A phononic crystal waveguide signal high-precision detection method and device based on a diamond quantum sensor

Abstract

The application relates to a phononic crystal waveguide signal high-precision detection method and device based on a diamond quantum sensor. The method comprises the following steps: obtaining a vibration signal transmitted through a phononic crystal waveguide through a diamond quantum sensor; sequentially performing frequency domain transformation on single-frame signals in the vibration signal to obtain single-frame frequency domain features; inputting the single-frame frequency domain features into a trained detector model, obtaining long-range dependency relationships between frequency features corresponding to different frequency components and context information between the frequency features through a self-attention module in the detector model, and correcting the frequency features corresponding to the related frequency components according to the long-range dependency relationships and the context information; identifying aggregated features containing the corrected frequency features through a classification layer in the detector model to obtain predicted binary values corresponding to the single-frame frequency domain features; the long-range dependency relationship refers to the correlation relationship between the amplitude information and / or phase information of two frequency components exceeding a preset threshold frequency component. The method can improve the signal detection precision.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of communication technology and quantum sensing, and in particular to a method, apparatus, computer equipment, computer-readable storage medium and computer program product for high-precision detection of phonon crystal waveguide signals based on diamond quantum sensors. Background Technology

[0002] With the rapid development of quantum technology and micro / nano fabrication, diamond quantum sensors, especially those based on nitrogen-vacancy (NV) color center defects, are becoming a core technology for next-generation high-performance physical signal detection due to their superior sensitivity, broadband response, atomic-scale spatial resolution, and ability to operate at room temperature. Utilizing the quantum coherent response of NV color centers to magnetic fields, electric fields, temperature, and strain (correlated with phonon waves), diamond quantum sensors offer unprecedented opportunities for detecting and analyzing phononic crystal waveguide signals in on-chip communication systems.

[0003] Phononic crystal waveguides, as an emerging on-chip communication medium, can manipulate acoustic / elastic waves at the subwavelength scale, possessing the potential for high integration and resistance to electromagnetic interference. However, the transmission characteristics of acoustic waves in phononic crystal waveguides are complex, exhibiting strong nonlinear dispersion and frequency selectivity, leading to waveform broadening, distortion, and inter-symbol interference in the transmitted signal, severely degrading the quality of the received signal. Traditional signal detection methods, such as equalization techniques based on linear assumptions, have limited effectiveness in handling this highly nonlinear and strongly frequency-selective channel model mismatch. In recent years, although deep learning models have improved signal detection performance, their sequential processing structure has bottlenecks in computational parallelism, and their ability to model long-range dependencies across signal lengths caused by strong dispersion remains insufficient.

[0004] This application aims to combine the unique advantages of diamond quantum sensors with advanced signal processing algorithms to overcome the limitations of traditional methods. By integrating a highly sensitive diamond quantum sensor into the receiver of a phononic crystal waveguide signal, we can more accurately capture vibrational signals carrying quantum information. Furthermore, utilizing the self-attention mechanism of the Transformer architecture, this application proposes an innovative signal detection method to overcome the challenges of existing technologies in handling complex nonlinear distortion and long-range dependence in phononic crystal waveguides, thereby achieving unprecedented high detection accuracy, strong environmental robustness, and efficient processing capabilities. This provides key technological support for next-generation quantum information processing and communication systems based on phononic crystal waveguides and diamond quantum sensors. Summary of the Invention

[0005] Based on the aforementioned technical challenges and the unique advantages of diamond quantum sensors, this application provides a high-precision detection method, apparatus, computer device, computer-readable storage medium, and computer program product for phononic crystal waveguide signals based on diamond quantum sensors, capable of high-precision detection of phononic crystal waveguide signals. This application overcomes the limitations of traditional methods in handling complex physical distortions in phononic crystal waveguides, especially in scenarios with high nonlinearity, strong dispersion, and long-range dependence.

[0006] The technical solution provided in this application combines the physical advantages of quantum sensing with the algorithmic advantages of artificial intelligence, offering a novel and efficient technical path for solving the problem of communication signal detection under extreme channel conditions. Firstly, this application provides a high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor, comprising:

[0007] Vibration signals transmitted through a phononic crystal waveguide are acquired using a diamond quantum sensor; the diamond quantum sensor has high sensitivity and broadband response characteristics, and the vibration signal is obtained after the original transmitted signal undergoes a channel impulse response with the phononic crystal waveguide, and the original transmitted signal carries a binary data stream;

[0008] The single-frame signal in the vibration signal is sequentially transformed in the frequency domain to obtain the single-frame frequency domain features; the single-frame frequency domain features include the frequency features of different frequency components, and the frequency features include amplitude information and phase information.

[0009] The single-frame frequency domain features are input into a pre-trained detector model. The self-attention module in the detector model obtains the long-range dependencies between frequency features corresponding to different frequency components and the contextual information between frequency features corresponding to different frequency components. Based on the long-range dependencies and the contextual information, the frequency features corresponding to the relevant frequency components are corrected. The classification layer in the detector model identifies the aggregated features containing the corrected frequency features to obtain the predicted binary value corresponding to the single-frame frequency domain features.

[0010] The long-range dependency refers to the frequency difference between two related frequency components exceeding a preset threshold, and the amplitude information and / or phase information between the two frequency components being correlated.

[0011] In one embodiment, acquiring the vibration signal transmitted through the phononic crystal waveguide using a diamond quantum sensor includes: probing the strain field generated by elastic vibration in the phononic crystal waveguide using nitrogen-vacancy color centers inside the diamond quantum sensor; detecting the fluorescence response change of the nitrogen-vacancy color centers by laser excitation and microwave modulation; the fluorescence response change includes fluorescence intensity change and / or resonance frequency drift, which directly reflects the energy level change of the NV color centers caused by the strain field, thereby carrying the quantum information of the phonon wave; and converting the fluorescence response change into a continuous time-domain electrical signal as the vibration signal.

[0012] In one embodiment, the long-range dependency includes the correlation between phase distortion information of a first frequency feature and amplitude attenuation information of a second frequency feature; the frequency corresponding to the second frequency feature is greater than the frequency corresponding to the first frequency feature.

[0013] The step of correcting the frequency characteristics corresponding to the relevant frequency components based on the long-range dependency and the context information includes:

[0014] The second frequency feature is compensated based on the association and the context information;

[0015] The first frequency feature is modified based on the correlation and the context information.

[0016] In one embodiment, the self-attention module is a multi-head self-attention module, comprising a first feature subspace, a second feature subspace, and a third feature subspace, wherein:

[0017] The first feature subspace is used to focus on the overall shape of the spectrum corresponding to the single-frame frequency domain feature, ignoring minute phase jitter;

[0018] The second feature subspace is used to focus on the correlation between high-frequency attenuation and low-frequency delay, so as to obtain the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature.

[0019] The third feature subspace is used to focus on randomly occurring, unrelated isolated frequency points, identify them as noise, and suppress them.

[0020] Before identifying the aggregated features containing the corrected frequency features through the classification layer in the detector model, the method further includes:

[0021] The aggregated features are obtained by aggregating the outputs of the first feature subspace and the third feature subspace with the output of the second feature subspace, which includes the corrected frequency features.

[0022] In one embodiment, before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes:

[0023] Frequency information is added to the frequency features of different frequency components in the single-frame frequency domain features so that the position of different frequency features in the spectrum can be identified by the detector model.

[0024] In one embodiment, before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes:

[0025] The single-frame frequency domain features are subjected to min-max normalization to scale their numerical range to a preset interval.

[0026] In one embodiment, the step of sequentially performing frequency domain transformation on the single-frame signals in the vibration signal includes:

[0027] The discrete cosine transform is performed sequentially on the single-frame signals in the vibration signal;

[0028] Alternatively, perform Fast Fourier Transform on the single-frame signals in the vibration signal sequentially.

[0029] Secondly, this application also provides a high-precision detection device for phonon crystal waveguide signals based on a diamond quantum sensor, comprising:

[0030] The signal acquisition module is used to acquire vibration signals transmitted through a phononic crystal waveguide via a diamond quantum sensor; the diamond quantum sensor has high sensitivity and broadband response characteristics, and the vibration signal is obtained after the original transmitted signal undergoes a channel impulse response with the phononic crystal waveguide, wherein the original transmitted signal carries a binary data stream;

[0031] The frequency domain transformation module is used to sequentially perform frequency domain transformation on the single-frame signal in the vibration signal to obtain single-frame frequency domain features; the single-frame frequency domain features include frequency features of different frequency components, and the frequency features include amplitude information and phase information.

[0032] The prediction module is used to input the single-frame frequency domain features into a pre-trained detector model, obtain the long-range dependencies between frequency features corresponding to different frequency components and the contextual information between frequency features corresponding to different frequency components through the self-attention module in the detector model, and correct the frequency features corresponding to relevant frequency components based on the long-range dependencies and the contextual information; identify the aggregated features containing the corrected frequency features through the classification layer in the detector model to obtain the predicted binary value corresponding to the single-frame frequency domain features; wherein, the long-range dependency refers to the frequency difference between two related frequency components exceeding a preset threshold, and the amplitude information and / or phase information between the two frequency components having a correlation relationship.

[0033] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in the first aspect.

[0034] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0035] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0036] The aforementioned high-precision detection method, apparatus, computer equipment, computer-readable storage medium, and computer program product based on diamond quantum sensors for phononic crystal waveguide signals, specifically addressing the distorted signals of this complex physical channel, phononic crystal waveguides, transforms the time-domain convolutional mixing problem into a frequency-domain feature distortion problem through frequency domain transformation. This provides a more suitable input format for subsequent feature learning and pattern recognition using deep learning models. This avoids directly processing complex inter-symbol interference in the time domain. Furthermore, by utilizing a self-attention mechanism to globally and parallelly model long-range frequency domain dependencies, it is possible to accurately learn and compensate for the unique nonlinear and strong dispersion distortion of phononic crystal waveguides, improving detection accuracy and significantly reducing the bit error rate. This application provides a revolutionary solution for quantum information processing, quantum communication, and other high-precision sensing applications based on phononic crystal waveguides. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating a high-precision detection method for phononic crystal waveguide signals based on a diamond quantum sensor in one embodiment.

[0039] Figure 2 This is a flowchart illustrating the process of correcting frequency characteristics in one embodiment;

[0040] Figure 3 This is a schematic diagram of a phononic crystal waveguide communication system architecture based on a diamond quantum sensor in another embodiment.

[0041] Figure 4 This is a schematic flowchart of the signal processing flow in one embodiment;

[0042] Figure 5 This is a schematic diagram of the structure of a Transformer-based detector model in one embodiment;

[0043] Figure 6 This is a schematic diagram of the phononic crystal waveguide band structure and transmission spectrum used in a simulation experiment in one embodiment.

[0044] Figure 7 This is a comparison chart of the bit error rate (BER) of the method in this application and the conventional method under different signal-to-noise ratio conditions in one embodiment;

[0045] Figure 8 This is a comparison chart of the bit error rate (BER) of the method of this application and the conventional method under different transmission rate conditions in one embodiment;

[0046] Figure 9 This is a structural block diagram of a high-precision phononic crystal waveguide signal detection device based on a diamond quantum sensor in one embodiment.

[0047] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] It should be noted that the terms "first," "second," etc., used in this application may be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more.

[0050] In one embodiment, such as Figure 1 As shown, a high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor is provided. It is understood that this method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method includes the following steps:

[0051] Step S102: Obtain the vibration signal transmitted through the phononic crystal waveguide using a diamond quantum sensor. The diamond quantum sensor has high sensitivity and broadband response characteristics. The vibration signal is obtained after the original transmitted signal undergoes a channel impulse response with the phononic crystal waveguide. The original transmitted signal carries a binary data stream.

[0052] In this context, a phononic crystal waveguide can refer to a channel introduced into an artificial periodic structure (phononic crystal) to guide and confine the propagation of sound waves or elastic waves at subwavelength scales. Its characteristics are determined by the geometric and physical parameters of the periodically arranged materials (such as steel columns / epoxy resin). The vibration signal can refer to the physical vibration of sound waves or elastic waves propagating in the phononic crystal waveguide. At the receiving end, this physical vibration is captured by a sensor (such as a diamond quantum sensor, which has high sensitivity and broadband response characteristics) and converted into an electrical signal. The original transmitted signal can refer to the baseband waveform formed after the binary data stream to be transmitted is modulated (such as by on / off keying OOK), used to drive the excitation source at the transmitting end to generate the initial vibration signal. The channel impulse response can refer to a function describing the channel characteristics of the phononic crystal waveguide; it is a response in the time domain that typically has a long tail and complex oscillation pattern, characterizing the waveguide's dispersion (different propagation speeds of different frequency components) and multipath effects, which are the root cause of signal distortion.

[0053] For example, a computer device uses a diamond quantum sensor to acquire severely distorted continuous physical vibrations transmitted from a phonon crystal waveguide channel and converts them into a continuous time-domain electrical signal. This signal can be mathematically modeled as the convolution of the original transmitted signal and the channel impulse response, superimposed with environmental noise. The original transmitted signal carries the binary data stream to be recovered. Subsequently, this continuous electrical signal can be sampled to obtain a discrete-time-domain digital signal sequence for subsequent digital signal processing.

[0054] Step S104: Perform frequency domain transformation on the single-frame signal in the vibration signal in sequence to obtain the single-frame frequency domain features; the single-frame frequency domain features include the frequency features of different frequency components, and the frequency features include amplitude information and phase information.

[0055] In this context, a single-frame signal can refer to a segment extracted from a continuously sampled signal according to a time window. In this embodiment, the duration of this segment typically corresponds to the transmission time of one symbol. Frame segmentation is used to transform the continuous detection task into an independent decision task for each individual symbol. Frequency domain transformation is a mathematical tool that converts a signal from a time-dimensional representation to a frequency-dimensional representation. Its purpose is to transform the complex convolutional distortion in the time domain (manifested as waveform broadening and inter-symbol interference) into the relatively more structured amplitude attenuation and phase distortion in the frequency domain, facilitating subsequent model identification of channel influence patterns. Frequency characteristics can refer to the representation of a signal at specific frequency components. For complex transformations (such as the Fast Fourier Transform, FFT), it is a complex number containing amplitude and phase (angle); for real transformations (such as the Discrete Cosine Transform, DCT), it is typically a real coefficient that primarily reflects energy intensity (related to amplitude).

[0056] For example, the computer device first divides the discrete time-domain digital signal sequence into frames to obtain a single-frame time-domain signal, and then performs a frequency domain transformation on the single-frame time-domain signal. The vector obtained after the transformation is the single-frame frequency domain characteristic.

[0057] Step S106: Input the single-frame frequency domain features into the pre-trained detector model. The self-attention module in the detector model obtains the long-range dependencies between frequency features corresponding to different frequency components and the context information between frequency features corresponding to different frequency components. Based on the long-range dependencies and context information, the frequency features corresponding to the relevant frequency components are corrected. The classification layer in the detector model identifies the aggregated features containing the corrected frequency features to obtain the predicted binary value corresponding to the single-frame frequency domain features.

[0058] Among them, long-range dependency refers to the frequency difference between two related frequency components exceeding a preset threshold, and the amplitude information and / or phase information between the two frequency components are correlated.

[0059] The detector model can be a mathematical model trained on a deep learning architecture to infer the original transmitted symbols from distorted frequency domain features; its core is a structure capable of processing sequence data and modeling global relationships between elements. The self-attention module can be a neural network mechanism capable of calculating the correlation weights (attention weights) between any two elements in the input sequence, regardless of their relative distance. In this embodiment, the elements of the input sequence are frequency features of different frequency components. Long-range dependencies can refer to the inherent correlation between two components that are far apart in the spectrum (large frequency difference) due to the unified physical dispersion law of phononic crystal waveguides. For example, abnormal attenuation of high-frequency components may always be accompanied by a specific phase shift of low-frequency components. Contextual information can refer to the comprehensive influence information of all other frequency components in the sequence on the current frequency component to be processed, summarized through the self-attention mechanism. Correction can refer to the mathematical operation of "dedistorting" or "compensating" the input features based on the learned channel distortion pattern, making them closer to the feature performance under ideal channel conditions. Aggregated features refer to the aggregation of frequency components, which have already contained global information, after being processed through multiple layers within the model (such as self-attention or feedforward networks), into a single, fixed-dimensional feature vector using a certain method (such as global average pooling), for the final classification decision. The classification layer refers to the final output layer of the detector model, typically a fully connected layer combined with a Softmax (multi-class) or Sigmoid (binary class) activation function, used to map the aggregated features to the probabilities of different symbols (such as "0" or "1").

[0060] For example, a computer device inputs a single-frame frequency domain feature vector into a detector model based on a Transformer encoder architecture. The model may first project the input into a high-dimensional space through a linear embedding layer and add positional encoding to enable the model to perceive frequency locations. Subsequently, the feature vector enters the core self-attention module (e.g., a multi-head self-attention mechanism module). This module computes attention weights between all pairwise frequency components in the spectrum in parallel. For example, it can discover a strong correlation weight between the phase feature of the 10th frequency component (low frequency) and the amplitude feature of the 50th frequency component (high frequency), which is a kind of "long-range dependency." The model uses these attention weights to perform a weighted summation of the values ​​of all frequency features, so that the new representation of each frequency component incorporates the "contextual information" of the entire spectrum, essentially "correcting" it according to physical laws. After deep feature extraction through multiple such encoder layers, the model aggregates the features of all frequency locations into a "aggregated feature" through a global average pooling layer. Finally, the aggregated feature is fed into the classification layer, which outputs a binary probability distribution, such as P('0')=0.1, P('1')=0.9, which means the current frame is classified as binary data '1'.

[0061] In the aforementioned high-precision detection method for phononic crystal waveguide signals based on diamond quantum sensors, the distortion signal specific to the complex physical channel of phononic crystal waveguides is transformed into a frequency domain feature distortion problem through frequency domain transformation. This provides a more suitable input format for subsequent feature learning and pattern recognition using deep learning models. This avoids directly processing complex inter-symbol interference in the time domain. Furthermore, by utilizing a self-attention mechanism to globally and parallelly model long-range frequency domain dependencies, the method can accurately learn and compensate for the nonlinear and strong dispersion distortion unique to phononic crystal waveguides, improving detection accuracy and significantly reducing the bit error rate.

[0062] In an exemplary embodiment, acquiring vibration signals transmitted through a phononic crystal waveguide using a diamond quantum sensor includes: probing the strain field generated by elastic vibrations in the phononic crystal waveguide using nitrogen-vacancy color centers within the diamond quantum sensor; detecting changes in the fluorescence response of the nitrogen-vacancy color centers through laser excitation and microwave modulation; the fluorescence response changes include changes in fluorescence intensity and / or resonant frequency drift, which directly reflect changes in the energy levels of the NV color centers caused by the strain field, thereby carrying quantum information of the phonon waves; and converting the fluorescence response changes into a continuous time-domain electrical signal as the vibration signal.

[0063] Signal acquisition is fundamental to the entire detection process. A diamond quantum sensor is configured at the receiving end of the communication system. The original transmitted signal (carrying a binary data stream) is modulated and converted into acoustic / elastic wave pulses by an excitation source, which are then injected into the phononic crystal waveguide channel. After transmission through the channel, the distorted acoustic / elastic wave signal reaches the receiving end. The diamond quantum sensor utilizes its internal nitrogen-vacancy (NV) color centers to highly sensitively detect the strain field generated by elastic vibrations (phonon waves) in the phononic crystal waveguide. When the phonon wave passes through the sensor region, it causes lattice deformation of the NV color center. This strain field couples to the electron spin Hamiltonian of the NV color center, resulting in a slight change in its spin energy level, which in turn affects its resonant absorption at a specific microwave frequency. By applying a stable green laser and a frequency-sweeping microwave, and monitoring the intensity of the red fluorescence emitted by the NV color center, the sensor can accurately read out this resonant frequency shift or fluorescence intensity change caused by the phonon wave. This process is the application of optically probed magnetic resonance or its variants in strain sensing. This quantum sensing mechanism can capture minute vibrational details that are difficult for traditional macroscopic sensors, such as piezoelectric transducers, to distinguish, especially in low signal-to-noise ratio environments, significantly improving the quality of signal acquisition. The fluorescence response changes output by the sensor are converted into a continuous time-domain electrical signal by a photodetector. This electrical signal is then digitized by an analog-to-digital converter to form a discrete time-domain signal sequence for subsequent frequency domain transformation and model-based detection.

[0064] In one exemplary embodiment, such as Figure 2 As shown, the correction of the frequency characteristics corresponding to the relevant frequency components based on long-range dependencies and contextual information includes steps S202 to S204. Wherein: the long-range dependency includes the correlation between the phase distortion information of the first frequency characteristic and the amplitude attenuation information of the second frequency characteristic; the frequency corresponding to the second frequency characteristic is greater than the frequency corresponding to the first frequency characteristic.

[0065] Step S202: Compensate the second frequency feature based on the correlation and context information.

[0066] Step S204: Correct the first frequency feature based on the correlation and context information.

[0067] For example, for the second frequency feature, the model uses contextual information from the first frequency feature (low frequency) and other frequency components to compensate for its undue amplitude attenuation, i.e., to appropriately increase its amplitude. For the first frequency feature, the model uses contextual information from the second frequency feature (high frequency) and other frequency components to correct its abnormal phase value, causing it to revert to a normal value.

[0068] In this embodiment, the model is clearly defined as being able to capture and compensate for a specific and critical physical distortion mechanism in phononic crystal waveguides, which is consistent with the underlying physical principles and enhances the interpretability and technical persuasiveness of the solution.

[0069] In an exemplary embodiment, the self-attention module is a multi-head self-attention module, including a first feature subspace, a second feature subspace, and a third feature subspace, wherein: the first feature subspace is used to focus on the overall shape of the spectrum corresponding to the frequency domain features of a single frame, ignoring minor phase jitter; the second feature subspace is used to focus on the correlation between high-frequency attenuation and low-frequency delay, so as to obtain the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the third feature subspace is used to focus on randomly occurring isolated frequency points without correlation, identify them as noise, and suppress them; before recognizing the aggregated features containing the corrected frequency features through the classification layer in the detector model, the module further includes: aggregating the output of the first feature subspace and the output of the third feature subspace with the output of the second feature subspace containing the corrected frequency features to obtain aggregated features.

[0070] In this context, a multi-head self-attention module can refer to extending self-attention computation to multiple "heads," each with independent linear projection parameters, allowing the model to learn different types of dependencies in parallel across different representation subspaces. The feature subspace can correspond to different "heads" in the multi-head attention, with each head learning to focus on different aspects of the input features. High frequency and low frequency are relative to the passband or operating frequency band of the waveguide. High frequency typically refers to the relatively high frequency region within the passband, near the upper edge; low frequency typically refers to the relatively low frequency region within the passband, near the lower edge. High-frequency attenuation refers to the loss of amplitude (energy); at the high-frequency end of the passband, due to structural design or inherent material properties, elastic waves are more easily dissipated or scattered out of the waveguide during propagation, resulting in a significant weakening of the amplitude of the high-frequency components in the signal. For example, a steep pulse that originally contained rich high-frequency components will become smoother and have a trailing effect after its high-frequency portion is "eaten up." Low-frequency delay can refer to phase (time) lag; this is a direct manifestation of the dispersion effect. Depending on the waveguide's band structure, the group velocity of low-frequency components (i.e., the speed at which wave packet energy propagates) may be lower than that of high-frequency components; the low-frequency components in the signal are "late" relative to the high-frequency components. This can lead to severe distortion of the signal waveform. For example, if the low-frequency portion of a pulse arrives late and overlaps with the high-frequency portion of subsequent pulses, it can cause severe intersymbol interference (ISI).

[0071] For example, in a multi-head self-attention layer, suppose there are 8 heads (8 feature subspaces). During training: the parameters of some heads (such as the first feature subspace) evolve to focus on capturing the global envelope shape of the spectrum, becoming insensitive to detailed phase changes. The parameters of other heads (such as the second feature subspace) specifically learn to recognize specific coupling patterns between low-frequency phase and high-frequency amplitude. The parameters of still others (such as the third feature subspace) become adept at recognizing abrupt feature points that have no stable correlation with any other component, attributing them to noise and thus assigning them very low weights during weighted summation. The outputs of all heads are concatenated and then fused through a linear transformation layer. This fusion process is called "aggregation," which integrates the specialized analysis results of each subspace to form a more robust and comprehensive feature representation.

[0072] In this embodiment, through a multi-head mechanism, the model can simultaneously analyze spectral features from multiple perspectives, focusing on overall trends, capturing specific physical correlations, and suppressing noise, thereby obtaining richer and more accurate feature representations than single-head attention. Parallel learning of multiple modes enhances the model's expressive power and adaptability to different distortion conditions.

[0073] In an exemplary embodiment, before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes: adding frequency information to the frequency features of different frequency components in the single-frame frequency domain features, so as to identify the position of different frequency features in the spectrum by the detector model.

[0074] The addition of frequency information can be achieved through positional encoding. Since the Transformer's self-attention mechanism is position-independent, it is necessary to explicitly inform the model of the order or position of each element in the sequence by adding frequency information.

[0075] In this embodiment, by adding frequency information to the frequency features of different frequency components in the single-frame frequency domain features, the detector model can distinguish between components of different frequencies such as high frequency and low frequency, and can accurately learn the frequency selective channel characteristics.

[0076] In an exemplary embodiment, before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes: performing min-max normalization on the single-frame frequency domain features to scale their numerical range to a preset interval.

[0077] Among them, min-max normalization can refer to a data standardization method that linearly transforms data to a fixed numerical range such as [0,1] or [-1,1].

[0078] In this embodiment, by performing min-max normalization on the single-frame frequency domain features, the scale of the input data is unified, avoiding the difficulty in optimizing the model due to excessive differences in the order of magnitude of feature values, thus accelerating the training process of the model and improving the numerical stability of the model.

[0079] In an exemplary embodiment, frequency domain transformation is performed sequentially on single-frame signals in the vibration signal, including: sequentially performing discrete cosine transform on single-frame signals in the vibration signal; or, sequentially performing fast Fourier transform on single-frame signals in the vibration signal.

[0080] The Fast Fourier Transform (FFT) can transform a time-domain frame by calling the FFT algorithm library to obtain a complex spectrum, simultaneously acquiring amplitude and phase information. The Discrete Cosine Transform (DCT) can transform a time-domain frame by calling the DCT algorithm library to obtain a sequence of real coefficients. This sequence has concentrated energy and mainly reflects the amplitude / energy spectrum of the signal.

[0081] This embodiment provides flexible implementation options. FFT can provide complete frequency domain information (amplitude + phase), which may help with higher precision compensation, but the calculation result is a complex number. DCT has high computational efficiency (real number operations) and good energy concentration characteristics, which may be more effective for signals whose main energy is concentrated in low frequencies. Alternative solutions are provided for application scenarios with different computing resources and accuracy requirements.

[0082] In one exemplary embodiment, this application also provides a method for intelligent detection of phonon crystal waveguide signals based on a diamond quantum sensor and the Transformer attention mechanism, which may specifically include the following steps:

[0083] Step 1: Acquire the vibration signal transmitted by the phonon crystal waveguide communication system.

[0084] like Figure 3 As shown, the phononic crystal waveguide communication system used in the above-mentioned intelligent detection method for phononic crystal waveguide signals based on the diamond quantum sensor and the Transformer attention mechanism mainly consists of three parts: a transmitter, a phononic crystal waveguide channel, and a receiver.

[0085] At the transmitter end, the binary data stream {b_k} to be transmitted first passes through a modulator (e.g., using on-off keying (OOK) modulation) and is converted into a baseband signal x(t). This signal drives an excitation source (such as a piezoelectric ceramic sheet) to generate a series of acoustic or elastic wave pulses at the entrance of the phononic crystal waveguide channel.

[0086] Phononic crystal waveguide channels are the core medium for communication. The propagation of waves within them can be described by the channel impulse response h(t). Due to the dispersion effects of materials and structures, h(t) typically exhibits as a long pulse with severe tailing and complex oscillations. Therefore, the continuous-time-domain signal y(t) received by the diamond quantum sensor at the receiver end can be represented as the convolution of the transmitted signal x(t) and the channel impulse response h(t), superimposed with additive white Gaussian noise n(t):

[0087]

[0088] The receiver first samples y(t) to obtain a discrete-time sequence y[n]. Due to severe inter-symbol interference, the waveform of y[n] is vastly different from the pulse shape of the original x(t), rendering traditional decision methods completely ineffective. The core task of the method in this embodiment is to accurately recover the original symbol sequence {b_k} from the severely distorted y[n].

[0089] Signal acquisition is fundamental to the entire detection process. For example... Figure 3As shown, at the receiving end of the communication system, a diamond quantum sensing module is configured. It uses a nitrogen-vacancy (NV) color center array integrated on a diamond substrate to obtain the vibration signal transmitted through the phononic crystal waveguide. The diamond quantum sensing module can accurately detect the response of the NV color center quantum state to the strain induced by the phonon wave through optical and microwave technologies, and convert the weak vibration signal of the phononic crystal waveguide into an electrical signal, thereby achieving high-sensitivity and low-noise signal front-end acquisition. Exemplarily, the diamond quantum sensing module may include: a diamond substrate on which a nitrogen-vacancy (NV) color center array (one or more nitrogen-vacancy color center diamonds for responding to the strain field generated by the vibration signal) is integrated; a supporting optical readout system for laser exciting the NV color center array and collecting fluorescence signals; and a microwave regulation system for applying a microwave field to the NV color centers. Specifically, the original transmitted signal (carrying a binary data stream) is converted into acoustic / elastic wave pulses through modulation and an excitation source and injected into the phononic crystal waveguide channel. After transmission through the channel, the distorted acoustic / elastic wave signal reaches the receiving end. The diamond quantum sensor uses the nitrogen-vacancy (NV) color centers inside it to highly sensitively detect the strain field generated by the elastic vibration (phonon wave) in the phononic crystal waveguide. When the phonon wave passes through the sensor area, it causes the lattice of the NV color centers to deform. This strain field is coupled to the electron spin Hamiltonian of the NV color centers, resulting in a slight change in its spin energy level, which in turn affects its resonance absorption at a specific microwave frequency. By applying a stable green laser and a swept-frequency microwave and monitoring the intensity of the red fluorescence emitted by the NV color centers, the sensor can accurately read out this resonance frequency shift or fluorescence intensity change caused by the phonon wave. This process is the application of optically detected magnetic resonance or its variant technologies in strain sensing. This quantum sensing mechanism can capture weak vibration details that are difficult to distinguish by traditional macroscopic sensors such as piezoelectric transducers, especially in a low signal-to-noise ratio environment, significantly improving the quality of signal acquisition. Then, through a set of optical readout systems supporting the diamond quantum sensing module, the fluorescence response change of the nitrogen-vacancy color centers is detected and converted into an electrical signal. That is, the fluorescence response change output by the sensor is converted into a continuous time-domain electrical signal by a photodetector, and this electrical signal is subsequently digitally processed by an analog-to-digital converter to form a discrete time-domain signal sequence for subsequent frequency-domain transformation and Transformer-based detection.

[0090] Step two, signal preprocessing and feature extraction.

[0091] As Figure 4 shown, the signal processing flow of this embodiment starts with the preprocessing and feature extraction of the received signal. It can be understood that the steps of signal preprocessing and feature extraction in this embodiment can be processed by the server or the terminal.

[0092] Step S11: Sampling and Framing. The received continuous signal y(t) is sampled at a rate f_s that satisfies the Nyquist sampling theorem to obtain a discrete time sequence y[n]. Subsequently, y[n] is divided into frames of length L, each frame corresponding to the duration of one or more symbols.

[0093] Step S12: Frequency Domain Transform. Apply Discrete Cosine Transform (DCT) to each signal frame y[n] (where n=0,1,...,L-1) to obtain its frequency domain feature vector Y[k] (where k=0,1,...,L-1):

[0094]

[0095] Here, w[k] is the weighting factor. DCT is chosen because it has good energy concentration characteristics, concentrating the main energy of the signal on a few low-frequency coefficients, and it is a real-number transformation with low computational complexity. Alternatively, Fast Fourier Transform (FFT) can be used to obtain complex frequency domain features containing amplitude and phase information.

[0096] Step S13: Normalization. To improve the stability and convergence speed of neural network training, the obtained frequency domain feature vector Y[k] is subjected to minimum-maximum normalization, scaling its numerical range to the [0, 1] interval:

[0097]

[0098] This normalized vector Y_norm will be used as input to the subsequent Transformer detector.

[0099] Step 3: Perform detection using a Transformer-based signal detector.

[0100] Among them, such as Figure 5 The diagram shows the structure of a signal detector based on a Transformer encoder architecture. Traditional Transformers are often used for processing text or images, while we transform the signal from the time domain to the frequency domain and use a standard encoder to process the 'frequency domain feature vector'. Physical mechanism matching: Utilizing the mature 'self-attention mechanism' in the standard architecture, we directly address the unique 'frequency-selective fading' and 'long-range dispersion interference' problems in phononic crystals. Therefore, using a standard architecture ensures the maturity and stability of the algorithm.

[0101] The detector receives a normalized frequency domain feature vector Y_norm as an input sequence and outputs the predicted probability of the original symbol.

[0102] Step S21, Input Embedding and Position Encoding. The input Y_norm is first passed through a linear layer (input embedding layer), mapping it from L dimensions to a higher-dimensional model internal representation space of dimension d_model. Since the Transformer itself does not have the ability to process sequence order, position information must be explicitly introduced. Therefore, a pre-computed position encoding vector is added to the embedding vector. This position encoding is usually generated using sine and cosine functions of different frequencies, enabling the model to distinguish the position of different frequency components k in the spectrum. Here, we transform the time-domain signal into a frequency-domain sequence, so each 'position' in the Transformer input sequence physically corresponds directly to a certain 'frequency value' in the spectrum. Position is frequency identity: the first position in the sequence represents the lowest frequency, and the Nth position represents the highest frequency. Because phononic crystal waveguides have drastically different dispersion and attenuation effects on signals of different frequencies (e.g., low-frequency pass, high-frequency block), without position encoding, the model only knows the magnitude of the value, but not whether the value belongs to 'low frequency' or 'high frequency', and therefore cannot specifically learn the specific distortion characteristics at that frequency.

[0103] Step S22, encoder processing. (e.g.) Figure 5 As shown, the main body of the model consists of N identical encoder blocks stacked together. Each encoder block contains two core sub-layers:

[0104] First, there's Multi-Head Self-Attention (MHSA): this is the core mechanism of the Transformer. It allows the model to simultaneously attend to all other frequency domain components in the input sequence while processing each frequency domain component. Specifically, the input sequence is linearly projected into multiple sets (i.e., multiple "heads") of query (Q), key (K), and value (V) vectors. For each head, an attention weight is obtained by calculating the dot product of Q and K, which determines how much attention each V should be given when synthesizing the output. The calculation formula is as follows:

[0105]

[0106] The physical significance of this mechanism lies in its ability to enable the model to autonomously learn complex correlations between frequency domain components. For example, the dispersion of a phonon crystal may lead to a strong correlation between phase distortion at 10 kHz and amplitude attenuation at 50 kHz, and MHSA can directly capture this long-range dependency. The multi-head mechanism allows the model to learn different correlation patterns in parallel across different representation subspaces.

[0107] Mathematically, this subspace is constructed from a specific set of linear projection matrices (i.e., model parameters $W_Q, W_K, W_V$). Physically, the original frequency domain data contains all amplitude and phase information, mixed together. When the data is mapped to this specific subspace, its representation is reorganized. From this perspective, "amplitude attenuation at 50kHz" and "phase distortion at 10kHz," which were originally far apart in the spectrum, are mapped to two mathematically highly similar (consistent vector directions) eigenvectors; MHSA directly captures this long-range dependency. Furthermore, initially, the model is unaware of these relationships. However, as backpropagation progresses, the model discovers that if 'amplitude attenuation' and 'phase distortion' occur simultaneously without relating them, an incorrect judgment will be made. To reduce the error rate, the model automatically adjusts the projection parameters entering this subspace. The ultimate effect is that when such specific physical coupling exists in the input signal (e.g., coupling caused by phonon crystal dispersion), the dot product of the query and key vectors generated in this subspace becomes very large. In other words, the model learns through training to transform 'physical causality' into 'geometric parallelism within the subspace,' thereby activating high-weight attention.

[0108] Because we used a 'multi-head' mechanism, the model has several different 'eyes,' each focusing on different aspects of the signal. In phonon crystal waveguide communication, these might focus on: Subspace A (amplitude envelope): specifically the overall shape of the spectrum, such as determining whether the signal energy is mainly concentrated in the passband center or at the edges, ignoring minute phase jitter. Subspace B (dispersion compensation): specifically seeking the complementary relationship between 'high-frequency attenuation' and 'low-frequency delay' to correct waveform broadening caused by nonlinear dispersion. Subspace C (noise suppression): specifically focusing on isolated frequency points that appear randomly and have no correlation evidence, treating them as noise and giving them very low weight (suppression). Through this set of parallel observations from multiple 'views,' the model can ultimately piece together the complete picture of the signal.

[0109] Second, there is the Feed-Forward Network (FFN): After the self-attention layer, the output at each location independently passes through a fully connected feed-forward network. This network typically consists of two linear layers and a ReLU activation function, used to perform non-linear transformations on the features, enhancing the model's expressive power.

[0110] Here, each position refers to a specific frequency value (e.g., a component at 10kHz) in the sequence, resulting from the frequency domain transformation performed at the input. The output at each position refers to the vector generated after processing by the self-attention layer, which now contains aggregated information from the entire spectrum. It no longer simply represents 'my amplitude at 10kHz,' but rather 'the characteristic state I should have at 10kHz after incorporating 50kHz attenuation and 30kHz phase.' In short, it's a frequency feature representation 'corrected by contextual information,' which the FFN then further processes nonlinearly so that the final classification layer can understand it.

[0111] The output of each sublayer (MHSA and FFN) employs residual connections and layer normalization (Add & Norm) to prevent gradient vanishing and accelerate training.

[0112] Step S23, Classification Output. After processing by N encoder blocks, the resulting output sequence is aggregated into a fixed-size vector through a global average pooling layer. Finally, this vector is fed into a classification layer (usually a linear layer followed by a Softmax or Sigmoid activation function), which outputs the probability of each possible transmitted symbol. The symbol with the highest probability is the final detection result.

[0113] For example, the scenario is set up as follows: a data sequence [0,1,0] is transmitted, and we assume that the data segment we want to detect is [0,1,0] (i.e., 0s at the beginning and end, with a pulse of 1 in the middle). Our goal is to enable the receiver to accurately determine which moment in the middle was when a "1" was transmitted.

[0114] Phase 1: "Disruption" of the physical layer (correlated using the frequency domain).

[0115] Transmitter: When transmitting 1, the transmitter generates a perfect rectangular pulse or Gaussian pulse. In the frequency domain, this perfect pulse corresponds to a standard spectral shape (e.g., a smooth Sinc function envelope), with energy distribution across the entire frequency band.

[0116] Through a phononic crystal waveguide: Dispersion effect: The band structure of the waveguide causes high-frequency components to travel slower and low-frequency components to travel faster (or vice versa). Frequency-selective fading: Certain specific frequencies (such as $f_{50}$) happen to fall near the waveguide's "cutoff frequency," and energy is significantly absorbed; while the $f_{10}$ frequency flows unimpeded. The result is that the originally perfect pulse is broadened and tailed in the time domain (energy leaks to positions that should originally be 0).

[0117] At the receiving end: the received time-domain waveform is no longer a towering peak, but a "flattened and distorted" waveform. The problem is that if you only look at the time-domain amplitude, the amplitude of this waveform may become very small due to attenuation, even lower than some noise. Traditional "threshold detection" will misjudge it as "0".

[0118] Phase Two: Data Processing (How Transformer Works).

[0119] Step 1: Frequency Domain Transform (DCT).

[0120] We perform a DCT transform on the received "distorted signal" frame to obtain a frequency domain vector. Assume that $Y$ has 64 components ($k=0$ to $63$). Current characteristics: Due to channel corruption, the originally smooth spectrum is now riddled with holes. Energy is high at $k=10$ (low frequency). Energy is almost zero at $k=50$ (high frequency) (possibly swallowed by the channel). Phase is completely reversed at $k=30$.

[0121] Step 2: Attention mechanism.

[0122] At this point, the model input is $Y$. The model needs to answer: "Is this riddled spectrum a corrupted 1, or pure noise 0?" The Transformer's multi-head self-attention mechanism begins calculating the correlation weights between components:

[0123] The "knowledge" learned by the model: During the training phase, the model has learned the physical properties of phononic crystals—that is, "when the energy at $k=10$ is high and the energy at $k=50$ is missing, this usually means that a 1 is sent, but $k=50$ is eaten up by dispersion."

[0124] Calculation process:

[0125] Query: The model queries the component $k=50$: "Why is your energy so low?".

[0126] Key: $k=10$ component answer: "Because I am high energy, according to waveguide properties, if you are part of 1, I must be high when you are low."

[0127] Attention Weight: The model finds a very strong complementary relationship (strong correlation) between $k=10$ and $k=50$. The weight is increased.

[0128] The significance of long-range dependence: $k=10$ and $k=50$ are far apart in the spectrum, but they together constitute a fingerprint feature that is "actually 1". If it were just random noise (which is 0), this specific complementary pattern would not appear between the two.

[0129] Step 3: Feature reconstruction and compensation.

[0130] Based on the calculated weights, the Transformer's feedforward network (FFN) performs a weighted summation and nonlinear transformation on the feature vector. It's similar to a brain-filled spectrum repair: using a strong signal at k=10 as evidence to compensate for the missing signal at k=50. Output: After multiple layers of processing, the Transformer outputs an aggregated feature vector. This vector no longer represents the original physical frequency, but rather the "semantics of the signal." Its mathematical meaning tends to be: this is a high-confidence pulse signal.

[0131] Step 4: Classification

[0132] Finally, this aggregated feature vector is fed into a fully connected layer (classification layer). Mapping: The classification layer maps the abstract features to two output neurons: $P(0)$ and $P(1)$. Result: Since the Transformer successfully identified the "pulse fingerprint" corroborated by $k=10$ and $k=50$, the classifier outputs: $P(0)=0.05$ and $P(1)=0.95$. Final code: The code with the highest probability is output as 1.

[0133] For example, this embodiment is a process of mapping from the "physical feature space" to the "semantic category space," that is, the model identifies the "shape" of the frequency domain vector and labels this shape. In this process, the model performs feature aggregation and classification decisions. Simply put, the Transformer model acts as a "waveform translator" here. It does not treat the input frequency domain vector as a simple number, but rather as a "fingerprint image." Its task is to identify whether this fingerprint belongs to "category 0" or "category 1." To ultimately obtain a binary sequence similar to [0,1,0,0,1,1], the data stream undergoes the following key morphological transformations:

[0134] 1. Physical Slicing (Continuous Stream → Unit Frame). Action: First, we slice the received continuous long signal into small segments over time. State: Each segment corresponds to a time window of a "potential symbol". Although the final output is a sequence, the model processes each time window individually (or in batches). Purpose: To discretize the continuous signal in preparation for individual decision-making.

[0135] 2. Feature Extraction (Time Domain → Frequency Domain Fingerprint). Action: Perform DCT / FFT transformation on this segment. State: Obtain the input vector $Y$ (e.g., 64 frequency points). Meaning of Data: The data at this point represents the 'energy distribution'. For example: 'high energy at low frequencies, low energy at high frequencies'. This is like the 'sound spectrum' of this symbol.

[0136] 3. Semantic Condensation (Frequency Domain Distribution → Abstract Feature Vector). Key Step: This is a crucial turning point in data dimensionality. The output sequence after processing by the Transformer encoder still possesses a frequency dimension (e.g., 64 feature vectors). Action: Use a global average pooling layer. Process: The pooling layer averages or compresses these 64 frequency position feature vectors, fusing them into a single high-dimensional vector (e.g., a vector of length 256). Meaning of Data: At this point, the data no longer represents "frequency" but rather "the essence of the feature." This vector might contain the information: "This is a signal with typical impulse characteristics and has undergone specific dispersion distortion."

[0137] 4. Logical Decision (Abstract Feature → Probability → Bit). Action: This unique vector is fed into the classification layer (fully connected layer + Softmax). Process: The classification layer is like a dictionary lookup. It calculates using the parameters obtained during training: "What is the probability that a signal with this feature vector belongs to logic 0, and what is the probability that it belongs to logic 1?" Output: Assume the output probability is $P(1)=0.95$. Result: The decision is 1.

[0138] 5. Sequence Generation (Single Point → Sequence). Finally: As the receiver continuously receives signals, the above process is performed continuously for each time window. Window 1 → Feature Determination → Output 0. Window 2 → Feature Determination → Output 1... Final Combination: By concatenating the results of these consecutive decisions, you obtain the binary data stream [0,1,0,0,1,1] that you see.

[0139] In conclusion, it's not that the frequency domain vector directly becomes 0 or 1, but rather that the model 'sees' that the frequency domain vector presents a certain shape, judges based on experience that this shape corresponds to 1, and thus outputs the predicted classification label.

[0140] In addition, to verify the validity of this application, a simulation platform was built to conduct simulation experiments and result analysis, and the performance of the proposed Transformer-based detector was compared with that of traditional methods.

[0141] The simulation settings include: Figure 6As shown, a phononic crystal waveguide model was established: a two-dimensional phononic crystal model was built using simulation software (such as COMSOL Multiphysics), consisting of steel pillars arranged periodically in an epoxy resin matrix. A W1-type line defect waveguide was formed by removing one row of steel pillars. Figure 6 The simulation in (a) shows the typical band structure of this phononic crystal waveguide. Figure 6 Figure (b) shows the waveguide transmission spectrum within a certain passband. A clear nonlinear dispersion relationship is evident from the band structure, and the transmission spectrum also exhibits frequency-selective fading, which are the physical causes of signal distortion. The channel impulse response h(t) is calculated using this model. Data generation: 10^6 bits are randomly generated and modulated using OOK. The modulated signal is convolved with the channel impulse response h(t), and Gaussian white noise of varying power is added to simulate different signal-to-noise ratio (SNR) environments.

[0142] Comparison method:

[0143] The method used in this application is: DCT + Transformer detector.

[0144] Traditional MMSE equalizer: a classic frequency domain linear equalization method.

[0145] Simple threshold detection method: a basic time-domain detection method that directly compares the amplitude of the received signal at the sampling point with a fixed threshold to make a decision.

[0146] Model parameters: The key parameters used in the simulation are summarized in Table 1 below.

[0147] Table 1

[0148]

[0149] Results analysis:

[0150] Experiment 1: such as Figure 7 As shown, the performance is compared under different signal-to-noise ratios. Figure 7 The curves showing the bit error rate (BER) of the three methods described above as a function of signal-to-noise ratio (SNR) are presented at a fixed symbol rate (20 kS / s).

[0151] from Figure 7As can be seen, under all tested signal-to-noise ratio (SNR) conditions, the proposed Transformer detector significantly outperforms the two traditional methods. The simple threshold detection method performs the worst due to not considering inter-symbol interference (ISI). The traditional MMSE equalizer, as a linear equalization method, shows some improvement, but its capabilities are limited under the combined effects of nonlinear distortion and noise. The BER curve of this application consistently lies below the other two methods, especially in the low SNR region (0-10 dB), where its advantage is more pronounced. This indicates that its self-attention mechanism can effectively extract key frequency domain features related to channel distortion patterns from a strong noise background, achieving more robust signal detection.

[0152] Experiment 2: For example Figure 8 As shown, this compares the performance at different transmission rates. Figure 8 The figure shows the bit error rate (BER) curves of the three methods as a function of symbol rate at a fixed signal-to-noise ratio (SNR) of 15 dB. A higher symbol rate means a shorter symbol period Tb, resulting in more severe inter-symbol interference (ISI). The experimental results clearly show that the performance of all methods decreases with increasing symbol rate, but the magnitude of the decrease varies significantly. The performance of the simple threshold detection method and the MMSE equalizer deteriorates sharply, almost failing at high symbol rates, indicating their inability to handle severe ISI. In contrast, the Transformer detector proposed in this application exhibits superior robustness, with the flattest BER curve, maintaining a low level even under conditions of extremely severe ISI caused by high rates (e.g., 100 kS / s). This fully demonstrates the superiority of the Transformer's self-attention mechanism in modeling and compensating for long-term, complex channel memory caused by strong dispersion.

[0153] In summary, this application provides an intelligent detection method for phononic crystal waveguide signals based on a diamond quantum sensor and the Transformer attention mechanism. By transforming the signal to the frequency domain and utilizing the powerful global dependency modeling capability of the Transformer model, it effectively overcomes the severe channel distortion problem that is difficult to handle by traditional methods. It exhibits significant advantages in terms of accuracy, robustness, and high-speed adaptability, and provides a feasible key technical solution for realizing high-performance communication systems based on metamaterials such as phononic crystals.

[0154] Compared to traditional methods such as MMSE equalizers and simple threshold detection, the intelligent detection method for phonon crystal waveguide signals based on diamond quantum sensors and the Transformer attention mechanism has a key innovation: the integration of a diamond quantum sensor as the core signal acquisition method. Utilizing its internal NV color center defects, the diamond quantum sensor can detect weak vibrational signals transmitted in the phonon crystal waveguide with extremely high sensitivity and spatial resolution, and efficiently convert the vibrational information carried by these quantum states into electrical signals. This mechanism of direct signal sensing at the quantum level can capture fine features that are difficult to detect with traditional sensors, providing higher-quality raw data for subsequent signal processing, thereby significantly improving the overall accuracy of signal detection.

[0155] Another key innovation of this application lies in employing a Transformer-based detector model to process signals acquired by a diamond quantum sensor. This model utilizes its core self-attention mechanism to effectively capture the long-range dependencies between different frequency components in the signal spectrum caused by the dispersion effect of phononic crystal waveguides. By learning this global frequency correlation feature, the model can accurately correct and compensate for signal distortion, ultimately achieving high-precision recovery of the original binary data stream. This application leverages the powerful global feature capture capability of the Transformer model to more accurately learn and compensate for the nonlinear and strong dispersion distortion of phononic crystal waveguides, thereby achieving a significantly lower bit error rate (BER) at various signal-to-noise ratios and transmission rates, resulting in higher detection accuracy. This method learns channel characteristics directly from the data without requiring a precise prior channel model, exhibiting good adaptability and robustness to channel variations and noise. The parallel architecture of the Transformer model enables higher training and inference efficiency on modern hardware (such as GPUs / TPUs) compared to serially processed recurrent neural networks, making high-speed, real-time signal detection possible.

[0156] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0157] Based on the same inventive concept, this application also provides a high-precision detection device for phonon crystal waveguide signals based on a diamond quantum sensor, for implementing the high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor described above. The solution provided by this device is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the high-precision detection device for phonon crystal waveguide signals based on a diamond quantum sensor provided below can be found in the limitations of the high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor described above, and will not be repeated here.

[0158] In one exemplary embodiment, such as Figure 9 As shown, a high-precision detection device 900 for phonon crystal waveguide signals based on a diamond quantum sensor is provided, comprising: a signal acquisition module 901, a frequency domain transformation module 902, and a prediction module 903, wherein:

[0159] The signal acquisition module 901 is used to acquire the vibration signal transmitted through the phononic crystal waveguide by the diamond quantum sensor. The diamond quantum sensor has high sensitivity and broadband response characteristics. The vibration signal is obtained after the original transmitted signal and the phononic crystal waveguide undergo a channel impulse response. The original transmitted signal carries a binary data stream.

[0160] The frequency domain transformation module 902 is used to sequentially perform frequency domain transformation on the single-frame signal in the vibration signal to obtain the single-frame frequency domain features; the single-frame frequency domain features include the frequency features of different frequency components, and the frequency features include amplitude information and phase information.

[0161] The prediction module 903 is used to input the single-frame frequency domain features into a pre-trained detector model. The self-attention module in the detector model obtains the long-range dependencies between the frequency features corresponding to different frequency components and the contextual information between the frequency features corresponding to different frequency components. Based on the long-range dependencies and contextual information, the frequency features corresponding to the relevant frequency components are corrected. The classification layer in the detector model identifies the aggregated features containing the corrected frequency features to obtain the predicted binary value corresponding to the single-frame frequency domain features. The long-range dependency refers to the frequency difference between two related frequency components exceeding a preset threshold, and the amplitude information and / or phase information between the two frequency components are correlated.

[0162] In an exemplary embodiment, the signal acquisition module 901 is further configured to detect the strain field generated by elastic vibration in the phononic crystal waveguide using nitrogen-vacancy color centers inside the diamond quantum sensor; detect the fluorescence response change of the nitrogen-vacancy color centers by laser excitation and microwave modulation; the fluorescence response change includes fluorescence intensity change and / or resonance frequency drift, which directly reflects the energy level change of the NV color center caused by the strain field, thereby carrying the quantum information of the phonon wave; and convert the fluorescence response change into a continuous time-domain electrical signal as a vibration signal.

[0163] In an exemplary embodiment, the long-range dependency includes the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the frequency corresponding to the second frequency feature is greater than the frequency corresponding to the first frequency feature; the prediction module 903 is further configured to compensate the second frequency feature based on the correlation and context information; and to correct the first frequency feature based on the correlation and context information.

[0164] In an exemplary embodiment, the self-attention module is a multi-head self-attention module, including a first feature subspace, a second feature subspace, and a third feature subspace, wherein: the first feature subspace is used to focus on the overall shape of the spectrum corresponding to the frequency domain features of a single frame, ignoring minor phase jitter; the second feature subspace is used to focus on the correlation between high-frequency attenuation and low-frequency delay, so as to obtain the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the third feature subspace is used to focus on randomly occurring isolated frequency points without correlation, identify them as noise, and suppress them; the prediction module 903 is further used to aggregate the output of the first feature subspace and the output of the third feature subspace with the output of the second feature subspace containing the corrected frequency features to obtain aggregated features.

[0165] In an exemplary embodiment, the frequency domain transformation module 902 is further configured to add frequency information to the frequency features of different frequency components in the single-frame frequency domain features, so as to identify the position of different frequency features of the detector model in the spectrum.

[0166] In an exemplary embodiment, the frequency domain transformation module 902 is further configured to perform minimum-maximum normalization processing on the single-frame frequency domain features to scale their numerical range to a preset interval.

[0167] In an exemplary embodiment, the frequency domain transformation module 902 is further configured to sequentially perform discrete cosine transform on the single-frame signals in the vibration signal; or to sequentially perform fast Fourier transform on the single-frame signals in the vibration signal.

[0168] Each module in the aforementioned high-precision detection device for phonon crystal waveguide signals based on diamond quantum sensors can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware within or independently of the processor in a computer device, or stored in software within the memory of the computer device, allowing the processor to call and execute the corresponding operations of each module.

[0169] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network. When the computer program is executed by the processor, it implements a high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor.

[0170] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0171] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0172] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0173] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0174] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0175] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0176] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0177] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A high-precision detection method for phonon crystal waveguide signals based on a diamond quantum sensor, characterized in that, The method includes: The strain field generated by elastic vibration in a phononic crystal waveguide is detected using nitrogen-vacancy color centers within a diamond quantum sensor. The fluorescence response changes of these nitrogen-vacancy color centers are detected through laser excitation and microwave modulation. These fluorescence response changes include fluorescence intensity changes and / or resonant frequency shifts, directly reflecting the energy level changes of the NV color centers caused by the strain field, thus carrying quantum information of the phonon waves. The fluorescence response changes are converted into a continuous time-domain electrical signal as a vibration signal. The diamond quantum sensor exhibits high sensitivity and broadband response characteristics. The vibration signal is obtained after the original transmitted signal undergoes a channel impulse response with the phononic crystal waveguide, and the original transmitted signal carries a binary data stream. The single-frame signal in the vibration signal is sequentially transformed in the frequency domain to obtain the single-frame frequency domain features; the single-frame frequency domain features include the frequency features of different frequency components, and the frequency features include amplitude information and phase information. The single-frame frequency domain features are input into a pre-trained detector model. The self-attention module in the detector model obtains the long-range dependencies between frequency features corresponding to different frequency components and the contextual information between frequency features corresponding to different frequency components. Based on the long-range dependencies and the contextual information, the frequency features corresponding to the relevant frequency components are corrected. The classification layer in the detector model identifies the aggregated features containing the corrected frequency features to obtain the predicted binary value corresponding to the single-frame frequency domain features. The long-range dependency refers to a frequency difference between two related frequency components that exceeds a preset threshold, and a correlation exists between the amplitude information and / or phase information of the two frequency components; the long-range dependency includes the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the frequency corresponding to the second frequency feature is greater than the frequency corresponding to the first frequency feature. The step of correcting the frequency characteristics corresponding to the relevant frequency components based on the long-range dependency and the context information includes: The second frequency feature is compensated based on the association and the context information; The first frequency feature is modified based on the correlation and the context information.

2. The method according to claim 1, characterized in that, The self-attention module is a multi-head self-attention module, comprising a first feature subspace, a second feature subspace, and a third feature subspace, wherein: The first feature subspace is used to focus on the overall shape of the spectrum corresponding to the single-frame frequency domain feature, ignoring minute phase jitter; The second feature subspace is used to focus on the correlation between high-frequency attenuation and low-frequency delay, so as to obtain the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature. The third feature subspace is used to focus on randomly occurring, unrelated isolated frequency points, identify them as noise, and suppress them. Before identifying the aggregated features containing the corrected frequency features through the classification layer in the detector model, the method further includes: The aggregated features are obtained by aggregating the outputs of the first feature subspace and the third feature subspace with the output of the second feature subspace, which includes the corrected frequency features.

3. The method according to claim 1, characterized in that, Before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes: Frequency information is added to the frequency features of different frequency components in the single-frame frequency domain features so that the position of different frequency features in the spectrum can be identified by the detector model.

4. The method according to claim 1, characterized in that, Before inputting the single-frame frequency domain features into the pre-trained detector model, the method further includes: The single-frame frequency domain features are subjected to min-max normalization to scale their numerical range to a preset interval.

5. The method according to claim 1, characterized in that, The step of sequentially performing frequency domain transformation on the single-frame signals in the vibration signal includes: The discrete cosine transform is performed sequentially on the single-frame signals in the vibration signal; Alternatively, perform Fast Fourier Transform on the single-frame signals in the vibration signal sequentially.

6. A high-precision detection device for phonon crystal waveguide signals based on a diamond quantum sensor, characterized in that, The device includes: The signal acquisition module is used to detect the strain field generated by elastic vibration in the phononic crystal waveguide using nitrogen-vacancy color centers inside the diamond quantum sensor; it detects the fluorescence response changes of the nitrogen-vacancy color centers through laser excitation and microwave modulation; the fluorescence response changes include fluorescence intensity changes and / or resonant frequency drift, which directly reflect the energy level changes of the NV color centers caused by the strain field, thus carrying the quantum information of the phonon wave; the fluorescence response changes are converted into a continuous time-domain electrical signal as a vibration signal; the diamond quantum sensor has high sensitivity and broadband response characteristics, and the vibration signal is obtained after the original transmitted signal undergoes a channel impulse response with the phononic crystal waveguide, and the original transmitted signal carries a binary data stream; The frequency domain transformation module is used to sequentially perform frequency domain transformation on the single-frame signal in the vibration signal to obtain single-frame frequency domain features; the single-frame frequency domain features include frequency features of different frequency components, and the frequency features include amplitude information and phase information. The prediction module is used to input the single-frame frequency domain features into a pre-trained detector model, obtain the long-range dependencies between frequency features corresponding to different frequency components and the contextual information between frequency features corresponding to different frequency components through the self-attention module in the detector model, and correct the frequency features corresponding to the relevant frequency components based on the long-range dependencies and the contextual information; identify the aggregated features containing the corrected frequency features through the classification layer in the detector model to obtain the predicted binary value corresponding to the single-frame frequency domain features; wherein, the long-range dependency refers to the frequency difference between two related frequency components exceeding a preset threshold, and there is a correlation between the amplitude information and / or phase information of the two frequency components; the long-range dependency includes the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the frequency corresponding to the second frequency feature is greater than the frequency corresponding to the first frequency feature; the prediction module is also used to compensate the second frequency feature based on the correlation and the contextual information; and to correct the first frequency feature based on the correlation and the contextual information.

7. The apparatus according to claim 6, characterized in that, The self-attention module is a multi-head self-attention module, comprising a first feature subspace, a second feature subspace, and a third feature subspace, wherein: the first feature subspace is used to focus on the overall shape of the spectrum corresponding to the single-frame frequency domain feature, ignoring minor phase jitter; the second feature subspace is used to focus on the correlation between high-frequency attenuation and low-frequency delay, so as to obtain the correlation between the phase distortion information of the first frequency feature and the amplitude attenuation information of the second frequency feature; the third feature subspace is used to focus on randomly occurring, uncorrelated isolated frequency points, identify them as noise, and suppress them; the prediction module is further used to aggregate the output of the first feature subspace and the output of the third feature subspace with the output of the second feature subspace containing the corrected frequency features, to obtain the aggregated feature.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images