Tunnel construction environment noise suppression method and system applying deep learning algorithm

Abstract

The embodiment of the invention discloses a tunnel construction environment noise suppression method and system applying a deep learning algorithm, and the method comprises the steps: collecting an original noise signal set containing a plurality of time domain noise signals and construction equipment noise components contained in the time domain noise signals, and completely obtaining complex and diversified noise information in a tunnel construction environment; based on noise feature distribution obtained through noise feature extraction processing, noise features can be comprehensively described from multiple dimensions of noise spectrum features, noise source orientation features and noise energy propagation features; performing suppression processing on the features by using a deep learning noise suppression model, and intelligently analyzing and generating a targeted suppression control feature set; furthermore, a noise suppression device is driven to execute noise cancellation operation based on the set, and cancellation sound wave signals opposite to the time domain noise signals in phase can be accurately generated.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of noise data processing, and specifically to a method and system for suppressing noise in a tunnel construction environment using a deep learning algorithm. Background Art

[0002] During tunnel construction, environmental noise has always been a critical issue that needs to be addressed. Tunnel construction involves the simultaneous operation of multiple large-scale equipment. The noise generated not only harms the health of construction workers, but also can lead to occupational diseases such as hearing loss due to long-term exposure to high-intensity noise. It also affects construction efficiency and quality, interfering with communication between construction workers and increasing the risk of operational errors.

[0003] However, existing technologies for addressing noise in tunnel construction environments have numerous shortcomings. Some traditional methods simply wrap certain equipment with soundproofing materials. This approach can only partially attenuate the noise emitted by the equipment itself, but fails to comprehensively and effectively address the complex noise levels of the entire construction environment. Other technologies, while attempting to monitor noise, lack in-depth analysis of noise characteristics and cannot accurately grasp key information such as noise sources and propagation characteristics, making it difficult to develop targeted noise reduction strategies. Summary of the Invention

[0004] The embodiments of the present application provide a method and system for suppressing tunnel construction environmental noise using a deep learning algorithm, which are used to effectively and specifically reduce the intensity of environmental noise.

[0005] In a first aspect, an embodiment of the present application provides a method for suppressing tunnel construction environmental noise using a deep learning algorithm, the method comprising: collecting a set of original noise signals from a tunnel construction site, the set of original noise signals comprising multiple time-domain noise signals, each time-domain noise signal containing a noise component generated by at least one construction equipment; performing noise feature extraction processing on the set of original noise signals to obtain a noise feature distribution for each time-domain noise signal, the noise feature distribution comprising noise spectrum features, noise source orientation features, and noise energy propagation features; calling a deep learning noise suppression model to suppress the noise feature distribution to generate a suppression control feature set corresponding to each time-domain noise signal; and driving a noise suppression device to perform a noise cancellation operation based on the suppression control feature set to generate a cancellation sound wave signal having a phase opposite to that of the time-domain noise signal to reduce the intensity of the environmental noise.

[0006] In a second aspect, an embodiment of the present application provides a tunnel construction environment noise suppression system using a deep learning algorithm, comprising: processor; a storage device having a computer program stored thereon, When the computer program is executed by the processor, the processor implements any one of the tunnel construction environment noise suppression methods using deep learning algorithms.

[0007] An embodiment of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the tunnel construction environment noise suppression method using deep learning algorithms are implemented.

[0008] Thus, the embodiments of the present application have the following beneficial effects: By collecting an original noise signal set covering multiple time-domain noise signals and their included construction equipment noise components, complex and diverse noise information in the tunnel construction environment is completely obtained; then, based on the noise feature distribution obtained by noise feature extraction processing, the noise characteristics can be comprehensively characterized from multiple dimensions such as noise spectrum features, noise source azimuth features, and noise energy propagation features; then, using a deep learning noise suppression model to perform suppression processing on the above features, an intelligent analysis can be performed to generate a targeted suppression control feature set; further, based on this set to drive a noise suppression device to perform a noise cancellation operation, a cancellation sound wave signal with a phase opposite to each time-domain noise signal can be accurately generated.

[0009] Designed in this way, the environmental noise intensity is effectively and specifically reduced, the potential hazards caused by noise are reduced, and at the same time, the accuracy and efficiency of construction operations are improved, avoiding poor communication and operation errors caused by noise interference, and overall improving the quality and safety of the tunnel construction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a flowchart of a tunnel construction environment noise suppression method using deep learning algorithms provided by an embodiment of the present application.

[0011] Figure 2 It is a schematic diagram of the basic structure of a tunnel construction environment noise suppression system using deep learning algorithms provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] To make the above objects, features, and advantages of the present application more obvious and understandable, the embodiments of the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0013] See Figure 1 As shown, this figure is a flowchart of a tunnel construction environment noise suppression method using deep learning algorithms provided by an embodiment of the present application. This method can be implemented by a tunnel construction environment noise suppression system using deep learning algorithms. As Figure 1 shown, this method may include step 110 - step 140.

[0014] Step 110: Collect the original noise signal set of the tunnel construction site. The original noise signal set includes multiple time-domain noise signals, and each time-domain noise signal contains noise components generated by at least one construction equipment.

[0015] In the embodiments of the present application, in the scenario of tunnel construction, there are multiple construction equipments operating simultaneously at the construction site. For example, when an excavator is performing earth excavation work, the operation of its engine and the excavation action will generate noises within a corresponding frequency range; for another example, when a loader is transporting materials, the friction of its mechanical components and the operation of the power system will also emit noises; in addition, a concrete mixer will also generate corresponding noises during the mixing process. At this time, relevant audio collection devices are used to collect the noises of the entire construction site according to the preset sampling frequency and duration settings. In the collected original noise signal set, each time-domain noise signal mixes the noise components generated by different construction equipments. For example, in one of the collected time-domain noise signals, it may simultaneously contain the low-frequency roar of the excavator, the medium-frequency mechanical friction sound of the loader, and the high-frequency mixing sound of the concrete mixer.

[0016] Step 120: Perform noise feature extraction processing on the original noise signal set to obtain the noise feature distribution of each time-domain noise signal. The noise feature distribution includes noise spectrum features, noise source azimuth features, and noise energy propagation features.

[0017] Next, noise feature extraction work is carried out on the collected original noise signal set. This process aims to comprehensively understand the characteristics of the noise and provide an accurate basis for subsequent suppression processing.

[0018] Optionally, the performing noise feature extraction processing on the original noise signal set to obtain the noise feature distribution of each time-domain noise signal includes: Step 121: Perform multi-scale decomposition processing on the time-domain noise signal to generate multiple noise sub-signals, and each noise sub-signal corresponds to a noise component in a different frequency band.

[0019] Further, the performing multi-scale decomposition processing on the time-domain noise signal to generate multiple noise sub-signals includes: Step 1210: Obtain multiple adaptive filter banks with different scales. Each adaptive filter bank includes at least one band-pass filter, where the scale parameter of the adaptive filter bank is adjusted according to the average spectral density of the time-domain noise signal; input the time-domain noise signal into the multiple adaptive filter banks with different scales for parallel filtering processing to obtain initial noise sub-signals divided into multiple frequency bands; perform reconstruction error correction processing on the overlapping regions of the frequency bands of the initial noise sub-signals to eliminate the signal aliasing effect between adjacent frequency bands and generate the multiple noise sub-signals.

[0020] In an embodiment of the present application, for one of the time-domain noise signals, first, the scale parameter of the adaptive filter bank is determined according to its average spectral density. If the average spectral density of the time-domain noise signal shows that its energy is mainly concentrated in a relatively wide frequency range, an adaptive filter bank with a corresponding scale will be selected, and each filter bank includes band-pass filters with appropriate center frequencies and bandwidths. For example, there are three adaptive filter banks A, B, and C with different scales. The band-pass filter of filter bank A has a relatively low center frequency and a narrow bandwidth, which is suitable for capturing low-frequency noise components; the parameters of filter bank B are moderate and are used to process medium-frequency noise; filter bank C is for high-frequency noise. The time-domain noise signal is simultaneously input into these three adaptive filter banks for parallel filtering to obtain initial noise sub-signals corresponding to different frequency bands. However, due to the characteristics of the filters, there may be signal aliasing between adjacent frequency bands. Therefore, based on relevant common algorithms (such as signal processing algorithms like Wiener filtering algorithm), the reconstruction error correction processing of the overlapping region of the above initial noise sub-signals can be performed to determine the error in the overlapping region and make adjustments, and finally, multiple noise sub-signals divided into different frequency bands are obtained.

[0021] Step 122: Perform spectral analysis processing on the multiple noise sub-signals, and extract the frequency-domain energy distribution characteristics of each noise sub-signal as the noise spectral characteristics.

[0022] In a preferred embodiment, the performing spectral analysis processing on the multiple noise sub-signals and extracting the frequency-domain energy distribution characteristics of each noise sub-signal includes: Step 1220: Perform short-time Fourier transform processing on each noise sub-signal to generate a corresponding time-frequency spectrogram; perform peak detection processing on the time-frequency spectrogram to identify the set of energy peak points in the time-frequency spectrogram; construct an energy distribution feature set according to the distribution density, peak amplitude, and frequency interval of the set of energy peak points, and extract the statistical features of the energy distribution feature set as the frequency-domain energy distribution characteristics.

[0023] For each noise sub-signal obtained through multi-scale decomposition, taking one of the noise sub-signals as an example, perform short-time Fourier transform processing on it. By selecting an appropriate window function and window length, segment the noise sub-signal in the time domain, and then perform Fourier transform on each segment to generate the corresponding time-frequency spectrogram. On this time-frequency spectrogram, the energy distribution at different time and frequency points can be determined. Then, perform peak detection processing on the time-frequency spectrogram, and use the generalized cross-correlation algorithm (GCC) to find the set of energy peak points in the figure. For example, after detection, it is found that there are five energy peak points in this time-frequency spectrogram, which are located at different frequency positions and have different amplitudes. Then, construct an energy distribution feature set based on the distribution density, peak amplitude, and frequency interval of the above energy peak point set. For example, calculate the frequency interval between adjacent peak points, count the number of peak points within the unit frequency range as the distribution density, and record the amplitude of each peak point, etc. Finally, extract statistical features such as mean and variance from this energy distribution feature set as the frequency-domain energy distribution feature of this noise sub-signal, which is also a part of the noise spectrum feature of the entire time-domain noise signal.

[0024] Step 123: Perform spatial sound field modeling processing on the time-domain noise signal, determine the sound source azimuth angle and the acoustic wave propagation path characteristics of the time-domain noise signal in three-dimensional space, and fuse the sound source azimuth angle and the acoustic wave propagation path characteristics into the noise source azimuth feature.

[0025] In another optional embodiment, the performing spatial sound field modeling processing on the time-domain noise signal to determine the sound source azimuth angle and the acoustic wave propagation path characteristics of the time-domain noise signal in three-dimensional space includes: Step 1230: According to multiple deployed spatial microphone array nodes, collect the sound pressure signals of the time-domain noise signal at different spatial positions; calculate the spatial propagation vector of the time-domain noise signal according to the time difference of arrival and the phase difference of the sound pressure signals; perform direction-of-arrival estimation processing based on the spatial propagation vector to determine the sound source azimuth angle; construct a three-dimensional sound field propagation model according to the position coordinates of the spatial microphone array nodes and the sound source azimuth angle, and extract the path attenuation coefficient and the reflection path characteristics of the three-dimensional sound field propagation model as the acoustic wave propagation path characteristics.

[0026] At the tunnel construction site, multiple spatial microphone array nodes can be pre-deployed, and the above nodes are distributed at different positions. When one of the time-domain noise signals is collected, each microphone array node will receive the signal simultaneously. However, due to different positions relative to the sound source, there will be differences in the arrival time and phase of the signal. For example, there are three microphone array nodes M1, M2, and M3. M1 is closer to the sound source, while M2 and M3 are farther from the sound source and have different directions. M1 receives the signal earliest, and M2 and M3 receive the signal relatively later, and there is a phase difference between them. According to the arrival time difference and phase difference of the above sound pressure signals, the multiple signal classification algorithm is used to calculate the spatial propagation vector of the time-domain noise signal, and this vector represents information such as the propagation direction and speed of the sound wave in space. Based on this spatial propagation vector, through the method of direction of arrival estimation, the azimuth angle of the sound source in the three-dimensional space is finally determined. For example, it is determined that the sound source is on a certain side of the tunnel and forms a certain angle with the horizontal direction. Then, combined with the position coordinates of each spatial microphone array node and the determined azimuth angle of the sound source, a three-dimensional sound field propagation model is constructed. In this model, considering the attenuation and reflection of the sound wave during propagation, the path attenuation coefficient and the reflection path characteristics are extracted. For example, it is found that during the propagation of the sound wave, due to the absorption and reflection of the tunnel wall, the energy gradually attenuates, and there are multiple reflection paths, and the above information constitutes the sound wave propagation path characteristics. Finally, the azimuth angle of the sound source is fused with the sound wave propagation path characteristics to form the noise source azimuth characteristics of the time-domain noise signal.

[0027] Step 124: Perform energy attenuation analysis and processing on the time-domain noise signal, calculate the energy attenuation curve of the time-domain noise signal according to the acoustic wave propagation medium attribute and the propagation distance, and extract the attenuation gradient feature of the energy attenuation curve as the noise energy propagation feature.

[0028] In the embodiment of the present application, for the collected time-domain noise signal, considering the attributes of the acoustic wave propagation medium, the medium in the tunnel may include air, rock, etc. Different media have different absorption and scattering degrees of acoustic wave energy. At the same time, the propagation distances from the sound source to each receiving point are measured. According to the above information, using relevant acoustic principles and algorithms, the energy attenuation curve of the time-domain noise signal during propagation is calculated. For example, as the propagation distance increases, the energy gradually decreases, forming a curve. Then, this energy attenuation curve is analyzed, and its attenuation gradient feature is extracted. For example, the slopes of the curve at different points are calculated to obtain a series of values representing the attenuation speed, and the above values constitute the noise energy propagation feature.

[0029] Step 125: Perform joint coding processing on the noise spectrum feature, the noise source azimuth feature, and the noise energy propagation feature to generate the noise feature distribution of the time-domain noise signal.

[0030] Optionally, the previously extracted noise spectrum features, noise source azimuth features, and noise energy propagation features are jointly encoded. For example, using the Huffman coding algorithm, the statistical features in the noise spectrum features, the azimuth angle and propagation path information in the noise source azimuth features, and the attenuation gradient value in the noise energy propagation features are integrated and encoded. Through the above encoding method, the different types of feature information are transformed into a unified distribution representation that can comprehensively represent the characteristics of the time-domain noise signal, generating the noise feature distribution of the time-domain noise signal.

[0031] Step 130: Invoke the deep learning noise suppression model to perform suppression processing on the noise feature distribution, generating a set of suppression control features corresponding to each time-domain noise signal.

[0032] As an optional technical solution, the invoking the deep learning noise suppression model to perform suppression processing on the noise feature distribution and generating a set of suppression control features corresponding to each time-domain noise signal includes: Step 131: Input the noise feature distribution into the encoder module of the deep learning noise suppression model for feature compression processing, generating a noise-encoded feature vector.

[0033] Optionally, input the generated noise feature distribution into the encoder module of the deep learning noise suppression model. The role of the encoder module is to compress and abstract the input features and extract key information. For example, the encoder module can be composed of multiple convolutional layers and pooling layers. The noise feature distribution is input into the above layers in a set format. Through convolutional operations, feature extraction and transformation are performed on different feature dimensions, and then through pooling operations, the data dimension is reduced. Finally, a noise-encoded feature vector is generated, which contains important information in the noise feature distribution and has a lower dimension compared to the original feature distribution.

[0034] Step 132: Input the noise-encoded feature vector into the decoder module of the deep learning noise suppression model for suppression parameter reconstruction processing, generating an initial suppression parameter sequence.

[0035] Input the generated noise-encoded feature vector into the decoder module of the deep learning noise suppression model. The task of the decoder module is to reconstruct the suppression parameters based on the encoded feature vector. For example, the decoder module performs step-by-step decompression and transformation on the noise-encoded feature vector through a series of deconvolutional layers and fully connected layers. During this process, the model learns how to recover the parameter information related to noise suppression from the encoded vector, and finally generates an initial suppression parameter sequence, which contains some preliminary parameter values for subsequent noise suppression operations.

[0036] Step 133: Perform time-domain alignment processing on the initial suppression parameter sequence, and adjust the phase and amplitude of the initial suppression parameter sequence according to the timestamps of the time-domain noise signal to generate the suppression control feature set.

[0037] As another alternative technical solution, the performing time-domain alignment processing on the initial suppression parameter sequence, and adjusting the phase and amplitude of the initial suppression parameter sequence according to the timestamps of the time-domain noise signal to generate the suppression control feature set includes: Step 1331: Obtain the timestamp sequence of the time-domain noise signal, where the timestamp sequence contains the timing identifiers of each sampling point of the time-domain noise signal.

[0038] Before processing the initial suppression parameter sequence, first obtain the timestamp sequence of the time-domain noise signal. This timestamp sequence records the time order of each sampling point in the time-domain noise signal. For example, the timestamp sequence can be an array of time values arranged in sampling order, with each value corresponding to a sampling point, accurately representing the position of the sampling point on the time axis.

[0039] Step 1332: Perform timestamp synchronization processing on the initial suppression parameter sequence, and perform interpolation reconstruction on the initial suppression parameter sequence according to the sampling interval of the timestamp sequence to generate a timestamp-synchronized suppression parameter sequence.

[0040] Optionally, perform interpolation reconstruction on the initial suppression parameter sequence according to the sampling interval of the obtained timestamp sequence. For example, if the sampling interval of the timestamp sequence is fixed and the sampling interval of the initial suppression parameter sequence is inconsistent with that of the timestamp sequence, then through an interpolation algorithm, insert or delete some parameter values at appropriate positions so that the time points of the initial suppression parameter sequence correspond to the time points of the timestamp sequence, generating a timestamp-synchronized suppression parameter sequence.

[0041] Step 1333: Extract the phase reference curve of the time-domain noise signal, and calculate the set of phase offset amounts of the timestamp-synchronized suppression parameter sequence, where the set of phase offset amounts contains the absolute values of the phase differences of each sampling point.

[0042] Extract the phase reference curve from the time-domain noise signal, and this curve reflects the phase change law of the time-domain noise signal. Then, compare the timestamp-synchronized suppression parameter sequence with the phase reference curve, and calculate the absolute value of the phase difference between the suppression parameter sequence and the phase reference curve at each sampling point. The above absolute values of the phase differences constitute the set of phase offset amounts.

[0043] Step 1334: Perform phase compensation processing on the timestamp-synchronized suppression parameter sequence according to the set of phase offset amounts to generate a phase-aligned suppression parameter sequence.

[0044] Among them, according to the calculated set of phase offset amounts, phase compensation is performed on the suppression parameter sequence for timestamp synchronization. For example, for each sampling point, if the phase offset amount is positive, the phase of the suppression parameter sequence is increased by the corresponding value; if the phase offset amount is negative, the corresponding value is decreased, so that the phase of the suppression parameter sequence is aligned with the phase reference curve of the time-domain noise signal, and a phase-aligned suppression parameter sequence is generated.

[0045] Step 1335: Extract the amplitude envelope curve of the time-domain noise signal, and calculate the set of amplitude scaling factors of the phase-aligned suppression parameter sequence, where the set of amplitude scaling factors includes the amplitude proportionality coefficients of each sampling point.

[0046] Among them, the amplitude envelope curve is extracted from the time-domain noise signal, and this curve describes the variation of the amplitude of the time-domain noise signal with time. Then, by comparing the phase-aligned suppression parameter sequence and the amplitude envelope curve, the amplitude proportionality coefficients of the suppression parameter sequence and the amplitude envelope curve at each sampling point are calculated, and the above coefficients constitute the set of amplitude scaling factors.

[0047] Step 1336: Perform amplitude adjustment processing on the phase-aligned suppression parameter sequence according to the set of amplitude scaling factors to generate an amplitude-matched suppression parameter sequence.

[0048] Optionally, according to the set of amplitude scaling factors, amplitude adjustment is performed on the phase-aligned suppression parameter sequence. For example, if the amplitude scaling factor is greater than 1, the amplitude of the suppression parameter sequence at the corresponding sampling point is increased by the corresponding multiple; if the amplitude scaling factor is less than 1, the corresponding multiple is decreased, and finally an amplitude-matched suppression parameter sequence is generated.

[0049] Step 1337: Perform dimensional consistency processing on the amplitude-matched suppression parameter sequence to unify the units of the phase dimension and the amplitude dimension of the suppression parameter sequence.

[0050] It can be understood that since the phase and the amplitude may have different dimensions, through the normalization method, the units of the phase dimension and the amplitude dimension are unified. For example, if the unit of the phase is radians and the unit of the amplitude is volts, through the appropriate conversion relationship, they are converted into a unified comparable unit form.

[0051] Step 1338: Perform feature dimension splicing processing on the phase parameter and the amplitude parameter of the amplitude-matched suppression parameter sequence to generate the suppression control feature set.

[0052] Further, the phase parameters and amplitude parameters of the amplitude-matched suppression parameter sequence after dimension consistency processing are subjected to feature dimension splicing. For example, the phase parameters and amplitude parameters are combined into a new vector or matrix form according to a preset order, and this new representation form is the generated suppression control feature set.

[0053] In an alternative embodiment, the training process of the deep learning noise suppression model includes: Step 210: Obtain a historical noise signal sample set and corresponding suppression effect labels, where the suppression effect labels include the residual noise energy value after noise suppression.

[0054] In this step, a large number of historical noise signal sample sets are collected. The above samples come from different tunnel construction scenarios and cover the noise conditions generated by various combinations of construction equipment. For each noise signal sample, the corresponding suppression effect label is recorded at the same time. The residual noise energy value in the label is measured through actual noise suppression operations. For example, there is a historical noise signal sample, which is the noise generated when an excavator, a loader, and a concrete mixer are operating simultaneously during the construction of a certain tunnel. After being processed by a noise suppression device, the measured residual noise energy value is a quantization value, and this value is used as the suppression effect label of the sample.

[0055] Step 211: Obtain an initial neural network model, where the initial neural network model includes a convolutional encoding layer, an attention fusion layer, and a fully connected decoding layer.

[0056] Optionally, the convolutional encoding layer is composed of multiple parallel convolutional channels. Each convolutional channel corresponds to a convolutional kernel of a different scale and is used for feature extraction and compression of the input noise features; the attention fusion layer can automatically focus on the importance between different features and perform fusion processing; the fully connected decoding layer is responsible for reconstructing the suppression parameters according to the encoded features.

[0057] Step 212: Input the historical noise signal sample set into the initial neural network model for forward propagation processing to generate a predicted suppression parameter sequence.

[0058] Optionally, the set of historical noise signal samples is sequentially input into the initial neural network model. In the convolutional encoding layer, the noise feature distribution is respectively input into multiple parallel convolutional channels for feature mapping processing to generate intermediate feature maps of multiple scales. For example, different convolutional channels perform feature extraction on different degrees of noise spectrum features, noise source orientation features, etc., to obtain intermediate feature maps with multiple different scales and feature representations. Then, cross-channel feature fusion processing is performed on the above intermediate feature maps to generate a fused feature map. Next, max pooling processing is performed on the fused feature map to generate a noise-encoded feature vector. After this vector is processed by the attention fusion layer to highlight important features, it is input into the fully connected decoding layer, and finally a predicted suppression parameter sequence is generated.

[0059] Step 213: Calculate the mean square error loss based on the predicted suppression parameter sequence and the suppression effect label, and update the weight parameters of the initial neural network model based on the mean square error loss until convergence to obtain the deep learning noise suppression model.

[0060] In a possible embodiment, the convolutional encoding layer includes multiple parallel convolutional channels, and each convolutional channel corresponds to a convolutional kernel of a different scale. The processing process of the convolutional encoding layer includes: Step 310: Respectively input the noise feature distribution into the multiple parallel convolutional channels for feature mapping processing to generate intermediate feature maps of multiple scales; perform cross-channel feature fusion processing on the intermediate feature maps of multiple scales to generate a fused feature map; perform max pooling processing on the fused feature map to generate the noise-encoded feature vector.

[0061] In the embodiment of the present application, the predicted suppression parameter sequence is compared with the suppression effect label to calculate the mean square error loss. For example, for the predicted suppression parameter sequence of each sample and the residual noise energy value in the corresponding suppression effect label, the average value of the sum of the squares of their errors is calculated. According to this mean square error loss, the backpropagation algorithm is used to update the weight parameters of the initial neural network model. In each iteration, the connection weights between the neurons in the convolutional encoding layer, the attention fusion layer, and the fully connected decoding layer are adjusted so that the mean square error loss gradually decreases until the model converges, and finally the deep learning noise suppression model is obtained.

[0062] Step 140: Drive the noise suppression device to perform a noise cancellation operation based on the suppression control feature set, and generate a cancellation sound wave signal with a phase opposite to that of the time-domain noise signal to reduce the environmental noise intensity.

[0063] In a preferred embodiment, driving the noise suppression device to perform a noise cancellation operation based on the suppression control feature set and generate a cancellation sound wave signal with a phase opposite to that of the time-domain noise signal includes: Step 141: Generate a target amplitude curve according to the amplitude features in the suppression control feature set; generate a target phase offset according to the phase features in the suppression control feature set.

[0064] In the embodiments of the present application, amplitude features and phase features are extracted from the suppression control feature set. For example, the suppression control feature set is a vector set containing multiple data dimensions, where a part of the dimensions correspond to amplitude features and another part of the dimensions correspond to phase features. For the amplitude features, through relevant algorithms and data processing methods (polynomial fitting algorithm), operations such as smoothing the amplitude feature data and fitting curves are performed to generate a target amplitude curve, which describes the amplitude values that the cancellation acoustic signal should have at different time points. For the phase features, through statistical analysis methods, the target phase offset is determined. For example, by statistically analyzing the phase feature data and finding its variation law, a target phase offset that can make the cancellation acoustic signal have a phase opposite to that of the original time-domain noise signal is obtained.

[0065] Step 142: Modulate the reference acoustic signal based on the target amplitude curve and the target phase offset to generate the cancellation acoustic signal.

[0066] It can be understood that after having the target amplitude curve and the target phase offset, a reference acoustic signal is selected. This reference acoustic signal can be a simple acoustic signal with a preset standard frequency, amplitude, and phase. Then, the reference acoustic signal is modulated according to the target amplitude curve and the target phase offset. For example, at each time point, the amplitude of the reference acoustic signal is adjusted according to the target amplitude curve, and at the same time, the phase of the reference acoustic signal is changed according to the target phase offset. Through the above modulation method, the amplitude and phase of the reference acoustic signal change in the expected manner, and finally, a cancellation acoustic signal with a phase opposite to that of the original time-domain noise signal is generated.

[0067] Step 143: Directionally emit the cancellation acoustic signal through the speaker array of the noise suppression device, so that the cancellation acoustic signal and the time-domain noise signal undergo destructive interference in the spatial superposition region to reduce the environmental noise intensity.

[0068] At the tunnel construction site, the noise suppression device is equipped with a speaker array, and the above speakers are precisely arranged at adapted positions to ensure that cancellation sound wave signals can be effectively emitted. After the cancellation sound wave signals are generated, they are emitted through the speaker array in a set direction and angle. For example, according to the previously determined azimuth characteristics of the noise source, the speaker array is adjusted to a position where the cancellation sound wave signals can directly propagate in the direction of the noise source. The cancellation sound wave signals propagate in space and meet the original time-domain noise signals in a set superposition area. Since the cancellation sound wave signals are opposite in phase to the original time-domain noise signals, they undergo destructive interference in the superposition area. During the interference process, the amplitudes of the two sound waves cancel each other out, reducing the ambient noise intensity in this area. For example, the noise intensity in this area was originally a relatively high value, and after destructive interference, the noise intensity is reduced to an acceptable lower level.

[0069] As a non-limiting embodiment, after generating the cancellation sound wave signals that are opposite in phase to the time-domain noise signals to reduce the ambient noise intensity, it further includes: Step 400: Real-time collect the set of residual noise signals after cancellation. The set of residual noise signals includes the time-domain waveforms and spectral distributions of multiple residual noise components; perform energy threshold detection processing on the set of residual noise signals, and identify the residual noise components that exceed the preset energy threshold as target optimized noise components; extract the optimized characteristic parameters of the target optimized noise components. The optimized characteristic parameters include residual energy gradient, spectral offset, and phase residual deviation; perform correction processing on the set of suppression control characteristics according to the optimized characteristic parameters to generate a phase compensation amount and an amplitude compensation coefficient; update the waveform parameters of the cancellation sound wave signals based on the phase compensation amount and the amplitude compensation coefficient to achieve iterative suppression of the target optimized noise components.

[0070] After reducing the ambient noise intensity, use a preset acquisition device to real-time collect the set of residual noise signals after cancellation. For example, arrange acquisition devices at multiple positions in the tunnel to ensure that the residual noise signals in each area can be comprehensively collected. The above set of residual noise signals includes multiple residual noise components, and each component has its corresponding time-domain waveform and spectral distribution. Perform energy threshold detection processing on the collected set of residual noise signals, and preset an energy threshold. For example, this threshold is determined according to the actual noise control requirements and the ambient background noise level. Compare the energy of each residual noise component with this threshold, and identify the residual noise components whose energy exceeds the threshold as target optimized noise components.

[0071] For the above target optimized noise component, further extract its optimized characteristic parameters. For example, calculate the residual energy gradient, that is, the change rate of the residual noise energy over time; analyze the spectral offset, that is, the movement of the residual noise spectrum on the frequency axis compared with the original noise spectrum; measure the phase residual deviation, that is, the phase difference between the residual noise and the ideal cancellation state. According to the above optimized characteristic parameters, correct the suppression control feature set. For example, the phase compensation amount and the amplitude compensation coefficient can be calculated based on the residual energy gradient, the spectral offset, and the phase residual deviation. Finally, use the above phase compensation amount and amplitude compensation coefficient to update the waveform parameters of the cancellation acoustic signal, such as changing the amplitude size, adjusting the phase, etc., so as to emit the cancellation acoustic signal again to iteratively suppress the target optimized noise component and further reduce the environmental noise intensity.

[0072] As a non-limiting embodiment, after generating the cancellation acoustic signal with a phase opposite to that of the time-domain noise signal to reduce the environmental noise intensity, it further includes: Step 500: Obtain the set of operating state parameters of the construction equipment, where the set of operating state parameters includes equipment type, position coordinates, and working power; determine the priority weight coefficient according to the equipment type, and calculate the noise radiation influence range of each construction equipment in combination with the position coordinates; generate a multi-equipment noise coupling feature set based on the working power and the noise radiation influence range, where the coupling feature set characterizes the energy interference distribution in the spatial superposition region of different equipment noises; perform cooperative optimization processing on the suppression control feature set according to the coupling feature set, and adjust the emission timing and spatial pointing angle of the cancellation acoustic signal to reduce the superposition interference intensity of the multi-equipment noise.

[0073] During the tunnel construction process, a set of operating state parameters of construction equipment is obtained through an equipment monitoring system. For example, construction equipment includes excavators, loaders, concrete mixers, etc. The equipment monitoring system records in real time the type of each equipment, its position coordinates in the tunnel, and its current working power. Priority weight coefficients are determined according to the equipment type. Different types of equipment are given different weights due to their different noise characteristics and the degree of impact on the construction environment. For example, an excavator has a relatively high noise level and is important for the construction progress, so it may be given a relatively high weight coefficient; while some small auxiliary equipment has a relatively low weight coefficient. Combining the position coordinates of the equipment, the noise radiation influence range of each construction equipment is calculated using an acoustic model and algorithms. For example, considering factors such as the propagation distance of sound waves, the geometric shape of the tunnel, and the medium characteristics, the spatial area that can be affected by the noise of each equipment is determined. Based on the working power of the equipment and the noise radiation influence range, a multi-equipment noise coupling feature set is generated, which describes the energy interference distribution of different equipment noises in the spatial superposition area. For example, analyze the superposition effect of the noises of an excavator and a loader in a certain area, and determine the position and degree of energy enhancement or weakening. The suppression control feature set is co-optimized according to this coupling feature set. For example, by adjusting the emission timing of the cancellation sound wave signal, the cancellation sound wave signal can play a more effective role during the critical time period when different equipment noises are superposed; at the same time, adjust the spatial pointing angle of the speaker array to emit the cancellation sound wave signal more accurately to the noise superposition area, thereby reducing the superposition interference intensity of multi-equipment noise and further improving the noise condition of the tunnel construction environment.

[0074] As a non-limiting embodiment, after generating the cancellation sound wave signal with a phase opposite to that of the time-domain noise signal to reduce the environmental noise intensity, it further includes: Step 600: Perform sound field sampling processing on the superposition area of the cancellation sound wave signal and the time-domain noise signal to generate a sound pressure distribution map; extract the local sound pressure area coordinates and energy fluctuation frequencies in the sound pressure distribution map; adjust the spatial pointing angle of the speaker array of the noise suppression device according to the local sound pressure area coordinates, and generate frequency-domain equalization parameters based on the energy fluctuation frequencies; perform frequency band matching processing on the frequency-domain equalization parameters and the suppression control feature set, and reconstruct the spectral profile of the cancellation sound wave signal to cover the full frequency band of the energy fluctuation frequencies.

[0075] Optionally, a plurality of sound field sampling points are arranged in the superposition region of the counteracting acoustic wave signal and the time-domain noise signal, and relevant sound field sampling equipment is used to collect sound pressure data according to a preset time interval and spatial distribution. For example, a sampling point is set at a certain distance within the superposition region, and the sound pressure signal is continuously collected for a period of time. By processing and analyzing the collected sound pressure data, a sound pressure distribution map is generated, which intuitively shows the spatial distribution of sound pressure within the superposition region. From the sound pressure distribution map, local sound pressure regions, that is, regions with relatively high sound pressure values, are identified, and the coordinates of the above regions are recorded. At the same time, the energy fluctuation situation of the sound pressure signal is analyzed to determine the energy fluctuation frequency. For example, it is observed that within certain time periods, the energy of the sound pressure signal exhibits fluctuations corresponding to a certain frequency. According to the coordinates of the local sound pressure regions, the spatial pointing angle of the speaker array of the noise suppression device is adjusted, and the speaker is aligned with the above high-energy region to enhance the cancellation effect. Further, frequency-domain equalization parameters can be generated based on the energy fluctuation frequency, and the above parameters are used to adjust the amplitude and phase of the counteracting acoustic wave signal at different frequencies to achieve a better frequency response. The frequency-domain equalization parameters are subjected to band matching processing with the suppression control feature set. For example, the frequency-related parameters in the suppression control feature set are adjusted and optimized according to the frequency-domain equalization parameters. By the above method, the spectral profile of the counteracting acoustic wave signal is reconstructed so that it can cover the entire frequency band of the energy fluctuation frequency, thereby more effectively suppressing the noise in this region and further improving the acoustic quality of the tunnel construction environment.

[0076] Furthermore, when implementing the above technical solutions, those skilled in the art can perform systematic optimization processing based on the adaptive filter bank design, short-time Fourier transform algorithm, direction-of-arrival estimation method, deep learning model architecture, and acoustic interference principle in the prior art, so as to achieve the logical connection optimization, spatial positioning, and dimension unification of the filter bank and spectral analysis.

[0077] In addition, for multi-scale decomposition and spectral analysis, an improved filter bank joint analysis framework can be adopted. The transition bandwidth matching the short-time Fourier transform window function is preset in the adaptive filter bank design stage, and the overlap-save method is introduced to ensure the alignment of sub-signal boundaries. For the direction-of-arrival estimation processing in spatial sound field modeling, the generalized cross-correlation time-delay estimation and MUSIC algorithm can be combined to construct an azimuth calculation model based on eigen-subspace decomposition, and a tunnel wall reflection coefficient compensation mechanism is introduced to correct the propagation path analysis.

[0078] In terms of dimension consistency processing, the normalization method in the IEEE 1451 smart sensor standard can be adopted. The phase radian value and amplitude volt value are mapped to the [-1, 1] interval through Z-score standardization, and a phase-amplitude synchronous matching mechanism based on dynamic time warping is constructed.

[0079] Specifically, for the deep learning model architecture, the U-Net encoding and decoding structure can also be referred to. In the encoder part, parallel multi-scale convolutional kernels (such as 32×1, 64×3, 128×5) are used in combination with the channel attention mechanism. In the decoder part, transposed convolution and skip connection structures are used, and relevant activation functions (such as PReLU) and loss functions (weighted mean square error) are introduced.

[0080] By introducing an adaptive filtering algorithm (such as FxLMS) in active noise control, a propagation delay compensation module is added during the acoustic wave emission stage, and the cross-correlation method is used to calculate the propagation delay from the sound source to the speaker array in real time, which can ensure the spatio-temporal synchronization of the cancellation acoustic wave and the noise signal, thus eliminating the problem of cancellation interference failure caused by not considering the acoustic wave propagation speed.

[0081] In summary, in the embodiments of the present application, by collecting the original noise signal set covering multiple time-domain noise signals and their included construction equipment noise components, the complex and diverse noise information in the tunnel construction environment is completely obtained; then based on the noise feature distribution obtained by noise feature extraction and processing, the noise characteristics can be comprehensively described from multiple dimensions including noise spectrum features, noise source azimuth features, and noise energy propagation features; then the above features are suppressed by using the deep learning noise suppression model, and the targeted suppression control feature set can be intelligently analyzed and generated; further, based on this set, the noise suppression device is driven to perform noise cancellation operations, and the cancellation acoustic wave signals with phases opposite to those of each time-domain noise signal can be accurately generated.

[0082] Designed in this way, the environmental noise intensity is effectively and specifically reduced, the potential hazards caused by noise are reduced, and at the same time, the accuracy and efficiency of construction operations are improved, avoiding poor communication and operation errors caused by noise interference, and overall improving the quality and safety of the tunnel construction process.

[0083] See Figure 2 As shown in the figure, this figure is a schematic diagram of the basic structure of a tunnel construction environment noise suppression system 200 applying a deep learning algorithm provided by the embodiments of the present application. The tunnel construction environment noise suppression system 200 applying a deep learning algorithm includes: A processor 201; A storage device 202, on which a computer program 2020 is stored; When the computer program 2020 is executed by the processor 201, the processor 201 realizes any of the tunnel construction environment noise suppression methods applying a deep learning algorithm.

[0084] On this basis, a readable storage medium is provided. Programs or instructions are stored on the readable storage medium, and when the programs or instructions are executed by a processor, the steps of the above method are realized.

[0085] It should be noted that the various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the systems or devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For related parts, reference can be made to the description in the method section.

Claims

1. A method for suppressing tunnel construction environmental noise using deep learning algorithms, characterized in that, Including: Collecting an original noise signal set at the tunnel construction site, where the original noise signal set includes multiple time-domain noise signals, and each time-domain noise signal contains at least one noise component generated by a construction device; Performing noise feature extraction processing on the original noise signal set to obtain the noise feature distribution of each time-domain noise signal, where the noise feature distribution includes noise spectrum features, noise source azimuth features, and noise energy propagation features; Invoking a deep learning noise suppression model to perform suppression processing on the noise feature distribution to generate a suppression control feature set corresponding to each time-domain noise signal; Based on the suppression control feature set, driving a noise suppression device to perform a noise cancellation operation to generate a cancellation sound wave signal with a phase opposite to that of the time-domain noise signal to reduce the environmental noise intensity.

2. The method according to claim 1, wherein The performing noise feature extraction processing on the original noise signal set to obtain the noise feature distribution of each time-domain noise signal includes: Performing multi-scale decomposition processing on the time-domain noise signal to generate multiple noise sub-signals, and each noise sub-signal corresponds to a noise component in a different frequency band; Performing spectrum analysis processing on the multiple noise sub-signals to extract the frequency-domain energy distribution feature of each noise sub-signal as the noise spectrum feature; Performing spatial sound field modeling processing on the time-domain noise signal to determine the sound source azimuth angle and the sound wave propagation path feature of the time-domain noise signal in three-dimensional space, and fusing the sound source azimuth angle and the sound wave propagation path feature into the noise source azimuth feature; Performing energy attenuation analysis processing on the time-domain noise signal, calculating the energy attenuation curve of the time-domain noise signal according to the sound wave propagation medium attribute and the propagation distance, and extracting the attenuation gradient feature of the energy attenuation curve as the noise energy propagation feature; Performing joint coding processing on the noise spectrum feature, the noise source azimuth feature, and the noise energy propagation feature to generate the noise feature distribution of the time-domain noise signal.

3. The method according to claim 2, wherein The performing multi-scale decomposition processing on the time-domain noise signal to generate multiple noise sub-signals includes: Obtaining multiple adaptive filter banks with different scales, and each adaptive filter bank includes at least one band-pass filter, where the scale parameter of the adaptive filter bank is adjusted according to the average spectrum density of the time-domain noise signal; Inputting the time-domain noise signal into the multiple adaptive filter banks with different scales for parallel filtering processing to obtain multiple initial noise sub-signals divided by frequency bands; Performing reconstruction error correction processing on the overlapping frequency band regions of the initial noise sub-signals to eliminate the signal aliasing effect between adjacent frequency bands and generate the multiple noise sub-signals.

4. The method according to claim 2, wherein The performing spectrum analysis processing on the multiple noise sub-signals to extract the frequency-domain energy distribution feature of each noise sub-signal includes: Performing short-time Fourier transform processing on each noise sub-signal to generate a corresponding time-frequency spectrogram; Performing peak detection processing on the time-frequency spectrogram to identify the set of energy peak points in the time-frequency spectrogram; Construct an energy distribution feature set based on the distribution density, peak amplitude, and frequency interval of the set of energy peak points, and extract the statistical features of the energy distribution feature set as the frequency-domain energy distribution features.

5. The method according to claim 2, characterized in that The spatial sound field modeling process for the time-domain noise signal to determine the sound source azimuth angle and acoustic wave propagation path characteristics of the time-domain noise signal in three-dimensional space includes: Collect the sound pressure signals of the time-domain noise signal at different spatial positions according to multiple deployed spatial microphone array nodes; Calculate the spatial propagation vector of the time-domain noise signal according to the time difference of arrival and phase difference of the sound pressure signals; Perform direction-of-arrival estimation processing based on the spatial propagation vector to determine the sound source azimuth angle; Construct a three-dimensional sound field propagation model according to the position coordinates of the spatial microphone array nodes and the sound source azimuth angle, and extract the path attenuation coefficient and reflection path characteristics of the three-dimensional sound field propagation model as the acoustic wave propagation path characteristics.

6. The method according to claim 1, characterized in that, The calling of the deep learning noise suppression model to suppress the noise feature distribution to generate a set of suppression control features corresponding to each time-domain noise signal includes: Input the noise feature distribution into the encoder module of the deep learning noise suppression model for feature compression processing to generate a noise-encoded feature vector; Input the noise-encoded feature vector into the decoder module of the deep learning noise suppression model for suppression parameter reconstruction processing to generate an initial suppression parameter sequence; Perform time-domain alignment processing on the initial suppression parameter sequence, and adjust the phase and amplitude of the initial suppression parameter sequence according to the time stamp of the time-domain noise signal to generate the set of suppression control features.

7. The method according to claim 6, wherein The performing of time-domain alignment processing on the initial suppression parameter sequence, and adjusting the phase and amplitude of the initial suppression parameter sequence according to the time stamp of the time-domain noise signal to generate the set of suppression control features includes: Obtain the time stamp sequence of the time-domain noise signal, where the time stamp sequence contains the timing identifiers of each sampling point of the time-domain noise signal; Perform time stamp synchronization processing on the initial suppression parameter sequence, and perform interpolation reconstruction on the initial suppression parameter sequence according to the sampling interval of the time stamp sequence to generate a time stamp-synchronized suppression parameter sequence; Extract the phase reference curve of the time-domain noise signal, and calculate the set of phase offset amounts of the time stamp-synchronized suppression parameter sequence, where the set of phase offset amounts contains the absolute values of the phase differences of each sampling point; Perform phase compensation processing on the time stamp-synchronized suppression parameter sequence according to the set of phase offset amounts to generate a phase-aligned suppression parameter sequence; Extract the amplitude envelope curve of the time-domain noise signal, and calculate the set of amplitude scaling factors of the phase-aligned suppression parameter sequence, where the set of amplitude scaling factors contains the amplitude ratio coefficients of each sampling point; Perform amplitude adjustment processing on the phase-aligned suppression parameter sequence according to the set of amplitude scaling factors to generate an amplitude-matched suppression parameter sequence; Perform dimensional consistency processing on the amplitude-matched suppression parameter sequence to unify the units of the phase dimension and amplitude dimension of the suppression parameter sequence. Perform feature dimension splicing processing on the phase parameters and amplitude parameters of the suppression parameter sequence with the same amplitude to generate the suppression control feature set.

8. The method according to claim 6, wherein The training process of the deep learning noise suppression model includes: Obtain a historical noise signal sample set and corresponding suppression effect labels, where the suppression effect labels include the residual noise energy value after noise suppression; Obtain an initial neural network model, where the initial neural network model includes a convolutional encoding layer, an attention fusion layer, and a fully connected decoding layer; Input the historical noise signal sample set into the initial neural network model for forward propagation processing to generate a predicted suppression parameter sequence; Calculate the mean square error loss based on the predicted suppression parameter sequence and the suppression effect labels, and update the weight parameters of the initial neural network model based on the mean square error loss until convergence to obtain the deep learning noise suppression model; Among them, the convolutional encoding layer includes multiple parallel convolutional channels, each convolutional channel corresponding to a convolutional kernel of a different scale. The processing process of the convolutional encoding layer includes: inputting the noise feature distribution into the multiple parallel convolutional channels respectively for feature mapping processing to generate intermediate feature maps of multiple scales; performing cross-channel feature fusion processing on the intermediate feature maps of multiple scales to generate a fused feature map; performing max pooling processing on the fused feature map to generate the noise encoding feature vector.

9. The method according to claim 1, wherein Driving the noise suppression device to perform a noise cancellation operation based on the suppression control feature set to generate a cancellation sound wave signal with a phase opposite to that of the time-domain noise signal includes: Generating a target amplitude curve according to the amplitude feature in the suppression control feature set; Generating a target phase offset according to the phase feature in the suppression control feature set; Modulating a reference sound wave signal based on the target amplitude curve and the target phase offset to generate the cancellation sound wave signal; Directionally transmitting the cancellation sound wave signal through the speaker array of the noise suppression device, so that the cancellation sound wave signal and the time-domain noise signal undergo destructive interference in the spatial superposition region to reduce the environmental noise intensity.

10. A tunnel construction environment noise suppression system applying deep learning algorithms, characterized in that, Including: A processor; A storage device on which a computer program is stored. When the computer program is executed by the processor, the processor implements the tunnel construction environment noise suppression method using a deep learning algorithm as described in any one of claims 1-9.

Images