Technical Field

[0001 ] The present invention relates to Radio Frequency Identification (RFID). In particular, the present invention relates to a chipless RFID system and method operable in the extremely high frequency band.

Background of Invention

[0002] RFID, unlike optical barcodes which work in the visual light spectrum, uses electromagnetic waves, which eliminates the Line-of-Sight (LoS) requirement (associated with 2-D barcodes, for example) and makes reading possible at further distances.

[0003] RFID tags are utilized in many areas such as inventory management, animal and health tracking, intelligent and remote sensors, and the like.

[0004] Chipped tags can be either active, semi active or passive. These tags contain application specific integrated circuit (ASIC) chipsets, they are powered up by either an on-board battery in the active/semi-active tags, or interrogation signal energy in the passive tags. The chipped tags respond to the readers by their communication protocols and reveal their encoded data. The chipless RFID tags, on the other hand, are simply printed metallic patterns and contain no chipset. Due to the void of intelligence in the tag, it is the reader that performs all signal processing tasks for interrogation and decoding the tag ID. The chipless tags are acting as both backscatterers and data encoders simultaneously. The wide demand and use of Chipless RFID drives demand for providing a Chipless RFID reader and tag which is operable at the extremely high frequency band. This is driven because of the benefits of the band such as less frequency interference, smaller tag sizes and higher encoding capacities and the available free ISM spectrum.

[0005] Chipless RFID tag design generally requires tags which are orientation free, include an anti-collision mechanism, are easy to implement and print and with higher encoding capability in lower finger prints. [0006] However, there are problems with providing Chipless RFID at this extremely high frequency band, such as greater attenuation in the air and other substances, more complex circuitry and the like. The challenging aspects of this method also relate to the design and implementation of reader hardware, development of the interrogation and detection algorithms, and reading of tags with much lower Radar Cross Sections (RCS) compared to their counterparts at microwave frequency bands.

[0007] It would therefore be desirable to provide a Chipless RFID method and system which can operate in the extremely high frequency (EHF) band.

Summary of Invention

[0008] According to a first aspect, the present invention provides, a method of reading one or more RFID chipless tags in the extremely high frequency band, the method including the following: (a) determining background noise; (b) providing the tag in a reader, the tag providing a backscattered signal; (c) receiving the backscattered signal from tag at one or more locations; and (d) providing the backscattered signal to a machine learning system to determine the pattern.

[0009] It will be appreciated that step (c), the received signal may be stored, the one or more locations may include one or more different angles and the like. It will further be appreciated that receiving the backscattered signal from tag at one or more frequencies or transmitter power may also be provided.

[0010] It will be appreciated that at step (d), providing the backscattered signal may be along with other dimensions such as tag/reader distance and angle, transmitter power and receiver sensitivity.

[001 1 ] Advantageously, the present invention may operate in the extremely high frequency band and therefore provides less frequency interference, smaller tag sizes and the feasibility of using microwave image scanning techniques to enhance reading and prevent tag collisions. The present invention advantageously, provides a system for extremely high frequency hybrid chipless RFID. Advantageously, in combination with millimetre wave (i.e. EHF) spectrum image scanning and frequency scanning techniques, tags may be scanned in a few relative positions to each other in the available spectrum (scanning in a few positions or using the linear rail might not be needed if the number of tags are few, for example 10). Pattern recognition methods are then used to categorize the tag's radar cross section (RCS) for different tag combinations which eliminates the need for high-Q planar tag resonators (frequency response method) or expensive substrates (time techniques). Further, the requirements for high precision tag image processing/recognition are alleviated.

[0012] Preferably, at step (c), the backscattered signal is received from the tag at one or more locations. In an alternative or in addition, at step (c), the backscattered signal is received from the tag at one or more locations or angles. In an alternative or in addition at step (c), the backscattered signal is received from the tag at one or more power levels. In an alternative or in addition, at step (c), the backscattered signal is received at one or more frequency steps in the available spectrum.

[0013] Preferably, the reader captures the data via microwave scanning/remote sensing techniques. Microwave scanning may include SAR/iSAR (Synthetic Aperture Radar/inverse SAR), microwave radiography, etc.

[0014] The symbolic tag may be a chipless RFID tag or the like.

[0015] Preferably, at step (b) the reader generates a baseband (BB) signal to be modulated by transmitter, and reads the backscattered signal by measuring one or more of amplitude (and phase) of the received BB signals. The inputs can be up to 4 for In Phase Positive Baseband Input (BBJP), In Phase Negative Baseband Input (BBJM), Quadrature Positive Baseband Input (BB_QP) and Quadrature Negative Baseband Input (BB_QN.

[0016] Advantageously, use of four inputs allows for sharper spectrum with less phase noise in detected BB output.

[0017] At step (b), gain/phase could alternatively be utilized instead of l/Q amplitude detection. [0018] Preferably, scattering parameters associated with the tag and background noise are determined by deducing l/Q of the sent signal from the detected l/Q in the receiver. These deducted signals will be considered as backscatters.

[0019] The scattering parameters may be determined by deducing gain/phase of the sent signal from the detected gain/phase in the receiver (converting to real/imaginary parts and deduction sent and received from each other).

[0020] The scattering parameters are preferably read over a range of frequencies, number of different transmission powers and in a number of different positions (like strip mode SAR) or angles (like iSAR). Positions don't need to be linear, but the same pattern of positions/angles should be used for all tags, both in learning phase of machine learning and in decoding phase. The positions may be provided along a linear rail, for example, with D/2 steps, where D is the diagonal of the interrogator antenna, which is approximately 6mm in 60GHz for our horn antennas in the experiments and simulation, if SAR strip-mode image formation is in place.

[0021 ] It will be appreciated that a linear rail need not be provided. For example, if there is only a few numbers of tags (e.g. less than 10) there might be no need for tag scanning in different positions, as the machine learning can recognize the tags quite efficiently.

[0022] The present invention may detect a tag in two ways: either purely via machine learning (or Artificial Intelligence) or a combination of image-processing and Al. For the Al-only based method, no strict steps needed as far as tag data can be decoded using an already trained network.

[0023] Preferably the method further includes the step of, for each position, subtracting the scattering parameters in step (a) from the detected scattering parameters at step (c) thereby removing background noise from the scattering parameters.

[0024] Preferably, step (d) further includes creating an image based on the collected data using image formation techniques, and sending the symbols to a machine learning system for further recognition. Image formation techniques which may include for example iSAR. It will be appreciated that any microwave image formation technique may be utilized. An image is then sent to the machine learning system for learning and testing recognition.

[0025] Advantageously, tag decoding may be carried out in two ways. Either by way of a tag image being formed using collected scattering parameters and reader/tag geometric positions or by sending collected scattering and tag location data to a machine learning system without image reproduction. In this way the whole tag will be recognized as one tagID which results in a faster read of the tag.

[0026] In an alternative, if the number of patterns does not exceed the predetermined threshold, one or more of the scattering parameters, relative tag/antenna position/angle and frequency are provided to the machine learning system for learning/testing and recognition.

[0027] Preferably the machine learning system includes supervised/semi- supervised training, with deep-learning techniques.

[0028] According to a second aspect, the present invention provides a chipless RFID tag operable in the extremely high frequency band, including: an RFID reader circuit, and one or more printed symbols associated with the tag wherein the printed symbols produce a distinguishable reflection signal in response to interrogation by a reader.

[0029] Advantageously, the present invention provides a new tag with higher encoding bit rates per footprint based on printed symbols being readable less dependent of their orientation and being ideal for character recognition by machine learning systems.

[0030] Preferably, the printed symbols are chosen such that they provide a level of scattering reflection. A suitable font type may include peyote. The symbols may be further chosen such that they have maximum scattering properties, have different scattering patterns compared to each other and/or are better adapted to being distinguished by machine learning methods. [0031 ] According to a third aspect, the present invention provides a system for reading RFID symbolic tags operable in the extremely high frequency band, the system including: a micro controller adapted to control a reader for reading one or more tags and a machine learning system component for recognizing the tags, wherein the micro controller is adapted to: (a) determine background noise via the reader; (b) determine a backscattered signal provided by the tag; (c) receive the backscattered signal from the tag at one or more locations; and (d) provide the backscattered signal to a machine learning system to determine the pattern.

[0032] According to a fourth aspect, the present invention provides, one or more symbolic tags, including one or more characteristics such that they are adapted to being distinguished by machine learning methods.

[0033] Preferably the characteristics include one or more of: (a) the individual symbolic tags being arranged such that it has improved scattering properties (like changing the fonts, or using bold fonts while printing), (b) the symbolic tags arranged or printed such that they have different scattering patterns when compared to each other (such as the different scattering pattern for 123 tag compared to 132 tag).

Brief Description of Drawings

[0034] Figure 1 is a schematic diagram illustrating the chipless RFID system and method of the present invention;

[0035] Figure 2 is a flow diagram illustrating the method of the present invention in a first embodiment;

[0036] Figure 3 are examples of RFID tags of the present invention;

[0037] Figures 4a, 4b and 4c are further examples of RFID tags using different substrate of printing inks;

[0038] Figure 5 is a flow diagram illustrating the method of the present invention in a second embodiment;

[0039] Figure 6 is a chart which illustrates measured scattering parameters over frequency range in simulations; [0040] Figure 7 illustrates error correction in the machine learning system of the present invention (simulation phase using feed forward network);

[0041 ] Figure 8 is a chart illustrating that the system of the invention may adjust itself to new scenarios that it has not been trained for to recognize RFID tags (simulation phase);

[0042] Figure 9 is a chart illustrating the results of an experiment where a comparison between different classifiers is provided, showing the performance of the system and method of the present invention ;

[0043] Figure 10 is a snapshot showing magnitude (S21) measured with and without a tag using vector network analysis (VNA);

[0044] Figure 1 1 a is a diagram of a simulation configuration in electromagnetic simulation software;

[0045] Figure 1 1 b is a chart illustrating the S21 values against frequency after relative movement of the reader and tags in simulations;

[0046] Figure 12 is a chart showing the relative magnitude and phase of one measured RFID tag (in the lab experiments) in a particular distance with a fixed transmitter power; and

[0047] Figures 13 and 14 show example hardware utilised in reading tags. Detailed Description

[0048] Figure 1 is a schematic diagram illustrating the chipless RFID system and method of the present invention which is operable in the extremely high frequency band (i.e. 30 to 300 GHz). In practice, the extremely high frequency band may be 57 to 64 GHz (this is the Global industrial, scientific and medical (ISM) band which is from 57 to 64 GHz, although in some jurisdictions, it might be 59-64 GHz (for Radiolocation applications). The system 100 includes one or more chipless RFID tags 105 which may be electromagnetically coupled to a reader 1 15 which has receiving and transmitting antennas 120R and 120T (sometimes one antenna can be used both as transmitter and receiver, but here we used cross-polarized antennas) if, such that when the tag 105 is placed in proximity with the reader 1 15 the tag may be read by the reader 1 15. It will be appreciated that receiving and transmitting antennas can be integrated into one to reduce circuit complexity. Also provided is a computing component 125 which receives data from the reader 1 15 in order to identify the tag 105. Also included is a machine learning system component 130 which allows learning of the tag microwave scattering parameters 105.

[0049] The system 100 of the present invention provides a tag 105 which allows for higher encoding bit rates per footprint, based on alphanumeric or symbolic characters which are less orientation dependent and additionally optimised for character recognition by the machine learning system component 130. Microwave scanning techniques are used for scanning of the tag 105. Raw data from the tag 105 is post-processed via the machine learning system component 130 for recognition of the tag 105. It will be appreciated that infrastructure on which the present invention operates and may be a local or cloud based infrastructure.

[0050] Advantageously, our current utilization of the alphanumeric or symbolic characters together with the machine learning system component 130 is for the microwave scanning at extremely high frequencies.

[0051 ] Figure 2 is a flow diagram illustrating the method 200 of the present invention. The method may be carried out on the system 100 in Figure 1 . In operation, the method 200 includes utilizing the reader 1 15 and at step 205 taking a measurement of the background noise present in the area of the tag reader so as to make the process of reading the backscattered signal from the tag 105 more accurate.

[0052] Control then moves to step 210 in which the tag 105 to be read is provided to the reader 1 15 and the tag 105 provides a back scattered signal which is then read by the reader 1 15. Control then moves to step 215 where the system 100 receives and determines the backscattered signal before control moves to step 220 where backscattering parameters are provided to a machine learning system (where learning testing and recognition is carried out) or to step 225 where an image processing step is carried out before being provided to a machine learning system. [0053] As noted above, tag detection may be carried out in two ways, either via a machine learning system only at step 220 or by step 225 and 230 which further includes the step of image formation (step 220) and then utilizing the machine learning system for image recognition (step 230).

[0054] The machine learning system may take any form such as supervised machine learning systems (which includes feed forward networks, K- Nearest Neighbors (K-NNs), Support Vector Machines (SVMs) and the like). Additional steps may be carried out to train the system 100 for all combination of tags 105 or symbols and for example if there is an untrained pattern or tag 105 it can be put into the reader 1 15 and in turn the machine learning system 130 to add it to the trained system 100. Advantageously the present invention provides a library of identified tags for a user to utilize and be comfortable that their tags will be able to be read at high frequency and with less orientation problems because of the recognition method, which is machine learning.

[0055] Figure 3 illustrates an example 300 of a tag 105 (such as that shown in Figure 1 ), but further illustrating the tag being printed with alphanumeric or symbolic characters. The alphanumeric or symbolic characters are less dependent on orientation. The example tags 300 may be thought of as symbolic tags which are utilized in the present invention for reading in the extremely high frequency band.

[0056] Advantageously the tags associated with the present invention are sensitive to orientation problem, such that they can be recognized even with slight orientation difference by a reader 1 15. As a result of using machine learning techniques, the example tags 300 are less sensitive to printing imperfections. For example tag 305 may be printed with a conventional Sato™ printer (or ink printing techniques using plastic substrate and conductive inks, such as M-Creative or InkTec) and as it can be seen, despite possible printing errors and the like, can be distinguished from one character to another and also can be distinguished with slight orientation of the tag. The less orientation dependency nature of the tags 300 allows the machine learning system component 130 (with training) to correctly identify the tag 300. [0057] In another example addition, tag 310 (fig 10/13A) may be printed directly to a printed circuit board and is readable to a machine learning system 130 with training. For example, the system can be utilized for different tags on different substrates as shown in Figure 4. Tag 400 shown in Figure 4a is a tag printed on FR4 (a glass- reinforced epoxy laminate material) with a permittivity ε _Γ of 4.3. Tag 410 shown in Figure 4b is a tag printed using conductive InkTec™ and tag 420 shown in Figure 4c is using M-creative ink, last two are printed on a Ι ΟΟμιτι Mylar™ polyester film.

[0058] Returning to Figure 3, the symbol selected to print on tags 305, 310 may include a font or symbols. The font can be read either by the human eye or via the machine learning system 130. Very close symbols, such as 1 , 7, L or 0 and O can be avoided (not used) or replaced by different symbols. 9 and 6 are ok as far as they are used with other characters and not alone.

[0059] Any suitable set of symbols may be used but for example a font such as the Peyote alphabet may be provided. Any set of symbols which can produce a distinguishable reflection signal is suitable. In addition, to enhance backscattering effects, some symbols might be replaced, such as the symbol @ may be used to represent 1 , as it produces more reflection.

[0060] A further advantage of the tags 300 of the present invention is their encoding capacity since their tag size may be in the order of 10mm by 10mm with building blocks (bricks) being 30mil by 30mil. Therefore, the free space between two alphabets for example may mean the tag size can be 1 cm by 1 cm, so it is possible to put 56 alphabets (lowercase and uppercase) plus 10 numbers (0 to 9) in combination which results in 66 symbols/cm ² (which almost equates to 6 bits/cm ² ). Advantageously this offers much more encoding capacity compared to presently known tags, which are typically in the order of 2 bits/ cm ² .

[0061 ] A further advantage of the present invention is the utilization of the machine learning system component 130 to recognize the alphabet symbols makes the tags further readable compared to typical 1 -0 lines which represent single bits. [0062] Further since the tags have components in both and / directions they are less sensitive to orientation and miss-oriented tags 105 are more readable via the reader, as enough backscattering is guaranteed in both x and y coordinates.

[0063] In a further advantage, error detection/correction methods may be used to further assist in noisy environments or collision preventing. In noisy environments or in anti-collision cases, certain errors may be prevented since it is known that a certain combination of characters in the tags is yet to be used. Alternatively, redundant characters as check bits/symbols may be provided.

[0064] Figure 5 is a flow diagram illustrating the method of the present invention in preferred embodiment.

[0065] The method 500 initialises at step 505 where a reader 1 15 is provided for reading one or more tags 105. At step 510 via a micro-controller associated with the reader 1 15 and transmit and receive antennas 120R, 120T associated with the reader 1 15 are initialised, as is a VCO BB Generator (with 1 output, or 2 or 4 l/Q signals). The tag reader 1 15 is also initialised to an origin or x=0 point. The reader via the BB generated signal reads the backscattered signal by measuring amplitude and phase. Alternatively, l/Q signals might be used to make the measurements easier, as amplitude and phase of transmitted/received signals can be calculated based on l/Q amplitudes only.

[0066] Improved results with less phase noise may be provided by utilizing four inputs namely In Phase Positive Baseband (BB_IP), In Phase Negative Baseband (BBJM), Quadrature Positive Baseband (BB_QP) and Quadrature Negative Baseband (BB_QN). However one or two inputs may also enough. Having 2 or 4 signals improves phase noise and provides more accurate results (this is specific to the hardware used in the experiments).

[0067] A control unit (such as a microcontroller unit, MCU) may set carrier frequency for the transmitting and receiving antennas 120R, 120T. The frequencies may be adjusted so as to operate from 57 GHz to 64 GHz with up to 1 .8 GHz bandwidth for BB. Step frequencies may be provided with BB VCO. At this step since it is an initialization step the tag 105 is removed from the system 200 to collect background noise. Scattering parameters (S21 ) may be collected by deducing amplitude/phase of the sent signal from the detected amplitude/phase (using real/imaginary calculations) in the receiver. If l/Q signals used, they can then be converted to amplitude and phase of the scattering parameters.

[0068] Having read the scattering parameters over a whole range of frequencies, in the number of positions (the reader moving) and in different range of TX power/RX sensitivity, the system then reads the background data in a number of steps for a position of the reader 1 15 (for example D/2 mm steps, where D is the diagonal of antenna) if the strip mode SAR is in use. Then the entire frequency range can be swept (for example in 10 MHz steps) and this information may be saved as initial data. It will be appreciated that when solely utilizing a machine learning system according to the invention it is not necessary to follow an SAR strip-mode image formation technique like the above. Fewer points are enough in this scenario if the machine learning system can reconstruct and decode the tag.

[0069] The number of generated samples depends on the step frequency, bandwidth, number of measurement in the rail, power steps of TX and sensitivity steps in RX. It will be appreciated that non-linear scanning may be provided (such as scanning in different angles, while the distance of tag/reader is kept constant). This can be similar to iSAR techniques, while the radar target is not behaving in a predicated way.

[0070] Control then moves to step 520 where the tag 105 to be read is put in the reader 1 15. At step 525, the tag is scanned and the scanning phase may be repeated across a predetermined distance in a one or more directions (for example, in the case of a linear direction, a 30cm rail) to gather backscattered signal from the tag. In case of iSAR like scanning, as mentioned, tags/reader distance might also kept constant, but the reading angle can be changed. Control then moves to step 530 in which the detected scattering parameters in each position are subtracted from the scattering parameters where there was no tag present as a first de-noising step, essentially removing background noise from the scattering parameters so as to allow the machine learning system 130 to more accurately identify the tag. [0071 ] This data is then at step 535 provided to the computer 125 (It can be a remote or cloud computer). Based on number of tags and their RCS, at step 540 we may decide whether or not the tag 105 can be considered as just one pattern or if the tag has to be image processed. Choosing between these two detection methods involves a predetermined threshold such that if the number of different patterns detected is limited to a predetermined amount (for example less than 50) it is optimal to proceed straight to the machine learning system step 545 whereas if it is determined that the number of distinguished patterns exceeds a predetermined amount and then additional steps 565 and 570 are carried out.

[0072] At step 565, since it has been determined that the number of patterns is exceeds the predetermined threshold, an image is created based on the collected data using microwave image formation technique. Image formation techniques may include for example SAR/iSAR. It will be appreciated that any microwave image formation technique may be utilized.

[0073] Control then moves to step 570 in which the image created in step 565 is sent to the machine learning system 130 for learning and testing recognition.

[0074] Control then moves to step 550 where it is determined whether or not the system has been trained for all possible tags or symbols. If it is determined that the system has been trained, control then moves to step 560 where the method ends (meaning that the tag has been recognized and decoded correctly) otherwise control moves to step 555 where the tag is put back into the reader to add it the trained system and control returns to step 525.

[0075] If at step 540 the entire tag 105 can be considered as one pattern, control moves directly step 545 where it is determined that the entire tag may be recognised using the machine learning system 130. In this step, scattering parameters are sent to the machine learning system 130 for learning/testing and recognition along with the tag position and frequency.

[0076] When the machine learning system 130 is in the learning phase and since the network may, for example, be a K-NN network and learning is supervised - the actually tag data is also present in the machine learning system 130. In training phase, raw collected scattering data is first shuffled randomly and then sent to the machine learning system. It will be appreciated that any arrangement of machine learning system may be provided. Based on number of tags utilized, different version of supervised NN can be used. Feed forward may be used for simple models, whereas deep learning techniques (Support Vector Machines (SVMs), K-Nearest neighbors (KNNs) and Ensembles are also applicable for more complicated environments (higher tags)).

[0077] Once the training is done and the method of recognition decided, scanned tags can be put into the rail. If tag detection is successful, the system shows encoded data in the END step. If not successful, more training is needed (or machine learning topology, including layers and cells should be adjusted) or tag should be rescanned.

[0078] Sometimes the system 100 needs to be trained several times with different combinations of the layers/parameters to reach the desired error rate. For example, the error may be Mean Square Error (MSE). In the present invention both phase and angle of the scattering parameters are utilised for training.

[0079] Figure 6 is a chart which illustrates where measured scattering parameters (S21 ) are provided in different locations over a whole frequency band between 57 GHz and 65 GHz. The absolute value of the S21 parameter is shown, read over entire frequency band. Different lines show different S21 parameter values once the tag moves in the linear rail. From this and the tag/reader geometrical information the tag's shape can be reconstructed.

[0080] Tag data capacity depends on the requirements, but, for example, 8 symbols may occupy 8cm x 1 cm of space (which may be the size of a credit card's magnetic strip for example) will lead to 8x6 = 42 bits of data which in theory results in 2 2 (i.e. 4.40e12) different products.

Machine learning system Training

[0081 ] The machine learning system 130 requires the right input and outputs and the correct amount of hidden layer neurons for training of the machine learning system (this part is simulation phase, and a feed forward network is used). In the present invention, we can have 4 parameters as input to the machine learning system 130, for example, (1 ) the relative position of the sensor and tag, (2) scanning frequency, (3) scatter parameter amplitude and (4) scatter parameter phase and one output. The other parameters can be added to increase the detection rate, such as angle of reading, power level of transmitter, sensitivity level of receiver (by changing the LNA and IF stages) and tag substrate material. The output in this case being integers associated with each tag ID.

[0082] As shown in Figure 7, the raw data is preferably shuffled first in order for training to work most effectively. In training multilayer networks, the data is generally divided into three subsets - the first being a training set, the second being a validation set (where the error on the validation set is monitored during the training process) and the third being a test set. As noted by Mira, J (Bioinspired Applications in Artificial and Natural Computation: Third International Work-Conference on the Interplay Between Natural and Artificial Computation, IWINAC 2009, Santiago de Compostela, Spain, June 22-26, 2009, Proceedings), "the validation error normally decreases during the initial phase of training, as does the training set error. However, when the network begins to overfit the data, the error on the validation set typically begins to rise. The network weights and biases are saved at the minimum of the validation set error. The test set error is not used during training, but it is used to compare different models. It is also useful to plot the test set error during the training process. If the error on the test set reaches a minimum at a significantly different iteration number than the validation set error, this might indicate a poor division of the data set." Chart 710 illustrates an error histogram with twenty Bins while 720 illustrates performance parameters including gradient, μ, and the number of validation fails during the training process. From the charts it is clear that the gradient trends become quite low so that further training will not contribute to further improvements. In the simulations with feed-forward network, the size of the networks cannot be bigger than a few hundred neurons otherwise conventional backpropagation techniques do not help with network training.

[0083] The present invention has been tested with a number of hidden layer neurons and with the various scenarios which were not included in the training. For example in one scenario a tag was rotated a 45 degrees and another one at 180 degrees (i.e. reversed). In this simulations, tag could be recognized by more than 60% even if it was rotated.

[0084] As shown in Figure 8, chart 800 illustrates that the reader 1 15 may adjust itself to new scenarios that it has not been trained with (once the neurons in the hidden layer are large enough). In this situation rotated tags are considered as new scenarios where the system has not yet been trained. The higher detection rates come with a cost, which is that there is exponential growth in training time versus the number of neurons in the hidden layer (this is for feed-forward network, and just in simulations).

[0085] In the lab experiments, It will be appreciated that one of a number of techniques may be utilized for classification and machine learning. Using larger training sets provides models which perform better at generalizing new data. At the same time, parameters such as training speed, memory usage, the accuracy of prediction and interpretability must also be considered. In one embodiment, the system and method of the present invention may run all available classifications, acquire initial results, and then fine-tune those results with good accuracy/Area Under Coverage (AUC). Accuracy may be interpreted as the percentage of correct matches for verification data (it will be appreciated that the data used in the verification sets are not used in training sets).

[0086] Training sets may be used in numerous classifiers with different configurations, and the process repeated numerous times. For example, depending on the data set the best performing classifiers may be one or more of fine Gaussian SVM, fine and weighted KNN, and Bagged Trees and Subspace KNN Ensembles.

[0087] Figure 9 is a chart 900 which illustrates a comparison between different classifiers showing the performance of the system and method of the present invention. This is the lab experiment data from actual tags and reader (for 3 tag/reader positions).

[0088] In an example mode of operation, and it will be appreciated that the present invention is not limited to this example mode of operation. System can be used for different tags on different substrates. For example, tag 400 shown in Figure 4a is a tag printed on FR4 (a glass-reinforced epoxy laminate material) with a permittivity ε _Γ of 4.3. Tag 410 shown in Figure 4b is a tag printed on conductive InkTec™ and tag 420 shown in Figure 4c is an M-creative, last two are printed on a 100μηπ Mylar™ polyester film.

[0089] Advantageously, there are no restrictions on use of tags substrates with the present invention, provided there is sufficient detectable backscattering signal for the reader.

[0090] Figure 10 shows a chart 1000 illustrating the S21 measured using vector network analysis (VNA), which shows a 5 to 10 dB difference denoted "A" in Figure 1 for tag presence in the 60GHz free spectrum.

[0091 ] Each of the tags in Figure 4a, 4b and 4c have the following characteristics: i) they have components in x and y-planes, ii) they are based on 30 mil x 30 mil block building bricks (each symbol made of these building blocks occupies 1 cm ² , iii) they are less sensitive to printing imperfections because of detection method utilized in the present invention (pattern recognition) and iv) increased data capacity is provided in comparison to a microwave counterpart since in each symbol position of 1 cm ² space, A possibility of at least 56 different alphabetical symbols (upper and lower case alphabets) + 10 numbers/symbols can be inserted. The result is 66 symbols/cm ² (almost 6 bits/cm ² ) capacity. Existing tag arrangements are largely limited to a maximum of 2 bits/ cm ² , as noted previously.

[0092] The present invention makes the tags 400, 410 and 420 less sensitive to orientation issues compared to meander-line image bar tags or the like. There is sufficient cross-polarization to guarantee reading in any direction. The RCS, σ, of a back scatterer, is given by the radar range equation:

[0093] ¾ _{= ff} i ^[_J_] ² _{(1 )}

[0094] where Ri and R2 are the tag to the transmitter and receiver distances, P _? and P _r are transmit and receive powers, GAnt and G^^ are transmitter and receiver antenna gains in a frequency corresponding to wavelength, respectively. [0095] The forward transmission coefficient (S21 vs frequency) used as an RCS indicator, as they are exchangeable by S21 _tat -S21 _str \

[0096] a _tgt = a _str . 10 " ) ₍₂ )

[0097] where σ is the RCS, S ₂ i is the scattering parameter, _tgt and _sfr indicate target and supporting structure respectively. In the present example, the average tag RCS was calculated to be 0.003 dBsm ² .

[0098] Figure 1 1 a shows a schematic tag simulation scenario (utilizing electromagnetic simulation software) with some of the magnitude S21 in different frequencies and different tag/reader distances are results shown in the chart of Figure 12b. In the simulations, the magnitude and phase of S21 were used as the tag related signature. In the decoding phase, a feedforward network is utilized to classify different tags based on those characteristics. The simulations may be based on 3-character tags combinations in electromagnetic simulation software for example (figure 1 1 a). Antennas for reading the tag are set in a cross-polarization configuration with set distance from the tags. The distance may be 50mm for example.

[0099] The tags may be made of thin aluminium. For example 17 m thickness may be suitable since this is comparable to Sato™ printer tags. The power of excitation signals in electromagnetic simulation is chosen to be the same as the output power of the vector network analyzer, namely -5 dBm.

[0100] Movement steps (these may be on a linear rail for example, but need not be) are based on the concept of SAR, as said, in which the best upper bound resolution for a single antenna is D/2, where D is the diagonal length of the horn antenna. In the case of a linear rail, tag combinations were moved in 6mm steps horizontally, as D = 12 mm for the horn antennas used.

[0101 ] Training of the system is via the steps of Figure 5, where at each step of tag-radar positioning, 7 GHz frequency span is scanned to get the magnitude and phase of S21 of any tag under test. After S21 is deducted from background noise, its phase and magnitude along with tag/reader position and frequency, are provided as inputs to an FF (feed-forward) machine learning system (steps 545 and 550 of Figure 5). The output of the FF network was then set to be the recognized pattern ID.

[0102] In the lab experiment, 5-symbolic patterns tags were utilized (tags 400, 410 and 420 as shown in Figures 4a, 4b and 4c), in contrast to 3 symbol tags in simulations. The size of the combination used is in the order 5 cm by 1 cm. This size is closely comparable to that of an optical barcode. Tags 400, 410 and 420 have a capacity of 6 bits/tag x 5 tags which is 30 bits. It has been found that if there is sufficient backscatter signals available, the substrate of a tag does not affect/influence the outcome when using this method.

[0103] In the present experiment, PCB fabricated on FR4 substrate (ε _Γ = 4.4), with different inks printed on a Ι ΟΟμιτι Mylar™ Polyester film were used to produce the different tags, as discussed with reference to Figure 4a, 4b and 4c. In this example, all tags detected despite the printing method or substrate used.

[0104] Figure 12 shows a chart 1200 which is an example of the collected magnitude and phase for a tested tag (in this case Tag 410 in Figure 4b). It is important to note this is the result after deduction from background structure mode. This is what we consider as tag fingerprint in a particular TX/RX power setup and in one position of Tag/Reader.

[0105] Figures 13 and 14 show example hardware utilised in reading tags as described with reference to the examples above. In Figure 13, the hardware 1300 includes a microcontroller unit (MCU) 1310, a rail controller 1320 and BB VCO and gain/phase detector 1330. In Figure 14, the hardware 1400 includes a tag 1410, linear rail 1420 and evaluation board/frequency transmitter 1430.