Model training method and apparatus, communication device, and readable storage medium

Abstract

This application discloses a model training method, apparatus, communication device, and readable storage medium, belonging to the field of wireless technology. The model training method in this application includes: a receiving end acquiring training data transmitted through a channel, wherein the training data is obtained by modulating a first bitstream generated by a pilot sequence generation algorithm at a transmitting end; generating training data tags using the pilot sequence generation algorithm; and retraining an artificial intelligence (AI) model used for demodulation processing based on the training data and the training data tags. This reduces the retraining overhead of the AI ​​model at the receiving end.

Description

Technical Field

[0001] This application belongs to the field of wireless technology, specifically relating to a model training method, apparatus, communication device, and readable storage medium. Background Technology

[0002] In related technologies, to improve transmission performance, artificial intelligence (AI) models can be used at the receiver for nonlinear compensation. A typical approach is to use AI models to perform channel estimation, interpolation, equalization, and demodulation of the demodulation reference signal (DMRS).

[0003] For AI model-based receivers, when there are significant differences in channel characteristics between the training and testing scenarios, or significant differences in nonlinearity between the transmitter and receiver, the AI ​​model may experience a significant performance degradation due to poor generalization. To improve the generalization of the AI ​​model at the receiver, it is typically necessary to collect the bitstream from the transmitter as labels for training data and then retrain the AI ​​model. This results in a large overhead for data collection at the receiver, leading to a significant overhead for retraining the AI ​​model. Summary of the Invention

[0004] The purpose of this application is to provide a model training method, apparatus, communication device, and readable storage medium to solve the problem of high retraining overhead of AI models at the receiving end in related technologies.

[0005] To solve the above-mentioned technical problems, this application is implemented as follows:

[0006] Firstly, a model training method is provided for application at the receiving end, including:

[0007] The receiving end acquires training data transmitted through the channel, which is obtained by modulating the first bit stream generated by the sending end using the pilot sequence generation algorithm;

[0008] The receiving end uses the pilot sequence generation algorithm to generate training data labels;

[0009] The receiving end retrains the artificial intelligence (AI) model used for demodulation processing based on the training data and the training data labels.

[0010] Secondly, a model training method is provided for application at the sending end, including:

[0011] The transmitting end uses a pilot sequence generation algorithm to generate the first bit stream;

[0012] The transmitting end modulates the first bit stream to obtain training data;

[0013] The transmitting end sends the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the training data and the training data labels generated by the pilot sequence generation algorithm.

[0014] Thirdly, a model training device is provided for use at the receiving end, including:

[0015] The acquisition module is used to acquire training data transmitted through the channel. The training data is obtained by modulating the first bit stream generated by the pilot sequence generation algorithm at the transmitting end.

[0016] The first generation module is used to generate training data labels using the pilot sequence generation algorithm;

[0017] The training module is used to retrain the artificial intelligence (AI) model used for demodulation processing based on the training data and the training data labels.

[0018] Fourthly, a model training device is provided for use at the transmitting end, comprising:

[0019] The second generation module is used to generate the first bit stream using a pilot sequence generation algorithm;

[0020] The processing module is used to modulate the first bit stream to obtain training data;

[0021] The second transmitting module is used to transmit the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the received training data and the training data labels generated by the pilot sequence generation algorithm.

[0022] Fifthly, a communication device is provided, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect, or the steps of the method described in the second aspect.

[0023] In a sixth aspect, a readable storage medium is provided, on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect, or the steps of the method described in the second aspect.

[0024] In a seventh aspect, a computer program product is provided, including computer instructions that, when executed by a processor, implement the steps of the method described in the first aspect, or the steps of the method described in the second aspect.

[0025] According to the solution of this application embodiment, when retraining the AI ​​model in the receiving end, the receiving end can reuse the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream of the sending end as the training data label, thereby reducing the data acquisition overhead and thus reducing the retraining overhead of the AI ​​model in the receiving end. Attached Figure Description

[0026] Figure 1 This is a flowchart of a model training method provided in an embodiment of this application;

[0027] Figure 2 This is a schematic diagram of the AI ​​model at the receiving end in an embodiment of this application;

[0028] Figure 3 This is a schematic diagram of the bitstream insertion method in an embodiment of this application;

[0029] Figure 4 This is a schematic diagram of the resource mapping method in a specific embodiment of this application;

[0030] Figure 5 This is a schematic diagram of the retraining process of the AI ​​model in the embodiments of this application;

[0031] Figure 6 This is a flowchart of another model training method provided in the embodiments of this application;

[0032] Figure 7 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application;

[0033] Figure 8 This is a schematic diagram of another model training device provided in an embodiment of this application;

[0034] Figure 9 This is a schematic diagram of the structure of a communication device provided in an embodiment of this application. Detailed Implementation

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0036] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0037] To facilitate understanding of this application, the following will be explained first.

[0038] Taking the uplink (UL) transceiver as an example, the functions of the uplink transceiver are as follows: For the transmitter, the bit stream encoded by Low Density Parity Check (LDPC) is I / Q modulated into a constellation diagram, then DMRS pilots are inserted and time-frequency resource mapping is performed. An Inverse Fast Fourier Transform (iFFT) is then performed to transform it into an Orthogonal Frequency Division Multiplexing (OFDM) waveform. Finally, a Cyclic Prefix (CP) is inserted for transmission. For the receiver, after passing through a noisy and interference-prone channel, the received time-frequency domain data is processed. After removing the CP and performing a Fast Fourier Transform (FFT), the DMRS pilots are extracted and estimated and interpolated to obtain the channel matrix in the entire time-frequency domain. Then, based on the estimated channel information, the data is equalized to remove the influence of the channel. Finally, the constellation diagram is demodulated and the channel is decoded to recover the original bit stream.

[0039] The transmitted signal is distorted due to the nonlinearity of the transmitter components and the non-stationary time-frequency fading channel. To improve transmission performance, nonlinearity compensation can be performed using AI models at the receiver. A typical approach is to use AI models to implement channel estimation, interpolation, equalization, and demodulation in DMRS.

[0040] The model training method, apparatus, communication device, and readable storage medium provided in this application will be described in detail below with reference to the accompanying drawings and through specific embodiments and application scenarios.

[0041] Please see Figure 1 , Figure 1This is a flowchart of a model training method provided in an embodiment of this application. This method is applied at the receiving end, such as... Figure 1 As shown, the method includes the following steps:

[0042] Step 11: The receiving end acquires the training data transmitted through the channel. The training data is obtained by modulating the first bit stream generated by the sending end using the pilot sequence generation algorithm.

[0043] Step 12: The receiving end uses the pilot sequence generation algorithm to generate training data labels;

[0044] Step 13: The receiving end retrains the AI ​​model used for demodulation processing based on the training data and the training data labels.

[0045] In this embodiment, the receiving end can be a network-side device or a terminal, and correspondingly, the sending end can be a terminal or a network-side device. For example, in an uplink transmission scenario, the receiving end is a network-side device and the sending end is a terminal; or, in a downlink transmission scenario, the receiving end is a terminal and the sending end is a network-side device. In other words, the solution in this application can be applied to the retraining of AI models in both uplink (UL) and downlink (DL) receiving ends.

[0046] The pilot sequence generation algorithm is specifically a pilot sequence generation algorithm, such as a DMRS generation algorithm. The pilot sequence may be, for example, a Gold sequence. The aforementioned modulation processing of the first bitstream may include modulating the first bitstream, inserting pilot symbols, and mapping the modulated bitstream and pilot symbols to corresponding time-frequency resources, etc. For example, the position where the transmitting end generates the first bitstream (i.e., the position where the first bitstream is inserted) can be as follows: Figure 3 As stated above.

[0047] The training data can be symbol-level data received by the receiving end (such as network-side devices or terminals). The training data label is specifically the label corresponding to the training data received by the receiving end, such as a 01 bit stream generated by the receiving end using a pilot sequence generation algorithm.

[0048] Optionally, the AI ​​model is specifically used for at least one of the following: channel estimation, channel interpolation, channel equalization, demodulation, and decoding using pilot signals (such as DMRS). This application does not limit the specific implementation methods of channel estimation, channel interpolation, channel equalization, demodulation, and decoding, which can be determined based on the actual situation.

[0049] For example, the AI ​​model can be used to implement DMRS channel estimation, channel interpolation, channel equalization, demodulation, and decoding. That is, the aforementioned demodulation processing can include DMRS channel estimation, channel interpolation, channel equalization, demodulation, and decoding. Figure 3 As shown. The retraining of the AI ​​model can be performed using supervised learning.

[0050] Optionally, the structure of the AI ​​model can be implemented by combining model structures such as Convolutional Neural Networks (CNN), Residual Networks (ResNet), and / or self-attention mechanisms. For example, the structure of the AI ​​model can be as follows: Figure 2 As shown, the model input includes at least the training data received by the receiver, the known DMRS pilot pattern, and the channel estimate at the pilot. The model output is a 0 / 1 bitstream probability sequence. By comparing this 0 / 1 bitstream probability sequence with the training data labels, the error of the model output can be determined, and the model parameters can be adjusted based on this error to achieve the model training process. The DMRS pilot pattern is the mapping of the pilot symbols and data bits (e.g., data bits set to 0) at the transmitter onto the time-frequency resource grid.

[0051] For example, the AI ​​model can be pre-trained offline by a wireless access network AI model training system. This AI model training system can be deployed on a centralized unit (CU) and / or a distributed unit (DU) on the network side, or it can be deployed on a logical entity across CUs on the network side, without limitation.

[0052] Retraining of the AI ​​model in the receiving end can be triggered when the performance of the AI ​​model deteriorates significantly due to scenarios such as large channel differences between the training scenario and the application scenario, or large nonlinear differences between the transmitting and receiving ends.

[0053] According to the solution of this application embodiment, when retraining the AI ​​model in the receiving end, the receiving end can reuse the pilot sequence generation algorithm to generate training data labels, without having to collect the bitstream from the transmitting end as the training data labels, thereby reducing data acquisition overhead and thus reducing the retraining overhead of the AI ​​model in the receiving end. Furthermore, by reusing the pilot sequence generation algorithm to generate training data labels, the overhead of signaling used to indicate the algorithm and seed for generating training data labels can also be reduced, and the terminal avoids calling multiple sets of algorithms to generate pilot sequences and bitstream sequences as training data labels, reducing terminal resource consumption, avoiding switching between different algorithms, and facilitating terminal management.

[0054] Optionally, before acquiring the training data transmitted through the channel, the model training method in this embodiment further includes: when the AI ​​model cannot meet the requirements of demodulation processing, the receiving end initiates retraining of the AI ​​model. This allows for timely optimization of the AI ​​model to meet the requirements of demodulation processing, effectively improving model performance and generalization.

[0055] Optionally, the above-mentioned generation of training data labels using the pilot sequence generation algorithm includes: the receiving end generating a second bit stream according to the pilot sequence generation algorithm and the configured initial parameters of the pilot sequence, and using the second bit stream as the training data label. That is, the bit stream used as the training data label does not need to be modulated.

[0056] In some alternative embodiments, if the pilot sequence generation algorithm is not used to generate training data tags, the network-side device can configure the sequence generation algorithm for generating training data tags to the terminal through Radio Resource Control (RRC) signaling, so that the terminal / network-side device uses a sequence generation algorithm different from the pilot sequence generation algorithm to generate training data tags.

[0057] Optionally, if the receiving end is a network-side device and the sending end is a terminal, the model training method in this embodiment may further include:

[0058] The network-side device sends first configuration information to the terminal; wherein, the first configuration information may include, but is not limited to, at least one of the following:

[0059] (1) First identifier, the first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; in this way, when the network-side device starts the retraining of the AI ​​model, the first identifier can indicate that the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels, that is, to indicate that the pilot sequence generation algorithm is used to generate training data labels, thereby reusing the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream of the sending end as the label of the training data, thereby reducing the retraining overhead of the AI ​​model in the receiving end;

[0060] (2) Configuration information of time-frequency resources for sending the training data; In this way, with the help of this configuration information, time-frequency resources for sending training data can be obtained, thereby realizing the transmission of training data based on these time-frequency resources;

[0061] (3) Second identifier, the second identifier indicates that the time-frequency resources configured by the first configuration information are used for the retraining of the AI ​​model, that is, no longer used for the transmission of business data; the second identifier is also called the retraining identifier; this can more clearly indicate the time-frequency resources used for the retraining of the AI ​​model;

[0062] (4) Modulation and Coding Scheme (MCS) information of the first bitstream; for example, the modulation scheme indicated by the MCS information includes, but is not limited to, 16-QAM. This makes it easier for the transceiver to obtain the corresponding MCS information of the training data, thereby accurately realizing the transmission and reception of the training data.

[0063] Optionally, the first identifier can be represented using a 1-bit flag to save signaling overhead.

[0064] Optionally, the second identifier can be represented using a 1-bit flag to save signaling overhead.

[0065] It should be noted that the first configuration information can be carried by one or more signaling methods. For example, the network-side device can send the first configuration information to the terminal via RRC signaling or Downlink Control Information (DCI). Alternatively, the network-side device can send the first identifier to the terminal via RRC signaling, and when the first identifier indicates that the training data labels are generated using the same sequence generation algorithm as the pilot sequence generation algorithm, it can send at least one of (2) to (4) above to the terminal via DCI (e.g., DCI 0-1).

[0066] Optionally, if the receiving end is a terminal and the sending end is a network-side device, the model training method in this embodiment may further include:

[0067] The terminal receives second configuration information sent by the network-side device; wherein the second configuration information may include, but is not limited to, at least one of the following:

[0068] (a) A third identifier, which is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; in this way, when the retraining of the AI ​​model is started, the third identifier can be used to indicate that the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels, that is, to indicate that the pilot sequence generation algorithm is used to generate training data labels, thereby reusing the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream at the transmitting end as the labels of the training data, thereby reducing the retraining overhead of the AI ​​model at the receiving end;

[0069] (b) Configuration information for time-frequency resources used to receive the training data; thus, with the aid of this configuration information, time-frequency resources used to receive the training data can be obtained, thereby enabling the reception of training data based on these time-frequency resources.

[0070] (c) A fourth identifier, which indicates that the time-frequency resources configured by the second configuration information are used for the retraining of the AI ​​model, i.e., no longer used for the transmission of business data; the second identifier is also called the retraining identifier; this more clearly indicates the time-frequency resources used for the retraining of the AI ​​model;

[0071] (d) The MCS information of the first bitstream; for example, the modulation scheme indicated by the MCS information includes, but is not limited to, 16-QAM. This facilitates the transceiver in obtaining the corresponding MCS information of the training data, thereby accurately realizing the transmission and reception of the training data.

[0072] Optionally, the third identifier can be represented using a 1-bit flag to save signaling overhead.

[0073] Optionally, the fourth identifier can be represented using a 1-bit flag to save signaling overhead.

[0074] It should be noted that the second configuration information can be carried by one or more signaling methods. For example, the network-side device can send the second configuration information to the terminal via RRC signaling or DCI. Alternatively, the network-side device can send the third identifier to the terminal via RRC signaling, and when the third identifier indicates that the training data labels are generated using the same sequence generation algorithm as the pilot sequence generation algorithm, it can send at least one of the above (b) to (d) to the terminal via DCI (e.g., DCI 0-1).

[0075] It should be noted that, taking DMRS pilots as an example, if the sequence generation algorithm for DMRS and the sequence generation algorithm for training data labels use the same algorithm (such as the Gold sequence generation algorithm), the difference in the generation process is as follows: after the 0-1 sequence of DMRS is generated, it is modulated using a fixed Quadrature Phase Shift Keying (QPSK); while after the training data labels are generated, the generated 0-1 sequence needs to be segmented according to the modulation order of the MSC and modulated according to the MCS modulation order set on the network side. This difference is detailed below.

[0076] Taking the DMRS that uses Gold sequences to generate the Physical Uplink Shared Channel (PUSCH) as an example, two sequences were confirmed: sequence x1 and sequence x2. Sequence x2 is generated first, followed by c. init Then according to c init Generate an x² sequence. The initial values ​​of the sequence are as follows:

[0077] The sequence x1 is: x1(0) = 1, x1(n) = 0, n = 1, 2, ..., 30;

[0078] x2 sequence:

[0079]

[0080] Where l is the OFDM symbol number within the time slot, It is the slot number within the frame. It is the number of symbols in each time slot. It is configured by high-level parameters. Equal to 0 or λ (as defined by the CDM group), Determined by higher-level parameters; mod represents the remainder sign. In summary, both the initial x1 sequence and the initial x2 sequence are 0-1 sequences with a length of 31.

[0081] Starting from sequence n=31, we have the following recursive sequence:

[0082] x1(n+31)=(x1(n+3)+x1(n))mod 2; n=0,1,…

[0083] x2(n+31)=(x2(n+3)+x2(n+2)+x2(n+1)+x2(n))mod 2

[0084] After generating sequences from x1 and x2, if the segmentation value N is taken... C =1600, then the calculated sequence is as follows:

[0085] c(n)=(x1(n+N C )+x2(n+N C ))mod 2

[0086] Here, c(n) is 0 or 1, serving as the bitstream of training data labels. No modulation is required; the sequence corresponding to c(n) can be taken directly, where n = 0, 1, ...

[0087] For DMRS, c(2n) and c(2n+1) need to be modulated in pairs to generate the modulated signal r(n):

[0088]

[0089] Network-side devices, such as base stations, send DCI (Digital Information Chaining) to terminals, carrying configuration information for PUSCH time-frequency resources. Using one slot and one resource block (RB), assuming 16-QAM modulation is used for the training data labels, each symbol carries 4 bits of information; while if QPSK modulation is used for DMRS, each symbol carries 2 bits of information. The number of bits required to generate the Gold sequence does not exceed 14 * 12 * 4 = 672 bits.

[0090] For example, considering a slot and an RB, assuming the number of layers is 1, the number of antenna ports is 4, DMRS ports 0 and 1 (P0 / P1DMRS) use code division multiplexing, and DMRS ports 2 and 3 (P2 / P3) use code division multiplexing, then as follows Figure 4 As shown, since one column in the time-frequency resources is used for mapping pilots and the other columns are used for mapping data, and the modulation order of 16-QAM is 4 and the modulation order of QPSK is 2, if there are no other configurations in this slot and RB, the transmitter needs to generate a total of 13*12*4+1*12*2=648 bits, where “13*12*4” represents the number of bits for transmitting data and “1*12*2” represents the number of bits for transmitting pilots.

[0091] For example, regarding the time-frequency resources configured for a base station, the 0 / 1 bitstream used as the training data label (corresponding to...) Figure 4 The first column in the table) and DMRS (corresponding to Figure 4 The mapping of the distribution of column 3 in the time-frequency resources can be as follows: Figure 4 As shown. It should be noted that the 0 / 1 bitstream used as training data labels needs to be modulated first and then mapped onto... Figure 4 The resource cell corresponding to the first column.

[0092] Optionally, based on the above c init The formula, if l=0, n = 24, Then we can get c init =7471124; For the first symbol (l=0) in the time slot corresponding to the configured time and frequency resources, the following training data label bit stream c(n) can be generated according to the gold sequence, which contains 48 bits: 011010000011010001100111100100101010101011100100.

[0093] If the bitstream c(n) above is modulated using 16-QAM, it can be grouped into sets of four, namely (0110)(1000)(0011)(0100)(0110)(0111)(1001)(0010)(1010)(1010)(1110)(0100). Furthermore, when mapping to time-frequency resources, the first symbol of the time slot can be filled from bottom to top, such as... Figure 4 As shown in column 1. Similarly, bit sequences for other time slots are generated.

[0094] It should be noted that both the network side and the terminal can generate training labels c(n) based on the above sequence generation algorithm and configuration information, effectively reducing data acquisition overhead. The above sequence generation algorithm takes the generation of uplink DMRS using Gold sequences as an example, but this application is not limited to this. This application can also use Gold sequences to generate downlink pilots, or use other sequence forms to generate uplink or downlink pilots.

[0095] The following is combined Figure 5 This application will be described in detail.

[0096] like Figure 5 As shown, taking an AI-based UL receiver as an example, the specific communication process includes:

[0097] S1: The wireless access network AI model training system (such as the CU and / or DU deployed on the network side, or deployed on a logical entity across CUs) pre-collects the data required for model training and trains the AI ​​model D in the UL receiver offline. The AI ​​model D is synchronized to the network side. The AI ​​model D is used to implement DMRS channel estimation, channel interpolation, channel equalization and demodulation, etc.

[0098] S2: The terminal accesses the network and enters the RRC-CONNECTED state;

[0099] S3: The network side sends the DMRS-UplinkConfig parameter set to the terminal via RRC signaling to configure the terminal demodulation reference signal generation sequence (such as Gold sequence, Low-PAPR sequence, etc. for PUSCH reference signal sequence), and configures the initial DMRS sequence parameters via DCI 0-1;

[0100] S4: The terminal generates a DMRS based on the configuration information and sends the modulated data stream and DMRS to the network side;

[0101] S5: The network side uses an AI receiver for demodulation. If the AI ​​model D cannot meet the system requirements (such as a high bit error rate for a long time), it switches to the traditional demodulation process and starts retraining of the AI ​​model D.

[0102] S6: The network side sends a first identifier to the terminal via RRC signaling to indicate whether to use the same sequence generation algorithm as the pilot sequence generation algorithm to generate training data labels, i.e., whether to use the pilot sequence generation algorithm to generate training data labels; if the same sequence generation algorithm is used, the network side can send configuration information via DCI, such as time and frequency resource configuration information required for retraining, retraining identifier (which can be identified by a 1-bit flag, indicating that the configured time and frequency resources are used for retraining and no longer for the transmission of service data), and MCS information;

[0103] S7: The terminal generates a bit stream by multiplexing the sequence generation algorithm of the demodulation reference signal (DMRS) in the PUSCH for uplink transmission according to the configuration information, modulates the bit stream, inserts DMRS pilot symbols, maps the modulated bit stream and pilot symbols to the corresponding time and frequency resources respectively, and sends them to the network side.

[0104] S8: The network side receives symbol-level data (i.e., training data) on the corresponding time-frequency resources from the terminal, and generates training data labels and pilot symbols according to the sequence generation algorithm for generating DMRS and the configuration information of the pilot reference signal generation sequence. This received training data is not submitted to higher layers. Afterwards, the network side obtains the channel estimate at the pilot based on the pilot symbols, and performs supervised training on AI model D based on the received training data, the corresponding training data labels, the pilot symbols of the demodulation reference signal, and the channel estimate at the pilot, until the trained AI model D meets the system requirements after testing, and then switches to using the retrained AI model D for demodulation processing.

[0105] Please see Figure 6 , Figure 6 This is a flowchart of a model training method provided in an embodiment of this application. This method is applied at the receiving end, such as... Figure 6 As shown, the method includes the following steps:

[0106] Step 61: The transmitting end uses a pilot sequence generation algorithm to generate the first bit stream;

[0107] Step 62: The transmitting end modulates the first bit stream to obtain training data;

[0108] Step 63: The transmitting end sends the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the received training data and the training data labels generated by the pilot sequence generation algorithm.

[0109] In this embodiment, the receiving end can be a network-side device or a terminal, and correspondingly, the sending end can be a terminal or a network-side device. For example, in an uplink transmission scenario, the receiving end is a network-side device and the sending end is a terminal; or, in a downlink transmission scenario, the receiving end is a terminal and the sending end is a network-side device. In other words, the solution in this application can be applied to the retraining of AI models in both uplink (UL) and downlink (DL) receiving ends.

[0110] The pilot sequence generation algorithm is specifically a pilot sequence generation algorithm, such as a DMRS generation algorithm. The pilot sequence may be, for example, a Gold sequence. The aforementioned modulation processing of the first bitstream may include modulating the first bitstream, inserting pilot symbols, and mapping the modulated bitstream and pilot symbols to corresponding time-frequency resources, etc.

[0111] The training data label is specifically the label corresponding to the training data received by the receiving end, such as a 01 bit stream generated by the receiving end using a pilot sequence generation algorithm.

[0112] Optionally, the AI ​​model is specifically used for at least one of the following: channel estimation, channel interpolation, channel equalization, demodulation, and decoding using pilot signals (such as DMRS). This application does not limit the specific implementation methods of channel estimation, channel interpolation, channel equalization, demodulation, and decoding, which can be determined based on the actual situation.

[0113] For example, the AI ​​model can be used to implement DMRS channel estimation, channel interpolation, channel equalization, demodulation, and decoding. That is, the aforementioned demodulation processing can include DMRS channel estimation, channel interpolation, channel equalization, and demodulation. Figure 3 As shown. The retraining of the AI ​​model can be performed using supervised learning.

[0114] Optionally, the AI ​​model can be pre-trained offline by a wireless access network AI model training system. This AI model training system can be deployed on the CU and / or DU on the network side, or on a logical entity across CUs on the network side; there is no limitation on this.

[0115] Retraining of the AI ​​model in the receiving end can be triggered when the performance of the AI ​​model deteriorates significantly due to scenarios such as large channel differences between the training scenario and the application scenario, or large nonlinear differences between the transmitting and receiving ends.

[0116] According to the solution of this application embodiment, when retraining the AI ​​model in the receiving end, the receiving end can reuse the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream of the sending end as the training data label, thereby reducing the data acquisition overhead and thus reducing the retraining overhead of the AI ​​model in the receiving end.

[0117] Optionally, if the receiving end is a network-side device and the sending end is a terminal, the model training method in this embodiment may further include:

[0118] The terminal receives first configuration information sent by the network-side device; wherein, the first configuration information may include, but is not limited to, at least one of the following:

[0119] (1) First identifier, the first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; in this way, when the network-side device starts the retraining of the AI ​​model, the first identifier can indicate that the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels, that is, to indicate that the pilot sequence generation algorithm is used to generate training data labels, thereby reusing the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream of the sending end as the label of the training data, thereby reducing the retraining overhead of the AI ​​model in the receiving end;

[0120] (2) Configuration information of time-frequency resources for sending the training data; In this way, with the help of this configuration information, time-frequency resources for sending training data can be obtained, and the transmission of training data can be realized based on these time-frequency resources;

[0121] (3) Second identifier, the second identifier indicates that the time-frequency resources configured by the first configuration information are used for the retraining of the AI ​​model, that is, no longer used for the transmission of business data; the second identifier is also called the retraining identifier; this can more clearly indicate the time-frequency resources used for the retraining of the AI ​​model;

[0122] (4) Modulation and Coding Scheme (MCS) information of the first bitstream; for example, the modulation scheme indicated by the MCS information includes, but is not limited to, 16-QAM. This makes it easier for the transceiver to obtain the corresponding MCS information of the training data, thereby accurately realizing the transmission and reception of the training data.

[0123] Optionally, the first identifier can be represented using a 1-bit flag to save signaling overhead.

[0124] Optionally, the second identifier can be represented using a 1-bit flag to save signaling overhead.

[0125] Optionally, if the receiving end is a terminal and the sending end is a network-side device, the model training method in this embodiment may further include:

[0126] The network-side device sends second configuration information to the terminal; wherein the second configuration information includes at least one of the following:

[0127] (a) A third identifier, which is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; in this way, when the retraining of the AI ​​model is started, the third identifier can be used to indicate that the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels, that is, to indicate that the pilot sequence generation algorithm is used to generate training data labels, thereby reusing the pilot sequence generation algorithm to generate training data labels, without having to collect the bit stream at the transmitting end as the labels of the training data, thereby reducing the retraining overhead of the AI ​​model at the receiving end;

[0128] (b) Configuration information for time-frequency resources used to receive the training data; with the help of this configuration information, time-frequency resources used to receive the training data can be obtained, thereby enabling the reception of training data based on these time-frequency resources.

[0129] (c) A fourth identifier, which indicates that the time-frequency resources configured by the second configuration information are used for the retraining of the AI ​​model, i.e., no longer used for the transmission of business data; the second identifier is also called the retraining identifier; this more clearly indicates the time-frequency resources used for the retraining of the AI ​​model;

[0130] (d) The MCS information of the first bitstream; for example, the modulation scheme indicated by the MCS information includes, but is not limited to, 16-QAM. This facilitates the transceiver in obtaining the corresponding MCS information of the training data, thereby accurately realizing the transmission and reception of the training data.

[0131] Optionally, the third identifier can be represented using a 1-bit flag to save signaling overhead.

[0132] Optionally, the fourth identifier can be represented using a 1-bit flag to save signaling overhead.

[0133] It should be noted that the model training method provided in this application embodiment can be executed by a model training device or a control module within that model training device for executing the model training method. This application embodiment uses the execution of the model training method by a model training device as an example to illustrate the model training device provided in this application embodiment.

[0134] Please see Figure 7 , Figure 7 This is a schematic diagram of a model training device provided in an embodiment of this application. The device is applied at the receiving end, such as... Figure 7 As shown, the model training device 70 includes:

[0135] The acquisition module 71 is used to acquire training data transmitted through the channel, wherein the training data is obtained by modulating the first bit stream generated by the pilot sequence generation algorithm at the transmitting end.

[0136] The first generation module 72 is used to generate training data labels using the pilot sequence generation algorithm;

[0137] The training module 73 is used to retrain the artificial intelligence AI model used for demodulation processing based on the training data and the training data labels.

[0138] Optionally, the AI ​​model is specifically used for at least one of the following:

[0139] Channel estimation using pilot signals;

[0140] Channel interpolation;

[0141] Channel equalization;

[0142] demodulation;

[0143] decoding.

[0144] Optionally, if the receiving end is a network-side device and the sending end is a terminal, the model training device 70 further includes:

[0145] A first sending module is configured to send first configuration information to the terminal; wherein the first configuration information includes at least one of the following:

[0146] The first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels;

[0147] Configuration information for time-frequency resources used to send the training data;

[0148] The second identifier indicates that the time-frequency resources configured in the first configuration information are used for the retraining of the AI ​​model;

[0149] The modulation and coding scheme (MCS) information of the first bit stream.

[0150] Optionally, the first identifier is represented using a 1-bit flag; and / or, the second identifier is represented using a 1-bit flag.

[0151] Optionally, if the receiving end is a terminal and the sending end is a network-side device, the model training device 70 further includes:

[0152] A first receiving module is configured to receive second configuration information sent by the network-side device; wherein the second configuration information includes at least one of the following:

[0153] The third identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels;

[0154] Configuration information for time-frequency resources used to receive the training data;

[0155] The fourth identifier indicates that the time-frequency resources configured in the second configuration information are used for the retraining of the AI ​​model;

[0156] The MCS information of the first bit stream.

[0157] Optionally, the third identifier is represented using a 1-bit flag; and / or, the fourth identifier is represented using a 1-bit flag.

[0158] Optionally, the model training device 70 also includes:

[0159] The startup module is used to initiate the retraining of the AI ​​model when the AI ​​model cannot meet the requirements of demodulation processing.

[0160] Optionally, the first generation module 72 is specifically used to: generate a second bit stream according to the pilot sequence generation algorithm and the configured pilot sequence initial parameters, and use the second bit stream as the training data label.

[0161] The model training device 70 of this application embodiment can achieve the above-mentioned... Figure 1 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0162] Please see Figure 8 , Figure 8 This is a schematic diagram of a model training device provided in an embodiment of this application. The device is applied at the transmitting end, such as... Figure 8 As shown, the model training device 80 includes:

[0163] The second generation module 81 is used to generate the first bit stream using a pilot sequence generation algorithm;

[0164] Processing module 82 is used to modulate the first bit stream to obtain training data;

[0165] The second transmitting module 83 is used to transmit the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the received training data and the training data labels generated by the pilot sequence generation algorithm.

[0166] Optionally, the AI ​​model is specifically used for at least one of the following:

[0167] Channel estimation using pilot signals;

[0168] Channel interpolation;

[0169] Channel equalization;

[0170] demodulation;

[0171] decoding.

[0172] Optionally, if the receiving end is a network-side device and the sending end is a terminal, the model training device 80 further includes:

[0173] The second receiving module is configured to receive first configuration information sent by the network-side device; wherein the first configuration information includes at least one of the following:

[0174] The first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels;

[0175] Configuration information for time-frequency resources used to send the training data;

[0176] The second identifier indicates that the time-frequency resources configured in the first configuration information are used for the retraining of the AI ​​model;

[0177] The MCS information of the first bit stream.

[0178] Optionally, the first identifier is represented using a 1-bit flag; and / or, the second identifier is represented using a 1-bit flag.

[0179] Optionally, if the receiving end is a terminal and the sending end is a network-side device, the model training device 80 further includes:

[0180] The third sending module is used to send second configuration information to the terminal; wherein the second configuration information includes at least one of the following:

[0181] The third identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels;

[0182] Configuration information for time-frequency resources used to receive the training data;

[0183] The fourth identifier indicates that the time-frequency resources configured in the second configuration information are used for the retraining of the AI ​​model;

[0184] The MCS information of the first bit stream.

[0185] The model training device 80 of this application embodiment can achieve the above-mentioned... Figure 6 The various processes of the method embodiments shown can achieve the same technical effect, and will not be described again here to avoid repetition.

[0186] Optional, such as Figure 9As shown, this application embodiment also provides a communication device 90, including a processor 91, a memory 92, and a program or instructions stored in the memory 92 and executable on the processor 91. When the program or instructions are executed by the processor 91, they implement the various processes of the above-described model training method embodiment and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0187] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they can implement the various processes of the above-described model training method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0188] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they can implement the various processes of the above-described model training method embodiments and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0189] Computer-readable media include both permanent and non-permanent, removable and non-removable media, which can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0190] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0191] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0192] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a service classification device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0193] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A model training method, characterized in that, include: The receiving end acquires training data transmitted through the channel, which is obtained by modulating the first bit stream generated by the sending end using the pilot sequence generation algorithm; The receiving end uses the pilot sequence generation algorithm to generate training data labels; The receiving end retrains the artificial intelligence (AI) model used for demodulation processing based on the training data and the training data labels.

2. The method according to claim 1, characterized in that, The AI ​​model is specifically used for at least one of the following: Channel estimation using pilot signals; Channel interpolation; Channel equalization; demodulation; decoding.

3. The method according to claim 1, characterized in that, If the receiving end is a network-side device and the sending end is a terminal, the method further includes: The network-side device sends first configuration information to the terminal; The first configuration information includes at least one of the following: The first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; Configuration information for time-frequency resources used to send the training data; The second identifier indicates that the time-frequency resources configured in the first configuration information are used for the retraining of the AI ​​model; The modulation and coding scheme (MCS) information of the first bit stream.

4. The method according to claim 3, characterized in that, The first identifier is represented using a 1-bit flag; and / or, the second identifier is represented using a 1-bit flag.

5. The method according to claim 1, characterized in that, If the receiving end is a terminal and the sending end is a network-side device, the method further includes: The terminal receives the second configuration information sent by the network-side device; The second configuration information includes at least one of the following: The third identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; Configuration information for time-frequency resources used to receive the training data; The fourth identifier indicates that the time-frequency resources configured in the second configuration information are used for the retraining of the AI ​​model; The MCS information of the first bit stream.

6. The method according to claim 1, characterized in that, Before the receiving end acquires the training data transmitted through the channel, the method further includes: When the AI ​​model cannot meet the demodulation processing requirements, the receiving end initiates retraining of the AI ​​model.

7. The method according to claim 1, characterized in that, The receiving end uses the pilot sequence generation algorithm to generate training data labels, including: The receiving end generates a second bit stream based on the pilot sequence generation algorithm and the configured initial parameters of the pilot sequence, and uses the second bit stream as the training data label.

8. A model training method, characterized in that, include: The transmitting end uses a pilot sequence generation algorithm to generate the first bit stream; The transmitting end modulates the first bit stream to obtain training data; The transmitting end sends the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the received training data and the training data labels generated by the pilot sequence generation algorithm.

9. The method according to claim 8, characterized in that, The AI ​​model is specifically used for at least one of the following: Channel estimation using pilot signals; Channel interpolation; Channel equalization; demodulation; decoding.

10. The method according to claim 8, characterized in that, If the receiving end is a network-side device and the sending end is a terminal, the method further includes: The terminal receives the first configuration information sent by the network-side device; The first configuration information includes at least one of the following: The first identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; Configuration information for time-frequency resources used to send the training data; The second identifier indicates that the time-frequency resources configured in the first configuration information are used for the retraining of the AI ​​model; The MCS information of the first bit stream.

11. The method according to claim 8, characterized in that, If the receiving end is a terminal and the sending end is a network-side device, the method further includes: The network-side device sends second configuration information to the terminal; The second configuration information includes at least one of the following: The third identifier is used to indicate whether the same sequence generation algorithm as the pilot sequence generation algorithm is used to generate training data labels; Configuration information for time-frequency resources used to receive the training data; The fourth identifier indicates that the time-frequency resources configured in the second configuration information are used for the retraining of the AI ​​model; The MCS information of the first bit stream.

12. A model training device, characterized in that, include: The acquisition module is used to acquire training data transmitted through the channel. The training data is obtained by modulating the first bit stream generated by the pilot sequence generation algorithm at the transmitting end. The first generation module is used to generate training data labels using the pilot sequence generation algorithm; The training module is used to retrain the artificial intelligence (AI) model used for demodulation processing based on the training data and the training data labels.

13. A model training device, characterized in that, include: The second generation module is used to generate the first bit stream using a pilot sequence generation algorithm; The processing module is used to modulate the first bit stream to obtain training data; The second transmitting module is used to transmit the training data to the receiving end through the channel, and the receiving end retrains the AI ​​model used for demodulation processing based on the received training data and the training data labels generated by the pilot sequence generation algorithm.

14. A communication device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as claimed in any one of claims 1 to 7, or the steps of the method as claimed in any one of claims 8 to 11.

15. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 7, or the steps of the method as described in any one of claims 8 to 11.

16. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as claimed in any one of claims 1 to 7, or the steps of the method as claimed in any one of claims 8 to 11.

Images