METHOD FOR DETERMINING HEAT-SINK CONTAMINATION BY MEANS OF ARTIFICIAL INTELLIGENCE

Abstract

A method for determining soiling of a heat sink for cooling an electronic component is disclosed. A load curve controlled by the component and a temperature curve along a heat transfer chain from the at least one component to the heat sink are continuously captured. In a training phase, a decision function is determined, which is provided for application to a portion of the load curve and to at least one correspondingly captured portion of a temperature curve. In a classification phase, a non-soiled state of the heat sink is detected when the portion of the load curve presented to the decision function is similar to the portions of the load curve captured in the training phase and when the at least one corresponding portion of a temperature curve is similar to the portions of the temperature curve in question that were correspondingly captured in the training phase.

Claims

1.-12. (canceled)

13. A method for determining the contamination of a heat sink provided for cooling at least one electronic component, the method comprising: continuously detecting a load profile for the at least one electronic component; detecting at least one temperature profile of a temperature along a heat transfer chain from the at least one electronic component to the heat sink by at least one temperature sensor; determining learning sections in a training phase on an uncontaminated heat sink in such a way that the learning sections have sections of the load profile similar to one another and to a prototypical load profile section; detecting for each of the learning sections a section of the at least one temperature profile; determining a decision function from a plurality of sections assigned to the learning sections of the load profile and the at least one temperature profile, the decision function provided for application to a section of the load profile and to at least one section of a temperature profile detected in a manner corresponding to the section of the load profile; detecting with the decision function an uncontaminated state of the heat sink if a presented section of the load profile is similar to the sections of the load profile detected in the learning sections and if the at least one corresponding section of a temperature profile is similar to the sections of the respective temperature profile detected in the learning sections and otherwise detecting a contaminated state of the heat sink; determining in a subsequent classification phase, classification sections in such a way that the classification sections have sections of the load profile which are similar to at least one section of the load profile assigned to a learning section; presenting the sections of the load profile and of the at least one temperature profile respectively assigned to a classification section of the decision function determined in the training phase; and detecting a contamination state of the heat sink by the decision function.

14. The method of claim 13, further comprising detecting the load profile by an electrical load as a time profile of an output current controlled by the at least one electronic component.

15. The method of claim 13, further comprising determining the load profile as a thermal load profile based on at least one difference between a first temperature profile, which is detected with a first temperature sensor, and at least one further temperature profile, which is detected by a further temperature sensor.

16. The method of claim 13, further comprising detecting a plurality of temperature profiles by a plurality of temperature sensors.

17. The method of claim 13, further comprising forming the decision function by determining at least one prototypical temperature section from the plurality of sections of the at least one temperature profile assigned to the learning sections, and in the presentation of the sections of the at least one temperature profile assigned to a classification section, determining a similarity to the respectively assigned prototypical temperature section according to a predetermined similarity measure and, if the similarity lies below a threshold value, detecting a contaminated state of the heat sink.

18. The method of claim 13, further comprising determining the similarity inversely to a mean square deviation between the at least one temperature profile and the respectively assigned prototypical temperature section.

19. The method of claim 18, further comprising determining the similarity inversely to an amount of a maximum time deviation between the at least one temperature profile and the respectively assigned prototypical temperature section.

20. The method of claim 18, further comprising determining the similarity inversely to an exceeding of a maximum value derived from the prototypical temperature section in the at least one temperature profile.

21. The method of claim 13, further comprising determining the decision function by a method of machine learning.

22. The method of claim 21, further comprising determining the learning sections by clustering of sections of the load profile.

23. The method of claim 21, further comprising forming the decision function by training a deep convolutional neural network by a deep learning method based on feature vectors which are determined from the sections of the load profile and of the temperature profile respectively assigned to the learning sections.

24. The method of claim 13, further comprising determining improved heat dissipation of the heat sink by detecting, in a classification section with an assigned section of the load profile, a section of a temperature profile which lies at least partially below a section of a temperature profile which has been detected in a learning section with a similar section of the load profile.

Description

[0058] The above-described properties, features and advantages of this invention and the manner in which these are achieved, will become clearer and more fully understood in conjunction with the following description of exemplary embodiments, which will be explained in connection with the drawings. It is shown hereby in:

[0059] FIG. 1 a diagrammatic plan view of a printed circuit board with power electronic components and a heat sink,

[0060] FIG. 2 a diagrammatic plan view of a printed circuit board with power electronic components,

[0061] FIG. 3 a diagrammatic view of a calibration load profile and assigned calibration temperature profiles and

[0062] FIG. 4 a diagrammatic view of a load profile and assigned temperature profiles.

[0063] Parts corresponding to one another are provided with the same reference characters in the figures.

[0064] FIG. 1 shows a printed circuit board 1 which is connected to a heat sink 3 having power electronic components 2. The power electronic components 2 are arranged on a substrate 8 between the printed circuit board 1 and the heat sink 3 and are enclosed in a housing 11 which can be designed, for example, as a potting compound frame.

[0065] The power electronic components 2 are connected to the heat sink 3 in a heat-conducting manner. On its side facing the power electronic components 2, the heat sink 3 has a planar metallic heat-receiving plate 9 which extends over the entire base area and is illustrated and described in more detail in FIG. 2.

[0066] From the side facing away from the power electronic components 2, cooling fins 4 project vertically from the heat sink 3 and extend parallel to the direction of flow 5 of the air conducted through the cooling fin intermediate spaces 6 for cooling. The entirety of the cooling fin intermediate spaces 6 forms the cooling channel 7.

[0067] The substrate 8 is designed as a ceramic plate laminated with copper on both sides and bears the power electronic components 2 on one component side. Conductor tracks are etched free on the component side in accordance with the electrical connections provided between the power electronic components 2.

[0068] The thermal contact side of the substrate 8 opposite the component side is copper-laminated over its entire surface and soldered to the heat-receiving plate 9 of the heat sink 3, However, other connection techniques and arrangements are also possible which ensure a low heat transfer resistance between the power electronic components 2 and the heat sink 3, for example by screwing or by form-fitting connection using heat-conducting paste.

[0069] Depending on the function of the power electronics, the power electronic components 2 can be designed, for example as a bipolar transistor with an insulated gate bipolar transistor (IGBT) and/or as a diode and can be arranged on the substrate 8, as is illustrated in more detail in FIG. 2 in a diagrammatic plan view of the heat-absorbing surface 9.

[0070] The heat loss released by the power electronic components 2 is transferred through the ceramic layer of the substrate 8 to the heat sink 3 and discharged via its cooling fins 4 to the air flowing past.

[0071] In the present exemplary embodiment, two heat sinks 3 are arranged above a printed circuit board 1, In each case, one heat sink 3 is configured for the heat dissipation of a plurality of power electronic components 2 which are arranged on three substrates 8 of identical structure and design. A temperature sensor 10 is arranged on each of the substrates 8 in the immediate vicinity of the power electronic components 2.

[0072] The temperature sensor 10 measures the temperature of the respective substrate 8, which is approximately equal to the temperature of the heat-absorbing surface gin the region covered by the substrate 8.

[0073] A load change, in other words, a change in the output current controlled by the power electronic components 2 of a substrate 8, brings about a change in the substrate temperature detected by the respective temperature sensor 10, the temperature change generally being attenuated and delayed with respect to the load change.

[0074] The relationship between the load change and the temperature change includes, inter alia, the heat transfer resistance between the power electronic components 2 and the substrate 8, the heat transfer resistance between the substrate 8 and the heat sink 3 and the heat flow carried away via the cooling channel 7. This relationship is therefore significantly determined by the installation situation of the printed circuit board 1.

[0075] FIG. 3 shows a diagrammatic view of a calibration load profile I.sub.K(t) and a first and a second calibration temperature profile ϑ.sub.K,1(t), ϑ.sub.K,2(t) as a function of time t. The first calibration temperature profile ϑ.sub.K,1(t) is recorded for a printed circuit board 1 installed in an air-conditioned control cabinet when the calibration load profile I.sub.K(t) is applied. The second calibration temperature profile ϑ.sub.K,2(t) deviating therefrom is recorded for a printed circuit board 1 in an un-air-conditioned installation position with the same application to the calibration load profile I.sub.K(t).

[0076] The calibration temperature profile ϑ.sub.K,1(t), ϑ.sub.K,2(t) of a substrate 8 can depend not only on the calibration load profile I.sub.K(t) and on the installation situation of the printed circuit board 1 but also on the aging of the power electronic components 2, in particular on the load-dependent power loss which can be varied as a function of aging. In addition, the calibration temperature profile ϑ.sub.K,1(t), ϑ.sub.K,2(t) on a first substrate 8 is also determined by the load profile on other substrates 8 which are cooled via the same heat sink 3.

[0077] In addition, a load profile in an actual operation of the printed circuit board 1 typically deviates from the calibration load profile I.sub.K(t). As the power loss of a power electronic component 2, for example an IGBT transistor, also depends on its temperature, positive feedback effects occur during heating, which make the transfer of the calibration temperature profile ϑ.sub.K,1(t), ϑ.sub.K,2(t) determined for the calibration load profile I.sub.K(t) to a different load profile more difficult or impossible.

[0078] In general, it is therefore not possible to determine an air flow reduced by contamination in a cooling channel 7 or a heat transfer resistance impaired by contamination between a heat sink 3 and the circulating air solely from the calibration temperature profile ϑ.sub.K,1(t), ϑ.sub.K,2(t), which is measured by a temperature sensor 10 on a substrate 8, the components 2 of which are exposed to the calibration load profile I.sub.K(t).

[0079] FIG. 4 shows a diagrammatic view of a load profile I(t) during the operational operation of the power electronic device with an installation state of the printed circuit board 1, which is not known in more detail, as well as the profile of the temperature measured at a temperature sensor 10.

[0080] In a training phase T, the load profile I(t) and a synchronously measured first temperature profile ϑ.sub.1(t) are detected. The training phase T is selected and dimensioned in such a way that heat sinks 3 do not become contaminated during the training phase T, or only to such a small extent that the function of the heat sinks 3 is not impaired.

[0081] By means of an unsupervised machine learning method, learning sections T1 to T3 are segmented, in which the load profile I(t) is approximately equal to a typical, recurring load profile section P.

[0082] Learning sections T1 to T3 are preferably segmented in such a way that the assigned load profile section P occurs repeatedly particularly frequently and differs particularly well from other sections of the load profile I(t).

[0083] In this case, it is not necessary for the sections of the load profile I(t) assigned to the learning sections T1 to T3 to be identical to one another and to the load profile section P. It is sufficient if, after a predetermined distance or similarity metric, these sections have a significantly higher similarity to one another than other sections of the load profile I(t) selected at random, preferably a statistically significantly increased similarity.

[0084] From the plurality of such segmented learning sections T1 to T3, the prototypical load profile section P is determined, for example, by averaging or by means of a cluster method.

[0085] Segmentation and cluster methods for identifying learning sections T1 to T3 and for determining a prototypical load profile section P are known from the prior art.

[0086] Each learning section T1 to T3 is also assigned a section of the first temperature profile NO. Analogously to the determination of the prototypical load profile section, a prototypical temperature section ϑ is determined from the plurality of these sections, for example by averaging or clustering.

[0087] The training phase T is followed by a classification phase K, in which the contamination state of the heat sink 3 is initially unknown and is determined using the method according to the invention.

[0088] For this purpose, classification sections K1, K2 are segmented within the classification phase K in such a way that the load profile I(t) in each of these classification sections K1, K2 has a large, preferably statistically significant similarity to the load profile section P, which was determined in the training phase T.

[0089] Each of the classification sections K1, K2 is also assigned a section of a temperature profile ϑ.sub.2(t), ϑ.sub.3(t), which is determined, inter alia, by the load profile I(t) in and immediately before the classification section K1, K2 and by the heat dissipation of the heat sink.

[0090] By way of example, FIG. 4 shows a second temperature profile ϑ.sub.2(t) for slight contamination of the heat sink 3 and an alternative third temperature profile ϑ.sub.3(t) for severe contamination of the heat sink 3.

[0091] In the case of slight contamination, the second temperature profile ϑ.sub.2(t) does not deviate, or deviates only slightly, from the determined prototypical temperature section ϑ in a classification section K1, K2, in the sense of a predetermined distance measure, as the heat sink 3 is cooled approximately as well as in the training phase T.

[0092] In the case of severe contamination, however, the third temperature profile ϑ.sub.3(t) in a classification section K1, K2 deviates comparatively more strongly from the prototypical temperature section ϑ as the heat dissipation of the heat sink 3 has changed (decreased) compared to the training phase T.

[0093] For the determination of the deviation between a temperature profile ϑ.sub.2(t), ϑ.sub.3(t) in a classification section K1, K2 and the prototypical temperature section ϑ, for example, the time-averaged quadratic deviation or the absolute value of the time deviation is possible as a distance measure.

[0094] For example, the amount of the difference between the maximum temperature value in the classification section K1, K2 and the maximum temperature value of the prototypical temperature section a can also be used as a simple distance measure.

[0095] Derived distance measures are also possible, which are determined, for example, on the basis of time constants of decaying exponential functions which are adapted to sections of the temperature profile ϑ.sub.2(t), ϑ.sub.3(t) in a classification section K1 K2.

[0096] As shown in FIG. 4, in general neither the second nor the third temperature profile ϑ.sub.2(t), ϑ.sub.3(t) in a classification section K1, K2 coincide perfectly, that is to say: congruently, with the prototypical temperature section ϑ.

[0097] Minor differences between the second temperature profile ϑ.sub.2(t) and the prototypical temperature section ϑ can be attributed, for example, to the aging of the power electronic components 2, to the heat transfer from a substrate 8 to the heat-absorbing surface 9 of the heat sink 3 altered by aging or degradation of, for example, soldered connections.

[0098] On the other hand, strong contamination of the cooling fin intermediate spaces 6 brings about a comparatively greater deviation of the third temperature profile ϑ.sub.3(t) from the prototypical temperature section ϑ. In particular, such strong contamination can bring about a steeper increase in temperature in the event of a load increase, a temperature profile ϑ.sub.3(t) shifted upward (toward higher temperatures) and/or a delayed, flatter drop in temperature in the event of a load drop.

[0099] The deviation of a temperature profile ϑ.sub.2(t), ϑ.sub.3(t) in classification sections K1, K2 can thus be compared with a threshold value. If such a threshold value is exceeded, contamination of the heat sink 3 that must be eliminated is detected. It is also possible to determine statistical measures for exceeding the threshold value. Contamination can thus be detected when the deviation in a certain relative proportion of the classification sections K1, K2—for example in at least 10% of the detected classification sections K1, K2—exceeds a predetermined threshold value.

[0100] It is also possible to use classifiers for detecting and/or quantifying a degree of contamination of a heat sink 3 instead of or in addition to one such threshold value criterion. The decision function, that is to say: the distinction between a contaminated and an uncontaminated state on the basis of a section of a temperature profile ϑ.sub.2(t), ϑ.sub.3(t) and the corresponding sections of a load profile I(t), can be formed by adapting such a classifier to the sections of the load profile I(t) and the temperature profile ϑ.sub.1(t) respectively detected in the learning sections T1 to T3.

[0101] The learning sections T1 to 13 are transformed into feature vectors. A feature vector can be formed, for example, as a set of sample values of the load profile I(t) and sample values of the temperature profile ϑ.sub.1(t). A feature vector can also or additionally comprise spectral features, for example one or more band powers in predetermined frequency bands. A feature vector can also or additionally comprise statistical features, for example a mean value, a standard deviation and/or statistical moments of a higher order.

[0102] By means of feature vectors obtained in this way, which describe the learning sections T1 to T3, at least one classifier is adapted (trained) in that a learning error criterion is minimized by a training algorithm.

[0103] Such classifiers can be formed, for example, by adapting a linear or quadratic discriminant function, Methods of linear and quadratic discriminant analysis are known from the prior art.

[0104] Such classifiers can also be formed by adapting multilayer neural networks, Deep learning methods for adapting neural networks with a particularly large number of hidden layers, which are particularly suitable for the classification of signals and images, are likewise known from the prior art.

[0105] In the classification phase K, feature vectors are formed for the classification sections K1, K2 in the same manner as for the learning sections T1 to T3 in the training phase T. By means of a trained (adapted) classifier, each of these feature vectors is assigned a classification value which describes the probability and/or the degree of contamination of the heat sink 3.

[0106] The method was explained on the basis of a single prototypical load profile section P, which was determined from a plurality of learning sections T1 to T3 with a similar load profile I(T). In an analogous manner, however, it is also possible to determine further prototypical load profile sections (likewise not shown in detail in FIG. 4) from further learning sections (not shown in detail in FIG. 4) and to use them for evaluating the state of contamination of a heat sink 3.

[0107] Furthermore, the method was explained on the basis of a single load profile I(t) and a temperature profile ϑ.sub.1(t), ϑ.sub.2(t), ϑ.sub.3(t) detected by a single temperature sensor 10. However, it is also possible for a plurality of temperature profiles ϑ.sub.1(t), ϑ.sub.2(t), ϑ.sub.3(t) to be detected in parallel with a single substrate 8, which is acted upon by a load profile I(t), by measurement with a plurality of temperature sensors 10. Furthermore, it is also possible for various power electronic components 2 to be acted upon by independent load profiles I(t) and, in parallel with this, for a plurality of temperature profiles ϑ.sub.1(t), ϑ.sub.2(t), ϑ.sub.3(t) to be detected by measurement using a plurality of temperature sensors 10.

[0108] Such multivariate signals can also be converted into feature vectors, used for training classifiers and presented to a trained classifier for detecting a degree of contamination.

[0109] Although the invention has been illustrated and described in more detail by preferred exemplary embodiments, the invention is not limited by the disclosed examples and other variations may be derived therefrom by a person skilled in the art without departing from the scope of the invention.