A method, apparatus and medium for monitoring displacement of a slope based on quantum single photons

Abstract

The application discloses a kind of slope displacement monitoring method, equipment and medium based on quantum single photon, belong to displacement monitoring technical field.The application is by emitting rectangular light pulse according to cycle, utilizes single-photon detector to complete echo detection and constructs photon counting time sequence, carries out wavelet denoising and reconstruction processing to it, extracts main peak position shift, energy change and form distortion matches multi-class feature vector, then carries out feature fusion and solution by multi-source feature fusion BP neural network, finally outputs high-precision slope displacement result.The application combines quantum single-photon detection technology with wavelet denoising, neural network fusion algorithm, effectively suppresses noise interference, improves displacement feature extraction accuracy, and is suitable for high-precision, long-distance real-time monitoring of slope micro-deformation.

Description

Technical Field

[0001] This invention relates to the field of displacement monitoring technology, specifically to a method, device, and medium for slope displacement monitoring based on quantum single photons. Background Technology

[0002] Slopes, as important geological structures formed by the interaction of natural geological environment and engineering construction, are widely distributed around highways, railways, mines, and water conservancy projects. Their stability is related to the normal operation of infrastructure. Under the influence of external factors such as rainstorms, earthquakes, and extreme weather, slopes are prone to geological disasters such as creep and landslides. Therefore, carrying out accurate and efficient slope displacement monitoring and timely capturing slope deformation characteristics is of great significance for geological disaster early warning and engineering safety protection.

[0003] Currently, several mature monitoring technologies have been developed in the field of slope displacement monitoring, among which distributed optical fiber sensing monitoring technology is one of the most widely used existing technologies. This technology lays an optical fiber sensor network inside and on the surface of the slope, and uses the strain and loss changes of optical signals as they are transmitted in the optical fiber to invert the slope displacement. It can achieve long-distance, continuous monitoring and has been widely used in large-scale slope engineering projects.

[0004] However, this slope displacement monitoring technology based on distributed optical fiber sensing has significant technical problems: external environmental noise easily intrudes into the optical signal transmission process, causing a large amount of interference information to be mixed in with the monitored displacement-related data, making it difficult to accurately capture minute slope displacement characteristics, and thus affecting the accuracy of displacement monitoring. Therefore, there is a problem of low monitoring accuracy in slope displacement monitoring. Summary of the Invention

[0005] To address the aforementioned shortcomings in existing technologies, this invention provides a slope displacement monitoring method, device, and medium based on quantum single photons, which solves the problem of low monitoring accuracy in existing technologies.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a slope displacement monitoring method based on quantum single photons, comprising the following steps:

[0007] S1. Periodically emit rectangular light pulses;

[0008] S2. Echo detection is performed using a single-photon detector, and a photon counting time series for each cycle is constructed;

[0009] S3. Perform wavelet denoising and reconstruction on the photon counting time series of each cycle to obtain the denoised wavelet coefficients and denoised photon counting time series of each cycle.

[0010] S4. Based on the denoised wavelet coefficients and denoised photon count time series of each cycle, construct the main peak position shift feature vector, energy change feature vector, and morphological distortion matching feature vector;

[0011] S5. A multi-source feature fusion BP neural network is used to process the main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector to obtain the slope displacement result.

[0012] Furthermore, the specific process of S2 includes:

[0013] The echo is detected by a single-photon detector, and the time number corresponding to each detected echo photon within the period is recorded to obtain the echo photon time.

[0014] For each cycle, iterate through all detected echo photon moments;

[0015] Use an indicator function to determine whether each time point is equal to the echo photon time point. If it is, the indicator function value is 1; otherwise, it is 0.

[0016] Summing all indicator function values ​​within a period yields the number of photons at the corresponding moment within that period.

[0017] Arrange the photon counts at each moment within each cycle in chronological order to obtain the photon count time series.

[0018] Furthermore, the specific process of S3 includes:

[0019] Continuous wavelet transform is performed on the photon counting time series for each period to obtain wavelet coefficients;

[0020] Wavelet threshold denoising is performed on the wavelet coefficients to obtain the denoised wavelet coefficients;

[0021] Wavelet reconstruction is performed on the denoised wavelet coefficients to obtain the denoised photon counting time series.

[0022] Furthermore, the process of constructing the feature vector of the main peak position offset includes:

[0023] Find the moment corresponding to the maximum value in the denoised photon count time series of each cycle to obtain the peak moment;

[0024] The offset of the main peak position is obtained by subtracting the reference main peak position from the peak time of each cycle and taking the absolute value.

[0025] The ratio of the main peak position offset to the reference main peak position is used as the main peak position offset feature.

[0026] Arrange the peak position shift features of each cycle in chronological order to obtain the peak position shift feature vector.

[0027] Furthermore, the process of constructing the energy change feature vector includes:

[0028] Square all the denoised wavelet coefficients within a period and sum them up to obtain the wavelet coefficient energy for the corresponding period.

[0029] Subtract the wavelet coefficient energy of the current period from that of the previous period to obtain the energy difference of the current period;

[0030] Take the absolute value of the energy difference for the current cycle to obtain the absolute energy difference for the current cycle;

[0031] If the energy difference in the current cycle is less than 0 and the offset of the main peak position in the current cycle is greater than the offset threshold, the ratio of the absolute energy difference in the current cycle to the baseline energy is used as the energy change feature. Otherwise, the energy change feature is set to a preset value.

[0032] Arrange the energy change characteristics of each cycle in chronological order to obtain the energy change characteristic vector.

[0033] Furthermore, the process of constructing morphological distortion matching feature vectors includes:

[0034] The difference in half-width at half-maximum (WHM) of the current period is obtained by subtracting the full width at half-maximum (FWHM) of the main peak from that of the previous period.

[0035] The ratio of the absolute value of the half-width difference of the current period to the reference half-width is used as the morphological distortion matching feature of the current period.

[0036] The morphological distortion matching features of each period are arranged in chronological order to obtain the morphological distortion matching feature vector.

[0037] Furthermore, the multi-source feature fusion BP neural network includes: a first 1D temporal convolutional layer, a second 1D temporal convolutional layer, a third 1D temporal convolutional layer, a temporal feature enhancement unit, a feature mapping unit, and a BP neural network;

[0038] The input of the first 1D temporal convolutional layer is used to input the main peak position offset feature vector;

[0039] The input of the second 1D temporal convolutional layer is used to input the energy change feature vector;

[0040] The input of the third 1D temporal convolutional layer is used to input the morphological distortion matching feature vector;

[0041] The input of the temporal feature enhancement unit is connected to the output of the first 1D temporal convolutional layer, the output of the second 1D temporal convolutional layer, and the output of the third 1D temporal convolutional layer, respectively, and its output is connected to the input of the feature mapping unit.

[0042] The input of the BP neural network is connected to the output of the feature mapping unit, and its output serves as the output of the multi-source feature fusion BP neural network.

[0043] Furthermore, the processing steps of the multi-source feature fusion BP neural network include:

[0044] The main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector are all processed through a 1D temporal convolutional layer to obtain the main peak position offset encoding vector, energy change encoding vector, and morphological distortion matching encoding vector.

[0045] The temporal feature enhancement vector is obtained by multiplying the main peak position offset encoding vector, energy change encoding vector and morphological distortion matching encoding vector element by element through the temporal feature enhancement unit.

[0046] A feature mapping unit is used to perform deep feature mapping on the temporal feature enhancement vector to extract high-dimensional representation features;

[0047] By inputting the high-dimensional representation features into the BP neural network, the slope displacement results are obtained.

[0048] Furthermore, the feature mapping unit comprises, in sequence: a fourth 1D temporal convolutional layer, a first max pooling layer, a fifth 1D temporal convolutional layer, a second max pooling layer, a global average pooling layer, and a fully connected layer;

[0049] The parameters of the fourth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 16, activation function of ReLU, output feature size of 16×M, where M is the feature length;

[0050] The parameters of the first max pooling layer include: kernel size of 2, stride of 2, and output feature size of 16×M / 2.

[0051] The parameters of the fifth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 32, activation function of ReLU, and output feature size of 32×M / 2.

[0052] The parameters of the second max pooling layer are: kernel size of 2, stride of 2, and output feature size of 32×M / 4.

[0053] The global average pooling layer averages the values ​​for each channel, resulting in a channel-level feature output size of 32.

[0054] The output dimension of the fully connected layer is 64, and the activation function is ReLU.

[0055] A computer device includes a processor and a memory; the memory stores computer program instructions, which, when executed by the processor, implement a slope displacement monitoring method based on quantum single photons.

[0056] A computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement a slope displacement monitoring method based on quantum single photons.

[0057] The beneficial effects of this invention are as follows:

[0058] 1. This invention employs quantum single-photon detection technology, which utilizes the high sensitivity of single-photon detectors to accurately capture weak echo light signals. Simultaneously, through wavelet denoising and reconstruction processing, it can selectively filter out irrelevant information such as environmental noise and transmission interference, significantly reducing the impact of interference signals on monitoring data and ensuring the authenticity and stability of the photon counting time series after denoising.

[0059] 2. Based on noise reduction processing, this invention constructs three major feature vectors: main peak position shift, energy change, and morphological distortion matching. These vectors comprehensively capture subtle changes in the optical signal during slope displacement. The main peak position shift is directly related to the magnitude of the displacement, the energy change reflects the signal attenuation characteristics during the displacement process, and the morphological distortion matching reflects the signal waveform variation caused by the displacement. The multi-dimensional features work together to cover the core information of displacement monitoring, effectively solving the problem that traditional technologies cannot capture minute displacements, significantly improving the accuracy of slope displacement monitoring, and accurately identifying potential minor displacement hazards.

[0060] 3. This invention employs a multi-source feature fusion BP neural network to process the three major feature vectors. The BP neural network has powerful nonlinear fitting and feature fusion capabilities, which can effectively integrate multi-dimensional feature information, reduce the limitations and errors of single feature monitoring, and at the same time, the quantum single-photon detection technology itself has strong anti-electromagnetic interference and anti-ambient light interference capabilities. Compared with traditional distributed fiber optic sensing technology, it can better adapt to the complex field environment of slopes (such as strong light, electromagnetic interference, temperature fluctuations, etc.), reduce the impact of environmental factors on monitoring results, improve the stability and reliability of monitoring, and ensure the accuracy of long-term continuous monitoring. Attached Figure Description

[0061] Figure 1 A flowchart of a slope displacement monitoring method based on quantum single photons;

[0062] Figure 2 This is a schematic diagram of the structure of a multi-source feature fusion BP neural network;

[0063] Figure 3 This is a schematic diagram of the feature mapping unit. Detailed Implementation

[0064] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0065] Example 1: As Figure 1 As shown, a slope displacement monitoring method based on quantum single photons includes the following steps:

[0066] S1. Periodically emit rectangular light pulses;

[0067] S2. Echo detection is performed using a single-photon detector, and a photon counting time series for each cycle is constructed;

[0068] S3. Perform wavelet denoising and reconstruction on the photon counting time series of each cycle to obtain the denoised wavelet coefficients and denoised photon counting time series of each cycle.

[0069] S4. Based on the denoised wavelet coefficients and denoised photon count time series of each cycle, construct the main peak position shift feature vector, energy change feature vector, and morphological distortion matching feature vector;

[0070] S5. A multi-source feature fusion BP neural network is used to process the main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector to obtain the slope displacement result. In this embodiment, the expression for the rectangular light pulse is:

[0071] ,

[0072] in, For the first Within the first cycle A rectangular light pulse at a given time. For pulse peak power, The pulse width. The time within a single launch cycle. This is a rectangle function.

[0073] Rectangle function It can keep the peak power of the optical pulse stable within a single emission cycle, ensuring that the signal energy is concentrated and evenly distributed, which makes it easier for single-photon detectors to accurately capture the echo signal and reduce counting errors caused by irregular pulse shape.

[0074] In this embodiment, a semiconductor quantum dot single-photon source is used to emit a rectangular light pulse with a wavelength of 1550nm (low transmission loss, anti-interference). After being injected into an optical fiber, the pulse is transmitted to a quantum reflective film (reflection efficiency ≥95%) at the slope monitoring point. The single photon is reflected by the reflective film to form an echo signal, which returns to the single-photon detector along the optical fiber to complete the echo acquisition.

[0075] In this embodiment, S2 records the time of the echo photon using a single-photon detector and calculates the photon count at different time points within each period using an indicator function, thereby constructing a photon count time series. The specific process includes:

[0076] Echo detection is performed using a single-photon detector, and the time number corresponding to each detected echo photon within the period is recorded. The echo photon time was obtained. ;

[0077] For each cycle, iterate through all detected echo photon moments. ;

[0078] Use an indicator function to determine each time step. Is it equal to the echo photon moment? If yes, the indicator function value is 1; otherwise, it is 0.

[0079] For each moment The summation of all indicator function values ​​within one period yields the number of photons at the corresponding moment within that period.

[0080] Arrange the photon counts at each moment within each period in chronological order to obtain the photon counting time series. .

[0081] The formula for calculating the number of photons is:

[0082] ,

[0083] in, For the first Within the first cycle The number of photons at any given time For indicator functions, For the first Within the first cycle The moment when the echo photon is detected. The time number is used to identify the time within each period. For the first The total number of echo photons detected within a period, For the first The first detected within the cycle A photon echo, Used to determine the moment within the period. Is it equal to the first? Within the first cycle The arrival time of each echo photon If they are equal, the result is 1; otherwise, it is 0.

[0084] It is the time number used to identify time points on the horizontal axis in the photon counting time series.

[0085] A single-photon detector performs echo detection, recording the arrival time of each echo photon. The indicator function serves to "filter valid moments and quantify the number of photons"—a single moment. There may be multiple echo photons arriving, or there may be none. By assigning a binary value of "1 for matching and 0 for non-matching", the total number of photons at each moment can be quickly counted.

[0086] The detection construction process of this invention records the arrival time of echo photons, uses an indicator function to count the number of photons at each time step, and arranges them in time sequence to form a photon counting time series. This can transform discrete and random single-photon echo signals into regular and quantifiable time-series data, ensuring that the photon counting results are accurate and reliable, while completely preserving the time distribution characteristics of the echo signal.

[0087] In this embodiment, the specific process of denoising and reconstruction includes:

[0088] Continuous wavelet transform is performed on the photon counting time series for each period to obtain wavelet coefficients;

[0089] Wavelet threshold denoising is performed on the wavelet coefficients to obtain the denoised wavelet coefficients;

[0090] Wavelet reconstruction is performed on the denoised wavelet coefficients to obtain the denoised photon counting time series.

[0091] In this embodiment, the db4 mother wavelet is selected for wavelet transform, and a 4-level wavelet decomposition is performed.

[0092] The noise reduction formula is:

[0093] , ,

[0094] in, For the first Wavelet coefficients after denoising each cycle For the first Wavelet coefficients for each period, For the first Wavelet threshold for each period, The noise standard deviation of the wavelet coefficients. For the first Length of photon counting time series per cycle For periodic numbering, It is a logarithmic function.

[0095] Noise standard deviation of wavelet coefficients The global noise standard deviation is calculated using the median method (MAD), which is well-known in the field.

[0096] Denoising and reconstruction employs db4 mother wavelet for 4-level wavelet decomposition, and combines an adaptive threshold denoising formula to sequentially perform wavelet transform, threshold denoising, and signal reconstruction on the photon counting time series. This effectively suppresses environmental noise and random interference while preserving effective signal characteristics, improves the smoothness and signal-to-noise ratio of the photon counting time series, and completely preserves the signal change information corresponding to the small displacement of the slope.

[0097] In this embodiment, wavelet coefficients include two types: approximation coefficients and detail coefficients. The approximation coefficients are mainly based on the effective signal, while the denoising process targets the detail coefficients, preserving the approximation coefficients.

[0098] This invention uses wavelet denoising to ensure that displacement features, rather than noise features, are accurately extracted during subsequent feature extraction.

[0099] In this embodiment, the process of constructing the main peak position offset feature vector includes:

[0100] Find the moment corresponding to the maximum value in the denoised photon count time series of each cycle to obtain the peak moment;

[0101] The offset of the main peak position is obtained by subtracting the reference main peak position from the peak time of each cycle and taking the absolute value.

[0102] The ratio of the main peak position offset to the reference main peak position is used as the main peak position offset feature.

[0103] Arrange the peak position shift features of each cycle in chronological order to obtain the peak position shift feature vector.

[0104] In this embodiment, ,in, For the first The shift in the position of the main peak over each cycle. For the first The peak moment of each cycle, The reference main peak location; ,in, For the first The characteristics of the main peak position shift in each cycle.

[0105] This invention obtains the main peak position offset by subtracting the peak time, then compares the offset with the reference main peak position to obtain the main peak position offset feature, and arranges them in periodic order to form a feature vector. This can intuitively and stably reflect the temporal offset law of the photon counting echo signal under the change of the transmission path, and effectively extract the signal position change information caused by slope displacement.

[0106] The reference peak position is obtained by processing the denoised photon counting time series collected under the initial stable state: under the initial state where the slope has not shifted and is in stable operation, several periods of photon counting time series are collected and reconstructed by wavelet denoising. The period with stable signal and no noise interference is selected, and the peak time of its photon counting time series is extracted. This peak time is taken as the reference peak position.

[0107] In this embodiment, the process of constructing the energy change feature vector includes:

[0108] Square all the denoised wavelet coefficients within a period, including the approximation coefficients and the denoised detail coefficients, and sum the squares of all the wavelet coefficients within a period to obtain the wavelet coefficient energy for the corresponding period.

[0109] Subtract the wavelet coefficient energy of the current period from that of the previous period to obtain the energy difference of the current period;

[0110] Take the absolute value of the energy difference for the current cycle to obtain the absolute energy difference for the current cycle;

[0111] If the energy difference in the current cycle is less than 0 and the offset of the main peak position in the current cycle is greater than the offset threshold, the ratio of the absolute energy difference in the current cycle to the baseline energy is used as the energy change feature. Otherwise, the energy change feature is set to a preset value.

[0112] Arrange the energy change characteristics of each cycle in chronological order to obtain the energy change characteristic vector.

[0113] In this embodiment, the formula for calculating the energy change characteristics is:

[0114] ,

[0115] in, For the first Energy change characteristics of each cycle For the first Energy of wavelet coefficients per period For the first Energy of wavelet coefficients per period As the reference energy, And, For other cases, | represents the absolute value operation. For the first The shift in the position of the main peak over each cycle. This is the offset threshold. This represents the average energy of the wavelet coefficients over the past 10 periods.

[0116] In this embodiment, the offset threshold The time interval is set to 3 time units, meaning that when the main peak shifts by more than 3 time units, it is considered that a valid displacement has occurred.

[0117] This invention calculates the periodic energy by squaring and summing the wavelet coefficients, which accurately reflects the intensity distribution of the echo signal within each period. By calculating the energy difference between adjacent periods, it can sensitively capture the attenuation or fluctuation of the echo signal caused by slope displacement. Introducing the main peak position offset as a joint judgment condition can eliminate energy changes caused by non-displacement factors such as noise, temperature drift, and system disturbances, thus avoiding false features. Normalized energy features are only used when the energy decreases and the offset exceeds the threshold, which can enhance the signal changes corresponding to the true slope displacement and improve the reliability and recognizability of the features. In other cases, preset values ​​are used to prevent outliers from affecting the recognition accuracy of the subsequent neural network.

[0118] Actual displacement only causes energy to decrease (echo diffusion), while environmental disturbances (such as strong light) cause energy to surge—through " "Directly exclude the interference scenario of "sudden energy increase" and only retain the scenario of "energy decrease".

[0119] In some scenarios (such as slight shading), the energy may decrease, but there is no main peak shift (not due to displacement) – through " "Ensure that the 'energy drop' is caused by displacement."

[0120] In this embodiment, the process of constructing the morphological distortion matching feature vector includes:

[0121] The difference in half-width at half-maximum (WHM) of the current period is obtained by subtracting the full width at half-maximum (FWHM) of the main peak from that of the previous period.

[0122] The ratio of the absolute value of the half-width difference of the current period to the reference half-width is used as the morphological distortion matching feature of the current period.

[0123] The morphological distortion matching features of each period are arranged in chronological order to obtain the morphological distortion matching feature vector.

[0124] In this embodiment, the formula for calculating the morphological distortion matching feature is:

[0125] ,

[0126] in, For the first Morphological distortion matching features for each cycle, For the first The half-width at half-maximum of the main peak in each cycle For the first The half-width at half-maximum of the main peak in each cycle The base half-height and width are represented by | | for absolute value operations.

[0127] Reference half-width The average half-width and height over the past 10 periods.

[0128] The process of obtaining the full width at half maximum (FWHM) of the main peak includes: finding the two time points on either side of the main peak where the signal value is equal to half the peak value, and identifying the time point on the left. and right-hand moment Using the right-hand side time Subtract the time on the left The half-height and width of the main peak were obtained.

[0129] This invention constructs a morphological distortion matching feature vector, which can positively vectorize the degree of morphological distortion of photon echo signals. It effectively characterizes subtle changes such as echo signal broadening and distortion caused by slope displacement. The feature values ​​are positively correlated with the degree of displacement and maintain the same trend as the main peak position shift feature and energy change feature. In the subsequent multi-feature fusion process, it can achieve co-directional enhancement, effectively amplify the real displacement signal, suppress environmental interference signals, and significantly improve the accuracy, reliability and anti-interference ability of displacement identification.

[0130] like Figure 2 As shown, the multi-source feature fusion BP neural network includes: a first 1D temporal convolutional layer, a second 1D temporal convolutional layer, a third 1D temporal convolutional layer, a temporal feature enhancement unit, a feature mapping unit, and a BP neural network;

[0131] The input of the first 1D temporal convolutional layer is used to input the main peak position offset feature vector;

[0132] The input of the second 1D temporal convolutional layer is used to input the energy change feature vector;

[0133] The input of the third 1D temporal convolutional layer is used to input the morphological distortion matching feature vector;

[0134] The input of the temporal feature enhancement unit is connected to the output of the first 1D temporal convolutional layer, the output of the second 1D temporal convolutional layer, and the output of the third 1D temporal convolutional layer, respectively, and its output is connected to the input of the feature mapping unit.

[0135] The input of the BP neural network is connected to the output of the feature mapping unit, and its output serves as the output of the multi-source feature fusion BP neural network.

[0136] In this embodiment, the parameters of the first 1D temporal convolutional layer, the second 1D temporal convolutional layer, and the third 1D temporal convolutional layer are the same, including: a kernel size of 1×3, a stride of 1, padding of 1, an output channel of 1, an activation function of ReLU, and an output feature size of 1×M. The input and output lengths of the three 1D temporal convolutional layers are equal, and M is the length of the main peak position offset encoding vector, the energy change encoding vector, and the morphological distortion matching encoding vector.

[0137] The processing steps of a multi-source feature fusion BP neural network include:

[0138] The main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector are all processed through a 1D temporal convolutional layer to obtain the main peak position offset encoding vector, energy change encoding vector, and morphological distortion matching encoding vector.

[0139] The temporal feature enhancement vector is obtained by multiplying the main peak position offset encoding vector, energy change encoding vector and morphological distortion matching encoding vector element by element through the temporal feature enhancement unit.

[0140] A feature mapping unit is used to perform deep feature mapping on the temporal feature enhancement vector to extract high-dimensional representation features;

[0141] By inputting the high-dimensional representation features into the BP neural network, the slope displacement results are obtained.

[0142] The expression for the temporal feature enhancement unit is: ,in, For time-series feature enhancement vectors, This is the encoding vector for the offset of the main peak position. Encode energy change vectors, This is the morphological distortion matching encoding vector. This is element-wise multiplication.

[0143] This invention identifies slope displacement using a multi-source feature fusion BP neural network. First, three types of one-dimensional temporal features are encoded separately through one-dimensional temporal convolutional layers, extracting local temporal features while maintaining the temporal length and dimension. Then, a temporal feature enhancement unit performs element-wise multiplication and fusion of the three encoded vectors, ensuring that the responses of the true displacement signal to the three types of features are enhanced in the same direction, while interference signals are effectively suppressed due to feature inconsistencies. Finally, a feature mapping unit and the BP neural network complete high-dimensional feature learning and displacement prediction. This invention effectively integrates three complementary features—peak position shift, energy change, and morphological distortion—and uses element-wise multiplication to adaptively enhance displacement confidence. It boasts advantages such as strong anti-interference capability, high displacement identification accuracy, and good robustness, effectively improving the accuracy and reliability of slope micro-displacement monitoring in complex environments.

[0144] like Figure 3 As shown, the feature mapping unit includes the following layers connected in sequence: a fourth 1D temporal convolutional layer, a first max pooling layer, a fifth 1D temporal convolutional layer, a second max pooling layer, a global average pooling layer, and a fully connected layer.

[0145] The parameters of the fourth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 16, activation function of ReLU, output feature size of 16×M, where M is the feature length;

[0146] The parameters of the first max pooling layer include: kernel size of 2, stride of 2, and output feature size of 16×M / 2.

[0147] The parameters of the fifth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 32, activation function of ReLU, and output feature size of 32×M / 2.

[0148] The parameters of the second max pooling layer are: kernel size of 2, stride of 2, and output feature size of 32×M / 4.

[0149] The global average pooling layer averages the values ​​for each channel, resulting in a channel-level feature output size of 32.

[0150] The output dimension of the fully connected layer is 64, and the activation function is ReLU.

[0151] The global average pooling layer performs a global average operation on the 32 channels output by the second max pooling layer, and outputs a scalar feature for each channel, finally obtaining a channel-level feature vector with a dimension of 32.

[0152] The fully connected layer receives 32-dimensional channel-level feature vectors and improves the feature dimension to 64-dimensional through mapping, resulting in 64-dimensional high-dimensional deep features.

[0153] In this embodiment, the BP neural network includes an input layer, a first hidden layer, a second hidden layer, and an output layer; the input layer has 64 neurons; the first hidden layer has 128 neurons, and the activation function is ReLU; the second hidden layer has 64 neurons, and the activation function is ReLU; the output layer has 4 neurons, and each neuron outputs the "probability of belonging to the category". Finally, the category with the highest probability is taken as the final result.

[0154] The feature mapping unit, by sequentially setting a fourth 1D temporal convolutional layer, a first max-pooling layer, a fifth 1D temporal convolutional layer, a second max-pooling layer, a global average pooling layer, and a fully connected layer, can perform multi-scale, deep feature extraction and dimensionality mapping on the temporal feature enhancement vector while preserving effective temporal feature information. Two layers of 1D temporal convolutions progressively increase the number of channels to enhance local temporal features. The max-pooling layer then performs feature downsampling and key information filtering. A global average pooling layer yields compact and stable channel-level features, and finally, a fully connected layer maps the features into a high-dimensional representation vector.

[0155] The training process of the multi-source feature fusion BP neural network includes: First, collecting a large amount of slope monitoring sample data, extracting the main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector corresponding to each sample, and labeling the corresponding real slope displacement results as labels; Second, inputting the three types of feature vectors into the multi-source feature fusion BP neural network; Then, with the goal of minimizing the error between the predicted slope displacement results and the real labels, using the Adam optimizer to minimize the mean squared error loss function, setting appropriate learning rate, batch size, and training epochs during training, introducing the Dropout regularization mechanism to prevent model overfitting, and using an early stopping strategy (stopping training when the validation set loss does not decrease for several consecutive epochs) to ensure the model's generalization ability; Finally, after multiple rounds of iterative training until the model loss converges to a preset threshold, a trained multi-source feature fusion BP neural network that can be used for slope displacement prediction is obtained.

[0156] The slope displacement results include: no displacement, slight displacement, moderate displacement, and significant displacement.

[0157] Example 2: A computer device includes a processor and a memory; the memory stores computer program instructions, which, when executed by the processor, implement the slope displacement monitoring method based on quantum single photons as described in Example 1.

[0158] Example 3: A computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the slope displacement monitoring method based on quantum single photons as described in Example 1.

[0159] This invention employs quantum single-photon detection technology, which utilizes the high sensitivity of single-photon detectors to accurately capture weak echo light signals. Simultaneously, through wavelet denoising and reconstruction processing, it can selectively filter out irrelevant information such as environmental noise and transmission interference, significantly reducing the impact of interference signals on monitoring data and ensuring the authenticity and stability of the photon counting time series after denoising.

[0160] Based on noise reduction processing, this invention constructs three major feature vectors: main peak position shift, energy change, and morphological distortion matching. These vectors comprehensively capture subtle changes in the optical signal during slope displacement. The main peak position shift is directly related to the magnitude of the displacement, the energy change reflects the signal attenuation characteristics during the displacement process, and the morphological distortion matching reflects the signal waveform variation caused by the displacement. The multi-dimensional features synergistically cover the core information of displacement monitoring, effectively solving the problem that traditional technologies are unable to capture minute displacements, significantly improving the accuracy of slope displacement monitoring, and accurately identifying potential minor displacement hazards.

[0161] This invention employs a multi-source feature fusion BP neural network to process the three major feature vectors. The BP neural network has powerful nonlinear fitting and feature fusion capabilities, which can effectively integrate multi-dimensional feature information, reduce the limitations and errors of single feature monitoring, and at the same time, the quantum single-photon detection technology itself has strong anti-electromagnetic interference and anti-ambient light interference capabilities. Compared with traditional distributed fiber optic sensing technology, it can better adapt to the complex field environment of slopes (such as strong light, electromagnetic interference, temperature fluctuations, etc.), reduce the impact of environmental factors on monitoring results, improve the stability and reliability of monitoring, and ensure the accuracy of long-term continuous monitoring.

[0162] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A slope displacement monitoring method based on quantum single photons, characterized in that, Includes the following steps: S1. Periodically emit rectangular light pulses; S2. Echo detection is performed using a single-photon detector, and a photon counting time series for each cycle is constructed; S3. Perform wavelet denoising and reconstruction on the photon counting time series of each period to obtain the denoised wavelet coefficients and denoised photon counting time series of each period. S4. Based on the denoised wavelet coefficients and denoised photon count time series of each cycle, construct the main peak position shift feature vector, energy change feature vector, and morphological distortion matching feature vector; S5. A multi-source feature fusion BP neural network is used to process the main peak position offset feature vector, energy change feature vector and morphological distortion matching feature vector to obtain the slope displacement result; The multi-source feature fusion BP neural network in S5 includes: a first 1D temporal convolutional layer, a second 1D temporal convolutional layer, a third 1D temporal convolutional layer, a temporal feature enhancement unit, a feature mapping unit, and a BP neural network; The input of the first 1D temporal convolutional layer is used to input the main peak position offset feature vector; The input of the second 1D temporal convolutional layer is used to input the energy change feature vector; The input of the third 1D temporal convolutional layer is used to input the morphological distortion matching feature vector; The input of the temporal feature enhancement unit is connected to the output of the first 1D temporal convolutional layer, the output of the second 1D temporal convolutional layer, and the output of the third 1D temporal convolutional layer, respectively, and its output is connected to the input of the feature mapping unit. The input of the BP neural network is connected to the output of the feature mapping unit, and its output serves as the output of the multi-source feature fusion BP neural network. The processing steps of a multi-source feature fusion BP neural network include: The main peak position offset feature vector, energy change feature vector, and morphological distortion matching feature vector are all processed through a 1D temporal convolutional layer to obtain the main peak position offset encoding vector, energy change encoding vector, and morphological distortion matching encoding vector. The temporal feature enhancement vector is obtained by multiplying the main peak position offset encoding vector, energy change encoding vector and morphological distortion matching encoding vector element by element through the temporal feature enhancement unit. A feature mapping unit is used to perform deep feature mapping on the temporal feature enhancement vector to extract high-dimensional representation features; By inputting the high-dimensional representation features into the BP neural network, the slope displacement results are obtained.

2. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The specific process of S2 includes: The echo is detected by a single-photon detector, and the time number corresponding to each detected echo photon within the period is recorded to obtain the echo photon time. For each cycle, iterate through all detected echo photon moments; Use an indicator function to determine whether each time point is equal to the echo photon time point. If it is, the indicator function value is 1; otherwise, it is 0. Summing all indicator function values ​​within a period yields the number of photons at the corresponding moment within that period. Arrange the photon counts at each moment within each cycle in chronological order to obtain the photon count time series.

3. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The specific process of S3 includes: Continuous wavelet transform is performed on the photon counting time series for each period to obtain wavelet coefficients; Wavelet threshold denoising is performed on the wavelet coefficients to obtain the denoised wavelet coefficients; Wavelet reconstruction is performed on the denoised wavelet coefficients to obtain the denoised photon counting time series.

4. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The process of constructing the main peak position offset feature vector in S4 includes: Find the moment corresponding to the maximum value in the denoised photon count time series of each cycle to obtain the peak moment; The offset of the main peak position is obtained by subtracting the reference main peak position from the peak time of each cycle and taking the absolute value. The ratio of the main peak position offset to the reference main peak position is used as the main peak position offset feature. Arrange the peak position shift features of each cycle in chronological order to obtain the peak position shift feature vector.

5. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The process of constructing the energy change feature vector in S4 includes: Square all the denoised wavelet coefficients within a period and sum them up to obtain the wavelet coefficient energy for the corresponding period. Subtract the wavelet coefficient energy of the current period from that of the previous period to obtain the energy difference of the current period; Take the absolute value of the energy difference for the current cycle to obtain the absolute energy difference for the current cycle; If the energy difference in the current cycle is less than 0 and the offset of the main peak position in the current cycle is greater than the offset threshold, the ratio of the absolute energy difference in the current cycle to the baseline energy is used as the energy change feature. Otherwise, the energy change feature is set to a preset value. Arrange the energy change characteristics of each cycle in chronological order to obtain the energy change characteristic vector.

6. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The process of constructing morphological distortion matching feature vectors in S4 includes: The difference in half-width at half-maximum (WHM) of the current period is obtained by subtracting the full width at half-maximum (FWHM) of the main peak from that of the previous period. The ratio of the absolute value of the half-width difference of the current period to the reference half-width is used as the morphological distortion matching feature of the current period. The morphological distortion matching features of each period are arranged in chronological order to obtain the morphological distortion matching feature vector.

7. The slope displacement monitoring method based on quantum single photons according to claim 1, characterized in that, The feature mapping unit comprises, in sequence: a fourth 1D temporal convolutional layer, a first max pooling layer, a fifth 1D temporal convolutional layer, a second max pooling layer, a global average pooling layer, and a fully connected layer. The parameters of the fourth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 16, activation function of ReLU, output feature size of 16×M, where M is the feature length; The parameters of the first max pooling layer include: kernel size of 2, stride of 2, and output feature size of 16×M / 2. The parameters of the fifth 1D temporal convolutional layer include: kernel size of 1×3, stride of 1, padding of 1, output channels of 32, activation function of ReLU, and output feature size of 32×M / 2. The parameters of the second max pooling layer are: kernel size of 2, stride of 2, and output feature size of 32×M / 4. The global average pooling layer averages the values ​​for each channel, resulting in a channel-level feature output size of 32. The output dimension of the fully connected layer is 64, and the activation function is ReLU.

8. A computer device, characterized in that, include: Processor and memory; The memory stores computer program instructions, which, when executed by the processor, implement the slope displacement monitoring method based on quantum single photons as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the slope displacement monitoring method based on quantum single photons as described in any one of claims 1 to 7.

Images