TECHNICAL FIELD OF THE DISCLOSURE

The present disclosure relates to spectrometer calibration.

BACKGROUND

An optical spectrometer is an instrument used to measure properties of light over a specific portion of the electromagnetic spectrum. Spectrometers can be used, for example, to identify particular materials, as illuminating different materials will result in different reflection profiles. The variable measured is sometimes the light's intensity, with the independent variable being the wavelength of the light. Some spectrometers measure spectral regions in or near the visible part of the electromagnetic spectrum, although some spectrometers also may be able to measure other wavelengths, such as the infra-red (IR) or ultraviolet (UV) parts of the spectrum.

In reflectance spectrometers, the spectrometer measures the fraction of light reflected from a surface as a function of wavelength. Reflectance measurements can be used to determine, for example, the color of a sample, or examine differences between objects for sorting or quality control.

In some instances, spectrometers are manufactured as small, compact modules that contain the required optoelectronic components (e.g., light source and optical sensor) in a housing under a cover glass. Light produced by the light source is emitted from the module toward a sample under test. Light reflected by the sample under test is detected by the sensor.

Manufacturing processes for the spectrometer modules may result in variations in fabrication, tolerances, and variability of the multiple components of the system. Such variations can result in unintended variations from one module to the next, e.g. the reflection spectrum for a given sample and a given spectrometer module may appear shifted in wavelength compared to that for the same sample and a different spectrometer module. It is therefore usually necessary to perform a calibration of individual modules before they leave the factory, or possibly at a later stage but prior to using it to measure the reflectivity of unknown samples.

Optical spectrometers may employ silicon micro electro-mechanical systems (MEMs) technology, and in particular MEMs employing a tunable Fabry-Perot interferometer (FPI). These devices are referred to here as “spectral sensors”. MEMS based-FPIs typically include a vertically integrated structure composed of two mirrors separated by an air gap. Wavelength tuning is achieved by applying a voltage between the two mirrors, which results in an electrostatic force, which pulls the mirrors closer to one another. Calibration of such an optical spectrometer may involve directing narrowband light onto a detector of the spectrometer and varying the control voltage to determine the voltage at which the detector output is a maximum. Calibrated crosstalk and dark noise configuration values may then be determined. A full calibration record is then determined using this data and measuring the system response with a reference material having a known reflectivity response. The full calibration record is stored in memory coupled to the spectrometer. In some cases, these operations must be repeated to calibrate the sensor across a range of temperatures, e.g. using a climatic chamber.

The calibration of spectral sensors is a non-negligible part of the cost of producing sensors. In particular, the calibration of MEMS-based spectrometer sensors requires the measurement of various spectra at all points of operation, and recording of the corresponding look-up-table of driving voltage versus transmission wavelength into memory. Whilst this has not been a significant issue in the case of low volume production of essentially laboratory instrumentation, it does represent a barrier to the mass production of spectral sensors such as is now desirable, e.g. where such sensors are greatly miniaturized for deployment in consumer electronic devices such a smartphones.

SUMMARY

According to a first aspect of the present invention there is provided a method of preparing devices comprising respective spectrometer modules to perform one or more specific applications, the spectrometer modules being part of a batch of spectrometer modules. The method comprises, for the or each said application, performing optical measurements using a subset of said batch of spectrometer modules to generate respective sets of wavelength-versus-operating parameter data, and using said data to train a neural network to successfully perform the or each application. The trained neural network is stored in memories of each of said devices, wherein said devices are configured to perform the or each said application by generating spectral data using respective spectrometer modules and applying that spectral data or data derived therefrom to the stored neural network.

The or each said application of the method may be the recognition of one or more given material types to which a spectrometer module is exposed.

Each set of wavelength-versus-operating parameter data may comprise data collected from a plurality of samples matching an application and a plurality of samples not matching the application.

The spectrometer modules may be micro electro-mechanical systems (MEMs) modules, for example MEMs-based tunable Fabry-Perot interferometers (FPIs), whilst said neural network may be a Convolutional Neural Networks (CNNs), where the step of training the CNN comprises determining biases and weights for neurons of the CNN.

The devices may be portable computer devices such as smartphones.

The trained neural network may be stored into memories of the devices at a device production facility. The trained neural network may stored into memories of the devices via a wireless data transfer process, e.g. at said production facility or at a later stage.

The spectrometer modules of said batch may be associated by one or more of a common production site, production during a defined time period, and production conditions. According to a second aspect of the present invention there is provided a device comprising a spectrometer module and a memory and configured to perform one or more applications using said spectrometer module, the memory storing a trained neural network produced by the method of the above first aspect of the invention.

The or each said application may be the recognition of a material type of a sample to which the spectrometer module is exposed.

According to a third aspect of the present invention there is provided computer program stored on a non-transient computer storage medium and configured to use a trained neural network generated by the method of the above first aspect of the present invention to determine whether or not wavelength-versus-operating parameter data generated by a spectrometer module satisfies a condition of one or more applications.

Embodiments of the present invention may provide spectrometer modules and devices that require little or no calibration and which are not required to have unique identifiers. This avoids the need for the installation of calibration files into the modules and devices and may offer higher operating speeds as they avoid the need to load calibration files prior to operation. Memory storage requirements may be reduced. Furthermore, whilst the applications initially available for the spectrometer module or device may initially be limited to a small number. Further applications may be “shipped” and installed as software updates at a later time.

Other aspects, features and advantages will be readily apparent from the following detailed description, the accompanying drawings, and the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a flow diagram illustrating a method of training a neural network suitable for use with a spectrometer module for a given application or applications;

Figure 2 illustrating a method of using a device to perform an application; and Figure 3 illustrates schematically a user device comprising a spectrometer module and a trained neural network.

DETAILED DESCRIPTION

The following discussion concerns a method for producing spectrometer modules and devices containing such modules with particular emphasis on miniaturized MEMS-based spectrometers. An example of such a sensor module is that manufactured and sold by AMS AG, Premstaetten, Austria, under the product code AS7341. However, the method is also applicable to other spectrometers such as AS7265x. A primary object is the need to avoid having to individually calibrate spectrometer modules by providing each of them with unique calibration tables. Rather, a method is proposed that requires testing, e.g. in a manufacturing facility, of only a sub-set of a manufactured batch of spectrometer modules. In a non-limiting example, testing may be performed on only a few hundred modules out of a total manufactured batch of many thousands of modules.

The method proposed here takes advantage of the fact that the end-user application or applications of a spectrometer module is or are known a priori. In the case of multiple applications, these are likely, although not necessarily, to be relatively small in number. This makes possible the use of artificial neural network technology to perform the required application(s), where a neural network is trained using only a sub-set of a manufactured batch of spectrometer modules.

The skilled person will appreciate that artificial neural networks can be implemented using computer code (software) running on suitable hardware including a processor or processors and memory. Nonetheless, the use of partial or fully hardware implemented solutions is not excluded. In any case, once suitably trained, a neural network is able, with some desired level of confidence, to implement an application by recognizing whether or not a complex set of input data is or is not indicative of a result on which the network has been trained. A particular class of artificial neural networks is known as Convolutional Neural Networks (CNNs). As with other classes of neural networks, CNNs comprise interconnected layers of “neurons” where individual neurons compute an output value by applying a specific function to the input values coming from the previous layer (or input data for the lowermost layer). In a CNN layer, neurons receive inputs from only a restricted subarea of the previous layer. This makes them particularly well suited to vision applications as the structure mimics that of the visual cortex. The function that is applied to the input values is determined by a vector of weights and a bias (typically real numbers). Learning, in a neural network, progresses by making iterative adjustments to these biases and weights. Such training seeks to adjust the weights used by individual neurons to minimize the difference between the actual output and an expected output of the network. This process is run a number of times using different training sets. CNNs are considered well suited to the method described here, but other classes of neural networks may also be suitable.

Consider a manufacturing facility that produces a batch of 10,000 MEMS-based spectrometers. Of these, a sub-set, e.g. 100, are selected, possibly at random, to train a CNN for a given application. An exemplary application may be the recognition of a material as being plastic. For a first of the selected 100 spectrometers, the spectrometer is “shown” as an example of the plastic to be recognized. The control voltage is varied to generate a scan across the wavelength range of the spectrometer and the output of the detector collected. The control voltage versus output (intensity) is stored as a first data set. This may be repeated for a number of different examples of the plastic to be recognized, e.g. with different colours and surface textures. This in turn is repeated for each of the remaining 99 spectrometers. The result is numerous data sets taken across all of the sub-set of sensors and for one or more examples of the plastic. This process is also be performed for one or more material types other than the plastic to be recognized, e.g. fabric, metal, wood, etc.

The collected data is then used to train the CNN so that it is able to discriminate between the data sets collected for the plastic to be recognized and those for other materials. The result is a set of biases and weights for the CNN. It will be appreciated that a CNN configured using this data will likely be able to recognize a data set as being indicative of the plastic even if the data set is not identical to one of the (plastic) training sets. The CNN structure and the biases and weights can then be saved to memories of or associated with the spectrometers.

An alternative approach relies upon retaining a subset of spectrometers at the factory or other site such that this subset can be subsequently used to train for additional applications. It is also possible to use an inverse function to transform spectra obtained for a given material into new, virtual spectra for a different material. These virtual spectra can then be used to train a CNN to recognize the different material.

It is of course possible to train a single artificial network to perform multiple applications. For example, the neural network may be trained to recognize multiple different materials, e.g. plastic and metal, and to provide a result that distinguishes between these materials. In some cases, at the time of production, training data may be collected for multiple applications, with the neural network only being trained on a subset of these applications. This allows for the later release of an updated neural network with additional capabilities.

In a non-limiting example, the spectrometers may be incorporated into smartphones, with the CNN structure, biases and weights being stored in memories of the smartphones. This may be done at the smartphone production facility, e.g. using a knowledge of the manufacturing batch number of the spectrometers. Alternatively, the data may be stored in the smartphone memories at a later stage, for example when an end-user installs an app that wants to make use of the spectrometer, i.e. using a wireless data transfer process.

Figure 1 is a flow diagram illustrating the method described above. At step S100, a subset of a manufactured batch of spectrometer modules is selected for testing. At step S200, taking into account the particular application for which the neural network is to be trained, training data sets are obtained for each tested module. This might involve exposing each module to a number or different material types some of which satisfy a test condition, e.g. they are plastic, and some of which do not. If the network is being trained for a plurality of applications, this may be repeated for each application.

At step S300 the collected training data sets are then used to train a neural network, e.g. using a back propagation procedure. The result is a trained neural network that can determine, for each application, whether a condition is satisfied, e.g. is the input data indicative of a required material type. At step S400 the trained neural network record, i.e. its structure and operating parameters, is stored in a some central storage means. At step S500 the trained neural network record is distributed to devices utilizing the spectrometer modules of the batch in question and is saved to respective device memories.

Figure 2 is a flow diagram illustrating the method that is carried out at each of the devices in order to perform the application (or applications). At step SlOOa, a sample to be tested is exposed to the spectrometer module of the devices. This may of course involve activating the module using some graphical user interface or other interface of the device. The module will be caused to cycle through the control voltage and the output wavelength of the module. The resulting wavelength-versus-operating parameter data is collected for the sample. At step 300a the device (processor) retrieves the trained neural network record from its memory and runs a process to apply the collected data to the trained neural network of the record. At step 400a a result is obtained, i.e. the output of the trained neural network. It is then determined whether or not the output satisfied a condition of the application, e.g. is the collected data indicative of a given material type.

Figure 3 illustrates schematically a device 1 utilizing a spectrometer module 2 and configured to implement an application or applications of the type discussed above. The Figure also illustrates a sample 3 on which the application or applications is being performed. This sample might be, for example, a product for which a user wishes to determine a material type, e.g. is the material type plastic. The spectrometer module 2 is operated in an essentially known way by a processor 4 of the device, i.e. to obtain wavelength versus-operating parameter data. To do this the processor will retrieve code and operating parameters from a device memory 5. The collected data set is stored, e.g. in the device memory 5. The processor then retries the trained neural network record from the memory 5, along with code required to run the network of the record. The collected data is provided as input to the trained neural network and a result generated.

Typically, for example where the device is a smartphone, the device will implement through the processor, display (not shown) etc., a graphical user interface (GUI) 6 that allows a user to initiate the sample test and view the results. A result might indicate for example that the sample is a particular material or not.

Various modifications may be made to the foregoing implementations, and features described above in different implementations may be combined in the same implementation. Further, unless expressly stated or implicitly required, the various operations may be performed in a different order than set forth in the foregoing examples. Some implementations may omit some operations and/or may include additional operations. Thus, other implementations are within the scope of the claims.