Method for determining an unknown load state of a filter in a ventilation system, device, ventilation system, and computer program product

Abstract

The invention relates to a method (10) for determining an unknown load state of a filter in a ventilation system, wherein the method comprises the following steps: a) continuously or repeatedly acquiring at least one measured variable of the ventilation system over a period of time during an unknown load state of the filter; b) providing a trained neural network (11) for determining a measure of reconstruction; c) reducing the data acquired in step a) of the at least one measure variable to a latent space by means of the trained neural network (11); d) reconstructing the acquired data from the latent space by means of the trained neural network (11); e) determining the measure of reconstruction by comparing the acquired data with the reconstructed data; and f) determining a similarity value from the measure of reconstruction, wherein the similarity value quantifies a similarity to a defined load state of the filter.

Description

[0001] METHOD FOR DETERMINING AN UNKNOWN LOADING STATE OF A FILTER OF AN AIR CONDITIONING SYSTEM, DEVICE,

[0002] AIR CONDITIONING SYSTEM AND COMPUTER PROGRAM PRODUCT

[0003] The invention relates to a method for determining an unknown loading state of a filter of a ventilation system, a device for determining an unknown loading state of a filter of a ventilation system, a ventilation system with such a device and a computer program product for carrying out the method.

[0004] Known methods and devices do not address the determination of unknown filter loading conditions in an air-conditioning system. However, methods and devices for preventing heavily loaded filter loading conditions have been known in practice for years.

[0005] Filters installed in air-conditioning systems – for example, in fans – become clogged over time. Dust and dirt thus reduce the filter capacity over time. This leads to an increase in pressure resistance and energy consumption and / or a reduction in the air-conditioning system's flow rate. According to known methods and devices for preventing heavily loaded filters in air-conditioning systems, dirty filters are changed at defined intervals or operating times, regardless of the actual load state of the respective filter. Thus, the change is not carried out according to demand. The operation of existing air-conditioning systems is therefore not resource-efficient and can be improved in terms of energy efficiency.

[0006] The present invention is therefore based on the object of proposing a method and a device for determining an unknown load state of a filter in a ventilation system and designing it in such a way that the operation of the ventilation system is more resource-conserving and energy-efficient. Furthermore, a needs-based filter change should be enabled. Furthermore, an improved ventilation system with a correspondingly configured device and a computer program product should be specified.

[0007] According to the invention, the above object is achieved by the features of claim 1.According to this, the method in question for determining an unknown loading state of a filter of an air-conditioning system comprises the following steps: a) Continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during an unknown loading state of the filter, b) Providing a trained neural network for determining a reconstruction measure, c) Reducing the data of the at least one measured variable recorded in step a) to a latent space by means of the trained neural network, d) Reconstructing the recorded data from the latent space by means of the trained neural network, e) Determining the reconstruction measure by comparing the recorded data with the reconstructed data and f) Determining a similarity value from the reconstruction measure, wherein the similarity value quantifies a similarity to the defined loading state of the filter.

[0008] The loading state is a measure of how clogged the filter has become. An unloaded loading state corresponds to the state of a new filter with full filter capacity. A fully loaded loading state corresponds to the state of a filter at the end of its operating life with no remaining filter capacity. In addition to the completely unloaded loading state and the fully loaded loading state, there are other defined loading states that can be expressed, for example, as a percentage value from 0% to 100% loading, as a percentage value of a filter capacity from 100% to 0%, as a standardized value between 0 and 1 for the loading and / or for the filter capacity and / or in another known way. The - remaining - filter capacity corresponds to the counterpart to the - "partial" - loading of the filter.

[0009] The time periods mentioned in this description can have a wide variety of durations. They can, for example, also be discrete points in time or timestamps.

[0010] The neural network used to determine a reconstruction measure can, for example, be an autoencoder model in which data is compressed into a latent space. The latent space represents the data acquired in step a) – in a lower dimension.

[0011] The reconstruction measure is determined by reducing the acquired data to its latent space – encoding – and then reconstructing it again – decoding. The reconstruction measure can be a measure of the reconstruction error or, conversely, a measure of the reconstruction quality. Encoding and decoding are performed, for example, using a neural network designed as an autoencoder model. The resulting error or the resulting deviation between the acquired data and the reconstructed data can be referred to as the reconstruction error. The reconstruction error can be quantified by one or more variables / metrics, for example, by one or more numerical values, such as common machine learning metrics, such as cost functions. For example, the reconstruction error can correspond to the mean square error and / or the root of the mean square error.The reconstruction error is ideally minimal in the case of the data underlying the model training, so that the data underlying the model training and the reconstructed data are approximately identical.

[0012] The optimal reconstruction measure can be minimized as reconstruction error as possible. Conversely, the reconstruction measure can be maximized as reconstruction quality as possible. For the purposes of the present disclosure, an optimal reconstruction measure corresponds to a minimal reconstruction error.

[0013] The reconstruction error can be an absolute or relative value. Using the reconstruction error, a similarity value can be quantified to the defined filter loading state, which can be converted into a numerical value that is easy for humans to interpret using a suitable mapping rule.

[0014] The reconstruction measure “translated” into the similarity value can, for example, be expressed as a filter loading state and / or as a filter capacity in percent and / or be an absolute indication of the loading state, for example a weight fraction, molar fraction, volume fraction, density, specific gravity, etc.

[0015] In accordance with the invention, it was first recognized that the proposed method for determining an unknown filter load in an air-conditioning system allows for filter replacement as needed. This demand-based replacement takes into account both resource-conserving and energy-efficient operation. Thus, a filter change can be recommended even though the filter could continue to be loaded. Due to the increased flow resistance, energy efficiency may decrease. Both models can independently make this trade-off between energy and maintenance costs.

[0016] Thus, a method is provided for determining the unknown loading state of a filter in an air-conditioning system, which allows for more resource-saving and energy-efficient operation of the air-conditioning system. Furthermore, filter replacement as needed is possible.

[0017] The similarity value can be identical to the unknown load state to be determined. According to one possible embodiment, the similarity value is not identical to the unknown load state to be determined. The similarity value can be converted into the load state. The method can preferably comprise the following additional step: g) Determining the unknown load state from the similarity value.

[0018] According to an advantageous development of the method, determining the unknown loading state from the similarity value according to step g) can also include converting the similarity value into a numerical value, preferably a percentage value for the loading state of the filter. Such a percentage value is a numerical value that is easy for humans to interpret.

[0019] With regard to a particularly needs-based filter change, the method may additionally comprise the following step: h) initiating a filter change depending on the determined unknown loading condition, preferably with additional consideration of energy, filter and / or maintenance costs.

[0020] Additional information on current energy, filter, and maintenance costs can be provided to optimize filter replacement as needed. In other words, a filter that is not yet fully loaded can be replaced if, taking these costs into account, the replacement appears preferable to continued operation with the partially loaded filter in terms of resource-saving and energy-efficient operation.

[0021] To calibrate a customized diagnostic model to determine an unknown load state of a filter system in an air-conditioning system, measurements such as power, temperature, and speed can be used at the defined load state. Using this data, the neural network can be trained and / or adapted.

[0022] Preferably, the neural network can be trained with data that was also obtained by continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during a defined filter loading state or over several periods of time during several defined filter loading states. The measured variable or variables used to train the neural network for determining a filter loading state of an air-conditioning system can be the same measured variable or variables that are subsequently continuously or repeatedly recorded according to step a) during an unknown filter loading state.In particular, through a clever selection of the measured variables or combinations of measured variables, the method can ensure particularly resource-saving and energy-efficient operation of the air-conditioning system and / or enable a particularly needs-based filter change.

[0023] For training the neural network for determining the loading state of a filter of an air-conditioning system, which is provided according to step b), a method comprising the following steps is suitable:

[0024] 1 ) Continuous or repeated recording of at least one measured variable of the air-conditioning system over a period of time during a defined loading condition of the filter, and

[0025] 2) Training the neural network, preferably an autoencoder, with the data of the at least one measured variable acquired in step 1), wherein the training comprises:

[0026] 2a) Reducing the collected data to a latent space,

[0027] 2b) Reconstructing the acquired data from the latent space and

[0028] 2c) Determine a reconstruction measure by comparing the acquired data with the reconstructed data.

[0029] The process for training a neural network to determine the filter loading state of an air-conditioning system can be repeated multiple times during the training process. Applying the generated model or trained neural network to new measurement points—i.e., to data from at least one measured variable of the air-conditioning system over a period of time during an unknown filter loading state—results in a new reconstruction measure. This enables a diagnosis of the filter loading state.

[0030] According to an advantageous further development, the method for training a neural network may further comprise the following step:

[0031] 2d) Optimize the reconstruction measure, preferably by adapting the neural network.

[0032] In other words, during model training, optimization can be achieved by minimizing the reconstruction error. The optimization goal is to approximate a global minimum of the reconstruction error so that the input and output values—encoding and decoding—are as consistent as possible.

[0033] In this way, the reconstruction error determined during training for the defined loading condition can ideally be used to determine a very high similarity value. A good reconstruction value or a low reconstruction error indicates a high similarity between the acquired data and the reconstructed data.

[0034] From the reconstruction error for the defined loading state determined in the training, a similarity value can later be determined using the method for determining an unknown loading state of a filter of an air-conditioning system, wherein the similarity value according to step f) quantifies a similarity between the reconstruction error for the unknown loading state and the reconstruction error for the defined loading state of the filter.

[0035] The method for determining an unknown loading state of a filter in an air-conditioning system also enables a prediction of an unknown loading state of a filter in an air-conditioning system at a specific time. A corresponding method for predicting an unknown loading state of a filter in an air-conditioning system at a specific time can comprise the following steps: aa) determining several defined loading states for previous periods using the described method for determining an unknown loading state of a filter in an air-conditioning system, bb) extrapolating the values ​​for the defined loading states for the previous periods to the specific period, and cc) predicting the unknown loading state of the filter in the air-conditioning system at the specific time from the extrapolation.

[0036] In this way, the filter capacity at a specific time between 100% and 0% filter capacity—or conversely, the load level—can be queried. The previous periods can preferably be different from each other. For example, the previous periods may not overlap or may only partially overlap.

[0037] Preferably, before extrapolating the values ​​for the loading conditions at the previous periods to the specific period according to step bb), the following step can be carried out: ab) interpolating the values ​​for the loading conditions at the different periods.

[0038] The larger the number of defined loading states at previous time periods and the longer the histones covered by the time periods in total, the more accurate the forecast can be.

[0039] The method for determining an unknown loading state of a filter in an air-conditioning system also enables a prediction of a time at which a fully loaded loading state of a filter in an air-conditioning system occurs at a specific time between 100% and 0% filter capacity.A corresponding method for predicting a point in time at which a fully loaded loading state of a filter of an air-conditioning system occurs can comprise the following steps: i) determining a plurality of reconstruction measures for previous time periods using steps a) to e) according to the method (10) according to claim 1, ii) storing the reconstruction measures (12) for previous time periods with respective time stamps, iii) approximating a calculation rule for reconstruction measures as a function of time by means of machine learning methods and iv) extrapolating a remaining time until a predetermined value of the reconstruction measure is reached on the basis of the given time stamp.

[0040] Approximating a calculation rule for reconstruction measures may preferably comprise training a further neural network, namely an extrapolation model for the reconstruction measure based on the history of reconstruction measures for the previous time periods.

[0041] If a sequence of reconstruction measures or corresponding similarity values ​​is generated over a longer period of time, a prediction model can be trained over time. For this purpose, several reconstruction measures or similarity values ​​are saved with timestamps. From this information, a calculation rule is approximated for each data sequence using machine learning methods, which is a function of time.

[0042] A sequence can contain either a fixed or a variable number of reconstruction measures, or similarity values, with associated timestamps. A sequence can cover a constant period of time, e.g., every ten minutes.

[0043] For example, a sequence can have a constant length with ten data points, i.e., ten reconstruction measures or similarity values ​​at ten different timestamps. After acquiring a new, tenth data point and subsequently determining the reconstruction error or similarity value, the sequence can additionally include the nine most recently determined reconstruction errors or similarity values, including the associated timestamps. Using the generated sequence, machine learning methods can be used to approximate a calculation rule that maps the reconstruction error or similarity value as a function of time.

[0044] Subsequently, based on the given timestamps, a time delta is extrapolated, which remains until a specified maximum value of the reconstruction measure or a specified similarity value is reached. In the simplest case, the maximum value of the reconstruction error or the worst-case reconstruction measure corresponds to the fully loaded loading state and a remaining filter capacity of zero. However, the maximum value can also be adjusted, for example, so that the maximum value of the reconstruction error is reached at a loading state of 80% or a remaining filter capacity of 20%.

[0045] With regard to a particularly practical further development for the described methods, the at least one measured variable can, for example, comprise one or more of the following:

[0046] Speed ​​information, for example a rotation rate, a motor current and / or a motor power, vibration information, for example an acceleration in one, two or three spatial axes, sound information, for example structure-borne sound, a temperature, a sound pressure, a flow velocity, a torque, a pressure, in particular an operating or ambient pressure, a humidity value and / or a measured force.

[0047] With regard to the device for determining an unknown loading state of a filter in an air-conditioning system, the aforementioned object is achieved by a device comprising at least one sensor for detecting at least one measured variable of the air-conditioning system and a computing unit configured to carry out the method described above. The computing unit can be, for example, a fan controller or an external computing unit.

[0048] With regard to the improved ventilation system, the aforementioned object is achieved by a ventilation system that includes a described device for determining an unknown load state of a filter of the ventilation system. The ventilation system can be a fan, preferably an EC fan with a brushless DC motor.

[0049] With regard to the computer program product, the object mentioned above is achieved by a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method according to claim 1. Thus, according to the object, an improved air-conditioning system with a correspondingly configured device for determining an unknown loading state of a filter of the air-conditioning system and a computer program product are specified.

[0050] There are now various possibilities for advantageously embodying and developing the teaching of the present invention. For this purpose, reference is made, on the one hand, to the claims subordinate to claim 1 and, on the other hand, to the following explanation of preferred embodiments of the invention with reference to the drawing. In conjunction with the explanation of the preferred embodiments of the invention with reference to the drawing, generally preferred embodiments and developments of the teaching are also explained. The drawing shows:

[0051] Fig. 1 is a schematic flow diagram of an embodiment of the method according to the invention,

[0052] Fig. 2 is a schematic flow diagram of a further development of the method according to the invention, Fig. 3 is a graph in which a relative filter capacity of a filter is plotted against a standardized operating time,

[0053] Fig. 4 shows a graph in which a relative filter capacity of a filter is plotted against an exemplary operating time in real time and

[0054] Fig. 5 shows a graph in which a relative filter capacity of a filter is plotted against a possible maximum operating time.

[0055] Fig. 1 shows a general schematic flow diagram of a method. The middle part of the flow diagram depicts a method 10 for determining an unknown loading state of a filter in an air-conditioning system. According to step a), a new data sample of an unknown loading state is provided. The unknown loading state may, in particular, differ from the unloaded loading state. This is done according to step a) by continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during which the filter's loading state is unknown.

[0056] Subsequently, a trained neural network 11 is provided according to step b) to determine a reconstruction measure in the form of a reconstruction error 12 and is then used as follows. The neural network 11 reduces the data of the at least one measured variable acquired in step a) to a latent space according to step c) and then reconstructs the acquired data from the latent space of the trained neural network 11 according to step d). By comparing the acquired data with the reconstructed data, the reconstruction error 12 is determined according to step e). The reconstruction error 12 can be quantified, for example, in a numerical range from 0 to 1000.

[0057] From the reconstruction error 12, a similarity value can be determined according to step f), which quantifies a similarity to the unloaded loading state of the filter. In other words, in this embodiment, a mapping of the reconstruction error to the similarity value takes place. In step g), the unknown loading state can optionally be determined from the similarity value, provided that the similarity value does not already map the unknown loading state. In this case, the similarity value can be converted using a suitable mapping rule into a numerical value 13 that is easy to interpret for humans, for example to a range from 0% to 100% loading, to a filter capacity of 100% to 0%, and / or to a standardized value between 0 and 1 for the loading and / or to an absolute value, for example 5g filtrate.

[0058] The left-hand part of the flow diagram in Figure 1 shows a method 20 for training the neural network 11 to determine a loading state of a filter of an air-conditioning system. In step 1), a data sample of an unloaded loading state is provided. This is done according to step 1) by continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during a defined loading state, for example during a partially loaded loading state or during the unloaded loading state of the filter. According to step 2), the neural network 11 is trained using the data of the at least one measured variable recorded in step 1). The neural network 11 can preferably be an autoencoder.During training, the acquired data is reduced to a latent space according to step 2a) and then reconstructed from the latent space according to step 2b). By comparing the acquired data with the reconstructed data, the reconstruction error in the unloaded state is determined according to step 2c).

[0059] To optimize the reconstruction measure, the neural network is then optionally adapted in step 2d). The adaptation can occur upon specific events, such as limit overrides or reaching certain values ​​or time periods, but it can also be performed at each training iteration. The goal of the adaptation is to minimize the reconstruction error for the database underlying the model training and thus create a high degree of similarity between the input data and the reconstructed data.

[0060] The method for training a neural network 11 for determining the loading state of a filter of an air-conditioning system can be repeated multiple times during the training process, so that, for example, step 1) is performed multiple times and a total of multiple data samples are provided during the training process. Finally, the trained neural network 11 is ready. For continuous improvement and / or calibration, a neural network pre-trained as described can be retrained at specific times during use or continuously by repeating the described method, so that the neural network is further specified and can deliver better results.

[0061] The right-hand part of the flow diagram in Figure 1 shows a method 30 for predicting a point in time at which a filter of an air-conditioning system will reach a fully loaded state. First, according to step i), several reconstruction errors 12 or similarity values ​​for previous time periods are determined using steps a) to e). From such a sequence of reconstruction errors 12 or similarity values, a time profile of the filter's loading state results. The reconstruction errors 12 or similarity values ​​are stored with their respective time stamps according to step ii). From this sequence or from the time profile of the filter's loading state, a calculation rule for reconstruction errors as a function of time is then approximated in step iii) using machine learning methods.Using the approximated function, a remaining time until a predetermined maximum value of the reconstruction error is reached can then be extrapolated in step iv) based on the given timestamps. In other words, the result of the forecast is a time delta 31 until the maximum value of the reconstruction error is reached. The maximum value can correspond to a fully loaded filter load state. Optionally, it can also be adjusted according to step v). For example, the maximum value can be adjusted such that the maximum value of the reconstruction error is reached starting at a load state of 80% or a remaining filter capacity of 20%.

[0062] Fig. 2 shows a general schematic flow diagram according to a further development of the described method for determining an unknown loading state of a filter in an air-conditioning system. Fig. 2 shows a method for predicting an unknown loading state of a filter in an air-conditioning system at a specific point in time. According to step aa), a plurality of loading states for previous time periods are determined using the method according to claim 1. Step aa) essentially corresponds to step i) described in connection with the right-hand part of the flow diagram in Figure 1, wherein according to step aa), not only a plurality of reconstruction errors are determined, but also the similarity values ​​translated therefrom to previous time periods using the method according to the invention for determining an unknown loading state of a filter in an air-conditioning system.The values ​​for the loading conditions for previous periods can be extrapolated. This extrapolation is extended to the specific period according to step bb). Finally, according to step cc), the unknown loading condition of the air-conditioning system's filter at the specific time is predicted from the extrapolation.

[0063] Fig. 3 shows a graph plotting the relative filter capacity of a filter over a standardized operating time. This diagram illustrates the described method for predicting an unknown filter load of an air-conditioning system at a specific time according to Figure 2.

[0064] The normalized operating time of the filter is plotted on the abscissa 50, where the value 0 corresponds to the start of the filter operating time and the value 1 to the end of the filter operating time. A relative filter capacity is plotted on the ordinate 51, where a filter capacity of 1.0 corresponds to an unloaded load state and a filter capacity of 0.0 to a fully loaded load state. The operating time plotted on the abscissa 50 is normalized in such a way that the filter capacity - or complementarily the filter load - can be plotted as a linear function. Under real operating conditions, the filter capacity or the load state does not necessarily have to follow a linear function.

[0065] The solid line describes a target value 52, where the target value corresponds to the actual, real filter capacity. The plotted points correspond to predictions 53 for loading states and show the result of a mapping, or in other words, a transfer of reconstruction errors to loading states. The discrete-time predictions 53 may deviate from the actual loading according to target value 52.

[0066] Two extrapolation lines are shown in Fig. 3. To enable a forecast of the filter loading state, the history of time-discrete predictions determined according to step aa) is recorded. In the case shown, for the dashed line 54, all model predictions between 1.0 and 0.75 - i.e., between 100% and 75% - filter capacity are saved. A linear regression is performed using these data points. This results in a linear interpolation line between 1.0 and 0.75 - i.e., between 100% and 75% - filter capacity and a linear extrapolation line between 0.75 and 0.0 - i.e., between 75% and 0% - filter capacity.

[0067] If the history is viewed over a longer period, in the case of the dotted line 55, extrapolation using the predictions between 1.0 and 0.5—i.e., between 100% and 50%—filter capacity results in a line that approximates the target value 52. The accuracy of the prediction can be increased by adding more data points.

[0068] With the described procedure, the regression does not necessarily have to be linear. Extrapolation does not necessarily have to be performed using the entire histone sequence, but can also be achieved using a sliding window with a defined window length. This allows extrapolation to be performed with a suitable number of data points. For example, extrapolation can be performed with one, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, up to 20, up to 50, or even up to 100 or more previous data points.

[0069] Furthermore, the extrapolation can also be carried out using the history of data points over a suitable sequence or a suitable defined period of time, e.g. all data points within the last minute, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, up to 20, up to 50 or even up to 100 seconds / minutes / hours and / or days. In addition, the filter capacity / filter load can also be extrapolated whenever the prediction of the relative filter capacity has decreased by a certain value. For example, a new extrapolation can always be carried out when the capacity has decreased by one, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, up to 20, up to 50 or even up to 75%.

[0070] Fig. 4 shows a graph in which a relative filter capacity of a filter is plotted against an operating time in real time. Such a diagram illustrates the described method for predicting a point in time at which a fully loaded loading state of a filter of an air-conditioning system occurs, according to the right-hand part of Figure 1. The operating time of the filter is plotted in real time on the abscissa 50 with exemplary dates. The relative filter capacity is plotted on the ordinate 51, as in Fig. 3, where a filter capacity of 1.0 corresponds to an unloaded loading state and a filter capacity of 0.0 to a fully loaded loading state.

[0071] Fig. 4 illustrates how a non-normalized point in time can be determined at which the relative filter capacity is 0.0 or 0%. For this purpose, the values ​​of the predictions 53 are recorded together with timestamps (or time differences between the predictions). A linear regression is performed using the history of past values ​​of the prediction 53. This results in a linear interpolation line 60 for past measurement points and a linear extrapolation line 61 for future measurement points. The extrapolation line extends over a time interval that shows the remaining time difference before the filter is fully loaded. The remaining time difference is offset against the time of the most recent prediction. This results in the remaining time or the non-normalized point in time at which the relative filter capacity is 0%, according to step iv).The regression, as well as the interpolation and extrapolation, do not necessarily have to be linear.

[0072] Finally, Fig. 5 shows a graph in which a relative filter capacity of a filter is plotted against a possible maximum operating time.

[0073] The abscissa 50 represents a possible maximum operating time—also called service life—of the filter in hours. The ordinate 51 represents a relative filter capacity, with a filter capacity of 1.0 corresponding to an unloaded condition and a filter capacity of 0.0 corresponding to a fully loaded condition.

[0074] To the left of the operating time value "0," relative predictions 53, or forecasts of the relative filter loading, are shown as points. The predictions 53 correspond to the displayed points and are model predictions determined using the method for determining an unknown loading state. The points or predictions are similarity values ​​or filter loading states at discrete points in time. The predictions 53 represent the respective relative filter loading at the respective discrete points in time.

[0075] Forecast values ​​74 represent the total of all forecasts for the remaining hours until a filter capacity of 0.0 is reached. Forecast values ​​74 correspond to the vertical lines shown on the abscissa 50.

[0076] The forecast values ​​74 are determined by extrapolation from the points 53, namely first over the first two model predictions, if a third model prediction is present, then over the first three model predictions, then over the first four model predictions, etc. The extrapolations result in the family of lines 71 and thus the forecast values ​​74.

[0077] Thus, from the relative predictions 53, a range of absolute filter loading states to the right of the value "0" can be determined by extrapolation, which is shown as a family of straight lines 71. At the bottom of the abscissa 50, this extrapolation results in forecast values ​​74 for the expected maximum operating time of the air-conditioning system. The probability that the air-conditioning system has already reached its individual maximum operating time after the respective operating time follows a distribution function 75, for example a probability density function. This distribution function 75 can be generated across all forecasts. The distribution function can also depict the temporal development of the forecasts. The reliability of the forecasts can be derived from the distribution function.The data from the distribution function 75 can in turn be transferred to a survival function 76, which indicates the survival probability of the filter and / or the air-conditioning system as a function of operating time. With respect to the survival function, the relative filter capacity corresponds to the survival probability after maximum operating time according to the abscissa 50.

[0078] With regard to further advantageous embodiments of the device according to the invention, reference is made to the general part of the description and to the appended claims in order to avoid repetition.

[0079] Finally, it should be expressly pointed out that the embodiments of the device according to the invention described above serve only to explain the claimed teaching, but do not limit it to the embodiments.

[0080] List of reference symbols

[0081] 10 Procedures a), b), c), d), e), f), g), h) Steps 11 Neural network 12 Reconstruction error 13 Numerical value 20 Procedures 1), 2), 2a), 2b), 2c) Steps 30 Procedures i), ii), iii), iv), v) Steps 31 Time delta 40 Procedures aa), ab), bb), cc) Steps 41 Time delta 50 Abscissa 51 Ordinate 52 Target value 53 Predictions 54 Dashed line 55 Dotted line 60 Interpolation line 61 Extrapolation line 71 Family of lines 74 Forecast values ​​75 Distribution function 76 Survival function

Claims

A n s p r ü c h e 1.Method (10) for determining an unknown loading state of a filter of an air-conditioning system, the method comprising the following steps: a) continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during an unknown loading state of the filter, b) providing a trained neural network (11) for determining a reconstruction measure, c) reducing the data of the at least one measured variable recorded in step a) to a latent space by means of the trained neural network (11), d) reconstructing the recorded data from the latent space by means of the trained neural network (11), e) determining the reconstruction measure by comparing the recorded data with the reconstructed data and f) determining a similarity value from the reconstruction measure, the similarity value quantifying a similarity to a defined loading state of the filter.

2. The method (10) according to claim 1, wherein the method further comprises the following step: g) determining the unknown loading state from the similarity value, wherein preferably the determining of the unknown loading state from the similarity value comprises converting the similarity value to a numerical value, for example to a percentage value for the loading state of the filter.

3. The method (10) of claim 1 or claim 2, wherein the method further comprises the following step: h) Initiating a filter change depending on the determined unknown load condition, preferably taking into account energy, filter and / or maintenance costs.

4. Method (10) according to one of claims 1 to 3, wherein the neural network (11) has been trained with data obtained by continuously or repeatedly recording at least one measured variable of the air-conditioning system over a period of time during a defined loading state of the filter.

5. A method (20) for training a neural network (11) for determining a loading state of a filter of an air-conditioning system, the method comprising the following steps: 1 ) Continuous or repeated recording of at least one measured variable of the air-conditioning system over a period of time during a defined loading condition of the filter, and 2) Training the neural network (11), preferably an autoencoder, with the data of the at least one measured variable acquired in step 1), wherein the training comprises: 2a) Reducing the collected data to a latent space, 2b) Reconstructing the acquired data from the latent space and 2c) Determine a reconstruction measure by comparing the acquired data with the reconstructed data.

6. Method (40) for predicting an unknown loading state of a filter of an air-conditioning system at a specific time, the method comprising the following steps: aa) determining a plurality of defined loading states at previous time periods using the method (10) according to claim 1, bb) extrapolating the values for the defined loading states at the previous time periods to the specific time period and cc) predicting the unknown loading state of the filter of the air-conditioning system at the specific time from the extrapolation.

7. The method (40) according to claim 6, wherein the method comprises the following step before extrapolating the values for the loading conditions at the previous time periods to the specific time period according to step bb): ab) interpolating the values for the loading conditions at the different time periods.

8. Method (30) for predicting a point in time at which a fully loaded loading state of a filter of an air-conditioning system occurs, the method comprising the following steps: i) determining a plurality of reconstruction measures for previous time periods using steps a) to e) according to the method (10) according to claim 1, ii) storing the reconstruction measures (12) for previous time periods with respective time stamps, iii) approximating a calculation rule for reconstruction measures as a function of time by means of machine learning methods and iv) extrapolating a remaining time until a predetermined value of the reconstruction measure is reached on the basis of the given time stamp.

9. Method (10, 20, 30, 40) according to one of claims 1 to 8, wherein the at least one measured variable comprises one or more of the following: Speed information, for example a rotation rate, a motor current and / or a motor power, vibration information, for example an acceleration in one, two or three spatial axes, sound information, for example structure-borne sound, a temperature, a sound pressure, a flow velocity, a torque, a pressure, in particular an operating or ambient pressure, a humidity value and / or a measured force.

10. A device for determining an unknown loading state of a filter of a ventilation system, comprising: at least one sensor for detecting at least one measured variable of the ventilation system and a computing unit configured to carry out the method according to one of claims 1 to 8.

11. A ventilation system comprising a device according to claim 10.

12. A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the method according to any one of claims 1 to 9.

Images