Joint decoding HARQ feedback and UE scheduling request bits using ML techniques

Abstract

Embodiments of a method in a base station for decoding Scheduling Request (SR) bits and Hybrid Automatic Repeat Request (HARQ) feedback bits are disclosed. In some embodiments, the method comprises receiving an uplink radio signal from a user equipment (UE), the uplink radio signal including SR bits and Hybrid Automatic HARQ feedback bits, and analyzing the received uplink radio signal using at least one neural network model to jointly decode the SR bits and HARQ feedback bits.

Description

P112081W001JOINT DECODING HARQ FEEDBACK AND UE SCHEDULING REQUEST BITS USING ML TECHNIQUESTechnical Field

[0001] The present disclosure relates to radio link adaptation, and in particular to joint decoding Hybrid Automatic Repeat Request (HARQ) feedback and UE Scheduling Request (SR) bits using machine learning (ML) techniques.Background

[0002] When a 5G User Equipment (UE) needs to transmit uplink data but does not have any available resources, it can request scheduling from the network by sending a Scheduling Request (SR) on the Physical Uplink Control Channel (PUCCH). This SR signals the intention of the UE to transmit data, but it does not specify the amount. To allocate the necessary resources, the network requires information about the amount of data in the UE's buffer, which is provided in the Buffer Status Report (BSR). The network then allocates resources, including enough resources for the UE to send BSRs, based on its implementation.

[0003] The network needs to detect SR reports on the allocated SR resources, as it does not know when the UE will require uplink resources. Only one SR is necessary, regardless of the number of uplink carrier units used. The SR can only be sent when the UE is in the RRC CONNECTED state and maintains uplink synchronization and is used exclusively for new data, not retransmitted data.

[0004] The SR is sent on the PUCCH because the UE does not have any available PUSCH resources. The network can allocate a dedicated SR resource for each UE, which appears once every "n" subframes. The cycle of SR is configured through the sr-Configlndex field in the ScheduleRequestConfig Information Element (IE). The network is aware of the specific correspondence between SR resources and UEs, so the UE does not need to provide its own ID (C-RNTI). The location of the SR resources,P112081W001configured through the sr-PUCCH-Resourcelndex field in the ScheduleRequestConfig IE, is used by the network to determine which UE is requesting uplink resources.

[0005] Since SR resources are dedicated to UEs and allocated by the gNodeB (gNB), each SR resource corresponds to a specific UE, and the gNB is aware of this correspondence. As a result, the UE does not need to specify its own ID (C-RNTI) when sending SR information. The gNB is able to determine which UE is requesting uplink resources by knowing the location of the SR resource.

[0006] Hybrid automatic repeat request (HARQ) is an essential component of 5G NR. gNB sends Downlink Control Information (DCI) on PDCCH which carries scheduling information for uplink and downlink. It provides a UE with the necessary information for proper reception and decoding of the downlink data on PDSCH as well as transmitting the uplink data and UCI on PUSCH or UCI on PUCCH, which contains the HARQ feedback information. gNB also needs to decode HARQ feedback for Downlink (DL) transmissions.

[0007] The DL transmission and reception process involves transmitting data, receiving ACK or NACK, and performing retransmissions if necessary. The whole procedure (transmission, reception, and ACK / NACK reporting, retransmission) is done by a special process called the HARQ process, which will be discussed in more detail below. Hybrid Automatic Repeat Request (HARQ) plays a critical role in 5G NR. The gNB transmits Downlink Control Information (DCI) on the Physical Downlink Control Channel (PDCCH), providing scheduling information for uplink and downlink. This information provides User Equipment (UE) with the necessary details for correctly receiving and decoding the downlink data on the Physical Downlink Shared Channel (PDSCH), while also transmitting the uplink data and Uplink Control Information (UCI) on the Physical Uplink Shared Channel (PUSCH) or the Physical Uplink Control Channel (PUCCH), which contains the HARQ feedback information. Additionally, the gNB is responsible for decoding HARQ feedback for Downlink (DL) transmissions.P112081W001

[0008] There remain certain challenges. The detection of Scheduling Request (SR) signals and Hybrid Automatic Repeat Request (HARQ) feedback on the Physical Uplink Control Channel (PUCCH) typically relies on fixed thresholds that are established through physical layer simulations. However, this approach can result in a high rate of false SR detections, leading to the issuance of un-necessary high-priority grants by the gNodeB (gNB) and the wastage of valuable resources. This issue becomes more pronounced as the number of active users increases, negatively impacting the user experience and causing unnecessary consumption of limited Physical Downlink Control Channel (PDCCH) and Physical Uplink Shared Channel (PUSCH) resources.

[0009] Additionally, the current method performs SR detection and HARQ feedback decoding independently from history information, failing to fully utilize the relationship between HARQ and SR, as well as other relevant information such as Signal-to-Interference-plus-Noise Ratio (SINR). Consequently, this leads to higher SR and HARQ feedback decoding errors, resulting in increased delay and reduced cell throughputs.Summary

[0010] Aspects of the present disclosure provide a method implemented in a radio node of a radio access network. The method comprises: receiving an uplink radio signal from a user equipment (UE), the uplink radio signal including Scheduling Request (SR) bits and Hybrid Automatic Repeat Request (HARQ) feedback bits; and analysing the received uplink radio signal using at least one neural network model to jointly decode the SR bits and HARQ feedback bits.

[0011] In some embodiments, the uplink radio signal comprises a Physical Uplink Control Channel (PUCCH) signal.

[0012] In some embodiments, analyzing the received uplink radio signal comprises: processing the received uplink radio signal to obtain feature data; and supplying the feature data to an input of the at least one neural network model.P112081W001

[0013] In some embodiments, the feature data comprises data indicative of a plurality of features of the received uplink radio signal, the plurality of features including any one or more of:• whether or not a current slot of the received uplink radio signal corresponds with an SR opportunity;• a number of HARQ feedback bits that the radio node expects to receive from the UE;• Real (Re) and Imaginary (Im) values of a demodulated HARQ feedback symbol in the received uplink radio signal;• Real (Re) and Imaginary (Im) values of a demodulated SR symbol in the received uplink radio signal; and• estimated SINR of SR and HARQ symbols of the UE.

[0014] In some embodiments, supplying the feature data to an input of the at least one neural network model comprises, for at least one feature:• applying a normalization function to the feature to compute a normalized feature value; and• supplying the normalized feature value to the input of the at least one neural network model.

[0015] In some embodiments, the at least one neural network model is selected from a set of predetermined neural network models including:• Model 1 characterised by a linear activation function and a Mean Squared Error cost function;• Model 2 characterised by a Rectified Linear Unit (ReLU) activation function and a Mean Squared Error cost function;• Model 3 characterised by a Sigmoid activation function and a Mean Squared Error cost function;• Model 4 characterised by a TanH activation function and a Mean Squared Error cost function;P112081W001• Model 5 characterised by a Sigmoid activation function and a CrossEntropy cost function; and• Model 6 characterised by a SoftMax activation function and a CrossEntropy cost function.

[0016] In some embodiments, analyzing the received signal comprises:• generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits;• selecting, from the plurality of candidate values, most likely SR and HARQ values, based at least in part on a detected Signal-to-Interference- plus-Noise Ratio (SINR) of the received uplink radio signal and one or more similarity thresholds.

[0017] In some embodiments, generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:• computing, by a first neural network model trained using Single-value target data, first predicted values for each of the SR and HARQ bits; • if the first predicted value of the SR bit is greater than 0, classifying the predicted SR bits as SR positive, and otherwise classifying the predicted SR bit as SR negative;• if the first predicted values of the HARQ bits are less than or equal to a predetermined first threshold, then the predicted HARQ is classified as DTX;• if the first predicted values of the HARQ bits are greater than the first threshold, then the predicted HARQ is classified as ACK; and • first predicted values of the HARQ bits are less than a predetermined second threshold, then the predicted HARQ is classified as NACK.

[0018] In some embodiments, generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:P112081W001• computing, by a second neural network model trained using One-hot encoded target data, second predicted values for each of the SR and HARQ bits;• if a SR-Positive predicted value (srPositivePred) is greater than a SR- Negative predicted value (srNegativePred), then classifying the predicted SR bit as SR positive, and otherwise classifying the predicted SR bit as SR negative;• if a HARQ=DTX predicted value (HqBitx DtxPred) is greater than either one of a HARQ=NACK predicted value (HqBitx NackPred) or a HARQ=ACK predicted value (HqBitx AckPred), then classifying the predicted HARQ feedback as DTX, and otherwise:• if the HARQ=NACK predicted value (HqBitx NackPred) is greater or equal to the HARQ=ACK predicted value (HqBitx AckPred), then classifying the predicted HARQ feedback as NACK, and otherwise classifying the predicted HARQ feedback as ACK.

[0019] In some embodiments, generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:• computing, by a third neural network model trained using Classified target data, third predicted values for each of the SR and HARQ bits;• if a SR-Positive predicted value (srPositivePred) is greater than a SR- Negative predicted value (srNegativePred), then classifying the predicted SR bit as SR positive, and otherwise classifying the predicted SR bit as SR negative;• if a HARQ=DTX predicted value (HqBitx DtxPred) is greater than either one of a HARQ=NACK predicted value (HqBitx NackPred) or a HARQ=ACK predicted value (HqBitx AckPred), then classifying the predicted HARQ feedback as DTX, and otherwise:• if the HARQ=NACK predicted value (HqBitx NackPred) is greater or equal to the HARQ=ACK predicted value (HqBitx AckPred), thenP112081W001classifying the predicted HARQ feedback as NACK, and otherwise classifying the predicted HARQ feedback as ACK.

[0020] In some embodiments, the classified target data comprises, for each one of a plurality of classes, respective values of the SR bit and HARQ feedback bits, and wherein the third predicted values for each of the SR and HARQ bits comprises:• the SR-Positive predicted value (srPositivePred), comprising a maximum predicted value of a subset of classes corresponding to a positive SR bit value;• the SR-Negative predicted value (srNegativePred), comprising a maximum predicted value of a subset of classes corresponding to a negative SR bit value;• the HARQ=DTX predicted value (HqBitx DtxPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of DTX;• the HARQ=NACK predicted value (HqBitx NackPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of NACK; and• the HARQ=ACK predicted value (HqBitx AckPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of ACK.

[0021] In some embodiments, the method further comprises dynamically adjusting the one or more similarity thresholds based at least in part on a predetermined target SR false alarm error rate (FAER).

[0022] To enhance the precision of SR detection and HARQ feedback decoding, the present disclosure provides an algorithm based on Neural Networks to analyze the features of the received signal and jointly decode the SR and HARQ feedback bits. An essential aspect of effective ML data-driven algorithms is selecting informative, discriminating, and independent features and defining the targets appropriately.P112081W001Embodiments of the described algorithm utilize six pairs of activation and cost functions to train the neural network models for the joint decoding of SR and HARQ feedback bits.

[0023] Multiple targets are used to train the Neural Network model and correspondingly leveraging multiple output values for SR detection and HARQ feedback decoding. Signal-to-Interference-plus-Noise Ratio (SINR) and similarity thresholds are introduced to reduce the rates of false alarms in SR detection and Ack False (DTX / NACK to ACK) in HARQ feedback. Instead of relying on fixed thresholds, the threshold is dynamically adjusted based on the current network conditions, such as the SR false alarm error rate (FAER). We can employ the On-The-Fly online training method to refine the model on a per gNB or per cell basis.

[0024] While specific embodiments are described in the context of 5GNR, the disclosed techniques are also applicable to other wireless networks, such as LTE. Furthermore, the On-The-Fly online training can be implemented in the cloud.

[0025] Specific embodiments of the methods disclosed herein may comprise any one of more of the following features:• Selection of Features and Targets Data for SR Detection and HARQ Feedback Decoding• Utilization of six pairs of activation and cost functions, including Sigmoid function combined with Cross-Entropy Loss• Incorporation of multiple target values to train neural network models and their corresponding output values for SR detection and HARQ feedback decoding• Implementation of SINR and Similarity Thresholds to Mitigate False Alarm Error Rate and Ack False Rate• Adaptive Threshold Adjustment based on Current Network Conditions, such as FAERP112081W001On-the-fly online Training Method for Fine-Tuning Neural Network Models on a per gNB or Cell basis

[0026] Specific embodiments of the methods disclosed herein may provide any one or more of the following technical advantages:• Through the implementation of neural network-based algorithms, our proposed solution demonstrates a substantial improvement in SR detection and HARQ feedback decoding performance, surpassing that of our current products.• Furthermore, the inclusion of SINR and Similarity Thresholds contributes to a reduction in SR false alarm error rates and ACK / NACK false rate, resulting in a more robust SR detection and HARQ feedback decoding. • The dynamic adjustment of thresholds in our solution ensures that SR false alarm error rates remain at an acceptable level, thereby enhancing the reliability of SR detection.• The On-The-Fly online training method is a valuable feature of our proposed solution, enabling continuous refinement of the model on a per gNB or per cell basis, leading to ongoing enhancements in SR detection and HARQ feedback decoding performance

[0027] Embodiments of a base station, communication system, and a method in a communication system are also disclosed.Brief Description of the Drawings

[0028] The accompanying drawing figures incorporated in and forming a part of this specification illustrate several aspects of the disclosure, and together with the description serve to explain principles of the disclosure.

[0029] Figure 1 is a message flow diagram illustrating a typical SR function in the wireless access network;P112081W001

[0030] Figure 2 is a message flow diagram illustrating a HARQ function in the wireless access network;

[0031] FIG. 3 is a table illustrating SR and HARQ feedback bit values using QPSK transmission;

[0032] FIG. 4 illustrates a typical QPSK symbol constellation and corresponding 2-bit symbol values;

[0033] FIG. 5 is a table illustrating example feature data (inputs) selected for SR detection and HARQ feedback decoding;

[0034] FIGs. 6A-6C show respective example target data usable in embodiments of the present disclosure;

[0035] FIG. 7 is a block diagram schematically illustrating the general structure of an example neural network useable in embodiments of the disclosure;

[0036] FIG. 8 is a table illustrating example activation functions usable useable in embodiments of the disclosure;

[0037] FIG. 9 is a table illustrating example pairs of activation and cost functions which may be used for training neural network models for SR detection and HARQ feedback decoding in embodiments of the disclosure; and

[0038] FIG. 10 is a table illustrating example target data classes usable in embodiments of the disclosure.Detailed Description

[0039] The embodiments set forth below represent information to enable those skilled in the art to practice the embodiments and illustrate the best mode of practicing the embodiments. Upon reading the following description in light of the accompanying drawing figures, those skilled in the art will understand the concepts of the disclosure and will recognize applications of these concepts not particularly addressed herein. It should be understood that these concepts and applications fall within the scope of the disclosure.P112081W001

[0040] At least some of the following abbreviations and terms may be used in this disclosure.• 2D Two Dimensional• 3 GPP Third Generation Partnership Project• 5G Fifth Generation• AAS Antenna Array System• AoA Angle of Arrival• AoD Angle of Departure• ASIC Application Specific Integrated Circuit• BF Beamforming• BLER Block Error Rate• BW Beamwidth• CPU Central Processing Unit• CSI Channel State Information• dB Decibel• DCI Downlink Control Information• DFT Discrete Fourier Transform• DSP Digital Signal Processor• eNB Enhanced or Evolved Node B• FIR Finite Impulse Response• FPGA Field Programmable Gate Array• gNB New Radio Base Station• ICC Information Carrying Capacity• IIR Infinite Impulse Response• LTE Long Term Evolution• MIMO Multiple Input Multiple Output• MME Mobility Management Entity• MMSE Minimum Mean Square ErrorP112081W001• MTC Machine Type Communication• NR New Radio• OTT Over-the-Top• PBCH Physical Broadcast Channel• PDCCH Physical Downlink Control Channel• PDSCH Physical Downlink Shared Channel• P-GW Packet Data Network Gateway• RAM Random Access Memory• ROM Read Only Memory• RRC Radio Resource Control• RRH Remote Radio Head• SCEF Service Capability Exposure Function• SINR Signal to Interference plus Noise Ratio• TBS Transmission Block Size• UE User Equipment• ULA Uniform Linear Array• URA Uniform Rectangular Array

[0041] Radio Node: As used herein, a “radio node” is either a radio access node or a wireless device.

[0042] Radio Access Node: As used herein, a “radio access node” or “radio network node” is any node in a radio access network of a cellular communications network that operates to wirelessly transmit and / or receive signals. Some examples of a radio access node include, but are not limited to, a base station (e.g., a New Radio (NR) base station (gNB) in a Third Generation Partnership Project (3GPP) Fifth Generation (5G) NR network or an enhanced or evolved Node B (eNB) in a 3GPP Long Term Evolution (LTE) network), a high-power or macro base station, a low-power base station (e.g., a micro base station, a pico base station, a home eNB, or the like), and a relay node.P112081W001

[0043] Core Network Node: As used herein, a “core network node” is any type of node in a core network. Some examples of a core network node include, e.g., a Mobility Management Entity (MME), a Packet Data Network Gateway (P-GW), a Service Capability Exposure Function (SCEF), or the like.

[0044] Wireless Device: As used herein, a “wireless device” is any type of device that has access to (i.e., is served by) a cellular communications network by wirelessly transmitting (and / or receiving) signals to (and / or from) a radio access node. Some examples of a wireless device include, but are not limited to, a User Equipment device (UE) in a 3 GPP network and a Machine Type Communication (MTC) device.

[0045] Network Node: As used herein, a “network node” is any node that is either part of the radio access network or the core network of a cellular communications network / system.

[0046] Cell: As used herein, a “cell” is a combination of radio resources (such as, for example, antenna port allocation, time and frequency) that a wireless device may use to exchange radio signals with a radio access node, which may be referred to as a host node or a serving node of the cell. However, it is important to note that beams may be used instead of cells, particularly with respect to 5G NR. As such, it should be appreciated that the techniques described herein are equally applicable to both cells and beams.

[0047] Note that references in this disclosure to various technical standards (such as 3GPP TS 38.211 V15.1.0 (2018-03) and 3GPP TS 38.214 V15.1.0 (2018-03), for example) should be understood to refer to the specific version(s) of such standard(s) that is(were) current at the time the present application was filed, and may also refer to applicable counterparts and successors of such versions.

[0048] The description herein focuses on a 3 GPP cellular communications system and, as such, 3GPP terminology or terminology similar to 3GPP terminology is oftentimes used. However, the concepts disclosed herein are not limited to a 3 GPP system.P112081W001

[0049] As used herein, the term neural network model refers to a computational representation of a neural network as a network of interconnected nodes (or “neurons”), each of which computes a respective output signal based on a given activation function and a received input signal. In the some nodes (neurons), the received input signal may be a sum of signals output from a plurality of other nodes (neurons).

[0050] Systems and methods are disclosed herein that provideThe process of uplink data transmission in 5G NR

[0051] The conventional process of uplink data transmission in 5G New Radio (NR) is depicted in FIG. 1. The process of FIG. 1 includes:

[0052] Step 1 (at 102): The User Equipment (UE) signals its intention to transmit data by sending a Scheduling Request (SR) on the Physical Uplink Control Channel (PUCCH) Format 1 channel to request resources.

[0053] Step 2 (at 104): Upon successful detection of the SR, the gNodeB (gNB) sends a Downlink Control Information (DCI) via the Physical Downlink Control Channel (PDCCH), which includes an uplink grant to allocate the necessary resources.

[0054] Step 3 (at 106): Upon successful receipt of the uplink grant, the UE then sends a Buffer Status Report (BSR) on the Physical Uplink Shared Channel (PUSCH) in the allocated resources, indicating to the gNB the amount of data in its buffer that needs to be transmitted.

[0055] Step 4 (at 108) Based on the received BSR, the gNB assigns the appropriate resources and Modulation and Coding Scheme (MCS) for the UE, and sends this information to the UE in an Uplink Grant on PDCCH.

[0056] Step 5 (at 110): Upon successful receipt of the uplink grant, the UE transmits its uplink data on PUSCH using the resources and MCS allocated by the gNB.The DL transmission and reception process

[0057] The DL transmission and reception process encompasses the transmission of data, reception of ACK or NACK responses, and conducting retransmissions as needed.P112081W001This entire sequence, including transmission, reception, acknowledgment / negative acknowledgment reporting, and retransmission, is facilitated by a specialized process known as the HARQ process. FIG. 2 illustrates an example embodiment of a DL transmission process including typical HARQ operation. In the example of FIG. 2:

[0058] Step 1 (at 202): The gNB (base station) transmits PDSCH with a givenHARQ revision (RV) number(e.g. rv 0) to the UE over the PDSCH.

[0059] Step 2 (at 204): The UE receives the data and performs error detection using the CRC. In the illustrated example, the CRC fails, indicating a decoding failure.

[0060] Step 3 (at 206): The UE sends a NACK message to the gNB through the PUCCH or PUS CH, signaling a decoding failure. The UE also stores the received data for later use.

[0061] Step 4 (at 208): Upon receiving the NACK, the gNB retransmits the data with a different RV number (e.g., rv=2) to enhance the likelihood of successful decoding.

[0062] Step 5 (at 210): The UE receives the retransmitted data, combines it with the previously stored data in the buffer, and performs error detection using the CRC.However, in this case, the CRC fails again, indicating another decoding failure.

[0063] Step 6 (at 212): The UE sends another NACK message to the gNB, indicating a second decoding failure, and stores the received data for later use.

[0064] Step 7 (at 214): Based on the second NACK, the gNB retransmits the data again with a different RV (e.g., RV=3) to further improve the likelihood of successful decoding.

[0065] Step 8 (at 216): The UE receives the retransmitted data, combines it with the previously stored data (from the two previous decoding attempts at 204 and 210), and performs error detection using the CRC. This time, the CRC is successful, indicating a successful decoding.P112081W001

[0066] Step 9 (at 218): The UE sends an ACK message to the gNB, signaling successful reception after the third transmission attempt.PUCCH Resource Configuration and Scheduling Request and ACK / NACK Transmission

[0067] Typically, each User Equipment (UE) is configured with nine (for example) Physical Format 1 (PF1) resources. Of these, eight (for example) are utilized for Hybrid Automatic Repeat Request (HARQ) acknowledgments (ACK / NACK), which may be shared among multiple UEs, and one (for example) is reserved for a Scheduling Request (SR), which is unique to each UE (although the same SR resource can be used by different UEs in different slots). Each PF1 resource for ACK / NACK can transmit up to two HARQ feedback bits. Each UE is also assigned two (for example) Physical Format 3 (PF3) resources, which may be shared among multiple UEs, and each PF3 resource can transmit between three to twelve (for example) HARQ feedback bits. When the gNodeB (gNB) schedules more than two HARQ feedback bits, PF3 will be used. Otherwise, PF1 will be utilized. If the Physical Uplink Shared Channel (PUSCH) is scheduled in the same time slot, the HARQ feedback bits will normally be transmitted on PUSCH.

[0068] The Scheduling Request (SR) resource is established periodically, appearing in every "n" subframes, and its offset is associated with the selected entry in the SR mapping matrix. The SR is transmitted using the dedicated SR resource. The ACK / NACK is sent using the resource designated in the Downlink Control Information (DCI). In cases where one or two ACK / NACK bits and SR need to be transmitted simultaneously in the same Uplink (UL) slot, the ACK / NACK bits will be transmitted using the SR resource if the SR is positive. Otherwise, they will be sent using the ACK / NACK resource. If the Physical Uplink Shared Channel (PUSCH) is scheduled in the same slot, the ACK / NACK bits will normally be transmitted on PUSCH.OPSK Modulation

[0069] The SR and ACK / NACK bits are commonly transmitted using QPSK modulation symbols, as shown in the table of FIG. 3. FIG. 4 illustrates a typical QPSK symbol constellation and corresponding 2-bit symbol values. For FR1, if SR isP112081W001positive, the ACK / NACK bits are transmitted using the SR resource; if SR is negative, the ACK / NACK bits are transmitted using the ACK / NACK resource; if there is no ACK / NACK transmission, a positive SR is transmitted using the dedicated SR resource.Features and Targets’ Data Selection for SR Detection

[0070] As noted above, conventional methods use a pre-defined threshold to decode the SR and HARQ feedback bits, which may only be optimum in some channel conditions and configurations, and consequently may introduce decoding errors. The techniques described herein use machine learning (ML) to improve performance by increase the SR detection and HARQ feedback decoding accuracy, thereby reducing the false alarm and Ack false rates.

[0071] A crucial step for effective ML data-driven-based algorithms is to choose informative, discriminating, and independent features and adequately define the targets.

[0072] Feature data: The measurable observations and characteristics that can be used to train the model and to infer (e.g. predict) the expected or desired results using the model. Example feature data (inputs) selected for SR detection and HARQ feedback decoding are detailed in the Table of FIG. 5, and include:• Whether or not a current slot corresponds with an SR opportunity;• The number of HARQ feedback bits that gNB expects to receive from a UE;• Real (Re) and Imaginary (Im) values of the demodulated PUCCH HARQ feedback symbol of the UE;• Real (Re) and Imaginary (Im) values of the demodulated SR symbol of the UE;• estimated SINR of SR and HARQ symbols of the UE (e.g. in dB).

[0073] Target data: The desired result. Target data can be used to train the model and measure its performance, such as SR and HARQ bit detection accuracies, falseP112081W001alarm and Ack false rates, etc. In some embodiments, the neural network model can be trained using three different sets of target data. For example:• Single-value target data used for decoding SR and HARQ feedback bits, as shown in FIG. 6A• One-hot encoded target data for decoding SR and HARQ feedback bits as shown in FIG. 6B• Classified target data for decoding SR and HARQ feedback bits as shown in FIG. 6CGeneral Neural Network Structure for Model Training and SR & HARQ Bits’ Decoding

[0074] Fig 7 shows the general structure of an example neural network 700 useable for model training and SR and HARQ Bits decoding. The illustrated neural network 700 includes an input layer 702 configured to receive eight inputs as defined in FIG. 5, two hidden layers 704A-B each having a plurality of neurons 706, and an output layer 708. In an example embodiment, each hidden layer 704A-B of the neural network may include 64 neurons 706. The neurons 706 within the hidden layers may be fully interconnected in a manner known in the art.

[0075] FIG. 8 illustrates example activation functions that may be implemented in each neuron 706. In the example of FIG. 8, these activation functions include:• An identity (or Linear) function represented by the equation / (x) = x, where x is the neuron input signal and / (x) is the neuron output;• A rectified Linear Unit (ReLU) function represented by the equation:. (0 for x < 0f(x) = 1(x ffor x >n0• A Logistic (also referred to as a Soft Step) function represented by the equation:• A TanH function represented by the equation:P112081W001• A Softmax function represented by the equation:

[0076] In example embodiments, the respective output of each neuron 706 in a hidden layer 704 may use the ReLU activation function (FIG. 8). The outputs of the output layer 708 may, for example, use any of the activation functions shown in the table of FIG. 8.

[0077] In example embodiments, two cost functions may be used during the training phase. One usable cost function is Mean Squared Error (e.g. L2 loss), which may be defined as:

[0078] Another usable cost function is Cross-Entropy, which may be defined as:i S" i x(i, j)log (y(i,j))

[0079] Where N is the total number of input data samples; M is the number of features; and x(i,j) and y (i,j) are target and predicted values, respectively. Note that the Cross-Entropy cost function can only be applied to Sigmoid and SoftMax activation functions. The table of FIG. 9 shows six example pairs of activation and cost functions which may be used for training the neural network models for SR detection and HARQ feedback decoding.

[0080] As may be appreciated, the general structure of the example neural network 700 in combination with one or more selected activation functions and selected cost functions may be taken together to define a neural network model, combinations

[0081] Known normalization techniques (for example a scaling technique where the values are centered around the mean with a unit standard deviation) may be applied to any one or more of the following input features before the model training and inferencing steps are performed.P112081W001• pucchHqSymbol Re: Real (Re) value of a demodulated HARQ feedback symbol in the received uplink (e.g. PUCCH) radio signal;• pucchHqSymbol Im: Imaginary (Im) value of a demodulated HARQ feedback symbol in the received uplink (e.g. PUCCH) radio signal;• pucchSrSymbol Re: Real (Re) value of a demodulated SR symbol in the received uplink (e.g. PUCCH) radio signal;• pucchSrSymbol Im: Imaginary (Im) value of a demodulated SR symbol in the received uplink (e.g. PUCCH) radio signal;• sinr HqSymbol: estimated HARQ-symbol SINR of UE• sinr SrSymbol: estimated SR-symbol SINR of UE

[0082] An example normalization technique may be formulated as:>Where y(i) is the normalized value derived from input value x(i).

[0083] As may be appreciated, the general structure of the example neural network 700 in combination with one or more selected activation functions and selected cost functions may be taken together to define a neural network model. The table of FIG. 9 illustrates 6 possible neural network models, each of which can be defined as a specific activation function / cost function pair implemented in the neural network structure of FIG. 7.SR and HARQ Bits’ Decoding

[0084] To properly train the neural network, it is advantageous to apply specific rules for different slot types. Example rules include:• For SR-only slots (e.g. isSrSlot = 0 and noHqBits = 0), set pucchHqSymbol Re and pucchHqSymbol Im to 0. Both HqBitl and HqBit2 are considered as DTX.• For HARQ feedback-only slots (e.g. isSrSlot = 1 and noHqBits > 0), set pucchSrSymbol Re and pucchSrSymbol Im to 0. SrBit is considered as false.P112081W001• When noHqBits = 1 , set HqBit2 equal to HqBitl .

[0085] After training the neural network model, the decoding of SR and HARQ bits occurs during the inference phase. As mentioned previously, three sets of target features (data) can be used to train the neural network model. Each model has its own inference process, which will be further explained in the following three subsections.Single-value target training and inference processNeural Network Training

[0086] To effectively train the neural network model using single-value targets, it is beneficial to consider either model 1 or model 4 (FIG. 9), which differ in their activation functions. Model 1 employs a linear activation function, while model 4 uses a hyperbolic tangent activation function. Both models utilize the mean squared error cost function.

[0087] Appropriate selection of the target values can be important to ensure proper training of the neural network model. For example, in case of model 1 and model 4, the target values of SrBit, HqBitl, and HqBit2, may be adjusted as follows:• For SrBit, HqBitl, and HqBit2 initially 0, the corresponding target values may be set to 1.• For SrBit, HqBitl, and HqBit2 initially 1, the corresponding target values may be set to -1• In the case where the HARQ bits are designated as DTX (discontinuous transmission), the corresponding target values may be set to 0.SR and HARQ Feedback Bits’ Decoding

[0088] During the inference phase, after applying the input data (features) to the neural network model for each slot where gNB expects to receive the SR or HARQ feedback or both, we get three outputs, which correspond to the SR bit (SrBit), HARQ bit 1 (HqBitl), and HARQ bit 2 (HqBit2). The algorithm for decoding them is as follows:SR bit decoding:P112081W001o If SrbitPred > 0, the result predicted by the neural network is classified as SR positive. Otherwise, it is classified as SR negative.• HqBitl and HqBit2 decodingo If |HqBitlPred| (or |HqBit2Pred|) <= Threshold, HARQ is classified as DTXo If HqBitlPred (or HqBit2Pred) > Threshold, HARQ is classified as ACKo If HqBitlPred (or HqBit2Pred) < -Threshold, HARQ is classified as NACKOne-hot encoded target training and inference processNeural Network Training

[0089] When using one-hot encoded targets to train the neural network model, all six models can be utilized. When models 2, 3, 5, or 6 are employed, corresponding target values may be set as follows:• For the SR bit:o If the SR bit is 0 (SR positive); set SR_Positive = 1 and SR_Negative = 0.o If the SR bit is 1 (SR negative); set SR_Positive = 0 and SR_Negative = 1.• For the HARQ feedback bit x (where x=l or 2):o If it is NACK; set HqBitx NACK = 1, HqBitx DTX = 0 and HqBitx ACK = 0.o If it is DTX; set HqBitx DTX = 1, HqBitx ACK = 0 and HqBitx NACK =0.

[0090] When models 1 and 4 are utilized, the target values that are set =0 as described above for models 2, 3, 5, or 6 may be converted to = -1.P112081W001SR and Feedback Bits

[0091] SR bit decoding:• If srPositivePred > srNegtivePred, the result predicted by the neural network is classified as SR positive. Otherwise, it is classified as SR negative.

[0092] HqBitx (where x=l or 2) decoding:• If HqBitx DtxPred > max{HqBitx_NackPred, HqBitx AckPred}, it is classified as DTX. Otherwise, proceed to the next step.• If HqBitx NackPred >= HqBitx AckPred, it is classified as NACK.Otherwise, it will be classified as ACK.Classified target training and inferenceNeural Network Training

[0093] When employing classified targets to train the neural network model, all six models can be utilized. There may be 10 outputs, each corresponding to a specific class, as defined in the table of FIG. 10.

[0094] For example, if both HARQ bits are ACK, and the SR bit is 0 (SR positive), its corresponding class index is 3. If models 2, 3, 5, or 6 are employed, the target data corresponding to class index 3 is set = 1, and the remaining target data values are set =0. If models 1 and 4 are utilized, the target data corresponding to index 3 is still set =1, and the remaining target data values are set = -1.SR and HARQ Feedback Bits’ DecodingSR bit decoding:

[0095] Define the following:• srPositivePred = max { Class I Pred, Class3_Pred, Class5_Pred, Class7_Pred, Class9_Pred} ; andP112081W001• srNegtivePred = max{ClassO_Pred, Class2_Pred, Class4_Pred, Class_Pred6, Class8_Pred}

[0096] If srPositivePred > srNegtivePred, the result predicted by the neural network is classified as SR positive. Otherwise, it is classified as SR negative.HqBitl and HqBit2 decoding:

[0097] Define the following:• HqBitl DtxPred = HqBit2_ DtxPred = max{ClassO_Pred, Classl Pred} • HqBitl AckPred = max{Class2_Pred, Class3_Pred, Class4_Pred, Class5_Pred}• HqBitl NackPred = max{Class6_Pred, Class7_Pred, Class8_Pred, Class9_Pred}• HqBit2_AckPred = max{Class2_Pred, Class3_Pred, Class6_Pred, Class7_Pred}• HqBit2_NackPred = max{Class4_Pred, Class5_Pred, Class8_Pred, Class9_Pred}

[0098] Once the above-predicted values of the HARQ feedback are calculated from the class-predicted values, we can use the following algorithm to decode the HARQ Feedback Bits.• If HqBitl DtxPred > max{HqBitl_NackPred, HqBitl AckPred}, it is classified as DTX.Otherwise, proceed to the next step.• If HqBitl NackPred >= HqBitl AckPred, it is classified as NACK. Otherwise, it will be classified as ACK.• If HqBit2_DtxPred > max{HqBit2_NackPred, HqBit2_AckPred}, it is classified as DTX.Otherwise, proceed to the next step.• If HqBit2_NackPred >= HqBit2_AckPred, it is classified as NACK. Otherwise, it will be classified as ACK.

[0099] P112081W001Neural Network Inference Model Performance Evaluation for SR Detection and HARO Feedback DecodingComputing SR Detection Errors

[0100] When the UE sends an SR bit on the SR opportunity slot, but the neural network predicts no SR bit, it is an SR Detection Error. The SR detection error rate (%) is calculated as100

[0101] When the UE doesn't send the SR bit on the SR opportunity slot, but the neural network predicts there is an SR bit, it is a false alarm error. The false alarm error rate (%) is calculated as

[0102] The total error rate is calculated asNumber of SR Detection Errors + Number of False Alarm Errors - - - - - * 100 Number of SR Positive + Number of SR Negative

[0103] The SR detection error rate measures the percentage of SR positive instances that were incorrectly predicted as SR negative. This helps evaluate how well the model can detect the presence of an SR bit when it is actually present.

[0104] The false alarm error rate measures the percentage of SR negative instances that were incorrectly predicted as SR positive. This helps evaluate how often the model incorrectly predicts the presence of an SR bit when it is not actually present.

[0105] The total error rate takes into account both types of errors and provides an overall assessment of the model's performance.

[0106] By monitoring these error rates, we can gain insights into the model's strengths and weaknesses in detecting SR bits, and use this information to make improvements if necessary.P112081W001Errors• DTX to HARQ Error Rate (%) :Number of DTX bits classified as ACK or NACKT otal DTX bits in dataset• HARQ to DTX Error Rate (%) :Number of ACK or NACK bits classified as DTXT otal ACK or NACK bits in dataset• NAK to ACK Error Rate (%) :Number of NACK bits classified as ACKTotal NACK bits in dataset• Total HARQ Feedback Decoding Error Rate (%):Number of Incorrect HARQ Feedback bit detectionTotal HARQ Feedback bits in dataset

[0107] The DTX to HARQ Error Rate measures the percentage of DTX bits that were incorrectly classified as ACK or NACK. This helps evaluate how well the model can distinguish DTX bits from ACK or NACK bits.

[0108] The NACK to ACK Error Rate measures the percentage of NACK bits that were incorrectly classified as ACK. This helps evaluate how often the model incorrectly predicts ACK when NACK is actually present.

[0109] The Total HARQ Feedback Decoding Error Rate provides an overall assessment of the model's performance in decoding HARQ feedback bits.Use SINR and Similarity Thresholds to Improve the Overall System Performance

[0110] As previously discussed, evaluating the performance of SR detection and HARQ feedback decoding is crucial for assessing their impacts on the overall system performance. However, it's important to recognize that the effects of SR detection and false alarm error rates differ significantly.

[0111] For instance, a 20% increase in the SR detection error rate would lead to a corresponding 20% increase in the number of SRs sent by the UEs. This increase couldP112081W001potentially delay new uplink transmissions by 20%, affecting user experience. On the other hand, a 20% increase in the SR false alarm error rate would result in an 80% increase in wasted uplink resources due to the triggering of additional uplink transmissions. This would significantly impact the uplink cell throughput. In specific scenarios, it may be beneficial to reduce the SR false alarm rate, even if it means slightly increasing the SR detection error rate. Adjusting SINR and similarity thresholds could help achieve this trade-off.

[0112] When it comes to HARQ feedback decoding, errors such as DTX to ACK or NACK and NACK to ACK are particularly detrimental, as they may trigger RLC layer retransmissions, leading to increased delay in DL data transmission and reduced DL cell throughput. By using SINR and similarity thresholds, it may be possible to reduce these errors, even if it results in an increase in ACK or NACK to DTX errors and ACK to NACK errors.

[0113] Understanding these trade-offs and the potential impacts of error rate adjustments is essential for optimizing the performance of SR detection and HARQ feedback decoding in wireless communication systems, such as 5G.The SINR Threshold

[0114] As previously discussed, a scheduling request (SR) is transmitted using the dedicated SR resource, and acknowledgment (ACK / NACK) is sent using the resource indicated by the downlink (DL) DCI. If both the ACK / NACK bits and SR are transmitted simultaneously on the same uplink (UL) slot, the ACK / NACK bits are transmitted using the SR resource if the SR is positive.

[0115] In general, when the user equipment (UE) sends an SR on a UL slot, the estimated signal-to-interference-plus-noise ratio (SINR) of the SR symbol should be higher, regardless of whether the ACK / NACK bits are transmitted on the same UL slot. If the estimated SINR is below a certain threshold, it is likely that the UE does not send an SR on this slot. We can use the estimated SINR value on the SR symbol to minimize SR false alarm errors:P112081W001If sinrSrSymbol < sinr Threshold, the SR is classified as negative.

[0116] Likewise, we can utilize the estimated SINR values on both the SR and HARQ feedback symbols to reduce errors in discontinuous transmission (DTX) to ACK or NACK:• If max {sinrSrSymbol, sinrHqSymbol} < sinr Threshold, the hybrid automatic repeat request (HARQ) feedback bit will be classified as DTX.The Similarity Threshold

[0117] The errors made by the neural network in SR detection often occur when the difference between srPositivePred and srNegativePred is relatively small. In other words, when the values of srPositivePred and srNegativePred are very close, it becomes challenging for the neural network to make the correct decision. To reduce the false alarm error rate, the prediction result is classified as SR negative when the difference between srPositivePred and srNegativePred is less than a certain threshold, which we refer to as the similarity threshold.

[0118] For the cases of one-hot and classified target data, we can use the following formula to predict the SR detection result:• If (srPositivePred - srNegtivePred) > Similarity Thr, the result predicted by the neural network is classified as SR positive. Otherwise, it is classified as SR negative.

[0119] For the case of single- value target data, we can use the following formula to predict the SR detection result:• If srBitPred > Similarity Thr, it is classified as SR positive. Otherwise, it is classified as SR negative.

[0120] Similarly, for the cases of one-hot and classified target data, we can utilize the similarity threshold to reduce errors in DTX to ACK or NACK and NACK to ACK, as outlined in the following two steps:P112081W001• If HqBitl DtxPred (HqBit2_ DtxPred) + Similarity Thr >max {HqBitl NackPred (HqBit2_ NackPred), HqBitl AckPred (HqBit2_AckPred)}, it is classified as DTX. Otherwise, proceed to the next step.• If HqBitl NackPred (HqBit2_ NackPred) + Similarity Thr >HqBitl AckPred (HqBit2_AckPred), it is classified as NACK. Otherwise, it will be classified as ACK.

[0121] For the single- value target data case, we can use the following formula to predict the HARQ feedback bit decoding result:• If |HqBitlPred| (or |HqBit2Pred|) <= Threshold + Similarity Thr, it is classified as DTX• If HqBitlPred (or HqBit2Pred) > Threshold + Similarity Thr, it is classified as ACK• If HqBitlPred (or HqBit2Pred) < -(Threshold + Similarity Thr), it is classified as NACKThe Combined Thresholds

[0122] In the third scenario, we can use both the SINK and the Similarity Thresholds to reduce the false alarm error rate. For one-hot and classified target data cases:• If (((srPositivePred - srNegtivePred) > Similarity Thr) AND (sinrSrSymbol > sinrThr)), the result predicted by the neural network is classified as SR positive. Otherwise, it is classified as SR negative.

[0123] For the single- value target data case, we can use the following formula to predict the SR detection result:If (srBitPred > Similarity Thr) AND (sinrSrSymbol > sinrThr), it is classified as SR positive. Otherwise, it is classified as SR negative.P112081W001

[0124] Similarly, for one-hot and classified target data cases we can utilize both the SINK and the similarity threshold to reduce errors in DTX to ACK or NACK and NACK to ACK, as outlined in the following two steps (noting again that x=l or 2, e.g. HqBitx_DtxPred= HqBitl DtxPred or HqBit2_DtxPred):• If (HqBitx DtxPred + Similarity Thr > max{HqBitx_NackPred, HqBitx AckPred}) OR (max{sinrSrSymbol, sinrHqSymbol} < sinr Threshold), it is classified as DTX. Otherwise, proceed to the next step.• If (HqBitx NackPred + Similarity Thr) > HqBitx AckPred , it is classified as NACK. Otherwise, it will be classified as ACK.

[0125] For the single- value target data case, we can use the following formula to predict the HARQ feedback bit decoding result:• If (|HqBitlPred| (or |HqBit2Pred|) <= Threshold + Similarity Thr) OR (max{sinrSrSymbol, sinrHqSymbol} < sinr Threshold), it is classified as DTX. Otherwise, proceed to the next step.• If HqBitlPred (or HqBit2Pred) > Threshold + Similarity Thr, it is classified as ACK• If HqBitlPred (or HqBit2Pred) < -(Threshold + Similarity Thr), it is classified as NACKDynamic Threshold

[0126] In the preceding sections, we discussed the utilization of SINR and Similarity Thresholds as a means to reduce the false alarm error rate. Instead of employing a fixed threshold, we can dynamically adjust the threshold based on the prevailing network conditions, such as the false alarm error rate (FAER).

[0127] When the gNB detects an SR positive, it transmits an uplink grant to the UE, enabling it to transmit a Buffer Status Report (BSR) on the allocated PUSCH resources. Subsequently, if the PUSCH is successfully decoded, the gNB examines the BSR to ascertain whether the UE has uplink data to transmit. If the UE has no data to send, theP112081W001detected SR positive is deemed to be false. In the event that the PUS CH fails to decode after retransmissions, and the estimated SINR falls below the dtxThreshold, the UE is considered to be in Discontinuous Transmission (DTX) mode, and the detected SR positive is also classified as false. If the estimated SINR is above the dtxThreshold, the SR detection result is marked as "unknown."

[0128] When the gNB detects an SR positive and sends an uplink grant to the UE, the subsequent decoding and analysis of the Buffer Status Report (BSR) and other parameters determine whether the SR positive result is accurate, false, or “unknown”. Based on this determination, the Similarity Threshold can be adjusted as follows:• If the current slot is an SR opportunity slot and the predicted result is SR negative, the Similarity Threshold is decreased by an amount defined as DOWN STEP.SimilarityThrq) = SimilarityThr(I-1)— D0WN_STEP • If the predicted result is SR positive but is determined as false, the Similarity Threshold is increased by a UP STEP.SimilarityThrq) = SimilarityThr(i-1) + UP_STEP • If the predicted result is SR positive and is determined as "unknown," no action is taken to adjust the Similarity Threshold.

[0129] This dynamic adjustment of the Similarity Threshold based on the SR detection results allows the network to adapt to changing conditions and improve the overall system performance by minimizing false alarms.

[0130] The ratio of the UP STEP and DOWN STEP parameters, which govern the rate of increase or decrease of the Similarity Threshold, may be determined based on the desired false alarm error rate (FAER) target according to the formula:UP STEP / DOWN STEP = l / FAER TARGET - 1.P112081W001

[0131] The DOWN STEP value can be adjusted to achieve the desired convergence speed. The FAER target, which can be pre-determined (e.g., 3%), may be now defined with the formula:Number of times when the SR positive is FALSE Number of times when the SR positive is detected

[0132] To prevent over-adjustment of the Similarity Threshold, two additional thresholds, a high threshold value, THR HIGH and a low threshold value, THR LOW, may be defined to limit the Similarity Threshold to within a specified range. The adjustment algorithm may then be revised as follows:• If the predicted result is SR negative, the Similarity Threshold may be decreased by a DOWN STEP:SimilarityThr(j) = SimilarityThr(I-1)— D0WN_STEP • Additionally, if Similarity Thr falls below THR LOW, it is set to THR LOW:if SimilarityThr(j) < THR_L0W, SimilarityThr(i+1) = THR_L0W• If the predicted result is SR positive but is determined as false, the Similarity Threshold is increased by a UP STEP:SimilarityThr(j) = SimilarityThr(i-1) + UP_STEP• Similarly, if Similarity Thr exceeds THR HIGH, it is set to THR HIGH: ifSimilarityThr(0> THR_HIGH, SimilarityThr(i+1)= THR_HIGH

[0133] The values for THR LOW and THR HIGH can be determined through simulation and field test results. This approach ensures that the Similarity Threshold is adjusted based on SR detection results while preventing excessive over-adjustment by constraining it within the defined thresholds. The dynamic threshold can be adjusted on a per UE or per cell basis.P112081W001On-The-Fly Online Training

[0134] Traditionally, Neural Network models have been trained using simulated data or data collected from the field, which may not always be optimal for all gNBs or cells. To address this limitation, an On-The-Fly online training method can be used to refine the model on a per gNB or per cell basis.

[0135] An example approach involves collecting training data during high-traffic volume periods, with gNBs capturing input data and test UEs gathering target data. Subsequently, the collected target data and their corresponding features are stored. During low-traffic periods, typically at night, the Neural Network model is retrained using the accumulated data, effectively replacing the older model with an updated version.

[0136] It's worth noting that while our invention is initially described in the context of 5GNR, it is also applicable to other wireless systems, such as LTE. Additionally, the On-The-Fly online training method can be seamlessly implemented in the cloud, enhancing its scalability and accessibility.

[0137] While processes in the figures may show a particular order of operations performed by certain embodiments of the present disclosure, it should be understood that such order is representative, and that alternative embodiments may perform the operations in a different order, combine certain operations, overlap certain operations, etc.

[0138] Those skilled in the art will recognize improvements and modifications to the embodiments of the present disclosure. All such improvements and modifications are considered within the scope of the concepts disclosed herein.

Claims

P112081W001ClaimsWhat is claimed is:

1. A method implemented in a radio node of a radio access network, the method comprising:receiving an uplink radio signal from a user equipment (UE), the uplink radio signal including Scheduling Request (SR) bits and Hybrid Automatic Repeat Request (HARQ) feedback bits; andanalysing the received uplink radio signal using at least one neural network model to jointly decode the SR bits and HARQ feedback bits.

2. The method as claimed in claim 1 , wherein the uplink radio signal comprises a Physical Uplink Control Channel (PUCCH) signal.

3. The method as claimed in claim 1 or 2, wherein analysing the received uplink radio signal comprises:processing the received uplink radio signal to obtain feature data; and supplying the feature data to an input of the at least one neural network model.

4. The method as claimed in claim 3, wherein the feature data comprises data indicative of a plurality of features of the received uplink radio signal, the plurality of features including any one or more of:whether or not a current slot of the received uplink radio signal corresponds with an SR opportunity;a number of HARQ feedback bits that the radio node expects to receive from the UE;Real (Re) and Imaginary (Im) values of a demodulated HARQ feedback symbol in the received uplink radio signal;P112081W001Real (Re) and Imaginary (Im) values of a demodulated SR symbol in the received uplink radio signal; andestimated SINR of SR and HARQ symbols of the UE.

5. The method as claimed in claim 4, wherein supplying the feature data to an input of the at least one neural network model comprises, for at least one feature:applying a normalization function to the feature to compute a normalized feature value; andsupplying the normalized feature value to the input of the at least one neural network model.

6. The method as claimed in claim 1 or 2, wherein the at least one neural network model is selected from a set of predetermined neural network models including:Model 1 characterised by a linear activation function and a Mean Squared Error cost function;Model 2 characterised by a Rectified Linear Unit (ReLU) activation function and a Mean Squared Error cost function;Model 3 characterised by a Sigmoid activation function and a Mean Squared Error cost function;Model 4 characterised by a TanH activation function and a Mean Squared Error cost function;Model 5 characterised by a Sigmoid activation function and a Cross-Entropy cost function; andModel 6 characterised by a SoftMax activation function and a Cross-Entropy cost function.P112081W0017. The method as claimed in any one of claims 1-6, wherein analysing the received signal comprises:generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits;selecting, from the plurality of candidate values, most likely SR and HARQ values, based at least in part on a detected Signal-to-Interference-plus- Noise Ratio (SINR) of the received uplink radio signal and one or more similarity thresholds.

8. The method as claimed in claim 7, wherein generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:computing, by a first neural network model trained using Single-value target data, first predicted values for each of the SR and HARQ bits; if the first predicted value of the SR bit (srBitPred) is greater than a predetermined first threshold, classifying the predicted SR bit as SR positive, and otherwise classifying the predicted SR bit as SR negative; if the first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are less than or equal to a predetermined second threshold, then the predicted HARQ is classified as DTX;if the first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are greater than the second threshold, then the predicted HARQ is classified as ACK; andif first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are less than a predetermined third threshold, then the predicted HARQ is classified as NACK.P112081W0019. The method as claimed in claim 8, wherein the predetermined first threshold has a value of zero (0).

10. The method as claimed in claim 8 or 9, wherein the predetermined first threshold comprises one of the one or more similarity thresholds.

11. The method as claimed in claim 8, wherein the predicted SR bit is classified as SR positive only if the first predicted value of the SR bit (srBitPred) is greater than the predetermined first threshold, and a signal to interference and noise ratio (SINR) of the received SR symbol (sinrSrSymbol) is greater than a first predetermined SINR threshold (sinrThr), and otherwise the predicted SR bit is classified as SR negative.

12. The method as claimed in any one of claims 8 - 11, wherein the predetermined third threshold is equal in magnitude and opposite in sign to the predetermined second threshold.

13. The method as claimed in any one of claims 8 - 12, wherein at least the predetermined second threshold comprises a sum of a predetermined value (Threshold) and one of the one or more similarity thresholds.

14. The method as claimed in claim 13, wherein the predicted HARQ is classified as DTX if either:the first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are less than or equal to the predetermined second threshold; or the greater one of a signal to interference and noise ratio (SINR) of the received SR symbol (sinrSrSymbol) and the SINR of the HARQ symbol (sinrHqSymbol) is less than a second predetermined SINR threshold (sinr Threshold).

15. The method as claimed in claim 14, wherein:P112081W001the predicted HARQ is classified as ACK if the first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are greater than the predetermined second threshold; andthe predicted HARQ is classified as NACK if the first predicted values of the HARQ bits (HqBitlPred and HqBit2Pred) are less than the predetermined third threshold.

16. The method as claimed in claim 7, wherein generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:computing, by a second neural network model trained using One-hot encoded target data, second predicted values for each of the SR and HARQ bits; if a SR-Positive predicted value (srPositivePred) is greater than a SR-Negative predicted value (srNegativePred), then classifying the predicted SR bit as SR positive, and otherwise classifying the predicted SR bit as SR negative; if HARQ=DTX predicted values (HqBitl DtxPred and HqBit2_DtxPred) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as DTX, and otherwise:if the HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) are greater than or equal to the HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as NACK, and otherwise classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as ACK.P112081W00117. The method as claimed in claim 16, wherein the predicted SR bit is classified as SR positive if a difference between the SR-Positive predicted value (srPositivePred) and the SR-Negative predicted value (srNegativePred) is greater than one of the one or more similarity thresholds, and otherwise classifying the predicted SR bit as SR negative.

18. The method as claimed in claim 16, wherein:if the sums of the HARQ=DTX predicted values and one of the one of more similarity thresholds (HqBitl DtxPred+SimilarityThr, and HqBit2_DtxPred+SimilarityThr) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as DTX, and otherwise:if the sums of the HARQ=NACK predicted values and the one of the one of more similarity thresholds (HqBitl NackPred+SimilarityThr, and HqBit2_NackPred+SimilarityThr) are greater than or equal to the HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as NACK, and otherwise classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as ACK.

19. The method as claimed in claim 16, wherein the predicted SR bit is classified as SR positive if both:a difference between the SR-Positive predicted value (srPositivePred) and the SR-Negative predicted value (srNegativePred) is greater than one of the one or more similarity thresholds, and otherwise classifying the predicted SR bit as SR negative; andP112081W001a signal to interference and noise ratio (SINR) of the SR symbol is greater than a predetermined SINR threshold (sinrSr)20. The method as claimed in claim 16, wherein:the predicted HARQ feedback bits (HqBitl and HqBit2) are classified as DTX if either one of:the sums of the HARQ=DTX predicted values and one of the one of more similarity thresholds (HqBitl DtxPred+SimilarityThr, and HqBit2_DtxPred+SimilarityThr) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred); andboth of a signal to interference and noise ratio (SINR) of the SR symbol (sinrSrSymbol) and the SINR of the HARQ symbol (sinrHqSymbol) are less than a predetermined second SINR threshold (sinr Threshold).

21. The method as claimed in claim 7, wherein generating, by the at least one neural network model, a plurality of candidate values for the SR bits and the HARQ bits comprises:computing, by a third neural network model trained using Classified target data, third predicted values for each of the SR and HARQ bits;if a SR-Positive predicted value (srPositivePred) is greater than a SR-Negative predicted value (srNegativePred), then classifying the predicted SR bit as SR positive, and otherwise classifying the predicted SR bit as SR negative; if HARQ=DTX predicted values (HqBitl DtxPred and HqBit2_DtxPred) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying theP112081W001predicted HARQ feedback bits (HqBitl and HqBit2) as DTX, and otherwise:if the HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) are greater or equal to the HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as NACK, and otherwise classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as ACK.

22. The method as claimed in claim 21, wherein the classified target data comprises, for each one of a plurality of classes, respective values of the SRbit and HARQ feedback bits, and wherein the third predicted values for each of the SR and HARQ bits comprises:the SR-Positive predicted value (srPositivePred), comprising a maximum predicted value of a subset of classes corresponding to a positive SR bit value;the SR-Negative predicted value (srNegativePred), comprising a maximum predicted value of a subset of classes corresponding to a negative SR bit value;the HARQ=DTX predicted value (HqBitx DtxPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of DTX;the HARQ=NACK predicted value (HqBitx NackPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of NACK; andthe HARQ=ACK predicted value (HqBitx AckPred), comprising a maximum predicted value of a subset of classes corresponding to a HARQ bit-x value indicative of ACK.P112081W00123. The method as claimed in claim 21, wherein the predicted SR bit is classified as SR positive if a difference between the SR-Positive predicted value (srPositivePred) and the SR-Negative predicted value (srNegativePred) is greater than one of the one or more similarity thresholds, and otherwise classifying the predicted SR bit as SR negative.

24. The method as claimed in claim 21, wherein:if the sums of the HARQ=DTX predicted values and one of the one of more similarity thresholds (HqBitl DtxPred+SimilarityThr, and HqBit2_DtxPred+SimilarityThr) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as DTX, and otherwise:if the sums of the HARQ=NACK predicted values and the one of the one of more similarity thresholds (HqBitl NackPred+SimilarityThr, and HqBit2_NackPred+SimilarityThr) are greater than or equal to the HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred), then classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as NACK, and otherwise classifying the predicted HARQ feedback bits (HqBitl and HqBit2) as ACK.

25. The method as claimed in claim 21, wherein the predicted SR bit is classified as SR positive if both:a difference between the SR-Positive predicted value (srPositivePred) and the SR-Negative predicted value (srNegativePred) is greater than one of the one or more similarity thresholds, and otherwise classifying the predicted SR bit as SR negative; andP112081W001a signal to interference and noise ratio (SINR) of the SR symbol is greater than a predetermined SINR threshold (sinrSr).

26. The method as claimed in claim 21, wherein:the predicted HARQ feedback bits (HqBitl and HqBit2) are classified as DTX if either one of:the sums of the HARQ=DTX predicted values and one of the one of more similarity thresholds (HqBitl DtxPred+SimilarityThr, and HqBit2_DtxPred+SimilarityThr) are greater than either one of HARQ=NACK predicted values (HqBitl NackPred and HqBit2_NackPred) or HARQ=ACK predicted values (HqBitl AckPred and HqBit2_AckPred); andboth of a signal to interference and noise ratio (SINR) of the SR symbol (sinrSrSymbol) and the SINR of the HARQ symbol (sinrHqSymbol) are less than a predetermined second SINR threshold (sinr Threshold).

27. The method as claimed in claim 7, further comprising dynamically adjusting the one or more similarity thresholds based at least in part on a predetermined target SR false alarm error rate (FAER).

28. The method as claimed in claim 27, wherein dynamically adjusting the one or more similarity thresholds comprises:decreasing, a given one of the one of more similarity thresholds by DOWN STEP, when a current slot is an SR opportunity slot and the predicted SR bit is classified as SR positive; andincreasing the given one of the one of more similarity thresholds by UP STEP, when the SR bit is classified as SR positive but is determined as false.P112081W00129. The method as claimed in claim 28, wherein a ratio of UP STEP and DOWN STEP, is determined based on a desired false alarm error rate (FAER) target.

30. The method as claimed in claim 29, wherein the FAER target indicates a ratio between a number of times when the SR positive classification is false and a number of times when the SR positive is detected.

Images