All-optical dual resonance and deep network gas feature decoupling detection method and system
By employing an all-optical dual-resonance and deep network gas feature decoupling detection method, the problems of low signal-to-noise ratio and baseline drift under harsh operating conditions are solved, achieving accurate extraction of gas features and stable concentration inversion, improving the signal-to-noise ratio and maintaining the stability of concentration readings.
Patent Information
- Application Number
- CN202611190165.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-06
- Publication Date
- 2026-11-10
AI Technical Summary
Under extremely harsh operating conditions, existing technologies inevitably couple external acoustic vibrations and electromagnetic noise into the photoacoustic spectroscopy detection system, resulting in low signal-to-noise ratio, severe baseline drift, and difficulty in achieving accurate extraction of gas characteristics and stable concentration inversion.
A gas feature decoupling detection method using all-optical dual resonance and deep network is adopted. By synchronously acquiring the main path photoacoustic signal and auxiliary physical parameter signals related to environmental vibration, the signal is aligned in the time domain and converted to the time-frequency domain. A physical prior guidance matrix is generated by combining real-time environmental parameters. Feature decoupling is performed using a deep neural network, including multimodal feature splicing, shared encoder, gas feature decoding branch and cross attention layer, to filter out non-gas components and output pure gas features.
It achieves accurate extraction of gas features and stable concentration inversion under extremely low signal-to-noise ratio, improves the model's generalization ability and anti-drift performance under complex conditions, significantly improves the signal-to-noise ratio and maintains the stability of concentration readings, and avoids frequency aliasing and baseline drift in traditional methods.
Smart Images

Figure CN123065971A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of gas spectral detection and deep learning technology, and in particular to a method and system for decoupling gas feature detection using all-optical dual resonance and deep network. Background Technology
[0002] Photoacoustic spectroscopy (PAS) is an ultra-sensitive trace gas detection technology based on the photoacoustic effect. When the target gas absorbs modulated laser energy, it undergoes periodic thermal expansion, which in turn excites an acoustic signal. This project employs an advanced "all-optical dual-resonance enhanced" photoacoustic cell and fiber optic acoustic sensor; however, under extremely harsh field conditions, external acoustic vibrations and electromagnetic noise will inevitably couple into the system.
[0003] The closest existing implementation schemes are: (1) Pure hardware vibration isolation and acoustic filtering: common mode environmental noise is canceled by adding thick sound insulation materials, vibration damping pads, and designing special acoustic resonant cavity shapes (such as differential photoacoustic cells). (2) Lock-in amplification and digital filtering (conventional software filtering): after acquiring the detector electrical signal, the harmonic signals with the same frequency and multiples (2f) of the laser modulation frequency are extracted by hardware or software lock-in amplifiers, or traditional signal processing methods such as Kalman filtering and wavelet denoising are used.
[0004] Pure hardware vibration isolation increases the size, weight, and cost of the equipment and cannot eliminate interference from low-frequency heavy mechanical vibrations coupled into the fiber optic sensor. Traditional lock-in amplification and digital filtering methods are "linear filtering." When the non-stationary mechanical vibration frequency in the industrial field happens to coincide with the resonant frequency (or harmonics) of the photoacoustic signal, traditional filters will mistakenly identify it as a valid signal and amplify it, resulting in a sharp drop in the signal-to-noise ratio. Summary of the Invention
[0005] This invention provides a method and system for decoupling gas feature detection using all-optical dual resonance and deep network, which solves the problems of low signal-to-noise ratio and severe baseline drift caused by the overlap of non-stationary noise and gas feature frequencies in the prior art, and achieves accurate extraction of gas features and stable concentration inversion under extremely low signal-to-noise ratio.
[0006] In a first aspect, the present invention provides a method for decoupling detection of all-optical dual resonance and deep network gas features, comprising: Simultaneously acquire the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration. After time-domain alignment of the two signals, convert them to the time-frequency domain to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram. Based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix that characterizes the theoretical form of the pure gas signal is dynamically generated. The main path photoacoustic time-frequency diagram, the auxiliary vibration time-frequency diagram, and the physical prior guidance matrix are input into a pre-trained deep neural network to obtain pure gas characteristics, and the target gas concentration is calculated based on the pure gas characteristics. The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer; The multimodal feature splicing layer is used to splice the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor; The shared encoder is used to encode the dual-channel input tensor and output a hybrid feature map; the gas feature decoding branch is used to decode and reconstruct the hybrid feature map and output gas feature components. The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output pure gas features.
[0007] Secondly, the present invention provides an all-optical dual-resonance and deep network gas feature decoupling detection system, comprising: In the preprocessing stage, the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration are acquired simultaneously. The two signals are time-domain aligned and then converted to the time-frequency domain to obtain the main optical-acoustic time-frequency map and the auxiliary vibration time-frequency map. In the prior stage, based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix characterizing the theoretical form of the pure gas signal is dynamically generated. In the network processing stage, the main path photoacoustic time-frequency diagram, the auxiliary vibration time-frequency diagram, and the physical prior guidance matrix are input into a pre-trained deep neural network to obtain pure gas features, and the target gas concentration is calculated based on the pure gas features. The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer; The multimodal feature splicing layer is used to splice the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor; The shared encoder is used to encode the dual-channel input tensor and output a hybrid feature map; the gas feature decoding branch is used to decode and reconstruct the hybrid feature map and output gas feature components. The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output pure gas features.
[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention compensates for the time delay in the physical transmission of mechanical vibrations by synchronously acquiring the main path photoacoustic signal and environmental vibration signal and aligning them in the time domain. This enables the subsequent network to accurately learn the spatiotemporal correlation between gas signals and noise, avoiding feature misalignment caused by phase mismatch in traditional methods. This lays the foundation for extracting weak gas signals under strong non-stationary interference. Furthermore, it utilizes real-time temperature and pressure parameters and a high-resolution transmission molecular absorption database. The HITRAN database dynamically generates theoretical absorption line shapes, enabling the physical prior matrix to adapt to changes in the field environment. This solves the problem that traditional fixed templates cannot adapt to drastic temperature and pressure fluctuations, providing a navigation template for deep networks and significantly improving the model's generalization ability and anti-drift performance under complex conditions. By splicing noisy photoacoustic signals and vibration reference signals along the channel dimension, high-dimensional information of phase and amplitude is preserved compared to traditional simple subtraction or filtering, providing richer discriminative features for subsequent decoupling. A multi-layer residual dilated convolution structure is employed, expanding the receptive field through progressively increasing dilation rates while maintaining high-resolution feature maps. This effectively extracts wideband, long-term time-dependent features of gas absorption peaks and non-stationary noise, overcoming the shortcomings of insufficient receptive field in traditional CNNs or the loss of low-frequency vibration details in ordinary convolutions. The method employs a series of transposed convolutional layers to gradually restore time-frequency resolution, specifically reconstructing gas absorption features and forming an independent subspace with the noise branch. Combined with orthogonal loss, this forces gas and noise features to be orthogonal in the latent space, fundamentally avoiding frequency aliasing and achieving high-fidelity gas signal extraction at extremely low signal-to-noise ratios. Physical priors are used as queries to filter gas features with attention, and the network only enhances activation regions highly similar to the theoretical absorption curve, completely eliminating spurious peaks caused by sudden vibrations, electromagnetic interference, and other non-stationary noise. By combining inverse short-time Fourier transform to restore the time-domain waveform and extract peak values, and using a pre-calibrated linear inversion formula to calculate concentration, the method avoids false alarms caused by baseline drift in traditional methods. During continuous monitoring, the concentration reading is stably locked near the true value, demonstrating significantly better repeatability and long-term stability than traditional lock-in amplification techniques. Attached Figure Description
[0009] Figure 1 This is a flowchart of the steps of the all-optical dual resonance and deep network gas feature decoupling detection method provided in the embodiments of the present invention; Figure 2 This is a waveform diagram of the original input signal under extremely harsh working conditions provided in an embodiment of the present invention; Figure 3 The time-domain waveform of the noise component output by the non-stationary noise decoding branch provided in the embodiments of the present invention; Figure 4 The time-domain waveform of pure gas features jointly output by the gas feature decoding branch and the cross-attention layer provided in this embodiment of the invention; Figure 5 Comparison chart of 24-hour online monitoring of trace concentrations in industrial settings provided for embodiments of the present invention. Detailed Implementation
[0010] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0011] In a first aspect, this invention provides a method for decoupling detection of gas features from all-optical dual resonance and deep network, see [link to relevant documentation]. Figure 1 The process includes the following steps S101 to S103.
[0012] S101, synchronously acquire the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration, and after time-domain alignment of the two signals, convert them to the time-frequency domain to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram; Specifically, in step S101, the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration are acquired simultaneously. After time-domain alignment of the two signals, they are converted to the time-frequency domain to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram, including the following steps S1011 to S1013.
[0013] S1011, Calculate the physical conduction delay based on the propagation speed of sound waves in the instrument housing and the equivalent geometric distance from the sound source to the sensor; S1012 uses physical conduction delay as the search center and cross-correlation function to compensate for the delay of the main optical acoustic signal and the auxiliary physical parameter signal, so that the two signals are aligned on the time axis. S1013, the aligned main optical-acoustic signal and the aligned auxiliary physical parameter signal are converted to the time-frequency domain by short-time Fourier transform to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram.
[0014] For example, in order to address the issue of physical time difference caused by mechanical vibration transmission leading to incorrect correspondences learned by the network, the present invention introduces a dynamic phase compensation mechanism for the synchronous acquisition and physical phase adaptive alignment of multimodal heterogeneous signals.
[0015] (1) Synchronous acquisition of multiple signals: In At the same sampling rate Three time series were acquired: the main path photoacoustic signal (weak and noisy). Triaxial mechanical vibration signals acquired by a piezoelectric accelerometer attached to the housing. ; Ambient temperature and pressure scalar sequence ,here, Indicates in The ambient temperature at that moment; Indicates in The constant environmental pressure.
[0016] (2) Calculation of microsecond-level physical phase delay: Assume the propagation speed of the sound wave in the metal casing of the instrument is... The equivalent geometric distance from the vibration source to the fiber optic sensor is Calculate the reference physical conduction delay .
[0017] (3) Cross-correlation optimization and alignment: based on Using the cross-correlation function as the search center, microsecond-level dynamic time delay compensation calculations are performed on the main and auxiliary path signals. The goal is to find the precise time delay compensation that maximizes the cross-correlation energy. Time-axis shift resampling of the main signal .
[0018] (4) Two-dimensional time-frequency diagram conversion: The strictly aligned main photoacoustic signal and vibration signal are mapped into two-dimensional time-frequency diagrams through Short-Time Fourier Transform (STFT), denoted as the main photoacoustic time-frequency diagram. Time-frequency diagram of auxiliary vibration This is to provide the network with time-frequency joint features for extraction.
[0019] S102, based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, dynamically generates a physical prior guidance matrix that characterizes the theoretical form of the pure gas signal. Specifically, in step S102, based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix characterizing the theoretical form of the pure gas signal is dynamically generated, including the following steps S1021 to S1024.
[0020] S1021, based on the real-time collected ambient temperature and pressure, calls the molecular spectral database to obtain the molecular transition parameters of the target gas; S1022, the Voigt function is used to calculate the theoretical absorption line shape of the target gas at the center wavenumber; S1023, perform second harmonic simulation of wavelength modulation spectrum on theoretical absorption line shape to obtain pure second harmonic time domain waveform; S1024 converts the pure second harmonic time-domain waveform into a time-frequency diagram using a short-time Fourier transform, and uses the time-frequency diagram as the physical prior guiding matrix.
[0021] For example, the traditional feature extraction and optimization process in deep learning is highly black-box and lacks explicit guidance from physical mechanisms. This invention uses spectral physics formulas to generate physical prior reference templates for the network in real time, thereby guiding the network to converge.
[0022] (1) Theoretical absorption line shape calculation: Based on real-time temperature T and pressure P from the ambient temperature and pressure scalar sequence, molecular transition parameters from the HITRAN database are used. The target gas at the central wavenumber... The absorption line shape at that point employs a Voigt function that combines Doppler broadening and collision broadening. Specifically, it is expressed as follows: ; in, Represents the integral variable; Indicates the Doppler broadening parameter; Indicates the frequency of the laser; Indicates the collision widening parameter; Represents a Gaussian linear function; This represents the Lorentz linear function.
[0023] (2) Wavelength Modulation Spectroscopy (WMS) Second Harmonic (2f) Physical Simulation and Mapping: Let the laser wavelength modulation frequency be... The modulation amplitude is Modulated signal Substitution The second harmonic (2f) time-domain waveform, which is a multiple of the modulation frequency, is extracted using a digital phase-locked loop algorithm. Represents a time variable; (3) Prior matrix construction: The generated pure second harmonic time-domain waveform is also transformed into a time-frequency diagram through short-time Fourier transform, which serves as the physical prior guiding matrix. Its mathematical essence is a high-dimensional tensor with a specific energy distribution topology.
[0024] S103, input the main path photoacoustic time-frequency diagram, auxiliary vibration time-frequency diagram and physical prior guidance matrix into the pre-trained deep neural network to obtain the pure gas characteristics, and calculate the target gas concentration based on the pure gas characteristics; The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer.
[0025] (1) Multimodal feature splicing layer.
[0026] The multimodal feature stitching layer is used to stitch the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor.
[0027] For example, the main path photoacoustic time-frequency diagram after the steps are strictly physically phase aligned and processed by short-time Fourier transform (STFT) is shown. Time-frequency diagram of auxiliary vibration Tensor splicing is performed in the channel dimension to fuse them into a dual-channel input tensor, thereby locking the original comparison features of cooling / pump valve interference and weak photoacoustic signals in the time-frequency joint domain.
[0028] (2) Shared encoder.
[0029] A shared encoder is used to encode a dual-channel input tensor and output a hybrid feature map.
[0030] Specifically, the shared encoder includes: a first residual dilated convolutional block, a second residual dilated convolutional block, a third residual dilated convolutional block, and a fourth residual dilated convolutional block cascaded in sequence; along the cascading direction, the dilation rate of the dilated convolutional layer in each residual dilated convolutional block increases progressively. A shared encoder is used to encode a two-channel input tensor and output a mixed feature map, including: (2.1) The photoacoustic time-frequency diagram of the main path and the time-frequency diagram of the auxiliary vibration are spliced together in the channel dimension to form a dual-channel input tensor; (2.2) Input the dual-channel input tensor into the first concatenated residual dilated convolution block, and perform two layers of dilated convolution operations with the first dilation rate in sequence to output the first intermediate feature map; (2.3) Input the first intermediate feature map into the second residual dilated convolution block; (2.4) In the second residual dilated convolution block, two layers of dilated convolution operations with a second dilation rate are executed in sequence. The second dilation rate is greater than the first dilation rate, and the second intermediate feature map is output. (2.5) Input the second intermediate feature map into the third residual dilated convolution block, and perform two layers of dilated convolution operations with a third dilation rate in sequence. The third dilation rate is greater than the second dilation rate, and the third intermediate feature map is output. (2.6) Input the third intermediate feature map into the fourth residual dilated convolution block and perform two layers of dilated convolution operations with a fourth dilation rate in sequence. The fourth dilation rate is greater than the third dilation rate, and output the mixed feature map.
[0031] For example, dual-channel input tensors are vertically injected downwards into a shared feature encoder layer. This encoder is constructed based on a multi-layer residual dilated convolutional network. Through layer-by-layer downsampling and adaptive widening of the receptive field, it deeply extracts high-dimensional latent space features that simultaneously couple trace gas spectral absorption features and non-stationary mechanical vibration features from harsh industrial environments, outputting a high-dimensional deep hybrid feature map. .
[0032] (3) Gas feature decoding branch.
[0033] The gas feature decoding branch is used to decode and reconstruct the mixed feature map, and output the gas feature components.
[0034] Specifically, the gas feature decoding branch includes: a first transposed convolutional layer, a second transposed convolutional layer, a third transposed convolutional layer, and a fourth transposed convolutional layer cascaded in sequence; Decode and reconstruct the mixed feature map to output gas feature components, including: The hybrid feature map output by the shared encoder is passed sequentially through the first transposed convolutional layer, the second transposed convolutional layer, the third transposed convolutional layer, and the fourth transposed convolutional layer. Transposed convolution operations are performed in each transposed convolutional layer, and the output of the fourth transposed convolutional layer is used as the gas feature component.
[0035] For example, during the usage phase, the gas feature decoding branch pairs the mixed feature map. Decode and reconstruct the gas characteristic components. .
[0036] The gas feature decoding branch consists of four cascaded transposed convolutional layers: a first transposed convolutional layer, a second transposed convolutional layer, a third transposed convolutional layer, and a fourth transposed convolutional layer. Each transposed convolutional layer has a 3×3 kernel size, a stride of 2, and channel numbers of 256, 128, 64, and 32, respectively. Each transposed convolutional layer is followed by a batch normalization layer and a ReLU activation function.
[0037] Hybrid Feature Map First, the first transposed convolutional layer is input, which doubles the resolution of the feature map and increases the number of channels to 256, outputting the first intermediate feature map. The first intermediate feature map is then input into the second transposed convolutional layer, which again doubles the resolution and reduces the number of channels to 128, outputting the second intermediate feature map. The second intermediate feature map is then input into the third transposed convolutional layer, which doubles the resolution and reduces the number of channels to 64, outputting the third intermediate feature map. Finally, the third intermediate feature map is input into the fourth transposed convolutional layer, which doubles the resolution and reduces the number of channels to 32, outputting the gas feature components. .
[0038] The characteristic components of the gas The input is then fed into a cross-attention layer, where the physical prior guidance matrix is used as a query term for attention enhancement. Finally, the pure gas features are output for concentration inversion.
[0039] (4) Cross-attention layer.
[0040] The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output pure gas features.
[0041] Specifically, guided by a physical prior guidance matrix, attention-weighted enhancement is applied to the gas feature components to filter out non-gas components and output pure gas features, including: (4.1) Use the physical prior guiding matrix as the query matrix and the gas characteristic components as the bond matrix and value matrix; (4.2) Calculate the attention weight matrix between the query matrix and the key matrix; (4.3) Use the attention weight matrix to perform a weighted summation of the value matrix to obtain the weighted feature matrix; (4.4) Perform residual connection between the weighted feature matrix and the gas feature components to output the pure gas features.
[0042] For example, the gas feature matrix The current flows vertically downwards to the physical prior-guided cross-attention layer. This layer changes the self-attention mechanism of traditional black-box networks, innovatively incorporating the physical prior-guided matrix from step S102, which combines real-time temperature and pressure parameters and is dynamically simulated using the Voigt function and the HITRAN database. The gas feature matrix will be used as the query term (Query, Q). The key (K) and value (V) matrices are used respectively for cross-attention weight calculation. This mechanism is equivalent to using the absolutely pure theoretical spectral line shape as a "visual navigation template" for pixel-level feature comparison, illuminating and enhancing only the activation regions that are highly similar to the standard theoretical shape, thereby completely eliminating spurious peak interference caused by non-stationary impacts or broadband vibrations, and outputting high-fidelity, extremely pure gas characteristics. .
[0043] During the online monitoring phase, the concentration of the target gas is calculated based on the characteristics of the pure gas, including: (1) The characteristics of the pure gas are restored to a one-dimensional time domain signal by inverse short-time Fourier transform, and the characteristic peaks of the one-dimensional time domain signal are extracted. (2) Substitute the characteristic peak value into the preset concentration inversion formula to calculate the target gas concentration; wherein, the preset concentration inversion formula is expressed as: ; in, Represents the system response coefficient. Indicates the background bias constant; This represents the characteristic peak-to-valley difference of a one-dimensional time-domain signal; This represents the target gas concentration obtained from the inversion calculation.
[0044] For example, during the online monitoring phase, the method provided by this invention outputs pure... By using the Inverse Short-Time Fourier Transform (ISTFT), the features are losslessly restored from the two-dimensional time-frequency domain to a one-dimensional second harmonic (2f) gas time-domain signal. The characteristic peaks (or peak-to-valley differences) are extracted and substituted into the gas concentration inversion formula. This enables adaptive output of highly stable, zero-drift, and high-precision trace gas concentration inversion results even under extreme conditions where the energy of strong mechanical vibration or electromagnetic interference noise far exceeds the signal energy (i.e., signal-to-noise ratio less than 1).
[0045] Here, the training process of a deep neural network includes: (1) Construct a training network; wherein, the training network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, a non-stationary noise decoding branch, and a cross attention layer; (2) Obtain the characteristics of pure gas signals and the characteristics of real industrial environmental noise, and generate physical prior features based on the characteristics of pure gas signals; (3) The pure gas signal features and environmental noise features are superimposed with different signal-to-noise ratios to generate mixed training samples; (4) Input the mixed training samples as the main path photoacoustic time-frequency map into the deep neural network to obtain the training gas feature components output by the gas feature decoding branch and the training noise feature components output by the non-stationary noise decoding branch. (5) Add the training gas feature components and the training noise feature components, calculate the mean square error between the sum and the mixed training samples, and use it as the reconstruction fidelity loss; (6) Calculate the Frobenius norm of the inner product between the training gas feature components and the training noise feature components, and use it as the absolute orthogonal loss. (7) Calculate the cosine similarity between the first derivative of the training gas feature components and the physical prior features, and use it as the topological loss. (8) The parameters of the deep neural network are updated by using the weighted sum of reconstruction fidelity loss, absolute orthogonality loss and topological morphology loss as the joint loss.
[0046] For example, since it is difficult to obtain absolutely pure gas signal tags in real industrial sites, this invention adopts a hybrid data augmentation strategy of "laboratory purification + on-site background noise superposition" to construct a training set: (1) Pure gas tag (Label) collection: Under the soundproof and vibration-proof constant temperature platform in the laboratory, standard target gases of different concentrations are introduced, and pure photoacoustic signal feature matrix without noise interference is collected as the absolute tag for supervised learning. (2) Blind collection of complex background noise: The instrument is placed in real harsh working conditions such as thermal power plants and chemical plants, and blind collection is carried out for a long period of time without the introduction of target gas to collect real non-stationary environmental noise matrix including compressor vibration, pump valve start and stop, etc. (3) Dynamic mixing and training: At the algorithm layer, "pure gas tags" and "real complex background noise" are nonlinearly and randomly superimposed and mixed according to different signal-to-noise ratios (SNR) decreasing from 10 to the extreme 0.1 to generate a large-scale synthetic training set. The mixed signal is input into the deep neural network PG-DDNet provided by this invention, and the dual-branch network is guided to perform backpropagation updates with the "pure gas label" as the target, thereby giving the network the ability to strip away real industrial noise under extreme signal-to-noise ratios.
[0047] The specific implementation of the training process for a deep neural network is as follows: (1) Construct the training network.
[0048] The training network uses the same hardware architecture as the online network, but includes an additional non-stationary noise decoding branch. Specifically, the shared encoder consists of four cascaded residual dilated convolutional blocks, each containing two dilated convolutional layers (dilation rates of 1, 2, 4, and 8) and a max-pooling layer with a stride of 2, with channel numbers of 64, 128, 256, and 512 respectively. The gas feature decoding branch consists of four cascaded transposed convolutional layers (3×3 kernels, stride 2, channel numbers 256→128→64→32), each followed by a batch normalization layer and a ReLU activation function. The non-stationary noise decoding branch consists of two cascaded transposed convolutional layers (3×3 kernels, stride 2, channel numbers 256→128), with no skip connections. The cross-attention layer uses single-head attention, with query, key, and value mapping dimensions all at 128.
[0049] (2) Obtain the characteristics of pure gas signals and environmental noise.
[0050] Under a soundproof, vibration-proof, and temperature-controlled laboratory platform, standard target gases of known concentrations (e.g., 0 ppb, 50 ppb, 100 ppb, 200 ppb) were introduced, and pure photoacoustic signals free from noise interference were collected. These signals were then converted into time-frequency maps via STFT as characteristic labels for the pure gas signals. Simultaneously, the instrument was placed in real, harsh working conditions such as thermal power plants and chemical plants. Without the introduction of the target gas, time-frequency maps containing non-stationary environmental noises, including compressor vibrations and pump / valve start-up and shutdown, were collected over extended periods to represent the noise characteristics of the real industrial environment.
[0051] (3) Generate mixed training samples.
[0052] The characteristics of pure gas signals and real industrial environmental noise are nonlinearly and randomly superimposed at different signal-to-noise ratios (SNR). For example, the SNR is gradually decreased from 10 dB to an extreme 0.1 dB to generate mixed training samples. ,in, Calculated based on the target SNR; Represents the noise characteristics of a real industrial environment; This represents the characteristics of pure gas signals. In each training batch, pure samples and noisy samples are randomly selected, and the SNR value is randomly chosen to enhance the model's generalization ability.
[0053] (4) Network forward computation.
[0054] Mix training samples The photoacoustic time-frequency map is used as the input to the training network, while the physical prior features (generated from pure gas signal features based on real-time temperature and pressure parameters through Voigt linear simulation) are used as query terms for the cross-attention layer. After the network forward propagates, the gas feature decoding branch outputs the training gas feature components. The non-stationary noise decoding branch outputs training noise feature components. The cross-attention layer outputs pure gas features (this output is only used for monitoring during this training phase and does not participate in loss calculation). During this process, an absolutely orthogonal loss constraint is introduced to train the gas feature components. and training noise feature components The Frobenius norm after calculating the inner product is used as an orthogonal penalty term, which forces the gas light absorption characteristics and sudden environmental vibration noise characteristics to be mapped to non-overlapping dimensions in the mathematical latent space, effectively avoiding frequency aliasing in the subspace.
[0055] (5)-(7) Calculate the combined loss.
[0056] Reconstruction fidelity loss :calculate Mixed samples with input The mean square error (MSE) between them.
[0057] Absolute orthogonal loss :calculate and The Frobenius norm of the inner product of , i.e. This forces the two to be orthogonal.
[0058] Topological loss :calculate With physical prior characteristics The cosine similarity of the first derivative, i.e. To preserve the peak-valley symmetry characteristics of the second harmonic (2f) signal.
[0059] (8) Update network parameters.
[0060] Joint loss is expressed as You can get it here. , , The Adam optimizer was used, with an initial learning rate of 0.001, which decayed to 0.9 times the original rate every 10 epochs, for a total of 200 epochs. Each batch size was 32. During training, the reconstruction error and orthogonality on the validation set were monitored, and training was stopped when there was no improvement for 10 consecutive epochs.
[0061] After training, the weights of the trained shared encoder, gas feature decoding branch, and cross-attention layer are saved, and the non-stationary noise decoding branch is discarded to obtain a deep neural network that can be deployed online.
[0062] Secondly, the present invention provides an all-optical dual-resonance and deep network gas feature decoupling detection system, comprising: In the preprocessing stage, the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration are acquired simultaneously. The two signals are time-domain aligned and then converted to the time-frequency domain to obtain the main optical-acoustic time-frequency map and the auxiliary vibration time-frequency map. In the prior stage, based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix characterizing the theoretical form of the pure gas signal is dynamically generated. In the network processing stage, the main path photoacoustic time-frequency map, auxiliary vibration time-frequency map, and physical prior guidance matrix are input into the pre-trained deep neural network to obtain the pure gas characteristics, and the target gas concentration is calculated based on the pure gas characteristics. The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer; The multimodal feature stitching layer is used to stitch the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor; The shared encoder is used to encode the dual-channel input tensor and output a mixed feature map; the gas feature decoding branch is used to decode and reconstruct the mixed feature map and output gas feature components. The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output pure gas features.
[0063] In a specific simulation provided by this invention, online monitoring and anti-interference verification of trace CH4 in blast furnace flue gas in a chemical industrial park are conducted.
[0064] 1. Implementation background and operating parameter settings.
[0065] The system of this invention was deployed at the blast furnace flue gas emission outlet of a chemical industrial park for online monitoring of trace methane (CH4). The on-site conditions were extremely harsh: Temperature field fluctuations: Affected by flue gas emissions, the ambient temperature fluctuates drastically between 15℃ and 55℃.
[0066] Non-stationary mechanical vibration: There is a strong low-frequency periodic oscillation with a main frequency of 50Hz in close proximity to the large induced draft fan, accompanied by irregular mechanical impacts from valve opening and closing (the measured transient acceleration impact peak reaches 2.5g).
[0067] Target gas state: CH4 standard gas with a nominal concentration of 200 ppb is introduced to simulate trace escape monitoring at extremely low concentrations.
[0068] 2. Performance of the traditional technology control group (failure conditions).
[0069] A conventional high-precision wavelength modulation spectroscopy (WMS-2f) detector (equipped with a conventional digital lock-in amplifier and hardware vibration isolation cotton) was used for comparative testing.
[0070] See Figure 2 The original noisy photoacoustic signal of the main path, collected on-site, is shown in the time domain waveform. Figure 3 This refers to the noise characteristic components output by the non-stationary noise decoding branch during the training phase, i.e., the environmental vibration interference components decoupled from the mixed signal. Figure 4 The output is the pure gas feature, which is reconstructed by the gas feature decoding branch and enhanced by the physical prior guidance matrix through the cross attention layer, and is the time domain restored signal.
[0071] Signal submersion: At the moment of startup of a large induced draft fan, the intense acoustic and mechanical coupling noise completely destroys the acoustic field resonance inside the photoacoustic cell. The second harmonic (2f) waveform extracted by the traditional lock-in amplifier is severely distorted, and the originally symmetrical characteristic peaks are completely submerged by the pseudo-peaks of non-stationary noise, causing the transient signal-to-noise ratio (SNR) to drop to 0.42 (the noise energy far exceeds the signal energy).
[0072] Baseline drift: During continuous 24-hour monitoring, traditional instruments suffer from baseline drift due to the inability to isolate background noise that varies with temperature and vibration. The system experiences drastic random drift, leading to frequent false alarms about exceeding limits.
[0073] 3. The implementation effect of the PG-DDNet system of the present invention.
[0074] After connecting to the "multimodal acquisition and PG-DDNet decoupling network" of this invention: Multi-source data synchronization and delay compensation: Piezoelectric sensors attached to the housing capture the characteristic timing of fan oscillations and impacts in real time. The system automatically compensated for the error through a cross-correlation algorithm. The physical transmission delay ensures that the photoacoustic signal and the vibration reference are strictly aligned in the time and frequency domain.
[0075] Latent Space Absolute Decoupling (Fidelity Reconstruction): Faced with a harsh input of SNR=0.42, PG-DDNet actively calls the CH4 theoretical absorption line type generated in real time from the HITRAN database based on the current T / P parameters as the Query in the cross-attention layer. The network forcibly "anchors" and extracts the true second harmonic (2f) peak and valley features of CH4 from the chaotic noisy signal, and assigns most of the mechanical vibration and impact interference to the "non-stationary noise feature branch".
[0076] Quantization performance indicators leap: (1) Improved signal-to-noise ratio: extremely pure output characteristics after decoupling After being restored to the time domain signal, its equivalent signal-to-noise ratio is significantly improved to 35.6 (an improvement of more than 80 times).
[0077] (2) Detection lower limit breach: Under extremely strong vibration conditions with the fan operating at full load, the system's CH4 detection lower limit ( It remains at an extremely high level, better than 15 ppb.
[0078] (3) Drift stability: See Figure 5 During 24-hour continuous online monitoring, the concentration reading output by the system of this invention remained extremely stable at [value missing]. Within the true value range of the standard gas, the zero-point drift rate is less than 1 / 15 of that of traditional equipment, completely eliminating false alarms and demonstrating revolutionary adaptability to complex working conditions.
[0079] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.
[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.
Claims
1. A method for decoupling detection of gas features from all-optical dual resonance and deep network, characterized in that, include: The main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration are acquired simultaneously. The main optical-acoustic signal and the auxiliary physical parameter signal are time-domain aligned and then converted to the time-frequency domain to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram. Based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix that characterizes the theoretical form of the pure gas signal is dynamically generated. The main path photoacoustic time-frequency diagram, the auxiliary vibration time-frequency diagram, and the physical prior guidance matrix are input into a pre-trained deep neural network to obtain pure gas characteristics, and the target gas concentration is calculated based on the pure gas characteristics. The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer; The multimodal feature splicing layer is used to splice the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor; The shared encoder is used to encode the dual-channel input tensor and output a hybrid feature map; The gas feature decoding branch is used to decode and reconstruct the hybrid feature map and output gas feature components; The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output the pure gas features.
2. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The synchronous acquisition of the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration is performed. After time-domain alignment of the two signals, they are converted to the time-frequency domain to obtain the main optical-acoustic time-frequency map and the auxiliary vibration time-frequency map, including: The physical conduction delay is calculated based on the propagation speed of sound waves in the instrument housing and the equivalent geometric distance from the sound source to the sensor. Using the physical conduction delay as the search center, the cross-correlation function is used to compensate for the delay of the main optical acoustic signal and the auxiliary physical parameter signal, so that the two signals are aligned on the time axis. The aligned main optical-acoustic signal and the aligned auxiliary physical parameter signal are converted to the time-frequency domain by short-time Fourier transform to obtain the main optical-acoustic time-frequency diagram and the auxiliary vibration time-frequency diagram.
3. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The process of dynamically generating a physical prior guidance matrix characterizing the theoretical form of a pure gas signal based on real-time environmental parameters and a spectral absorption physical model of the target gas includes: Based on the real-time collected ambient temperature and pressure, the molecular transition parameters of the target gas are obtained by calling the molecular spectral database; The theoretical absorption line shape of the target gas at the center wavenumber is calculated using the Voigt function; The second harmonic of the wavelength modulation spectrum of the theoretical absorption line shape was simulated to obtain the pure second harmonic time-domain waveform. The pure second harmonic time-domain waveform is converted into a time-frequency diagram by a short-time Fourier transform, and the time-frequency diagram is used as the physical prior guiding matrix.
4. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The shared encoder includes: a first residual dilated convolutional block, a second residual dilated convolutional block, a third residual dilated convolutional block, and a fourth residual dilated convolutional block cascaded sequentially; along the cascading direction, the expansion rate of the dilated convolutional layer in each residual dilated convolutional block increases progressively. The encoding of the dual-channel input tensor to output a hybrid feature map includes: The dual-channel input tensor is input into the cascaded first residual dilated convolution block, and two layers of dilated convolution operations with the first dilation rate are executed sequentially to output the first intermediate feature map. The first intermediate feature map is input into the second residual dilated convolution block, and two layers of dilated convolution operations with a second dilation rate are executed sequentially. The second dilation rate is greater than the first dilation rate, and the second intermediate feature map is output. The second intermediate feature map is input into the third residual dilated convolution block, and two layers of dilated convolution operations with a third dilation rate are executed sequentially. The third dilation rate is greater than the second dilation rate, and the third intermediate feature map is output. The third intermediate feature map is input into the fourth residual dilated convolution block, and two layers of dilated convolution operations with a fourth dilation rate are executed sequentially. The fourth dilation rate is greater than the third dilation rate, and the mixed feature map is output.
5. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The gas feature decoding branch includes: a first transposed convolutional layer, a second transposed convolutional layer, a third transposed convolutional layer, and a fourth transposed convolutional layer cascaded in sequence; The decoding and reconstruction of the hybrid feature map to output gas feature components includes: The hybrid feature map output by the shared encoder is sequentially passed through the first transposed convolutional layer, the second transposed convolutional layer, the third transposed convolutional layer, and the fourth transposed convolutional layer. A transposed convolution operation is performed in each transposed convolutional layer, and the output of the fourth transposed convolutional layer is used as the gas feature component.
6. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The step of using the physical prior guidance matrix as guidance to perform attention-weighted enhancement on the gas feature components, filtering out non-gas components, and outputting pure gas features includes: The physical prior guidance matrix is used as the query matrix, and the gas feature components are used as the bond matrix and value matrix; Calculate the attention weight matrix between the query matrix and the key matrix; The value matrix is weighted and summed using the attention weight matrix to obtain a weighted feature matrix; The weighted feature matrix is residually concatenated with the gas feature components to output the pure gas features.
7. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The training process of the deep neural network includes: Construct a training network; wherein the training network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, a non-stationary noise decoding branch, and a cross-attention layer; Acquire the characteristics of pure gas signals and the noise characteristics of real industrial environments, and generate physical prior features based on the pure gas signal characteristics; The pure gas signal features and the environmental noise features are superimposed at different signal-to-noise ratios to generate mixed training samples; The mixed training samples are used as the main path photoacoustic time-frequency map and input into the deep neural network to obtain the training gas feature components output by the gas feature decoding branch and the training noise feature components output by the non-stationary noise decoding branch. The training gas feature components and the training noise feature components are added together, and the mean square error between the sum and the mixed training samples is calculated as the reconstruction fidelity loss. Calculate the Frobenius norm of the inner product between the training gas feature components and the training noise feature components, and use it as the absolute orthogonality loss; The cosine similarity between the first derivative of the training gas feature components and the physical prior features is calculated and used as the topological morphology loss. The parameters of the deep neural network are updated using the weighted sum of the reconstruction fidelity loss, the absolute orthogonality loss, and the topological morphology loss as the joint loss.
8. The all-optical dual resonance and deep network gas feature decoupling detection method according to claim 1, characterized in that, The calculation of the target gas concentration based on the characteristics of the pure gas includes: The characteristics of the pure gas are restored to a one-dimensional time-domain signal by inverse short-time Fourier transform, and the characteristic peaks of the one-dimensional time-domain signal are extracted. The target gas concentration is calculated by substituting the characteristic peak value into a preset concentration inversion formula; wherein the preset concentration inversion formula is expressed as: ; in, Represents the system response coefficient. Indicates the background bias constant; This represents the characteristic peak-to-valley difference of a one-dimensional time-domain signal; This represents the target gas concentration obtained from the inversion calculation.
9. A decoupled detection system for all-optical dual resonance and deep network gas features, characterized in that, include: In the preprocessing stage, the main optical-acoustic signal and at least one auxiliary physical parameter signal related to environmental vibration are acquired simultaneously. The two signals are time-domain aligned and then converted to the time-frequency domain to obtain the main optical-acoustic time-frequency map and the auxiliary vibration time-frequency map. In the prior stage, based on real-time environmental parameters and according to the spectral absorption physical model of the target gas, a physical prior guidance matrix characterizing the theoretical form of the pure gas signal is dynamically generated. In the network processing stage, the main path photoacoustic time-frequency diagram, the auxiliary vibration time-frequency diagram, and the physical prior guidance matrix are input into a pre-trained deep neural network to obtain pure gas features, and the target gas concentration is calculated based on the pure gas features. The deep neural network includes: a multimodal feature splicing layer, a shared encoder, a gas feature decoding branch, and a cross-attention layer; The multimodal feature splicing layer is used to splice the main path photoacoustic time-frequency map and the auxiliary vibration time-frequency map in the channel dimension to form a dual-channel input tensor; The shared encoder is used to encode the dual-channel input tensor and output a hybrid feature map; the gas feature decoding branch is used to decode and reconstruct the hybrid feature map and output gas feature components. The cross-attention layer is used to enhance the gas feature components with attention weighting, guided by the physical prior guidance matrix, to filter out non-gas components and output pure gas features.
