The present invention is related to agricultural harvesters, in particular to monitoring the condition of the cutting knives of a combine harvester.
The knives of a combine harvester are part of a detachable header mounted at the front of the harvester. As the harvester advances through a field of crops, the crops are cut through a reciprocating motion of a knife bar relative to a set of stationary counterknives and subsequently collected by an auger or by moving belts towards the mid-section of the header. From there, harvested crops are drawn into the main body of the harvester which typically contains a threshing and cleaning arrangement for separating grains from residue plant material.
The condition of the knives has to be optimal in order to ensure the crop yield. Therefore it is important to monitor said condition and to replace damaged knives regularly. However a regular visual inspection is not the most efficient way of detecting knife damage quickly unless the inspection is done at uneconomically short intervals.
Systems have therefore been developed for detecting the knife condition during the operation of the harvester, and to alert the operator of potential damage so that the replacement of damaged knives can be done in a timely fashion without requiring unnecessarily frequent interruptions.
Such systems as they are known today are however not without a number of drawbacks. One known arrangement is described in US2021/0144917, in which a force sensor is incorporated in the drive train of the knives. A position sensor detects the stroke position of the knives, and the two signals are fed to a control unit for determining the knife condition.
A force sensor incorporated in the drive train is however not necessarily the best choice in terms of the sensitivity of the measurements and the capability of detecting and determining knife damage. Also, replacing such a force sensor is difficult as it requires interrupting the drive train.
The invention is related to an agricultural header, to a harvester equipped with said header, and to a method as described in the appended claims. The present invention is thus related to a header for a combine harvester wherein the header comprises one or more vibration sensors mounted on a knife drive train of the header. Such a knife drive train is defined as an assembly of movable parts including a rotatable header drive shaft, a mechanical knife drive, a transmission for transferring the rotation of the drive shaft to the knife drive and a support member with a set of knives attached thereto, the member being coupled to the knife drive so as to undergo a reciprocating movement in a direction that is transverse to the forward direction of a harvester when the header is attached thereto. According to preferred embodiments, the one or more sensors are mounted on the support member and/or on the housing of the knife drive. The sensors are configured to measure the vibration in the direction of said reciprocating movement. The latter feature is not limited to sensors mounted strictly parallel to the direction of said movement, but includes sensors mounted at a slight inclination relative to said movement, which still allow to measure a signal that is representative of the vibrations in the direction of the reciprocating movement. According to embodiments of the invention, the one or more sensors are mounted so as to measure the vibration at an angle between 0 and 10° or between 0 and 5° relative to the direction of the reciprocating movement.
The sensors may include accelerometers or strain gauges. A type of accelerometer that is applicable in the invention is a knock sensor. The header further comprises a control unit connected to the one or more sensors.
The method of the invention comprises the steps of acquiring a vibration signal from one of the sensors, deriving one or more features from the signal, comparing the one or more features to one or more thresholds, deriving from said comparison information on the condition of the knives, and communicating said information regarding the condition of the knives to an operator. According to particular embodiments, the signal is sampled or re-sampled and filtered according to one or more specific filters, while the extracted features are chosen from a specific group including for example the mean or the standard deviation determined for each cycle of the reciprocating movement of the knives.
According to an embodiment, the one or more thresholds applied for each feature consist of an upper control limit and a lower control limit and the condition of the knives is communicated as potentially damaged or damaged when at least one of said features is outside the range defined by said control limits.
The header and the method of the invention enable the monitoring of the condition of the knives using only vibration sensors mounted on an exterior surface of the header’s knife drive train. This is done reliably even with standard sensors like a knock sensor or a strain gauge. The use of these sensors represents an improvement over the prior art in terms of versatility : vibration sensors can be applied in a larger number of locations along the header drive train. Also, these sensors are easier to replace than sensors applied in the art for a similar purpose. Specific embodiments of the method of the invention enable to detect potential damage to the knives and to classify the damage according to the degree of said damage.
Preferred embodiments will now be described with reference to the drawings. The detailed description is not limiting the scope of the invention, which is defined only by the appended claims.
The knife drives 21 force the respective knife bars 20 to undergo a reciprocating movement perpendicular to the forward direction of the harvester, as indicated by the arrows 26 in
According to the invention, and as illustrated in
According to particular embodiments of the invention, the header further comprises a position sensor for measuring the position of the knives 2 during each cycle of the knives’ reciprocating movement. As illustrated in
As seen in
Before describing the method however, it is recapitulated that a header and a harvester according to the invention are characterized by the presence of one or more vibration sensors in the header’s knife drivetrain (preferably in all knife drivetrains if there is more than one), the sensors being mounted so as to measure vibration in the direction of the reciprocating knife movement. The term ‘vibration sensor’ as used in the present context includes any sensor that produces a signal that is representative of a vibration in said direction. This includes an accelerometer and also a strain gauge mounted on the knifebar, but it excludes for example a force sensor mounted between two rotating parts of the knife drivetrain, for example a force sensor mounted inside a wobble box. Vibration sensors mounted such as to measure the vibration in the direction of the reciprocating movement are necessarily mounted on exterior surfaces of the drive train, and are therefore more easily replaceable than such incorporated force sensors.
In the broadest sense, the method of the invention comprises the steps of:
Details of the method of the invention will now be described without any intention of limiting the protection scope, but rather as a way of describing specific embodiments and/or as a possible way of bringing the invention into practice.
First the signal is re-sampled by angular resampling in order to obtain the same number of samples for each cycle. This is done because the knife frequency is not perfectly constant so that the signal does not have the exact same number of samples per cycle. The angular resampling requires the knowledge of the knife position, which may be obtained from a position sensor, for example an incremental encoder 32 as illustrated in
The invention is not limited to any specific filter design, nor to the use of filtering as such. Feature extraction (see further) may be done on the basis of the non-filtered signal according to some embodiments. Also, a standard high-pass filter can be applied, configured to remove the knife frequency from the signal. Any of the filters described in this disclosure may be combined with a standard low pass filter for removing high-frequency noise above a given frequency.
However, four specific filters were developed in the framework of the invention, and have been proven to be particularly effective in given circumstances or in combination with other elements of the method.
Illustrated in
The synchronous average
Filtering the jth cycle xj is realized by subtracting the average
The FSAR filter thereby effectively removes the periodic component of the signal.
The moving synchronous average residual (MSAR) filter illustrated in
wherein :
The MSAR filter has the advantage that not only the periodic component is removed, but also any low frequent component due to changing crop conditions.
For the FSAR and MSAR filters, the number of cycles w in the regions 41 and 42 can be chosen arbitrarily or by training on a set of acquired data, for example by searching the value that minimizes the mean residual of a series of monitoring intervals. The maximal value of w may be restricted to limit the moving window to a given value, for example 3 seconds.
The exponentially weighted moving synchronous average residual (EWMSAR) filter illustrated in
wherein z1 = x1 and wherein λ is a value between 0 and 1 defined as a forgetting factor, as it determines the weight of each new cycle sample xi compared to the previous exponentially weighted moving average zi-1. The EWMSAR-filter residual is calculated as:
The forgetting factor λ determines the rate at which the weight of preceding cycles decreases, i.e. are ‘forgotten’, in the calculation of the moving average. This principle is symbolized by the shape of the window 43 in
The fourth filter (illustrated in
and with s1 = (x2 - z1)2
Where, as for the EWMSAR-filter,
and wherein the EWSNR-filter residual is as follows :
The EWSNR filter thereby corrects for amplitude differences between the different positions within a cycle sample. This guarantees an equal weight for each of the knife positions within the knife cycle sample during the feature calculation.
Said feature calculation is the next step of the method, and is applied to the filtered signal, or if no filter is applied, to the sampled or resampled but non-filtered signal. Returning to
The condition of the knives is then evaluated by comparing for each cycle the feature value to one or more thresholds. According to preferred embodiments, each feature is compared to an upper control limit and to a lower control limit, and it is concluded that damage or potential damage has occurred to the knives when a feature is outside the range defined by the control limits.
From tests performed within the framework of the invention, a number of combinations have been proven to be particularly effective in terms of accurately detecting potentially damaging events. These are the following, each combination being defined as a sequence of ‘sensor type and location’, filter, feature :
The piezo-electric accelerometer in the latter case is a high-end accelerometer. In tests performed by the inventors, a triaxial accelerometer of the type PCB356A02 from PCB Piezotronics® was mounted so that one measurement axis was parallel to the knifebar and, wherein only the signal was used that represents vibrations in the direction of the knife motion,
Knock sensor on the knife bar, FSAR, MSAR or EWMSAR, standard deviation,
Knock sensor on the knifedrive, EWMSNR, standard deviation,
For the last two cases, the knock sensor may be a standard available knock sensor type. The sensor used during the tests was a KS4-R from Bosch®.
The determination of the thresholds can be done in various ways and the invention is not limited to any particular approach. One possible way that has been proven to lead to good results is to apply statistical process control (SPC), which is a methodology known as such and therefore not described here in detail. This approach involves two sets of test data. The first set consists of monitoring intervals labelled as ‘in control’ (IC), of which it is known that no detrimental impacts on the knives have occurred. The second set consists of monitoring intervals of which at least some intervals comprise known ‘events’, i.e. impacts which are known to be potentially damaging to the knives. Such test data can be obtained by doing harvesting runs in a field wherein obstacles have been planted at predefined locations, such as metal rods or the like. The thresholds, such as the upper and lower control limit may be derived for example from the empirical distribution of the features extracted from the IC data. The event data of the second data set, which include spikes in the obtained signals when the knives encounter the planted obstacles, can be used to test the validity of the control limits derived from the IC date, and possibly to further update said control limits. Suitable learning algorithms known in the art may be applied for deriving the thresholds and other parameters of the method from such sets of test data.
The test data of the first and the second set can also be used for training the 4 specific filters described above, for example for determining the optimal number of cycles w in the windows 41, 42 and 43 and the optimal value of the forgetting factor λ in the EWMSAR and EWMSNR filters.
During a harvesting operation, the method of the invention takes place essentially in real time, which may be described as follows on the basis of a given signal, for example the signal obtained from an accelerometer mounted on a wobble box used as the knife drive 21. The following steps are valid for an embodiment of the method that uses a position sensor, for example an encoder, and wherein MSAR filtering is applied before extracting the mean value for a number of consecutive cycles of the knives.
The signal is acquired together with the signal from the encoder during the full duration of a monitoring interval (for example a harvesting run of several minutes along a straight stretch of the field). The vibration signal is resampled by angular resampling with the help of the encoder signal. On the re-sampled signal, MSAR filtering is applied on each cycle following the passage of a first window 42 at the start of the monitoring interval. After said first window, the average is based on a moving window 42 immediately preceding the cycle that is to be filtered. The mean value of the filtered cycle is compared to the upper and lower control limits. If the mean value exceeds the range defined by the control limits, the operator of the harvester is alerted that the knives are potentially damaged. The extraction of the features and the comparison to the control limits takes place essentially in real time : as the signal is acquired, consecutive cycles of the signal are processed and evaluated one by one quasi-immediately upon acquisition of the cycles, so that an alert may be given as soon as a potentially damaging impact has occurred.
According to particular embodiments, a form of ‘smart filtering’ is applied to the acquired signals. This means that either one of the ‘moving average’ filters described above is used, i.e. MSAR, EWMSAR or EWMSNR, wherein however the moving average is only based on the window immediately preceding a cycle when no potentially damaging event has been detected in a given number of cycles prior to that cycle. When such an event has been detected in a given cycle, that cycle is thereafter excluded from the calculation of the moving average. For example in the case of the MSAR filter with a window length w equal to 10 cycles : if the event is detected in the 20th cycle, the average used for filtering the 21st cycle is taken from cycles 10-19, i.e. the average is not updated for the first cycle following the cycle in which the event took place. For the 22nd cycle, the average is taken from cycles 11-19 and 21. For the 23d cycle, the average is taken from cycles 12-19, 21 and 22, and so on until cycle 31, at which point the average is again taken from the 10 preceding cycles 21-30.
According to an embodiment of the method, in addition to detecting a potentially damaging impact on the knives, the severity of the actually occurring damage is classified in one of a number of classes. Preferably two classes are applied, which may be called ‘no or limited damage’ and ‘severe damage’. The first class refers to impacts which cause essentially no damage or only limited damage, so that immediate replacement of the knife is not required. The second class refers to impacts which cause severe damage, i.e. damage which requires replacement of the impacted knife at the earliest convenience.
The limit between the classes may be decided on the basis of test data, i.e. a number of measured impact signals and the effective damage resulting from the measured impacts. Various classifier algorithms are available in the art for performing such a classification effort on the basis of a set of data, and the invention may use any suitable algorithm.
The classification, i.e. assessing whether an impact is to be classified into the ‘no/limited damage’ class or the ‘severe damage’ class, can be based on the value of the feature or features that were used to detect the impact, for example the mean or standard deviation of a cycle during which an impact occurs.
According to specific embodiments however, the classification is not based on the extracted features, but on an additional analysis of multiple cycles : the cycle in which the impact occurs and a number of subsequent cycles, for example two cycles, so that the total of analyzed cycles is three.
According to said specific embodiments, the samples of these three cycles, i.e. the measured amplitudes of these samples serve as the basis for classifying the impact. In practice this can be a very large number of samples, for example a few thousand, so that a classifier algorithm may have difficulty classifying the impacts of a test run taking into account all the samples.
It was found however that a very effective classification is possible on the basis of a reduced number of representative samples, determined by a statistical analysis. Using a principal component analysis (PCA) for example, it was possible to reduce the number of representative samples to about 12 for the example shown in
This probability reached about 85% by classifying the PCA-reduced indicators derived from the 3-cycle strain gauge signal using a non-linear support vector machine (SVM) classifier algorithm. The invention is however not limited to this particular approach. Other methods for reducing the number of indicative samples include for example auto-encoding. Besides non-linear SVM, other suitable classifier algorithms which can be used in the invention include discriminant analysis, naive bayes and k-nearest neighbor.
According to an embodiment, the method further comprises a step of visually verifying the condition of the knives after a harvesting run, i.e when the harvester is stopped, and updating one or more parameters of the method when the method is applied during a next harvesting run. The inspection may be done by the operator in between harvesting runs or after a series of runs. The parameter update may involve an update of the one or more thresholds and/or of the features extracted from the signals. However, other parameters may be updated as well, for example the type of filter that is applied, or parameters of the filters themselves, for example the number of cycles w in the windows 41, 42 and 43, or the forgetting factor λ applied in the EWMSAR or EWMSNR filters. In practice, visually determined knife damage may be registered in the control unit 34 by the operator, for example through a user interface that presents an image of the knife bar, allowing the operator to indicate which knives are damaged and how severe the damage is. This information is then used by the control unit 34 to re-calculate one or more of the parameters, based on a learning algorithm that may be similar to the learning algorithm used during the initial determination of the parameters on the basis of a set of test data.
According to an embodiment already referred to above, the header does not include a position sensor and/or the method includes the step of deriving the position of the knives and performing angular resampling based on a signal produced by one of the sensors mounted on the header drive train, without the help of a position-related signal.
This can be done in various ways, at least some of which are within the knowledge of a person skilled in the art of signal processing. One possible approach includes one or more of the following actions, performed on a raw sampled time signal, for example sampled at a sampling frequency of 10 kHz :
Number | Date | Country | Kind |
---|---|---|---|
22170900.9 | Apr 2022 | EP | regional |