An artificial intelligence micro-vibration detection method based on FMCW millimeter wave radar

Abstract

This invention relates to the field of building micro-vibration detection technology, providing an artificial intelligence-based micro-vibration detection method based on FMCW millimeter-wave radar. The method includes: dynamically adjusting the transmission signal parameters of the FMCW radar using an adaptive optimization strategy and acquiring the radar's intermediate frequency (IF) signal; preprocessing the acquired IF signal and simultaneously extracting its time-frequency domain physical features and deep learning abstract features, performing cross-domain feature fusion to construct a fused feature vector; and inputting the fused feature vector into a lightweight Transformer micro-vibration recognition model trained with a domain adaptive framework for processing to obtain micro-vibration detection and classification results. This invention achieves closed-loop dynamic optimization of transmission parameters through reinforcement learning, actively adapting to the environment and improving the signal-to-noise ratio of the micro-vibration signal, laying a high-quality data foundation for subsequent analysis. Furthermore, by fusing physical features and deep learning features, it balances model performance and interpretability.

Description

Technical Field

[0001] This invention belongs to the field of building micro-vibration detection technology, and particularly relates to an artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar. Background Technology

[0002] Structural health monitoring, especially the early monitoring of micro-deformation and micro-vibration in large infrastructure such as bridges and high-rise buildings, is a core technical means to prevent structural disasters and ensure public safety. Traditional monitoring methods, such as total stations, GPS positioning systems, or fiber optic sensors, are widely used but have significant limitations.

[0003] Therefore, in recent years, frequency modulated continuous wave millimeter wave radar technology has shown great potential in the field of micro-deformation monitoring due to its advantages such as high precision, high resolution, strong anti-interference ability and all-weather operation.

[0004] FMCW radar transmits continuous waves with linearly varying frequencies and receives the echoes reflected from targets. By mixing these echoes, it extracts the intermediate frequency signal containing range and velocity information, thus enabling precise measurement of minute displacement changes in targets. However, applying millimeter-wave radar to long-term, high-precision monitoring of urban infrastructure still faces a series of significant technical challenges.

[0005] First, at the hardware level, in order to reliably acquire weak signals over long distances, radar antennas need to have high gain and high isolation. Existing array antennas often suffer from excessively high sidelobe levels, causing the radar to easily receive interference signals from the sidelobes, which seriously affects the signal-to-noise ratio and accuracy of the monitoring data.

[0006] Secondly, to meet the requirements of convenient on-site deployment and environmental adaptability, radar systems need to develop towards miniaturization and integration. However, the high-frequency millimeter-wave band introduces numerous signal integrity, power integrity, and electromagnetic compatibility issues.

[0007] Finally, the high-density integration of multiple components such as processors, memory, and RF front-ends within a system-in-package can easily lead to severe mutual interference. Summary of the Invention

[0008] The purpose of this invention is to provide an artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar, which aims to solve the problem that traditional monitoring methods cannot accurately measure extremely small displacement changes of targets.

[0009] The present invention is implemented as follows: an artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar includes the following steps: S1: Adaptive optimization strategy is adopted to dynamically adjust the transmission signal parameters of FMCW radar and acquire radar intermediate frequency signals; S2: Preprocess the acquired radar intermediate frequency signal and simultaneously extract its time-frequency domain physical features and deep learning abstract features, perform cross-domain feature fusion, and construct a fused feature vector; S3: Input the fused feature vector into the lightweight Transformer micro-vibration recognition model trained by the domain adaptive framework for processing to obtain micro-vibration detection and classification results; S4: Based on the signal complexity indicated by the micro-vibration detection and classification results, dynamically adjust the inference accuracy and computing power configuration of the micro-vibration recognition model to achieve a balance between detection accuracy and real-time performance.

[0010] Preferably, step S1 includes the following steps: S11: Calculate the signal-to-noise ratio of the received radar signal in real time and use it as the environmental state observation value of the reinforcement learning agent; S12: Construct a reinforcement learning model with the slope, bandwidth, and period of the Chirp signal as the action space and the signal-to-noise ratio improvement and data validity as the reward function; S13: Through the interaction and iterative optimization of the policy network between the agent and the radar signal environment, the current optimal Chirp signal modulation parameters are output. S14: Control the radar radio frequency front end according to the modulation parameters to generate and transmit FMCW signals adapted to the current environment and target characteristics.

[0011] Preferably, step S2 includes the following steps: S21: Perform wavelet packet transform on the preprocessed radar signal to decompose it into a preset frequency band; then extract the instantaneous amplitude, instantaneous frequency and energy envelope of each frequency band signal to form a physical feature vector; S22: Input the same preprocessed radar signal into a lightweight one-dimensional convolutional neural network to extract its high-level, abstract semantic features and form an abstract feature vector. S23: A multi-head attention mechanism is used to interact and assign weights to the physical feature vector and the abstract feature vector to obtain a weighted fusion feature representation; S24: The weighted physical features and abstract features are concatenated and then dimensionality reduced and integrated through a fully connected layer to output the final fused feature vector.

[0012] Preferably, in S3, the training process of the lightweight Transformer micro-vibration recognition model trained by the domain adaptive framework includes: The source domain is collected, which includes a large amount of unlabeled generalized scene radar signal data, and the target domain is collected, which includes a small amount of labeled specific micro-vibration scene data. Construct a classification model that takes fused feature vectors as input and uses a lightweight Transformer as its core architecture; In model training, the standard classification loss function is used, and a domain discriminator loss is introduced. At the same time, the gradient reversal layer is used to enable the model to learn domain-invariant features. By jointly optimizing the classification task and the domain adaptation task, the model can reduce the difference in feature distribution between the source domain and the target domain while retaining its ability to distinguish micro-vibration signals.

[0013] Preferably, in S3, the lightweight Transformer core architecture includes: using a one-dimensional convolutional layer to model local dependencies and compress sequences in the input fused feature vector; in, Use a grouped self-attention mechanism instead of the standard multi-head self-attention mechanism to reduce computational complexity; A depthwise separable convolutional structure is used in the linear transformation layer to further reduce the number of model parameters; Introducing gated linear units into feedforward neural networks enhances the model's nonlinear expressive power.

[0014] Preferably, in step S4, dynamically adjusting the inference accuracy and computing power configuration of the micro-vibration recognition model includes: Based on the micro-vibration detection and classification results, the time-domain variance and frequency-domain entropy of the signal are calculated as evaluation indicators of signal complexity. Preset complexity thresholds, including a first threshold and a second threshold; When the evaluation metric is below the first threshold, the shallow sub-network obtained by knowledge distillation is used for reasoning. When the evaluation metric is between the first and second thresholds, the standard lightweight Transformer model is enabled. When the evaluation metric exceeds the second threshold, an enhanced model with more attention heads and a wider feedforward network is enabled for inference.

[0015] Preferably, the shallow sub-network and the standard lightweight Transformer model are compressed using a combination of model quantization and knowledge distillation, including: Training phase: Using the enhanced model as the teacher model and the shallow sub-network or standard model as the student model, knowledge transfer is achieved by softening the output distribution and minimizing the KL divergence. Deployment phase: The weights and activation values ​​of the student model are quantized using INT8 to reduce storage and computational overhead.

[0016] Preferably, it includes a signal modulation module, a feature fusion module, an intelligent analysis module, and a dynamic reasoning module; in, S1 is executed through the signal modulation module; S2 is executed through the feature fusion module; S3 is executed through the intelligent analysis module; S4 is executed through the dynamic inference module.

[0017] Compared with existing technologies, the artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar provided by this invention has the following advantages: This invention achieves closed-loop dynamic optimization of transmission parameters through reinforcement learning, actively adapts to the environment, fundamentally improves the signal-to-noise ratio of micro-vibration signals, and lays a high-quality data foundation for subsequent analysis. This invention balances model performance and interpretability by fusing physical features with deep learning features; it adopts a domain-adaptive training framework, which effectively solves the problem of model incompatibility in new scenarios, significantly improves practicality and deployment scope, and breaks through the traditional paradigm that high precision inevitably means high time consumption.

[0018] This invention achieves the optimal configuration of accuracy and speed under limited computing power through a signal complexity self-sensing and model dynamic switching mechanism, making it possible to run high-precision micro-vibration detection in real time on low-cost embedded devices.

[0019] This invention does not improve a single link in isolation, but rather carries out end-to-end collaborative innovation from radar waveforms, feature engineering, AI models to inference engines, forming a powerful technological synergy. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to the present invention. Figure 2 This is a schematic diagram illustrating the implementation process of adjusting the transmission signal parameters using the adaptive optimization strategy in this invention. Figure 3 This is a schematic diagram of the training process of the lightweight Transformer micro-vibration recognition model in this invention. Figure 4 This is a flowchart illustrating the logic of dynamically adjusting model inference accuracy and computing power configuration in this invention. Figure 5 This is a logic flowchart of the complete encoder-decoder structure of the lightweight Transformer in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0023] like Figure 1-5 As shown, this invention provides an artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar, comprising the following steps: S1: Adaptive optimization strategy is adopted to dynamically adjust the transmission signal parameters of FMCW radar and acquire radar intermediate frequency signals; S2: Preprocess the acquired radar intermediate frequency signal and simultaneously extract its time-frequency domain physical features and deep learning abstract features, perform cross-domain feature fusion, and construct a fused feature vector; S3: Input the fused feature vector into the lightweight Transformer micro-vibration recognition model trained by the domain adaptive framework for processing to obtain micro-vibration detection and classification results; S4: Based on the signal complexity indicated by the micro-vibration detection and classification results, dynamically adjust the inference accuracy and computing power configuration of the micro-vibration recognition model to achieve a balance between detection accuracy and real-time performance. Example

[0024] S1 specifically includes: S11: Calculate the signal-to-noise ratio of the received radar signal in real time and use it as the environmental state observation value of the reinforcement learning agent; S12: Construct a reinforcement learning model with the slope, bandwidth, and period of the Chirp signal as the action space and the signal-to-noise ratio improvement and data validity as the reward function; S13: Through the interaction and iterative optimization of the policy network between the agent and the radar signal environment, the current optimal Chirp signal modulation parameters are output. S14: Control the radar radio frequency front end according to the modulation parameters to generate and transmit FMCW signals adapted to the current environment and target characteristics.

[0025] In this embodiment S11: The signal-to-noise ratio (SNR) calculation is based on the intermediate frequency (IF) signal received by the radar. The received signal is first filtered (using a Butterworth low-pass filter, with the cutoff frequency set according to the highest frequency of the target's micro-vibration, usually 100Hz-1kHz) to remove high-frequency clutter interference.

[0026] The effective signal amplitude is determined by extracting the peak value of the filtered signal, and the noise amplitude is calculated by statistically analyzing the mean and standard deviation of the signal amplitude during idle periods (times when there is no target micro-vibration or the radar does not transmit signals). The signal-to-noise ratio is the ratio of the effective signal amplitude to the noise amplitude.

[0027] The signal-to-noise ratio (SNR) calculation cycle is synchronized with the Chirp signal cycle. That is, after each set of Chirp signals is transmitted and received, an SNR calculation is immediately performed to ensure the real-time nature of environmental condition observation.

[0028] In addition to the signal-to-noise ratio, the observed values ​​can be supplemented with the interference ratio (the ratio of the interference signal amplitude to the total signal amplitude) as auxiliary observation information to improve the accuracy of the agent's judgment of the environmental state. This auxiliary observation information does not change the core observation logic, but only enhances the comprehensiveness of environmental perception.

[0029] In this embodiment S12: Action space definition: The action space consists of three core modulation parameters of the Chirp signal: slope, bandwidth, and period. Each parameter is given a reasonable range of values ​​and adjustment granularity to ensure the feasibility and effectiveness of parameter adjustment.

[0030] The slope value ranges from 10MHz / μs to 100MHz / μs, and the adjustment granularity is 5MHz / μs, which can be flexibly adapted according to the hardware performance of the radar RF front end. The bandwidth range is 1GHz-5GHz, with an adjustment granularity of 500MHz, balancing distance resolution and signal coverage. The period range is 100μs-1ms, and the adjustment granularity is 50μs, balancing signal acquisition efficiency and data integrity.

[0031] Reward function design: The reward function adopts a weighted summation form, comprehensively considering the two core objectives of improving the signal-to-noise ratio and data validity. The specific design logic is as follows: Signal-to-noise ratio bonus: If the current signal-to-noise ratio (SNR) increases compared to the previous cycle, a positive reward is given; the greater the increase, the higher the positive reward value. If the SNR decreases, a negative reward is given; the greater the decrease, the higher the negative reward value. Data validity reward items: Data validity is assessed by determining whether the received signal is distorted (whether the signal amplitude exceeds the dynamic range of the radar receiver) and whether the interference ratio is lower than a preset threshold (usually set to 20%). If the signal is not distorted and the interference ratio is lower than the threshold, a positive reward is given; otherwise, a negative reward is given. Weighting: The signal-to-noise ratio (SNR) reward accounts for 70% of the weight, while the data validity reward accounts for 30%. The weights can be adjusted according to the priority requirements of the actual application scenario, but both indicators must be taken into consideration.

[0032] The reinforcement learning agent uses the proximal policy optimization (PPO) algorithm to construct the policy network. The network structure includes an input layer, two hidden layers (each hidden layer has 64 neurons) and an output layer. The activation function is the ReLU function, and the output layer is a softmax layer, which outputs the selection probability of each action (i.e., the combination of Chirp signal modulation parameters).

[0033] In this embodiment S13: In the initial stage of iterative optimization, the agent randomly selects a combination of modulation parameters in the action space, transmits a signal and receives feedback (signal-to-noise ratio and data validity), and calculates the reward value. Every 100 signal transmission-reception-feedback interaction cycles are completed, the policy network is updated once, with the update objective being to maximize the cumulative reward value. The convergence condition for iterative optimization is: within 50 consecutive interaction cycles, the fluctuation range of the cumulative reward value is less than 5%. At this time, it is considered that the policy network has learned the optimal policy that adapts to the current environment and target characteristics, the iteration stops and the network parameters are fixed; if the environment changes significantly (such as the signal-to-noise ratio changing by more than 30%), the iterative optimization process is restarted.

[0034] In this embodiment S14: The optimal Chirp signal modulation parameters (slope, bandwidth, period) output by the strategy network are transmitted to the signal generator module of the radio frequency front end through the radar's control interface (such as SPI interface, I2C interface); The RF front end generates the corresponding FMCW signal based on the received modulation parameters: first, a frequency sweep signal that meets the slope and bandwidth requirements is generated by a voltage-controlled oscillator (VCO), and then the transmission period of the signal is controlled by a timer to ensure that the signal parameters are completely consistent with the output of the policy network. During signal transmission, the radio frequency front-end provides real-time feedback on the transmission status (such as whether the signal is transmitted normally and whether the parameters are loaded accurately) to the reinforcement learning agent. If a transmission anomaly occurs, the agent immediately triggers re-optimization and outputs suboptimal modulation parameters to ensure the stability of the system. Example

[0035] Specifically, S2 includes: S21: Perform wavelet packet transform on the preprocessed radar signal to decompose it into a preset frequency band; then extract the instantaneous amplitude, instantaneous frequency and energy envelope of each frequency band signal to form a physical feature vector; S22: Input the same preprocessed radar signal into a lightweight one-dimensional convolutional neural network to extract its high-level, abstract semantic features and form an abstract feature vector. S23: A multi-head attention mechanism is used to interact and assign weights to the physical feature vector and the abstract feature vector to obtain a weighted fusion feature representation; S24: The weighted physical features and abstract features are concatenated and then dimensionality reduced and integrated through a fully connected layer to output the final fused feature vector.

[0036] In this embodiment S21: Wavelet packet transform parameter configuration: The db4 wavelet basis function is selected as the transform kernel, and the number of decomposition layers is set according to the frequency range of the target micro-vibration (usually 3-5 layers) to ensure that each frequency band after decomposition can completely cover the characteristic frequency band of the target micro-vibration (e.g., 0.1Hz-1kHz). During the decomposition process, the signal is bidirectionally split into high-frequency and low-frequency components in each layer, ultimately obtaining 2^N equal-width frequency bands (N is the number of decomposition layers), avoiding the omission of characteristic frequency bands.

[0037] Specific methods for extracting physical features: Instantaneous amplitude: The absolute value of the wavelet packet coefficients of each frequency band is extracted time-by-time as the instantaneous amplitude, reflecting the intensity change of the signal in that frequency band; Instantaneous frequency: The signal in each frequency band is analyzed by Hilbert transform to obtain the phase information of the signal. Then, the phase is differentiated and normalized to obtain the instantaneous frequency, which describes the fluctuation law of the signal frequency over time; Energy envelope: The instantaneous amplitude of each frequency band signal is smoothed by a sliding window (the window length is 5-10 times the signal sampling period), and the smoothed sequence is taken as the energy envelope to highlight the overall trend of signal energy change.

[0038] Physical feature vector integration: The instantaneous amplitude sequence, instantaneous frequency sequence, and energy envelope sequence of all frequency bands are spliced ​​together in frequency band order to form a one-dimensional physical feature vector. The vector length is the number of frequency bands × 3 × the number of signal sampling points, ensuring that the time-frequency domain physical information of each frequency band is completely preserved.

[0039] In this embodiment S22: Overall network structure: The network has a 4-layer architecture (in order: input layer, convolutional layer 1, convolutional layer 2, pooling layer and output layer), with the total number of parameters controlled within 100,000, adapting to the computing power constraints of embedded devices and meeting the requirements of lightweight design.

[0040] Specific configuration of each layer: Input layer: Receives the preprocessed one-dimensional radar signal sequence, and the input dimension is consistent with the number of signal sampling points (e.g., 1024 points). Convolutional layer 1: The number of convolutional kernels is 32, the kernel size is 3, the stride is 1, zero padding is used to keep the output dimension consistent with the input, and the activation function is ReLU, which is used to extract the local basic features of the signal; Convolutional layer 2: The number of convolutional kernels is 64, the kernel size is 5, the stride is 1, zero padding is also used, and the activation function is ReLU to further extract high-level combined features; Pooling layer: Max pooling is used with a pooling kernel size of 4 and a stride of 4. The output of the convolutional layer is downsampled to reduce parameter redundancy. Output layer: The output of the pooling layer is transformed into a fixed-dimensional one-dimensional vector (such as 256-dimensional) through global average pooling, which serves as an abstract feature vector containing high-level semantic pattern information of the signal.

[0041] Network training instructions: This lightweight 1D convolutional neural network is trained jointly with the subsequent micro-vibration recognition model, without separate pre-training, to ensure the adaptability of feature extraction to the final recognition task.

[0042] In this embodiment S23: Attention head configuration: Set up 4-8 independent attention heads, each responsible for capturing the correlation between physical feature vectors and abstract feature vectors in different dimensions, thereby improving the comprehensiveness of feature interaction.

[0043] Interaction and weight calculation logic: First, the physical feature vector and the abstract feature vector are mapped to the same dimensional space (e.g., unified to 256 dimensions) to ensure the feasibility of attention calculation; each attention head calculates the relevance score of each element in the two sets of feature vectors. The higher the relevance score, the greater the weight of the corresponding element, and vice versa; after weight allocation, the two sets of feature vectors are weighted and summed to obtain the weighted physical feature fragment and weighted abstract feature fragment corresponding to each attention head.

[0044] Multi-head output fusion: The output segments of all attention heads are concatenated along the channel dimension and then integrated through a lightweight linear layer to obtain a unified weighted fusion feature representation, which preserves the correlation information of each head and avoids feature redundancy.

[0045] In this embodiment S24: Feature concatenation: The weighted physical feature vector and the abstract feature vector are directly concatenated according to their dimensions to form a concatenated feature vector. The dimensions of the concatenated feature vector are the physical feature vector dimension + the abstract feature vector dimension (e.g., 1024 + 256 = 1280 dimensions).

[0046] Fully connected layer configuration: The fully connected layer consists of one hidden layer (128 neurons) and one output layer (64-128 neurons, set according to the input requirements of the subsequent recognition model). The activation function of the hidden layer is ReLU, and the output layer has no activation function.

[0047] Dimensionality reduction and integration effect: Through linear transformation and nonlinear mapping of fully connected layers, redundant information in spliced ​​features is eliminated, while the most critical core features for micro-vibration recognition are retained. The final output is a fusion feature vector with fixed dimensions and high recognizability, providing efficient input for subsequent model inference. Example

[0048] In S3, the lightweight Transformer micro-vibration recognition model trained using the domain adaptive framework includes the following training process: The source domain is collected, which includes a large amount of unlabeled generalized scene radar signal data, and the target domain is collected, which includes a small amount of labeled specific micro-vibration scene data. Construct a classification model that takes fused feature vectors as input and uses a lightweight Transformer as its core architecture; In model training, the standard classification loss function is used, and a domain discriminator loss is introduced. At the same time, the gradient reversal layer is used to enable the model to learn domain-invariant features. By jointly optimizing the classification task and the domain adaptation task, the model can reduce the difference in feature distribution between the source domain and the target domain while retaining its ability to distinguish micro-vibration signals.

[0049] In this embodiment: The data collection scenarios cover a variety of environments that may be involved in the micro-vibration of the target, including but not limited to: radar signals under different electromagnetic interference intensities (such as urban commercial areas, industrial areas, and suburbs), different temperatures and humidity (-10℃-40℃, 30%-80% humidity), and different target distances (1m-50m).

[0050] The data type is the preprocessed sequence of unlabeled raw intermediate frequency signals, with a total sample size of no less than 100,000 to ensure coverage of sufficiently rich environmental and target change patterns. The duration of each sample is set according to the micro-vibration characteristic period (e.g., 1s-5s), and the sampling rate is adapted to the radar transmission parameters (e.g., 10kHz-100kHz).

[0051] Target domain data: Data is collected for specific application scenarios (such as bridge micro-vibration monitoring and precision equipment vibration detection), with a sample size of 1000-5000 records (far fewer than the source domain), and each sample is accompanied by label information. Label types include: micro-vibration type (e.g., structural vibration, equipment resonance, environmental interference), vibration amplitude level (e.g., weak / moderate / strong), and vibration frequency range (e.g., low frequency / high frequency). The data acquisition environment is consistent with the actual deployment scenario to ensure that the features learned by the model directly adapt to application requirements.

[0052] Input and Structure Adaptation of Lightweight Transformer Classification Model - Input Layer: Receives fused feature vectors (dimensions such as 64-128), enhances sequence temporal information through positional encoding (using sine and cosine positional encoding to avoid absolute positional dependence), and forms the model input sequence (length equal to the feature vector dimension, such as 128).

[0053] Core architecture: It adopts the lightweight Transformer structure (including 1D convolutional local modeling, group self-attention, etc.), the output layer is a classification head, and the softmax activation function is used. The output dimension is consistent with the number of target domain label categories (e.g., if there are 4 categories of micro-vibrations, the output is 4-dimensional), directly predicting the probability distribution of micro-vibration categories.

[0054] The specific implementation of the domain discriminator loss and gradient reversal layer - Domain discriminator structure: It is a lightweight network independent of the classification model. The input is the intermediate features of the lightweight Transformer (i.e., the domain-independent feature candidates output by the feature extractor), and the output is the binary classification result (determining whether the feature comes from the source domain or the target domain). The network structure consists of two fully connected layers (64 neurons in the first layer and 1 neuron in the second layer), with ReLU and sigmoid activation functions respectively. The total number of parameters is kept below 10,000 to avoid excessive computational overhead.

[0055] The gradient inversion layer works as follows: It is embedded between the feature extractor and the domain discriminator. During forward propagation, the features are not processed (the features are directly passed to the domain discriminator). During backward propagation, the loss gradient of the domain discriminator is multiplied by a negative coefficient (e.g., -1) before being passed to the feature extractor. This mechanism forces the feature extractor to learn features that make it difficult for the domain discriminator to distinguish the source (i.e., domain-invariant features), while not affecting the classification head's ability to distinguish micro-vibration categories.

[0056] Domain discriminator loss calculation: The binary cross-entropy loss is used to compare the prediction results (source domain / target domain) of the domain discriminator with the true domain labels (source domain is marked as 0, target domain is marked as 1). The smaller the loss value, the more accurate the domain discriminator is in judging the source of the feature (conversely, the greater the loss value, the smaller the domain difference of the feature).

[0057] Joint optimization process for classification and domain adaptation tasks – loss function combination: The total loss is a weighted sum of the classification loss and the domain discriminator loss. The classification loss uses cross-entropy loss (calculating only the prediction error of labeled samples in the target domain), and the domain discriminator loss uses the aforementioned binary cross-entropy loss (calculating the domain judgment error of all source and target domain samples). The weight ratio is dynamically adjusted according to the training phase: In the initial phase (first 10 epochs), the classification loss weight is 0.7, and the domain discriminator loss weight is 0.3, ensuring that the model first masters basic classification ability; in subsequent phases (after 10 epochs), the weights are adjusted to 0.5 for the classification loss and 0.5 for the domain discriminator loss to enhance domain adaptation ability.

[0058] Training process: Randomly initialize the parameters of the lightweight Transformer (including feature extractor and classifier) ​​and the domain discriminator; in each iteration, randomly sample an equal number of samples (e.g., 32 samples each) from the source domain (unlabeled) and the target domain (labeled), and input them into the model after feature fusion; the feature extractor outputs features, which are fed into the classifier (to calculate the classification loss of the target domain samples) and the domain discriminator (to calculate the domain discriminator loss of all samples through a gradient inversion layer); calculate the total loss, and update the parameters of the feature extractor, classifier, and domain discriminator through backpropagation (where the feature extractor is driven by both the classification loss and the inverted domain discriminator loss); repeat the iteration until the total loss converges (loss fluctuation is less than 3% for 10 consecutive epochs), and save the model parameters. Example

[0059] In S3, the lightweight Transformer core architecture includes: One-dimensional convolutional layers are used to model local dependencies and compress sequences in the input fused feature vector; in, Use a grouped self-attention mechanism instead of the standard multi-head self-attention mechanism to reduce computational complexity; A depthwise separable convolutional structure is used in the linear transformation layer to further reduce the number of model parameters; Introducing gated linear units into feedforward neural networks enhances the model's nonlinear expressive power.

[0060] In this embodiment: Local dependency modeling and sequence compression of one-dimensional convolutional layers - Convolutional layer configuration: One-dimensional convolutional layer is used as the input preprocessing module of Transformer; The specific parameters are: the number of convolutional kernels is 32-64, that is, it is adapted according to the dimension of the fused feature vector, such as 64-dimensional features corresponding to 64 convolutional kernels; The kernel size is 3-5, meaning it covers the local correlations of 3-5 adjacent elements in the feature vector; Step size 2 is used to achieve sequence compression; The zero-padding method is set to "same" to ensure that edge features are not lost.

[0061] Local dependency modeling logic: By using the sliding window operation of the convolution kernel, the correlation between adjacent elements in the fused feature vector (such as the local trend of energy changes and frequency fluctuations at adjacent time points) is captured, replacing the inefficient modeling of local information by the Transformer and enhancing the ability to capture the local temporal features of micro-vibration signals.

[0062] Sequence compression effect: The input fusion feature vector has a length of L (e.g., 128). After convolution with a stride of 2, the output sequence length is compressed to L / 2 (e.g., 64), which directly reduces the computational cost of the subsequent attention mechanism. The computational complexity of the attention mechanism is proportional to the square of the sequence length, thus achieving length reduction and burden reduction.

[0063] The implementation of the grouped self-attention mechanism includes: Grouping method: The feature sequence (dimension D, such as 64) after convolutional layer processing is divided into groups according to channel dimension. The number of groups is set to 2-4 (such as 64 dimensions divided into 4 groups, each with 16 dimensions). Self-attention is calculated independently for each group, replacing the mode in standard multi-head self-attention where each head processes full-dimensional features.

[0064] The difference between standard multi-head attention and grouped self-attention is that in standard multi-head attention, each head needs to perform a linear transformation on all-dimensional features before calculating attention, while in grouped self-attention, each group only processes its own sub-dimensional features, avoiding redundant calculations across groups.

[0065] For example, the computational cost of attention for 4 groups of 16-dimensional features is only 1 / 4 of the computational cost for 4 heads of 64-dimensional features (because the dimension of each group is reduced to 1 / 4).

[0066] Attention calculation process: First, the query (Q), key (K), and value (V) vectors are calculated within each group (generated through a lightweight linear layer within the group). Then, the attention weights of the elements within the group are calculated (reflecting the correlation strength of the features within the group). Finally, the outputs of each group are concatenated by channel and output through an integrated linear layer to ensure the integrity of the features after grouping.

[0067] Depth-separable convolutional structures with linear transform layers include: Alternative logic: Traditional linear transformation layers in Transformer (such as Q / K / V generation and attention output integration layers) are implemented through fully connected layers, with the number of parameters being the input dimension × the output dimension, resulting in high computational overhead.

[0068] This implementation uses depthwise separable convolutions instead of fully connected layers, decomposing cross-channel mapping into independent transformations within channels and cross-channel aggregation, thus significantly reducing parameters.

[0069] In this embodiment: Deep convolution: Each feature group is equipped with one convolution kernel (kernel size 1, only applied to a single channel) to achieve linear transformation within the channel (e.g., a 16-dimensional feature group corresponds to 16 1×1 convolution kernels, processing each channel separately). Point convolution: Uses a 1×1 convolution kernel (the number is the same as the output dimension) to aggregate the output of depth convolution across channels, integrating the correlation information of each group of features.

[0070] Parameter simplification effect: Taking a 64-dimensional input and 64-dimensional output as an example, the number of parameters in a traditional fully connected layer is 64×64=4096; In depthwise separable convolution, the number of parameters for depthwise convolution is 64×1×1=64, and the number of parameters for pointwise convolution is 64×1×1=64, for a total of 128 parameters, which is only 3.1% of that of the traditional method.

[0071] Gated linear units in a feedforward neural network include: Unit Structure: The gated linear unit (GLU) is composed of parallel linear transformation branches and gated branches, specifically: Linear transformation branch: Performs a linear transformation on the input features (e.g., mapping 64-dimensional features to 128-dimensional features); the activation function is ReLU, which is responsible for the non-linear transformation of the features; Gated branch: Perform another linear transformation on the same input feature (with the same dimension as the transformation branch, such as 64-128), with the activation function being Sigmoid (outputting weights of 0-1), to generate a gated signal; Output: Multiply the output of the transform branch by the output of the gate branch element by element, retaining only the feature components with higher weights (e.g., >0.5) in the gated signal, and filtering out redundant information.

[0072] The principle of enhancing nonlinear representation: Compared with the linear transformation and single activation function mode in traditional feedforward networks, GLU dynamically adjusts feature transmission through gating signals, enabling the model to adaptively focus on key features (such as the sudden change time and characteristic frequency band of micro-vibration signals). At the same time, it introduces the combined nonlinearity of Sigmoid and ReLU to enhance the ability to characterize complex vibration modes. Example

[0073] In S4, the inference accuracy and computing power configuration of the micro-vibration recognition model are dynamically adjusted, including: Based on the micro-vibration detection and classification results, the time-domain variance and frequency-domain entropy of the signal are calculated as evaluation indicators of signal complexity. Preset complexity thresholds, including a first threshold and a second threshold; When the evaluation metric is below the first threshold, the shallow sub-network obtained by knowledge distillation is used for reasoning. When the evaluation metric is between the first and second thresholds, the standard lightweight Transformer model is enabled. When the evaluation metric exceeds the second threshold, an enhanced model with more attention heads and a wider feedforward network is enabled for inference.

[0074] In this embodiment: Threshold determination principle: The first threshold is: the low complexity critical value; The second threshold is: the high complexity critical value; The first and second thresholds need to be determined based on the distribution of micro-vibration characteristics in the specific application scenario. The complexity range of typical samples needs to be statistically analyzed through experiments to ensure that the thresholds can effectively distinguish between simple, medium, and complex signals.

[0075] In this embodiment: Experiment setup procedure: Collect typical micro-vibration signals in the target scene, including at least 1000 samples, covering known simple, medium and complex types; Calculate the time-domain variance and frequency-domain entropy of each sample, and plot a two-dimensional distribution scatter plot with variance on the horizontal axis and entropy on the vertical axis. Regions are divided based on the natural clustering of scatter points: The signals densely distributed in the lower left corner are simple signals, including those with low variance and low entropy. The densely distributed signals in the upper right corner are complex signals, including high variance and high entropy. The middle region represents a medium signal. The boundary values ​​of the clustering regions are used as thresholds. For example, the boundary between simple signals and medium signals is the first threshold, and the boundary between medium and complex signals is the second threshold. A buffer of 5%-10% is reserved to avoid frequent switching near the critical value.

[0076] Example scenario reference: In micro-vibration monitoring of precision equipment: The first threshold can be set as follows: time domain variance < 0.02 and frequency domain entropy < 1.5, to correspond to slight vibrations during stable operation of the equipment; The second threshold is set as follows: time domain variance > 0.1 and frequency domain entropy > 3.0, to correspond to abnormal multi-frequency vibrations of the equipment.

[0077] In bridge micro-vibration monitoring, due to greater environmental interference, the threshold can be appropriately relaxed. For example, the time domain variance can be <0.05 in the first threshold and >0.2 in the second threshold.

[0078] Activation logic and switching mechanisms for different models - model types and applicable scenarios: Shallow sub-network: A lightweight model containing only 2-3 convolutional layers and 1 fully connected layer, with fewer than 100,000 parameters. It is suitable for simple signals, i.e., the complexity is below the first threshold. The inference speed is 3-5 times faster than the standard model, which can meet the requirements of millisecond-level response, such as the monitoring of small displacements of static targets.

[0079] Standard lightweight Transformer model: With 300,000 to 500,000 parameters, it is suitable for medium-complexity signals (between the first and second thresholds), balancing accuracy and speed (such as stable vibration identification of conventional equipment operation). Enhanced model: Add 2-4 attention heads (e.g., from 4 to 8 heads) and widen the feedforward network dimension (e.g., from 128 to 256 dimensions) on the basis of the standard model, with 800,000 to 1,000,000 parameters. It is suitable for complex signals (above the second threshold) and ensures recognition accuracy through stronger feature modeling capabilities (e.g., multi-target mixed vibration, weak vibration under strong interference).

[0080] Real-time switching mechanism: After signal acquisition and preprocessing, the time domain variance and frequency domain entropy are calculated and compared with preset thresholds; based on the comparison results, the corresponding model is called from the model library for inference (the model library is preloaded into memory to avoid loading delay). To prevent frequent switching (e.g., more than 3 switching times within 1 second) caused by fluctuations in signal complexity around the threshold, a hysteresis judgment is set: if the current model is a shallow sub-network, the complexity must exceed the first threshold for 3 consecutive windows before switching to the standard model; similarly, when switching to the enhanced model or reverting to the low-complexity model, the condition must be met for 3 consecutive windows to ensure system stability. Example

[0081] In S4, shallow subnetworks and the standard lightweight Transformer model are compressed using a combination of model quantization and knowledge distillation, including: Training phase: Using the enhanced model as the teacher model and the shallow sub-network or standard model as the student model, knowledge transfer is achieved by softening the output distribution and minimizing the KL divergence. Deployment phase: The weights and activation values ​​of the student model are quantized using INT8 to reduce storage and computational overhead.

[0082] In this embodiment, a two-level compression strategy of knowledge distillation transfer discrimination capability and INT8 quantization simplifies storage and computation is adopted to significantly reduce the number of parameters and computational overhead while ensuring the recognition accuracy of shallow sub-networks and standard lightweight Transformer models (hereinafter collectively referred to as student models).

[0083] The core idea is to leverage the generalization ability and feature discrimination experience of the augmented model (teacher model) to guide the student model to learn key knowledge by softening the output distribution; Then, through numerical precision compression, the model is converted from floating-point format to low-precision integer format, ultimately enabling efficient deployment of the student model on embedded devices.

[0084] Knowledge distillation process during the training phase - the roles of the teacher model and the student model: The teacher model is an augmented model (800,000-1,000,000 parameters), which has been fully trained and has the ability to identify complex micro-vibration signals with high precision (accuracy ≥ 95%). Its output includes fine-grained discrimination information on micro-vibration categories (such as the probability distribution of different vibration modes).

[0085] The student model includes a shallow sub-network (less than 100,000 parameters) and a standard lightweight Transformer model (300,000-500,000 parameters). It needs to learn the discriminative experience of the teacher model through distillation to approximate the accuracy of the teacher model with fewer parameters.

[0086] The specific implementation of softened output distribution: The output of the teacher model (raw predicted probability) is softened by a temperature parameter, which ranges from 1 to 10. The temperature was adjusted according to the complexity of the student model: a higher temperature of 5-10 was used for shallow sub-networks, and a lower temperature of 1-5 was used for the standard model.

[0087] In the teacher model, the higher the temperature, the flatter the output probability distribution of the teacher model, which better reflects the subtle differences between different categories. This makes it easier for the student model to learn the teacher's discrimination logic for atypical samples, that is, the micro-vibration signal close to the boundary between the two classes.

[0088] For example, when the temperature is 1, the output might have a 90% probability for category A and a 10% probability for category B; when the temperature is 5, it might become: a 80% probability for category A and a 20% probability for category B.

[0089] In this embodiment: the training process for knowledge transfer: The teacher model parameters are fixed (and will not be updated); only the student model is trained. Input the same batch of fused feature vectors (from labeled data in the target domain and unlabeled data in the source domain), and obtain the softened output (processed with temperature parameters) of the teacher model and the original output of the student model respectively; Calculate the difference between the student model output and the teacher's softened output (measured by KL divergence; the smaller the difference, the closer the student is to the teacher's discrimination logic), while preserving the classification loss between the student model and the real label (to ensure that the student model does not deviate from the task objective). The total loss is a weighted sum of the KL divergence loss (70% weight) and the classification loss (30% weight), and the student model parameters are updated through backpropagation. Iterative training continues until the total loss converges (loss fluctuation <3% over 10 consecutive epochs). At this point, the student model has learned the key discriminative knowledge of the teacher model, and the accuracy can usually reach 90%-95% of that of the teacher model.

[0090] In this embodiment: INT8 quantization process during the deployment phase - quantization object and target: Quantization is used for the student model's weight parameters (fixed parameters learned during model training) and activation values ​​(intermediate results of input features calculated by each layer during inference), converting the original 32-bit floating-point numbers (float32) into 8-bit integers (int8). The goal is to compress the model's storage size to 1 / 4 of its original size (because int8 occupies 1 / 4 the space of float32) while keeping accuracy loss controllable (recognition accuracy decrease ≤3%), and at the same time reducing computational overhead (integer operations are 2-5 times faster than floating-point operations, adapting to the integer computing units of embedded devices).

[0091] Quantization Method and Calibration Process: An offline calibration quantization strategy is adopted. The specific steps are as follows: Prepare a calibration dataset (randomly select 100-500 typical samples from the target domain data, covering various micro-vibration scenarios); Perform pseudo-quantization simulation on the student model: Record the range of weights and activation values ​​(maximum and minimum values) during the inference process without changing the parameter values; Determine the quantization mapping relationship based on the value range: Linearly map the range of float32 values ​​to the [-128, 127] interval of int8 (e.g., if the weight range of a certain layer is [-2.0, 2.0], the mapping formula is int8 value = (float32 value / 2.0) × 127, ensuring that key values ​​do not overflow); Verify the accuracy of the quantized model using the calibration dataset. If the accuracy decreases by more than 3%, adjust the mapping range (e.g., narrow the range to retain more critical value intervals) and requantize until the requirements are met; Generate the quantized int8 model file and replace the original float32 model for deployment.

[0092] Quantized inference adaptation: The inference engine of embedded devices needs to support int8 calculation instructions (such as ARM's NEON instruction set and DSP's MAC unit) to convert convolution, fully connected operations, etc., of each layer of the model into integer multiplication and addition, avoiding the format conversion overhead during quantization and inference. For layers sensitive to quantization errors (such as the output layer), some float32 calculations can be retained (quantization only for intermediate layers) to balance accuracy and efficiency. Example

[0093] An AI-based micro-vibration detection method based on FMCW millimeter-wave radar includes a signal modulation module, a feature fusion module, an intelligent analysis module, and a dynamic reasoning module. S1 is executed through the signal modulation module; S2 is executed through the feature fusion module; S3 is executed through the intelligent analysis module; S4 is executed through the dynamic inference module.

[0094] In this embodiment, the signal modulation module, feature fusion module, intelligent analysis module, and dynamic reasoning module described above are all or partly implemented through software, hardware, firmware, or other arbitrary combinations. When implemented using software, the above embodiments can be implemented in whole or in part as a computer program product.

[0095] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0096] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0097] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0098] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An artificial intelligence micro-vibration detection method based on FMCW millimeter wave radar, characterized in that, Includes the following steps: S1: Adaptive optimization strategy is adopted to dynamically adjust the transmission signal parameters of FMCW radar and acquire radar intermediate frequency signals; S2: Preprocess the acquired radar intermediate frequency signal and simultaneously extract its time-frequency domain physical features and deep learning abstract features, perform cross-domain feature fusion, and construct a fused feature vector; S2 includes the following steps: S21: Perform wavelet packet transform on the preprocessed radar signal to decompose it into a preset frequency band; then extract the instantaneous amplitude, instantaneous frequency and energy envelope of the signal in each frequency band to form a physical feature vector; S22: Input the same preprocessed radar signal into a lightweight one-dimensional convolutional neural network to extract its high-level, abstract semantic features and form an abstract feature vector. S23: A multi-head attention mechanism is used to interact and assign weights to physical feature vectors and abstract feature vectors to obtain a weighted fusion feature representation; S24: The weighted physical features and abstract features are concatenated and then dimensionality reduced and integrated through a fully connected layer to output the final fused feature vector; S3: Input the fused feature vector into the lightweight Transformer micro-vibration recognition model trained by the domain adaptive framework for processing to obtain micro-vibration detection and classification results; S4: Based on the signal complexity indicated by the micro-vibration detection and classification results, dynamically adjust the inference accuracy and computing power configuration of the micro-vibration recognition model to achieve a balance between detection accuracy and real-time performance.

2. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 1, characterized in that, The S1 includes the following steps: S11: Calculate the signal-to-noise ratio of the received radar signal in real time and use it as the environmental state observation value of the reinforcement learning agent; S12: Construct a reinforcement learning model with the slope, bandwidth, and period of the Chirp signal as the action space and the signal-to-noise ratio improvement and data validity as the reward function; S13: The strategy network is iteratively optimized through interaction between the agent and the radar signal environment, and the current optimal Chirp signal modulation parameters are output. S14: Control the radar radio frequency front end according to the modulation parameters to generate and transmit FMCW signals adapted to the current environment and target characteristics.

3. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 2, characterized in that, In S3, the training process of the lightweight Transformer micro-vibration recognition model trained by the domain adaptive framework includes: The source domain is collected, which includes a large amount of unlabeled generalized scene radar signal data, and the target domain is collected, which includes a small amount of labeled specific micro-vibration scene data. Construct a classification model that takes fused feature vectors as input and uses a lightweight Transformer as its core architecture; In model training, the standard classification loss function is used, and a domain discriminator loss is introduced. At the same time, the gradient reversal layer is used to enable the model to learn domain-invariant features. By jointly optimizing the classification task and the domain adaptation task, the model can reduce the difference in feature distribution between the source domain and the target domain while retaining its ability to distinguish micro-vibration signals.

4. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 3, characterized in that, In S3, the lightweight Transformer core architecture includes: using one-dimensional convolutional layers to model local dependencies and compress sequences in the input fused feature vector; in, Use a grouped self-attention mechanism instead of the standard multi-head self-attention mechanism to reduce computational complexity; A depthwise separable convolutional structure is used in the linear transformation layer to further reduce the number of model parameters; Introducing gated linear units into feedforward neural networks enhances the model's nonlinear expressive power.

5. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 4, characterized in that, In S4, the inference accuracy and computing power configuration of the micro-vibration recognition model are dynamically adjusted, including: Based on the micro-vibration detection and classification results, the time-domain variance and frequency-domain entropy of the signal are calculated as evaluation indicators of signal complexity. Preset complexity thresholds, including a first threshold and a second threshold; When the evaluation metric is below the first threshold, the shallow sub-network obtained by knowledge distillation is used for reasoning. When the evaluation metric is between the first and second thresholds, the standard lightweight Transformer model is enabled. When the evaluation metric exceeds the second threshold, an enhanced model with more attention heads and a wider feedforward network is enabled for inference.

6. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 5, characterized in that, The shallow sub-network and the standard lightweight Transformer model are compressed using a combination of model quantization and knowledge distillation, including: Training phase: Using the enhanced model as the teacher model and the shallow sub-network or standard model as the student model, knowledge transfer is achieved by softening the output distribution and minimizing the KL divergence. Deployment phase: The weights and activation values ​​of the student model are quantized using INT8 to reduce storage and computational overhead.

7. The artificial intelligence micro-vibration detection method based on FMCW millimeter-wave radar according to claim 6, characterized in that, It includes a signal modulation module, a feature fusion module, an intelligent analysis module, and a dynamic reasoning module; in, S1 is executed through the signal modulation module; S2 is executed through the feature fusion module; S3 is executed through the intelligent analysis module; S4 is executed through the dynamic inference module.

Images