A language-guided i / q characterization method and system for wireless signal processing

Abstract

The application belongs to the technical field of wireless communication, and particularly relates to a language-guided I / Q characterization method and system for wireless signal processing. The data structure of the method comprises customized expert instance-level description text. For input data, the instance-level instruction text is encoded into an instruction vector by using a first text encoder. A lightweight controller is used to map the instruction vector into a control signal. The instance-level description text can interact with a one-dimensional encoder in the model training and inference stages. A contrast learning mechanism based on the DPO idea is used to complete accurate and stable semantic injection, and fine and full-process language guidance is achieved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, specifically to a language-guided I / Q representation method and system for wireless signal processing. Background Technology

[0002] Deep learning, which operates solely on raw in-phase and quadrature (I / Q) data, has become the dominant end-to-end paradigm in wireless signal processing, supporting tasks such as automatic modulation identification and RF fingerprinting. Under controlled settings, neural encoders can achieve good accuracy through efficient training pipelines. However, field deployments have exposed a persistent gap between algorithmic expectations and real-world performance: systems not only require large amounts of old data and offer little leverage for operator control, but also tend to degrade when operating conditions differ from those seen during training. These limitations have inspired a language-guided I / Q representation learning approach that allows readily available context to permeate the representation itself, rather than leaving it outside the target.

[0003] In actual workflows, practitioners possess background knowledge regarding data acquisition, equipment, and scenarios, such as acquisition time, transmitter identity, scenario labels, acquisition specifications, and labeling policies. However, mainstream models typically only accept I / Q samples and discrete labels, excluding this crucial information from the representation. This leads to three specific consequences: first, the model requires more data to achieve the target accuracy; second, the control interface provided to the operator is limited; and third, the interpretability across tasks and environments is weak. Therefore, robustness under different operating conditions and cross-domain transfers remains elusive.

[0004] In addition to raw I / Q data, most collection pipelines also record textual information, such as collection time and transmitter identifiers. Traditional methods rely solely on signals and ignore this text, thus limiting their ability to encode deployment context and adapt to distribution changes. Omitting context introduces two drawbacks. First, supervisory signals are often compressed into category labels, while sample-by-sample explanatory text expressing intent and requirements is rarely or completely ignored, effectively removing instance-level guidance from training to inference. Second, model performance degrades under varying conditions caused by channel variations, hardware mismatches, interference patterns, and data acquisition practices. This behavior has been documented in cross-sky RF fingerprinting and Wi-Fi sensing, driving research into domain-specific, principle-based evaluation and comparison protocols. Discarding additional textual information, such as collection time and transmitter identifiers, exacerbates these two problems.

[0005] Recent research on wireless language interfaces, including MMCLIP, WiTalk, and WiChat, has provided a useful foundation, but they typically retain textual information at the class level or provide application cues only after feature extraction. This design offers limited controllability within the encoder and blurs the distinction between runtime guidance and training-only knowledge. Similar problems and shortcomings persist in patents related to I / Q data representation, RF fingerprinting, and automatic modulation recognition in the field of wireless signal processing.

[0006] While existing patents have made numerous attempts in areas such as hardware calibration, feature extraction, and domain-adaptive optimization, they have failed to fundamentally address core pain points in practical deployments, including the model's high sensitivity to operational conditions, lack of controllable interfaces, and insufficient utilization of textual context. Furthermore, their exploration of cross-modal semantic fusion technologies has significant limitations. For example, patents related to RF fingerprint recognition and automatic modulation recognition, even when incorporating strategies such as domain adaptation and physical damage-based data augmentation to improve cross-domain performance, still only perform feature extraction and network architecture optimization on the original I / Q sequences, completely ignoring the prior value of textual context such as acquisition time and transmitter identity, leaving this crucial information as external annotations to the model. This not only leads to low model data efficiency but also fails to leverage textual priors to predict and mitigate distribution bias.

[0007] The very few wireless communication patents that attempt to integrate linguistic information merely use text as a category-level supervisory label to construct a signal-text shared embedding space, or perform simple text calibration after feature extraction. They fail to achieve instance-level text-based, layer-by-layer dynamic guidance of the encoder, nor do they construct a precise semantic injection mechanism. Such designs blur the distinction between runtime guidance and training-level knowledge, resulting in severely insufficient controllability within the encoder. While some patents optimizing inference efficiency reduce inference costs from the perspectives of resource scheduling and low-power network architecture, they do not integrate this with signal representation learning. Introducing external knowledge to improve performance still increases sequence length or computational complexity due to operations such as token concatenation and cross-modal attention, failing to achieve a balance between performance improvement and deployment efficiency.

[0008] Furthermore, the existing patented model designs provide operators with rather limited control interfaces, only accepting I / Q samples and discrete category labels as input. The reasoning process lacks flexible adjustment methods, the interpretation path across tasks and environments is weak, and it is difficult to adapt to the dynamic changes in actual data collection scenarios. Summary of the Invention

[0009] The purpose of this invention is to provide a language-guided I / Q representation method and system for wireless signal processing, so as to solve the problems of low data efficiency, lack of human-computer interaction control, and poor cross-domain robustness of existing wireless signal processing methods.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A language-guided I / Q representation method for wireless signal processing includes the following steps:

[0012] S1, Constructing the dataset: The dataset includes a training set and a test set, defined as follows:

[0013] ;

[0014] ;

[0015] in, Represents the I / Q time series; Indicates category label; Indicates instance-level instruction text; This indicates expert tips text that are only available during the training phase; This label represents the data collection conditions and can be either optional or empty.

[0016] S2, Instruction Encoding and Control Signal Generation: Using the first text encoder instance-level instruction text Encoded as instruction vector Expert suggestion for encoding: Use a second text encoder. Expert text Encode as feature vector Use a lightweight controller Will Mapped to control signals ;

[0017] S3, ICM modulation: In the control signal Guided by a one-dimensional encoder I / Q time series Mapping to representation vector During this process, control signals A affine modulation method applied to a one-dimensional encoder The intermediate activation values ​​of the feedforward network (FFN) for each Transformer block;

[0018] S4, Classification prediction output: represented by a vector For input, use the prediction head Prediction is performed to obtain classification results. ;

[0019] S5, Training Phase: Includes preference alignment and joint optimization steps, where:

[0020] S51, Preference Alignment: Preference alignment is an alignment method based on the Direct Preference Optimization (DPO) concept, namely DPO-SI preference alignment. The iterative optimization strategy model is defined as the student model, and the frozen reference baseline model is defined as the expert benchmark model. Its implementation method is based on feature vectors. Construct positive and negative sample pairs and calculate the DPO preference loss; where the positive samples are the preference samples. Negative samples are unbiased samples. DPO preference loss Used to measure the alignment between student representations and preference target representations built based on expert prompts;

[0021] S52, Joint Optimization: Constructing a joint objective function and using it on the text encoder. Lightweight controller One-dimensional encoder Predicting Head Text encoder And the gradient update of the parameters of the projection layer that constructs positive and negative sample pairs in step S51;

[0022] S6, Inference Phase: Remove the text encoder The module on which positive and negative sample pairs are constructed and the DPO preference loss is calculated; using the text encoder jointly optimized in step S52. Lightweight controller One-dimensional encoder and prediction head Taking the I / Q time series x and instance-level instruction text i as input, the classification results are output according to steps S2 to S4, without the need for expert prompt text.

[0023] Furthermore, the ICM modulation includes:

[0024] I / Q time series The sequence is converted into a token sequence using a one-dimensional patch embedder with stride, and class tokens are pre-appended to the head of the sequence.

[0025] In multiple Transformer blocks, for the feedforward network of each block, the control signals are generated by independently setting a FiLM generator for each layer of the network. Mapped to scaling factor and offset factor ;

[0026] Feedforward Network (FFN) consists of two linear transformation layers and an intermediate nonlinear activation layer. The input is subjected to dimensionality upscaling within the first linear transformation layer to obtain the intermediate activation value. For intermediate activation values The formula for performing element-wise affine modulation is as follows:

[0027] ;

[0028] in, The element-wise product of Hadamard, the modulated activation value After dimensionality reduction through the second-level linear transformation, the feature fusion is completed by connecting it with the original input residual.

[0029] After stacking L Transformer blocks, the resulting features are normalized in the final layer to output a token representation. .

[0030] Furthermore, in the DPO-SI preference alignment, based on feature vectors... The method for constructing positive and negative sample pairs is as follows:

[0031] The feature vector generated by the prompt mapper from the expert prompts corresponding to the current sample is used as the preference sample. ;

[0032] The feature vectors generated by the prompt mapper from expert prompts for other samples with different device IDs within the same data batch are used as non-preference samples. ;

[0033] During loss calculation, the expert-hinted feature vectors in the same data batch that have the same device identifier as the current sample but are different samples are masked to prevent them from participating in the close constraint and push-away constraint.

[0034] Furthermore, the DPO preference loss The calculation formula is:

[0035] ;

[0036] in, Indicates feature similarity calculation; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; This represents the set of unbiased samples.

[0037] Furthermore, the first text encoder Second text encoder Both are lightweight sentence encoders and text encoders. It is used in conjunction with preference alignment during the training phase and completely disabled during the inference phase. The lightweight sentence encoder is a sentence encoder with fewer than 100 million parameters.

[0038] Furthermore, the data used to construct the dataset all come from the open-source I / Q-text pairing corpus RF-Lang v1.1.

[0039] A language-guided I / Q representation system for wireless signal processing includes: a data construction module, an instruction encoding and controller module, an ICM encoder module, a prediction head module, an expert prompt encoding module, a preference alignment and joint optimization module, and an inference control module; the above modules implement a language-guided I / Q representation method for wireless signal processing.

[0040] By adopting the above technical solution, the present invention has the following beneficial effects:

[0041] 1. This invention proposes a language-guided I / Q representation system (LinguIQ), whose data structure includes customized expert instance-level descriptive text. For input data, a first text encoder is used. instance-level instruction text Encoded as instruction vector Use a lightweight controller Will Mapped to control signals It enables instance-level descriptive text to interact with the one-dimensional encoder during both model training and inference phases; and combines the contrastive learning mechanism based on the DPO concept to achieve accurate and stable semantic injection, realizing refined, end-to-end language guidance.

[0042] 2. The I / Q representation system of this invention can flexibly start and stop the control path without changing the tokenization process, sequence length and attention overhead, which is convenient for ablation experiments. It completes the layer-by-layer modulation of the one-dimensional Transformer (1-D Transformer) according to the instruction text. It is additionally configured with DPO-SI semantic injection which is only effective in the training phase. Based on contrastive learning, a loss function is constructed to optimize semantic alignment performance without increasing inference cost. This design is based on privileged information learning, teacher-student learning and DPO optimization theory.

[0043] 3. FiLM-related research demonstrates that lightweight controllers can generate affine transform modulation internal features and guide model computation; this type of conditional adjustment scheme is also suitable for handling one-dimensional models with long-range time-dependent dependencies. Therefore, this invention will control the signal... A affine modulation method applied to a one-dimensional encoder The intermediate activation values ​​of the feedforward network (FFN) in each Transformer block; compared with text token concatenation and cross-modal attention fusion, the feature modulation scheme does not bring sequence redundancy and computational burden, and has obvious advantages in semantic regulation and running efficiency.

[0044] Instruction manual illustrations

[0045] Figure 1 This is a flowchart of the language-guided I / Q representation system for wireless signal processing proposed in this invention. Detailed Implementation

[0046] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0047] Language supervision has been applied to improve the performance of wireless measurement data normalization and recognition. mmCLIP aligns millimeter-wave data with text embeddings to optimize human activity recognition; WiChat combines WiFi channel information to conduct research on large-scale prompting strategies; Zhang Zhenxi et al. improved modulation recognition generalization ability by aligning I / Q signals with text in a multimodal manner and introducing semantic priors. Related research on cross-modal alignment has also confirmed that language can constrain heterogeneous signal structures in a shared feature space, and existing schemes have achieved phono-image alignment, phono-text matching, and joint learning of phono-image-text three modalities. Currently, most wireless signal processing systems only use text for category-level supervision or introduce prompting learning after feature extraction, and semantic injection often uses simple distillation loss, resulting in poor cross-modal alignment performance.

[0048] In view of this, the present invention proposes a language-guided I / Q representation method for wireless signal processing, such as... Figure 1 As shown, it includes the following steps:

[0049] S1, Constructing the dataset: The dataset includes a training set and a test set, defined as follows:

[0050] ;

[0051] ;

[0052] in, Represents the I / Q time series; Indicates category label; Indicates instance-level instruction text; This indicates expert-level prompts that are available only during the training phase, and they complement instance-level instruction texts. This label represents the data collection conditions used to record hardware location, date, etc. It can be optional or empty.

[0053] To facilitate a clear understanding of instance-level instruction text and expert prompt text, a specific example is provided: The instruction text is "I need to know the correct ID to use for the entry of 3 / 1 / 2023" or "I need to know the correct ID used for entry on March 1, 2023". The corresponding expert prompt would be "The date is the 1st of March, 2023, and the transmitter is unit 0" or "The date is March 1, 2023, and the transmitter is device 0". The instruction text conveys the operator's intent to guide representation learning and remains usable during the testing phase; the expert prompt clarifies the context, such as date semantics and device identification, and is introduced only during the training phase to allow the model to internalize this information without increasing inference overhead.

[0054] In a specific example, the data used to construct the dataset was obtained from the open-source I / Q-text pairing corpus RF-Langv1.1. This corpus contains a two-channel I / Q sequence, category label, domain label, and two text fields for each sample. When used, the "qa" field in the corpus is mapped to instruction text, and the "Prompt" field is mapped to expert hints used only for training. In addition, the release of this corpus also includes dataset instruction cards, a license, and usage examples, in order to facilitate the above. Tuples can be precisely instantiated in practical applications.

[0055] S2, Instruction Encoding and Control Signal Generation: Using the first text encoder instance-level instruction text Encoded as instruction vector Using a second text encoder Expert text Encode as feature vector Use a lightweight controller Will Mapped to control signals .

[0056] S3, ICM modulation: In the control signal Guided by a one-dimensional encoder I / Q time series Mapping to representation vector During this process, control signals A affine modulation method applied to a one-dimensional encoder The intermediate activation values ​​of the feedforward network (FFN) in each Transformer block. Instruction-Controlled Modulation (ICM) can transform natural language guidance into compact control signals, enabling control of the one-dimensional encoder without altering the tokenization and attention topology. Full-depth control. Among them, the one-dimensional encoder... Specifically, a one-dimensional Transformer encoder architecture is adopted. The ICM modulation process includes:

[0057] (1) The instruction text is generated by a lightweight sentence encoder. (Parameters are) Embedding is performed to obtain a fixed-dimensional text vector. Subsequently, the focus shifted to the internet. (Parameters are) )Will Mapped to control vector Its dimensions It is independent of the I / Q sequence length, ensuring universality for signals of different lengths.

[0058] (2) Convert the original I / Q time series The sequence is converted into a token sequence using a one-dimensional patch embedder with stride (temporal patchsize=16), and a class token is pre-appended to the beginning of the sequence. Positional encodings are added to preserve temporal information. The self-attention sublayer remains unchanged to ensure the integrity of the sequence structure.

[0059] (3) In multiple Transformer blocks, for the feedforward network (FFN) of each Transformer block, an independent FiLM generator is set for each layer of the network. The FiLM generator consists of a single fully connected layer with the input dimension being the control vector dimension. The output dimension is set to Control signal Mapped to scaling factor and offset factor ;

[0060] A standard feedforward network (FFN) consists of two linear transformation layers and an intermediate nonlinear activation layer. The input is subjected to dimensionality upscaling within the first linear transformation layer to obtain the intermediate activation value. This embodiment is based on the FiLM modulation mechanism, performing element-wise affine modulation on the intermediate activation:

[0061] ;

[0062] in, The element-wise product of Hadamard, the modulated activation value After dimensionality reduction through the second-level linear transformation, the feature fusion is completed by connecting it with the original input residual.

[0063] After stacking L Transformer blocks, the resulting features are finally normalized (LN) to output a token representation. Token representation It is the only output of ICM, which is applied to the DPO preference alignment objective function in subsequent classifiers and training loops.

[0064] In practice, FiLM-style conditional tuning requires only a few additional parameters to achieve explicit and stable control (Perez et al, 2018) and is well compatible with Transformer architectures that process time-series I / Q data (Wen et al, 2023).

[0065] S4, Classification prediction output: represented by a vector For input, use the prediction head Prediction is performed to obtain classification results. ;

[0066] S5, Training Phase: This includes preference alignment and joint optimization steps, the details of which are as follows:

[0067] S51, Preference Alignment: This embodiment employs a Direct Preference Optimization Semantic Injection (DPO-SI) strategy, introducing expert knowledge from the training phase as privileged information to complete semantic injection during the representation learning process. In its specific implementation, a contrastive learning mechanism is constructed using Direct Preference Optimization (DPO). This mechanism only operates during the training process and does not modify the original computational pathway during the model inference phase. Its implementation methods include:

[0068] S511. Feature Vector Generation:

[0069] Define the iterative optimization strategy model as the student model, and the frozen reference baseline model as the expert benchmark model; for the expert benchmark model, use the parameters as... Lightweight text encoder Expert advice Encoding and embedding are performed to obtain fixed-dimensional text feature vectors. Then, a prompt mapper with parameters frozen throughout is used. Text feature vectors Mapping to the corresponding feature representation space yields aligned cross-space feature vectors. It should be noted that during this process, the mapper is prompted. parameters It does not participate in gradient updates throughout the entire training phase.

[0070] S512. Construct positive and negative sample pairs based on the generated feature vectors:

[0071] Positive sample definition: referring to preference samples Positive samples are feature vectors generated by the expert hints corresponding to the current sample through the hint mapper, and can be understood as query vectors. Corresponding preferred answers;

[0072] Negative sample definition: Non-preferred samples Negative samples are selected from other samples within the same batch whose device IDs are different. Their feature vectors, generated by the expert hint mapper, are used as the query vectors. Corresponding non-preference answers;

[0073] A masking rule is set based on positive and negative sample pairs: For expert hint feature vectors with the same device ID but different samples within the same batch, they belong to neither the preset positive sample pair nor the negative sample pair; therefore, they are masked during the loss calculation process to prevent them from participating in the gradient update of pulling closer or pushing apart, thus preventing gradient conflicts caused by such invalid matching and ensuring that the comparative loss is optimized only for valid positive and negative sample pairs.

[0074] For positive and negative sample pairs, a preference loss is constructed based on the DPO (Discrimination Point of Interest) concept. Preference loss The calculation formula is:

[0075] ;

[0076] in, Indicates feature similarity calculation, This is a temperature coefficient (used to adjust the smoothness of the similarity distribution). This represents the set of unpreferenced samples. Preference loss. Used for joint optimization with task loss during the training loop. During the inference phase, the DPO-SI module is completely shut down, incurring no additional computational overhead or memory usage.

[0077] S52, Joint Optimization:

[0078] This embodiment uses joint loss to complete model training. The overall optimization objective is a weighted combination of the supervised classification loss and the DPO preference alignment loss. The hyperparameter that balances the two losses is... Classification task loss Using cross-entropy loss Step S511 has completed the construction of positive and negative preference sample pairs and given the DPO preference loss LDPO expression. This loss is used to constrain the alignment between the sample class token representation and the expert preference target representation. In the previous matching logic, this class token representation was used as a query vector to participate in the preference measurement.

[0079] ICM modulation output The sample class token is represented and fed as input into the learnable linear prediction head. Generate category scores (logits) , This represents the total number of categories. The softmax function is defined as follows: ;

[0080] Let the unique heat label be Then the single-sample cross-entropy loss is:

[0081] ;

[0082] For a batch sample set B, the classification loss is weighted and fused with the DPO preference loss defined above to obtain the batch average joint loss:

[0083] ;

[0084] in, The DPO loss weights are used to balance the importance of classification tasks and preference alignment.

[0085] Token representation The gradient is the superposition of the discriminant term and the preference-driven term, that is:

[0086] ;

[0087] In the formula, The discriminative gradient obtained by backpropagation of cross-entropy loss is derived from the error between the classification prediction probability and the true one-hot label. This gradient acts backward on the class token representation and is further passed to the encoder to update the parameters of the entire layer, guiding the representation to learn class discrimination information. The gradient corresponding to the DPO preference loss is used as the query vector for preference matching. This gradient will drive the current representation to move closer to the expert's positive preference representation and away from the negative preference representation, thus achieving semantic preference alignment constraint.

[0088] Two gradients in token-like representation After merging, the representations are updated jointly: the classification gradient ensures the basic classification accuracy of the model, while the DPO-aligned gradient constrains the representation to fit expert preference semantics. This collaborative optimization mode not only alleviates the overfitting problem in small sample scenarios and improves the efficiency of sample data utilization, but also enhances the stability of feature distribution and the overall robustness of the model, which can explain the performance gain phenomenon in the experiments in this paper.

[0089] The training process aims to minimize the joint objective function, which is a weighted combination of the supervised loss term and the DPO preference alignment term. Its specific expression is defined as follows:

[0090] ;

[0091] In the formula:

[0092] λ>0 is used to control the trade-off between the two losses;

[0093] Cross-entropy loss for classification tasks;

[0094] DPO preference loss This is used to measure the alignment between student representations and preference target representations built based on expert prompts.

[0095] Using a joint objective function on the text encoder Lightweight controller One-dimensional encoder Predicting Head Text encoder And the parameters of the projection layer that constructs positive and negative sample pairs in step S51 are updated with gradients.

[0096] S6, Inference Phase: Remove the text encoder The text encoder, jointly optimized in step S6, is used to construct positive and negative sample pairs and compute the DPO preference loss. Lightweight controller One-dimensional encoder and prediction head The predictor takes I / Q sequence code x and instruction text i as input, and outputs the classification result according to steps S2 to S4, without requiring expert prompt text. Since the predictor only inputs... That is, only the I / Q sequence and instruction text are input, and the one-dimensional encoder is retained. Due to the limited sequence length and attention overhead, this embodiment is able to maintain computational and memory parity while preserving controllability through instruction text.

[0097] In summary, this embodiment reuses the same encoder and classifier in both the training and inference phases, and the instruction text continuously provides the operator with a stable and interpretable control interface.

[0098] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A language-guided I / Q representation method for wireless signal processing, characterized in that, Includes the following steps: S1, Constructing the dataset: The dataset includes a training set and a test set, defined as follows: ； ； in, Represents the I / Q time series; Indicates category label; Indicates instance-level instruction text; This indicates expert tips text that are only available during the training phase; This label represents the data collection conditions and can be either optional or empty. S2, Instruction Encoding and Control Signal Generation: Using the first text encoder instance-level instruction text Encoded as instruction vector Expert suggestion for encoding: Use a second text encoder. Expert text Encode as feature vector Use a lightweight controller Will Mapped to control signals ; S3, ICM modulation: In the control signal Guided by a one-dimensional encoder I / Q time series Mapping to representation vector During this process, control signals A affine modulation method applied to a one-dimensional encoder The intermediate activation values ​​of the feedforward network (FFN) for each Transformer block; S4, Classification prediction output: represented by a vector For input, use the prediction head Prediction is performed to obtain classification results. ; S5, Training Phase: Includes preference alignment and joint optimization steps, where: S51, Preference Alignment: Preference alignment is an alignment method based on the Direct Preference Optimization (DPO) concept, specifically DPO-SI preference alignment. The iterative optimization strategy model is defined as the student model, and the frozen reference baseline model is defined as the expert benchmark model. Its implementation method is based on feature vectors. Construct positive and negative sample pairs and calculate the DPO preference loss; where the positive samples are the preference samples. Negative samples are unbiased samples. DPO preference loss Used to measure the alignment between student representations and preference target representations built based on expert prompts; S52, Joint Optimization: Constructing a joint objective function and using it on the text encoder. Lightweight controller One-dimensional encoder Predicting Head Text encoder And the gradient update of the parameters of the projection layer that constructs positive and negative sample pairs in step S51; S6, Inference Phase: Remove the text encoder The module on which positive and negative sample pairs are constructed and the DPO preference loss is calculated; using the text encoder jointly optimized in step S52. Lightweight controller One-dimensional encoder and prediction head Taking the I / Q time series x and instance-level instruction text i as input, the classification results are output according to steps S2 to S4, without the need for expert prompt text.

2. The language-guided I / Q characterization method for wireless signal processing as described in claim 1, characterized in that, The ICM modulation includes: I / Q time series The sequence is converted into a token sequence using a one-dimensional patch embedder with stride, and class tokens are pre-appended to the head of the sequence. In multiple Transformer blocks, for the feedforward network of each block, the control signals are generated by independently setting a FiLM generator for each layer of the network. Mapped to scaling factor and offset factor ; Feedforward Network (FFN) consists of two linear transformation layers and an intermediate nonlinear activation layer. The input is subjected to dimensionality upscaling within the first linear transformation layer to obtain the intermediate activation value. For intermediate activation values The formula for performing element-wise affine modulation is as follows: ； in, The element-wise product of Hadamard, the modulated activation value After dimensionality reduction through the second-level linear transformation, the feature fusion is completed by connecting it with the original input residual. After stacking L Transformer blocks, the resulting features are normalized in the final layer to output a token representation. .

3. The language-guided I / Q characterization method for wireless signal processing as described in claim 1, characterized in that, In the DPO-SI preference alignment, based on feature vectors The method for constructing positive and negative sample pairs is as follows: The feature vector generated by the prompt mapper from the expert prompts corresponding to the current sample is used as the preference sample. ; The feature vectors generated by the prompt mapper from expert prompts for other samples with different device IDs within the same data batch are used as non-preference samples. ; During loss calculation, the expert-hinted feature vectors in the same data batch that have the same device identifier as the current sample but are different samples are masked to prevent them from participating in the close constraint and push-away constraint.

4. The language-guided I / Q representation method for wireless signal processing as described in claim 1, wherein the DPO preference loss The calculation formula is: ； in, Indicates feature similarity calculation; This is a temperature coefficient used to adjust the smoothness of the similarity distribution; This represents the set of unbiased samples.

5. The language-guided I / Q characterization method for wireless signal processing as described in claim 1, characterized in that, First Text Encoder Second text encoder Both are sentence encoders and text encoders. It is used in conjunction with preference alignment during the training phase and completely turned off during the inference phase.

6. A language-guided I / Q characterization method for wireless signal processing as described in any one of claims 1 to 5, characterized in that, The data used to construct the dataset all came from the open-source I / Q-text pairing corpus RF-Langv1.

1.

7. A speech-guided I / Q representation system for wireless signal processing, comprising: The system comprises a data construction module, an instruction encoding and controller module, an ICM encoder module, a prediction head module, an expert prompt encoding module, a preference alignment and joint optimization module, and an inference control module; the collaborative implementation of these modules achieves the language-guided I / Q representation method for wireless signal processing as described in claim 1.

Images