshing load, and avoid blockage. The present invention can make the harvester threshing load as stable as possible within the normal range, improve the diagnostic accuracy of the load state of the threshing and separation device of the combine harvester, improve work efficiency and reduce the failure rate, and change the previous status quo of relying on the operator's experience to subjectively judge the threshing load state.

[0170] It should be understood that although this specification is described according to various embodiments, not every embodiment contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

[0171] The series of detailed descriptions listed above are only specific descriptions of feasible embodiments of the present invention. They are not intended to limit the scope of protection of the present invention. All equivalent embodiments or changes that do not deviate from the technical spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for diagnosing low-load and blocking-load states of threshing, characterized in that: The threshing low load and blocking load state diagnosis system is characterized by comprising a signal acquisition module, a control module and an execution module; The signal acquisition module includes a damping vibration isolation unit and a signal acquisition conversion unit, wherein the damping vibration isolation unit is used to perform preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit is used to collect the preliminary filtered vibration signal, amplify it and convert it into a frequency signal for secondary filtering; The control module includes a signal storage unit and a signal analysis and processing unit. The signal storage unit is used to segment and store the vibration signal after secondary filtering to form a short-time feature set. The signal analysis and processing unit is used to analyze and process the short-time feature set and compare it with the preset working condition to identify the real-time load state. The execution module includes a display unit and an alarm unit, wherein the display unit is used to display the current load state, and the alarm unit is used to send an alarm signal according to the load state of the control module. When the load state is a low load state or a congestion state, the control unit controls the alarm unit to send an alarm signal; The method comprises the following steps: The damping vibration isolation unit of the signal acquisition module performs preliminary filtering on the original vibration signal generated by the threshing and separation device, and the signal acquisition conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency before performing secondary filtering; The signal storage unit of the control module divides and stores the vibration signal after secondary filtering into segments, constructs a short-time feature set, and the signal analysis and processing unit analyzes and processes the short-time feature set, and compares it with the preset working condition, so as to identify the real-time load state; the display unit of the execution module displays the current load state, and when the load state is a low load state or a congestion state, the control unit controls the alarm unit to send an alarm signal; The construction of the short-term feature set comprises the following steps: Step S1: Assume that the first group of time-series original vibration signals collected from the outer surface of the threshing and separation device is X 11 , X 12 …X 1n The second set of original vibration signals is X 21 , X 22 …X 2n , ... the original vibration signal of the Mth group of time series is X M1 , X M2 …X Mn , and so on, M groups of time series original vibration signals are collected cumulatively, and the high-frequency random noise part of the signal is filtered out by the damping vibration isolation unit. The initial filtered vibration signal is Y 11 , Y 12 …Y 1n , Y 21 , Y 22 …Y 2n , …Y M1 , Y M2 …Y Mn By analogy, there are M groups of preliminary filtered vibration signals; Step S2: The signal acquisition and conversion unit amplifies the collected preliminary filtered vibration signal and converts it into a frequency, and then performs secondary filtering to output a vibration signal Z. 11 , Z 12 …Z 1n , Z 21 , Z 22 …Z 2n , …Z M1 , Z M2 …Z Mn , and so on, there are M groups of secondary filtered vibration signals; Step S3: The filtered vibration signals of the M groups of original vibration signals enter the signal storage unit and are shifted to the short time window W1 to W m Divide into m time sequence segments; Step S4, the signal analysis processing unit analyzes the short time windows W1 to W m Extract time domain feature indicators, and use S1~S n Represented by, it constitutes m short-term feature sets; The samples in the short-term feature set include five different load states: no-load state, low-load state, high-load state, congestion state, and congestion state; The analysis and processing of the short-term feature set comprises the following steps: Step S1), building a convolutional neural network framework, for the signal characteristics of the original load state of different working conditions in the short-term feature set formed by the signal storage unit, through the convolutional neural network, realize feature extraction, feature processing and load state recognition; Step S2), construct a feature attention module of multi-core information attention, adaptively pay attention to the time domain features that affect decision performance, and adaptively weaken the redundant features that interfere with decision performance; Step S3), based on the convolutional neural network framework, the feature attention module of multi-core information attention is combined to perform threshing load state diagnosis.

2. The method of the threshing low load and blocking load state diagnosis system according to claim 1, characterized in that: The step S2) constructing the feature attention module of multi-core information attention specifically includes the following steps: Step S21), group channel-by-channel convolution: According to different convolution kernel branches F, channel-by-channel convolution is used to calculate the feature matrix X∈R H×W×C , the channel-by-channel convolution formula is: F:X→U∈R H×W×C Among them, R represents the feature matrix, H represents the matrix height, W represents the matrix width, C represents the number of channels, and U represents the elements in the branch convolution kernel. After the channel-by-channel convolution, batch normalization is applied to maintain the same distribution of convolution parameters and the nonlinear mapping of the ReLU activation function. The channel-by-channel convolution is fused by element summation to obtain U. 和 , U 和 =U1+U2+U3.............+U n ; Step S22), channel attention in order to obtain global feature information, while introducing global average pooling s ac To obtain channel attention, global average pooling s ac The calculation method is: Where s∈R C , Uc is the element in the branch convolution kernel of the Cth channel, (i, j) refers to the number of rows and columns, after average pooling, combined with ECA to achieve information exchange between channels, ECA's attention w ac The calculation method is as follows: In ac =δ(conv1d(s ac ,k)) δ is the Sigmoid function, k is the number of convolution kernels; Step S23), select feature kernels. In order to adaptively select different feature kernels, apply cross-channel soft attention, U nc The soft attention vector is: Among them, TR is a matrix, TR nc is the cth row of the nth branch representing the convolution kernel, TR nc ∈R C×d Finally, the feature matrix X calculated by the feature attention module of the multi-core information attention (n) It is expressed as: X (n) ∈R H×W×C Among them, X (n) represents the nth element in the feature matrix, U nc ∈U n .

3. The method of the threshing low load and blocking load state diagnosis system according to claim 1, characterized in that: The signal acquisition conversion unit in the signal acquisition module includes an acceleration sensor, an amplifier circuit, a signal conversion circuit and a filter circuit; The acceleration sensor is used to collect the vibration signal after preliminary filtering by the damping vibration isolation unit, the amplifier circuit is used to amplify the vibration signal after preliminary filtering, the signal conversion circuit is used to convert the amplified vibration signal into frequency, and the filter circuit sets the cutoff frequency for secondary filtering.

4. The method of the threshing low load and blocking load state diagnosis system according to claim 1, characterized in that: The length of each time segment is no longer than 2s, and the overlap rate between segments is no more than 50%.

5. A harvester, characterized in that: A method for controlling a threshing low load and blocking load state diagnosis system according to any one of claims 1 to 4.

Images