Substation personnel safety distance monitoring method and device

Abstract

The invention provides a transformer substation personnel safety distance monitoring method and device, and belongs to the technical field of transformer substations. A millimeter wave radar module transmits a linear frequency modulation continuous wave signal and receives a target echo signal, and a four-microphone array synchronously collects an environment sound field signal; carrying out compressed sensing reconstruction on echo signals to extract distance and speed information, carrying out non-negative matrix factorization on sound field signals to extract sound source azimuth angles and footstep sound features, and inputting radar features and acoustic features into a multipath interference suppression model; coarse-grained and fine-grained progressive classification is realized through a layered soft maximum classifier to output real position coordinates of a target, a safe distance is calculated according to the target position and the boundary of an equipment dangerous area, and when the distance is smaller than a preset threshold value, an LED display lamp is triggered to flicker and a loudspeaker is triggered to give a voice alarm. The technical problem of personnel position misjudgment caused by radar multipath interference in a transformer substation environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of substation technology, and more specifically, relates to a method and device for monitoring personnel safety distance in substations. Background Technology

[0002] Substation personnel safety distance monitoring typically employs millimeter-wave radar technology for non-contact target detection. This involves transmitting linear frequency modulated continuous wave signals and receiving target echo signals to extract distance and velocity information. An alarm is triggered when the distance between personnel and high-voltage equipment falls below a safety threshold. However, traditional millimeter-wave radar monitoring methods face severe multipath interference problems in the complex electromagnetic environment of substations. Radar signals are reflected by metal equipment cabinets, pipes, and the ground, generating false echoes that cause the system to misinterpret multipath reflections as real targets or fail to accurately locate personnel. Existing technologies use Doppler filtering or constant false alarm rate (CFAR) detection to suppress multipath interference, but these methods rely solely on single radar signal characteristics. They fail when Doppler characteristics are not obvious in scenarios involving stationary personnel or low-speed movement, and they cannot distinguish the physical sources of different types of multipath reflections. In other words, existing technologies suffer from the technical problem of radar multipath interference leading to misjudgment of personnel positions in substation environments. Summary of the Invention

[0003] In view of this, the present invention provides a method and device for monitoring the safe distance of personnel in substations, which can solve the technical problem of misjudgment of personnel position caused by radar multipath interference in the substation environment in the prior art.

[0004] This invention is implemented as follows: The first aspect of this invention provides a method for monitoring personnel safety distance in substations: It includes a millimeter-wave radar module transmitting a linear frequency modulated continuous wave signal and receiving a target echo signal; a four-microphone array synchronously acquiring ambient sound field signals; a positioning module acquiring the spatial coordinates of the device; undersampling reconstruction processing of the target echo signal based on compressed sensing to establish a sparse representation in the wavelet domain; solving the L1 norm minimization problem using an orthogonal matching pursuit algorithm to recover the complete echo signal; extracting target distance information and target radial velocity information; and performing non-negative matrix decomposition processing on the ambient sound field signal to decompose the time-frequency matrix into a sound source basis matrix and an activation coefficient matrix to extract footstep feature vectors. The speech feature vector is used to calculate the azimuth of the sound source. The target distance information, target radial velocity information, echo amplitude, sound source azimuth, and footstep sound feature vector are input into the multipath interference suppression model to output the true position coordinates of the target and the false multipath echo marker set. The shortest distance value of the target equipment is calculated based on the true position coordinates of the target and the boundary parameters of the dangerous area of ​​the equipment. When the shortest distance value of the target equipment is less than the preset safe distance threshold, the function execution module triggers the high-brightness LED indicator to flash and issues a voice warning through a 5W explosion-proof speaker. The communication module encapsulates the true position coordinates of the target, the shortest distance value of the target equipment, the sound source azimuth, and the timestamp into an alarm data packet and transmits it to the mobile terminal or PC terminal.

[0005] Specifically, the undersampling reconstruction process based on compressed sensing involves deskewing the received target echo signal to obtain an intermediate frequency (IF) signal, randomly undersampling the IF signal at a rate lower than the Nyquist sampling rate to obtain a measurement vector, performing wavelet transform on the IF signal to obtain a wavelet coefficient matrix, and constructing the product of the Gaussian random observation matrix and the wavelet transform matrix as the sensing matrix.

[0006] The orthogonal matching pursuit algorithm is solved iteratively. The residual vector is initialized as the measurement vector, the support set is initialized as an empty set, the inner product of each column of the perception matrix and the residual vector is calculated, the column index with the largest absolute value of the inner product is selected and added to the support set, and the columns corresponding to the support set in the perception matrix are extracted to form a submatrix for least squares fitting.

[0007] Before the undersampling reconstruction process based on compressed sensing, the process includes verifying that the equidistant constant of the sensing matrix is ​​less than 0.3, randomly selecting 500 different column combinations from the sensing matrix, calculating the singular values ​​of the submatrix formed by each column combination, and calculating the ratio of the maximum singular value to the minimum singular value of the submatrix as the condition number.

[0008] The sparsity condition of the wavelet coefficient matrix means that the proportion of coefficients greater than the energy threshold in the wavelet coefficient matrix is ​​less than 12%. The optimization objective function is to minimize the L1 norm of the wavelet coefficient vector, and the constraint is that the product of the sensing matrix and the wavelet coefficient vector minus the L2 norm of the measurement vector is less than the noise energy estimate.

[0009] Specifically, the non-negative matrix decomposition process involves performing short-time Fourier transforms on the signals from the four microphone channels to obtain time-frequency matrices, concatenating the amplitudes of the time-frequency matrices of the four channels into a joint time-frequency matrix, initializing the source basis matrix and activation coefficient matrix as random positive matrices, and establishing a non-negative matrix decomposition objective function using a multiplicative update rule for iterative optimization.

[0010] The objective function of the nonnegative matrix decomposition is expressed as minimizing the squared Frobenius norm of the product of the joint time-frequency matrix, the source basis matrix, and the activation coefficient matrix, plus an L1 norm sparsity penalty term and a time smoothing regularization term, wherein the coefficient of the L1 norm sparsity penalty term is 0.01 and the coefficient of the time smoothing regularization term is 0.05.

[0011] The extraction of the footstep sound feature vector involves identifying the corresponding column of the footstep sound from the decomposed sound source basis matrix, extracting the row vector of the corresponding activation coefficient matrix as the time-domain activation sequence, and extracting the peak interval, zero-crossing rate and energy envelope of the footstep sound time-domain activation sequence as the footstep sound feature vector.

[0012] A second aspect of this invention provides a substation personnel safety distance monitoring system, comprising a millimeter-wave radar module, a 4-microphone array, an external input module, a RISC-V main control chip, a storage module, a positioning module, a high-brightness LED indicator, a 5W explosion-proof speaker, a communication module, a function execution module, an external output module, a rechargeable lithium battery pack, a power supply module, and a battery management system. The RISC-V main control chip is electrically connected to the millimeter-wave radar module, the 4-microphone array, the external input module, the storage module, the positioning module, the high-brightness LED indicator, the 5W explosion-proof speaker, the communication module, the function execution module, the external output module, the rechargeable lithium battery pack, the power supply module, and the battery management system. The millimeter-wave radar module is used to transmit linear frequency modulated continuous waves. The system receives target echo signals and receives target echo signals. The four-microphone array is used to collect ambient sound field signals and locate the sound source. The external input module is used to receive target echo signals and ambient sound field signals and perform analog-to-digital conversion and signal preprocessing. The storage module is used to store instruction codes, model parameters, running data, and alarm records. The positioning module is used to obtain the spatial coordinates of the device and provide location information. The high-brightness LED display is used to output light warnings with different colors and flashing modes. The 5W explosion-proof speaker is used to play voice warnings. The communication module is used to transmit alarm data packets and running status to a mobile terminal or PC terminal. The function execution module is used to drive the high-brightness LED display and the 5W explosion-proof speaker to execute alarm outputs and trigger the communication module to transmit data. The power supply module is used to provide stable voltage for each hardware module.

[0013] The storage module stores instruction code, and the RISC-V main control chip executes the instruction code to perform the above-described method.

[0014] This invention performs multimodal fusion of millimeter-wave radar echo signals and microphone array sound field signals. It extracts target distance and velocity information by reconstructing radar signals through compressed sensing, and extracts sound source azimuth and footstep features through non-negative matrix factorization. A hierarchical soft-maximum classifier is constructed to progressively classify real targets and different types of false targets using coarse and fine granular methods. The complementarity of acoustic and radar features is utilized to eliminate the ambiguity of multipath interference. The multipath interference suppression model of this invention quickly distinguishes real and false targets using a coarse-grained classifier, and then accurately identifies the specific sources of metal reflections and ground reflections using a fine-grained classifier. The hierarchical structure avoids decision-making confusion when a single-layer classifier faces multiple categories. An adaptive temperature parameter adjustment mechanism dynamically adjusts the classification confidence based on signal quality, ensuring robustness under different signal conditions. In summary, this invention solves the technical problem mentioned in the background art of misjudging personnel positions due to radar multipath interference in substation environments. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a schematic diagram of the components of the substation personnel safety distance monitoring device involved in the present invention.

[0017] Figure 3 The distance spectrum diagram after reconstructing the target echo signal.

[0018] Figure 4 This is a schematic diagram of the hierarchical classification structure of the multipath interference suppression model.

[0019] Figure 5 A flowchart for monitoring data transmission and display.

[0020] Figure 6 This is an actual diagram of the substation personnel safety distance monitoring device involved in the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0022] like Figure 1 The diagram shown is a flowchart of a substation personnel safety distance monitoring method provided by the present invention. This method includes the following steps:

[0023] S1. The millimeter-wave radar module transmits linear frequency modulated continuous wave signals and receives target echo signals. The 4-microphone array synchronously collects ambient sound field signals. The positioning module obtains the spatial coordinates of the device. The storage module reads the preset safe distance threshold and the boundary parameters of the equipment's hazardous area.

[0024] S2. Perform undersampling reconstruction processing on the target echo signal based on compressed sensing, establish a sparse representation in the wavelet domain, and recover the complete echo signal by solving the L1 norm minimization problem through the orthogonal matching pursuit algorithm, and extract the target distance information and the target radial velocity information.

[0025] S3. Perform non-negative matrix decomposition on the environmental sound field signal, decompose the time-frequency matrix into the sound source basis matrix and activation coefficient matrix, extract the footstep sound feature vector and speech feature vector, and calculate the sound source azimuth angle.

[0026] S4. Input the target distance information, target radial velocity information, echo amplitude, sound source azimuth angle and footstep sound feature vector into the multipath interference suppression model. The multipath interference suppression model outputs the target's true position coordinates and the false multipath echo marker set.

[0027] S5. Calculate the shortest distance value of the target equipment based on the target's real location coordinates and the boundary parameters of the equipment's hazardous area. When the shortest distance value of the target equipment is less than the preset safe distance threshold, the function execution module triggers the high-brightness LED indicator to flash and issues a voice warning through a 5W explosion-proof speaker.

[0028] S6. Optionally, the communication module encapsulates the target's true location coordinates, the target device's shortest distance value, the sound source's azimuth angle, and the timestamp into an alarm data packet, which is then transmitted to the mobile terminal or PC via a communication protocol.

[0029] The specific steps of the undersampling reconstruction processing based on compressed sensing include: deskewing the received target echo signal to obtain an intermediate frequency (IF) signal; randomly undersampling the IF signal at a rate lower than the Nyquist sampling rate to obtain a measurement vector; performing wavelet transform on the IF signal to obtain a wavelet coefficient matrix, wherein the proportion of coefficients greater than the energy threshold in the wavelet coefficient matrix is ​​less than 12%, satisfying the sparsity condition; constructing the product of the Gaussian random observation matrix and the wavelet transform matrix as the sensing matrix, verifying that the limiting equidistant constant of the sensing matrix is ​​less than 0.3; and establishing an optimization objective function, which is expressed as follows: minimizing the small The L1 norm of the wavelet coefficient vector is determined by the constraint that the product of the sensing matrix and the wavelet coefficient vector minus the L2 norm of the measurement vector is less than the noise energy estimate. An orthogonal matching pursuit algorithm is used iteratively to solve the problem. In each iteration, the wavelet basis function with the highest correlation to the residual is selected and added to the support set. The residual is updated until the residual energy is below a set threshold or the maximum number of iterations is reached. The obtained sparse wavelet coefficients are subjected to inverse wavelet transform to reconstruct the complete intermediate frequency signal. A fast Fourier transform is performed on the reconstructed complete intermediate frequency signal to obtain the distance spectrum. Target distance information is extracted through peak detection, and target radial velocity information is extracted using a moving target detection algorithm.

[0030] The specific steps of the non-negative matrix decomposition process include: performing short-time Fourier transforms on the signals from the four microphone channels to obtain time-frequency matrices, where the number of rows in the time-frequency matrix represents the number of frequency points and the number of columns represents the number of time frames; concatenating the amplitudes of the time-frequency matrices from the four channels into a joint time-frequency matrix, the dimension of which is the number of frequency points multiplied by four times the number of time frames; initializing the source basis matrix and activation coefficient matrix as random positive value matrices, with the number of columns in the source basis matrix set to the preset number of source types; establishing a non-negative matrix decomposition objective function, which is expressed as follows: minimizing the squared Frobenius norm between the joint time-frequency matrix and the product of the source basis matrix and activation coefficient matrix, plus an L1 norm sparsity penalty term and a time smoothing regularization term, where the coefficient of the L1 norm sparsity penalty term is 0.01 and the coefficient of the time smoothing regularization term is 0.05; and iteratively optimizing using a multiplicative update rule, with the source basis matrix update formula being the current... The product of the source basis matrix multiplied by the joint time-frequency matrix and the transpose of the activation coefficient matrix is ​​divided by the sum of the products of the source basis matrix, the activation coefficient matrix, and the transpose of the activation coefficient matrix. The activation coefficient matrix update formula is the current activation coefficient matrix multiplied by the product of the transpose of the source basis matrix and the joint time-frequency matrix, divided by the sum of the products of the transpose of the source basis matrix, the source basis matrix, and the activation coefficient matrix. The number of iterations is set to 200. From the decomposed source basis matrix, the columns corresponding to footsteps and speech are identified, and the row vectors of the corresponding activation coefficient matrices are extracted as time-domain activation sequences. The peak interval, zero-crossing rate, and energy envelope are extracted as footstep feature vectors from the footstep time-domain activation sequence, and the first 13 dimensions of the Mel-frequency cepstral coefficients are extracted as speech feature vectors from the speech time-domain activation sequence. The arrival time difference of the signals from the four microphone channels is calculated using the generalized cross-correlation time delay estimation algorithm, and the azimuth of the source is calculated using the triangulation formula with an accuracy of 5 degrees.

[0031] The multipath interference suppression model is structured as follows: The input layer receives the spliced ​​model input feature vector, which includes target distance information, target radial velocity information, echo amplitude, sound source azimuth angle, and the first 5 dimensions of footsteps. The total dimension of the model input feature vector is 9. The first convolutional layer uses 32 convolutional kernels to perform one-dimensional convolution on the model input feature vector to extract local patterns. The first-level classifier performs global average pooling on the output of the first convolutional layer and connects it to a hierarchical soft-maximum classification head. The hierarchical soft-maximum classification head includes a coarse-grained classifier and a fine-grained classifier. The coarse-grained classifier classifies the target into three coarse categories: real target, metal reflection false target, and ground reflection false target, outputting a 3-dimensional coarse category probability vector. The second convolutional layer uses 64 convolutional kernels to further extract features from the output of the first convolutional layer. The second-level classifier further extracts features from the output of the second convolutional layer. The output of the convolutional layer is globally averaged and then connected to a fine-grained classifier. The fine-grained classifier further subdivides the coarse-class probability vector, classifying real targets into stationary and moving people, metal reflection false targets into equipment cabinet reflections and pipe reflections, and ground reflection false targets into near-range and far-range ground reflections, outputting a 6-dimensional fine-class probability vector. The third convolutional layer uses 128 convolutional kernels to extract high-level semantic features. The output layer performs global average pooling on the output of the third convolutional layer and then connects it to a fully connected layer. The fully connected layer outputs the two-dimensional corrected offset and confidence score of the target's true position coordinates. The loss function consists of three weighted parts: the first part is the coarse-grained classification cross-entropy loss, the second part is the fine-grained classification cross-entropy loss, and the third part is the smoothed L1 loss of coordinate regression, with a weight ratio of 0.3:0.3:0.4.

[0032] The steps for establishing the training dataset for the multipath interference suppression model specifically include: deploying millimeter-wave radar modules and a 4-microphone array in the substation environment to record the spatial coordinates of the devices and the boundary parameters of the hazardous areas of the equipment; arranging personnel to move within the monitoring area according to a preset trajectory, which covers different distances, angles, and motion states; synchronously acquiring target echo signals and environmental sound field signals for 80 hours; using a high-precision laser rangefinder and a visual positioning system to record the personnel's real position coordinates as labels; performing undersampling reconstruction processing based on compressed sensing on the acquired target echo signals to extract target distance information, target radial velocity information, and echo amplitude; performing non-negative matrix factorization processing on the environmental sound field signals to extract the sound source azimuth angle and footstep feature vectors; manually labeling the target category of each sample, including stationary and moving personnel in real targets, equipment cabinet reflections and pipe reflections in metal reflection false targets, and near-range and far-range ground reflections in ground reflection false targets; organizing the labeled model input feature vectors and personnel real position coordinate labels into training samples, with 120,000 training samples; and dividing the training set and validation set into an 8:2 ratio.

[0033] The training steps of the multipath interference suppression model specifically include: initializing the convolutional layer weights using the Kaiming initialization method and the fully connected layer weights using the Xavier initialization method; setting the batch size to 64, the initial learning rate to 0.001, and using the Adam optimizer; training the coarse-grained classifier in the first stage, freezing the fine-grained classifier and output layer parameters, and updating only the parameters of the first convolutional layer and the coarse-grained classifier, with 50 training epochs; training the fine-grained classifier in the second stage, freezing the parameters of the coarse-grained classifier, unfreezing the fine-grained classifier, and simultaneously updating the parameters of the first convolutional layer. The first stage involves training the parameters of the second convolutional layer and the fine-grained classifier for 80 epochs. The third stage involves joint training, where all parameters are unfrozen, and a hierarchical soft-maximum loss function is used to jointly optimize the coarse-grained classification, fine-grained classification, and coordinate regression tasks for 120 epochs. A cosine annealing strategy is used to dynamically adjust the learning rate, which decays to 10% of its initial value at the end of each training stage. Every 10 epochs, the model performance is evaluated on the validation set, and the accuracy of coarse-grained classification, fine-grained classification, and mean absolute error of coordinate regression are recorded. The model weights with the best overall performance on the validation set are selected and saved as the final model.

[0034] The implementation of the multi-scale classification optimization mechanism based on hierarchical soft maximum in the multipath interference suppression model is as follows: In the first-level classifier, the coarse-grained classifier performs global average pooling on the shallow features extracted by the first convolutional layer to obtain the first feature vector. The first feature vector is mapped to a 3D first logistic regression value through the first fully connected layer. The first logistic regression value is then normalized by soft maximum to obtain a coarse class probability vector. The soft maximum function calculates the exponent value of each class divided by the sum of the exponent values ​​of all classes. The coarse class probability vector and the true coarse class label are used to calculate the coarse-grained classification intersection. The cross-entropy loss for coarse-grained classification is the negative logarithm of the probability corresponding to the true class. In the second-level classifier, the fine-grained classifier receives deep features from the output of the second convolutional layer. These deep features contain richer discriminative information. The fine-grained classifier first reads the coarse class probability vector output by the coarse-grained classifier and weights the fine-grained classification task according to this vector by multiplying the fine-grained class probability by the corresponding coarse class probability. The fine-grained classifier then performs global average pooling on the deep features to obtain the second feature vector, which is then passed through the second fully connected layer. The layer is mapped to a 6-dimensional second logistic regression value. Soft-maximum normalization is applied to the second logistic regression value to obtain a fine-grained category probability vector. The fine-grained category probability vector is then used to calculate the fine-grained classification cross-entropy loss using the weighted true fine-grained category labels. During training, the coarse-grained and fine-grained classification cross-entropy losses are weighted and summed at a ratio of 0.3 to 0.3, guiding the model to first learn a coarse target type discrimination ability and then gradually learn a refined sub-category discrimination ability. During inference, the model first outputs a coarse category probability vector. When the maximum value of the coarse category probability vector exceeds... A coarse classification result of 0.7 is considered reliable, and the model proceeds to the fine-grained classification stage. The fine-grained classification stage outputs a fine-class probability vector, and calculates the joint confidence score by combining the coarse and fine-class probability vectors. The joint confidence score is the product of the corresponding elements in the coarse and fine-class probability vectors. When the joint confidence score exceeds 0.6, the classification result is considered reliable, and the final class label and two-dimensional correction offset are output. When the maximum value of the coarse-class probability vector is below 0.7 or the joint confidence score is below 0.6, the model outputs an uncertainty marker, triggering a manual review process.

[0035] The multi-scale classification optimization mechanism based on hierarchical soft maximum values ​​achieves progressive classification and uncertainty quantification from coarse to fine granularity through a hierarchical classifier structure. This brings significant technical benefits to the multipath interference suppression model and the overall monitoring scheme, manifested in the following aspects: Regarding improved classification accuracy, the coarse-grained classifier utilizes shallow features to quickly distinguish real targets from different types of false targets, avoiding confusion of multipath reflections of different properties in the early stages. This provides clear category boundaries for subsequent fine-grained classification. The fine-grained classifier further subdivides subcategories based on coarse classification, using deep features to capture subtle differences, achieving accurate identification of false targets reflecting metal and those reflecting ground. The hierarchical structure allows the model to gradually focus on the difficult points, avoiding the decision confusion that occurs when a single-layer classifier faces multiple categories. In terms of robustness enhancement, the probability distribution output by the soft-maximum function provides a flexible decision boundary rather than a hard class division, enabling the model to quantify the uncertainty of classification results. Uncertainty labeling is automatically triggered when the maximum value of the coarse class probability vector or the joint confidence score falls below a threshold, avoiding false alarms or missed alarms caused by misclassification. This mechanism is particularly important in the electromagnetic environment of substations, capable of identifying anomalies such as severely distorted signals or abnormally strong multipath interference. Regarding computational efficiency optimization, the hierarchical structure allows the model to use shallower network layers and fewer computational resources in the coarse classification stage, activating deeper networks only for samples requiring further discrimination. This avoids redundant computation of all samples through the complete network. The early decision-making capability of the coarse-grained classifier reduces the burden of subsequent processing, allowing the model to maintain high accuracy while reducing average inference time, meeting the response speed requirements of real-time monitoring. Regarding improvements in training stability, a phased training strategy first fixes some parameters to train a coarse-grained classifier, then gradually unfreezes the parameters to train a fine-grained classifier. This avoids gradient conflicts and convergence difficulties in the early stages of multi-task joint training. The weighting mechanism of the hierarchical loss function balances the learning progress of different granularity classification tasks, preventing one task from dominating the training process and causing performance degradation in other tasks. This allows the model to learn robust multi-scale feature representations with limited training data. In terms of improved generalization ability, the high-level semantic features learned by the coarse-grained classifier have strong scene adaptability, maintaining stable coarse classification capabilities even under unfamiliar equipment layouts or environmental conditions. The fine-grained classifier learns sub-category features under the constraints of coarse classification, reducing the risk of overfitting. This enables the model to adapt to the multipath characteristics caused by differences in metal structures and ground materials in different substation environments.

[0036] The temperature parameters of the hierarchical soft-maximum classifier in the multipath interference suppression model are determined based on the echo amplitude, the angular deviation between the sound source azimuth and the radar measurement angle, and the energy of the footstep sound feature vector. The temperature parameters adjust the smoothness of the soft-maximum function. The calculation formula for the temperature parameters is as follows: The temperature parameter equals the base temperature value multiplied by the product of the amplitude adjustment factor, the angle adjustment factor, and the energy adjustment factor. The base temperature value is 1.0. The amplitude adjustment factor is equal to the ratio of the echo amplitude to the standard echo amplitude plus 0.5, limited to the range of 0.7 to 1.3. The angle adjustment factor is equal to 1 plus the absolute value of the angle deviation divided by 90 degrees, limited to the range of 0.8 to 1.2. The energy adjustment factor is equal to the ratio of the footstep sound feature vector energy to the standard energy value plus 0.3, limited to the range of 0.6 to 1.4. The standard echo amplitude is set to the median amplitude statistically analyzed on the training set. The standard energy value is set to the median energy value statistically analyzed on the training set. When the echo amplitude is strong, the angular deviation is small, and the energy of the footstep feature vector is high, increasing the temperature parameter makes the soft-maximum function output a smoother probability distribution, reducing the model's overconfidence and preventing the model from ignoring potential multipath interference when the signal quality is good. When the echo amplitude is weak, the angular deviation is large, or the energy of the footstep feature vector is low, decreasing the temperature parameter makes the soft-maximum function output a sharper probability distribution, enhancing the model's ability to distinguish the dominant category and preventing the output of ambiguous classification results when the signal quality is poor. The adaptive adjustment mechanism of the temperature parameter enables the model to dynamically adjust its decision-making strategy according to the quality of the current observation data. When the signal reliability is high, it remains cautious to avoid misjudgment, and when the signal reliability is low, it makes decisive decisions to avoid hesitation, thereby improving the robustness and adaptability of the model under different signal conditions.

[0037] The iterative solution steps of the orthogonal matching pursuit algorithm are as follows: Initialize the residual vector as the measurement vector, initialize the support set as an empty set, and initialize the number of iterations to 0; calculate the inner product of each column of the sensing matrix with the residual vector, and select the column index with the largest absolute value of the inner product to add to the support set; extract the columns corresponding to the support set in the sensing matrix to form a submatrix, and perform least squares fitting on the submatrix to obtain the estimated wavelet coefficients corresponding to the support set; calculate the product of the submatrix and the estimated wavelet coefficients, and update the residual vector to the measurement vector minus the product result; calculate the L2 norm of the residual vector, and terminate the iteration when the L2 norm of the residual vector is less than the noise energy estimate or the number of iterations reaches the maximum number of iterations, otherwise increment the number of iterations by 1 and return to the second step; output the estimated wavelet coefficients at the corresponding positions in the support set, and set the wavelet coefficients at other positions to 0 to obtain the complete sparse wavelet coefficient vector.

[0038] The verification steps for the restricted equidistant constant are as follows: 500 different column combinations are randomly selected from the sensing matrix, with each group having a preset sparsity of columns; singular values ​​are calculated for the submatrix formed by each column combination, and the singular values ​​are obtained through singular value decomposition; the ratio of the maximum singular value to the minimum singular value of the submatrix is ​​calculated and recorded as the condition number; the maximum value of all condition numbers is counted, and the restricted equidistant constant is approximately the square of the maximum condition number minus 1, divided by the square of the maximum condition number plus 1; when the restricted equidistant constant is less than 0.3, the sensing matrix is ​​considered to satisfy the restricted equidistant property, ensuring the uniqueness and stability of signal reconstruction.

[0039] The specific steps for calculating the noise energy estimate are as follows: when the millimeter-wave radar module is not transmitting a linear frequency modulated continuous wave signal, the background noise of the receiver is collected, and the background noise of the receiver is undersampled to obtain a noise measurement vector; the L2 norm of the noise measurement vector is calculated as the noise energy estimate; the noise energy estimate is multiplied by a safety margin coefficient of 1.5 as the termination threshold of the orthogonal matched pursuit algorithm, and the safety margin coefficient is used to compensate for noise power fluctuations and environmental interference changes.

[0040] The moving target detection algorithm specifically comprises the following steps: performing a Fast Fourier Transform (FFT) on the reconstructed complete intermediate frequency signal along the slow time dimension to obtain the Doppler spectrum; calculating the energy distribution of the Doppler spectrum and extracting frequency points with energy exceeding three times the background noise energy as the target Doppler frequency; calculating the target radial velocity information based on the relationship between the target Doppler frequency, the radar carrier frequency, and the speed of light, wherein the target radial velocity information is equal to the target Doppler frequency multiplied by the speed of light divided by twice the radar carrier frequency; determining the target as a moving target when the absolute value of the target radial velocity information is greater than 0.2 m / s, and determining the target as a stationary target when the absolute value of the target radial velocity information is less than 0.2 m / s.

[0041] The generalized cross-correlation time delay estimation algorithm specifically includes the following steps: cross-correlation calculation is performed on the signals from the four microphone channels in pairs to obtain six sets of cross-correlation functions; phase transformation weighting is applied to each set of cross-correlation functions, where the phase spectrum of the cross-correlation function in the frequency domain is divided by the amplitude spectrum, and the weighted cross-correlation function has a stronger ability to suppress noise and reverberation; inverse Fourier transform is performed on the weighted cross-correlation function to obtain the generalized cross-correlation function; the time delay corresponding to the peak position of the generalized cross-correlation function is extracted, where the time delay is the time difference between the arrival of the sound wave at different microphones; based on the six sets of time delays and the geometric position of the microphone array, the three-dimensional coordinates of the sound source are solved using the least squares method, and the sound source azimuth is obtained by projecting it onto the horizontal plane.

[0042] The preset safety distance threshold is determined based on the equipment voltage level and national power safety regulations. The preset safety distance threshold is set to 0.7m when the equipment voltage level is 10kV, 1.0m when the equipment voltage level is 35kV, 1.5m when the equipment voltage level is 110kV, 3.0m when the equipment voltage level is 220kV, and 5.0m when the equipment voltage level is 500kV.

[0043] The boundary parameters of the hazardous area of ​​the equipment include the coordinates of the center point of the hazardous area and the radius of the hazardous area. The coordinates of the center point of the hazardous area are the spatial coordinates of the device, and the radius of the hazardous area is a preset safety distance threshold plus half of the physical size of the equipment. The physical size of the equipment is obtained by laser scanning or manual measurement.

[0044] The specific steps for calculating the shortest distance value of the target device are as follows: read the target's true location coordinates and the boundary parameters of the device's hazardous area; calculate the Euclidean distance between the target's true location coordinates and the center point coordinates of the hazardous area; the shortest distance value of the target device is equal to the Euclidean distance minus the radius of the hazardous area; when the shortest distance value of the target device is negative, it indicates that the target has entered the hazardous area, and when the shortest distance value of the target device is positive, it indicates that the target is outside the hazardous area.

[0045] The alarm data packet encapsulation format includes a data packet header, timestamp, device identifier, target real location coordinates, target device minimum distance value, sound source azimuth angle, alarm level, and data packet checksum. The data packet header is a fixed 4-byte identifier, the timestamp uses Unix timestamp format and occupies 8 bytes, the device identifier uses a 16-byte string, the target real location coordinates are represented by two floating-point numbers and occupy 8 bytes, the target device minimum distance value is represented by a floating-point number and occupies 4 bytes, the sound source azimuth angle is represented by a floating-point number and occupies 4 bytes, the alarm level is divided into 3 levels based on the ratio of the target device minimum distance value to a preset safe distance threshold and occupies 1 byte, and the data packet checksum is generated using a cyclic redundancy check algorithm and occupies 4 bytes.

[0046] The alarm level classification rules are as follows: when the shortest distance to the target device is less than 50% of the preset safe distance threshold, the alarm level is set to Level 1, indicating serious danger; when the shortest distance to the target device is between 50% and 80% of the preset safe distance threshold, the alarm level is set to Level 2, indicating moderate danger; when the shortest distance to the target device is between 80% and 100% of the preset safe distance threshold, the alarm level is set to Level 3, indicating slight danger; and when the shortest distance to the target device is greater than the preset safe distance threshold, the alarm level is set to Level 0, indicating a safe state.

[0047] The flashing mode of the high-brightness LED indicator is determined according to the alarm level. When the alarm level is 1, the high-brightness LED indicator flashes continuously at a frequency of 5Hz and displays red. When the alarm level is 2, the high-brightness LED indicator flashes intermittently at a frequency of 2Hz and displays orange. When the alarm level is 3, the high-brightness LED indicator flashes slowly at a frequency of 1Hz and displays yellow. When the alarm level is 0, the high-brightness LED indicator is constantly on and displays green.

[0048] The voice warning content is generated based on the alarm level and the shortest distance to the target device. When the alarm level is 1, the voice warning content is "Severe danger, please evacuate the danger area immediately." When the alarm level is 2, the voice warning content is "Moderate danger, please maintain a safe distance." When the alarm level is 3, the voice warning content is "Slight danger, please do not approach the device." The voice warning content is synthesized using a text-to-speech algorithm and stored in the storage module. The function execution module reads the corresponding voice file according to the alarm level and plays it through a 5W explosion-proof speaker.

[0049] The angular deviation is the absolute value of the difference between the azimuth angle of the sound source and the radar measurement angle, which is calculated from the phase difference of the target echo signal.

[0050] The echo amplitude is the amplitude value corresponding to the peak value in the distance spectrum of the intermediate frequency signal obtained after the target echo signal has undergone deskewing processing.

[0051] The energy of the footstep sound feature vector is the sum of the squares of all elements in the footstep sound feature vector.

[0052] The specific steps for calculating the radar measurement angle are as follows: The millimeter-wave radar module contains multiple receiving antennas, the spacing between the receiving antennas is half a wavelength, and there is a phase difference between the target echo signals received by the receiving antennas; the phase difference between the target echo signals received by adjacent receiving antennas is calculated; the radar measurement angle is equal to the result of the arcsine function acting on the phase difference multiplied by the wavelength divided by twice pi and then divided by the spacing between the receiving antennas.

[0053] The second aspect of the present invention provides a substation personnel safety distance monitoring device, including a millimeter-wave radar module, a 4-microphone array, an external input module, a RISC-V main control chip, a storage module, a positioning module, a high-brightness LED indicator, a 5W explosion-proof speaker, a communication module, a function execution module, an external output module, a rechargeable lithium battery pack, a power supply module, and a battery management system.

[0054] The millimeter-wave radar module is used to transmit linear frequency modulated continuous wave signals and receive target echo signals. The millimeter-wave radar module operates at a frequency of 77 GHz, has a bandwidth of 4 GHz, a range resolution of 0.04 m, and a maximum detection range of 50 m. The millimeter-wave radar module contains multiple receiving antennas with a spacing of half a wavelength between them.

[0055] The 4-microphone array adopts a square layout with a microphone spacing of 0.1m, a sampling rate of 48kHz, and a frequency response range of 20Hz to 20kHz. The 4-microphone array is used to collect ambient sound field signals and realize sound source localization.

[0056] The external input module is responsible for receiving the target echo signal and the ambient sound field signal. The external input module performs analog-to-digital conversion and signal preprocessing.

[0057] The RISC-V main control chip adopts a 64-bit architecture with a main frequency of 1.5GHz. The RISC-V main control chip integrates a floating-point arithmetic unit and a vector processing unit. The RISC-V main control chip is responsible for executing instruction code in the memory module, performing signal processing, feature extraction, model inference, and decision control tasks.

[0058] The storage module includes a program storage area and a data storage area. The program storage area has a capacity of 512MB and is used to store instruction code and model parameters. The data storage area has a capacity of 2GB and is used to store runtime data, log information, and alarm records.

[0059] The positioning module uses a dual-mode Beidou and GPS positioning chip with a positioning accuracy of 1m. The positioning module is used to obtain the spatial coordinates of the device and provide location information for alarm data packets.

[0060] The high-brightness LED display uses an RGB three-color LED array with a brightness of 10,000 cd and a viewing distance of 100m. The high-brightness LED display is used to output light warnings in different colors and flashing modes.

[0061] The 5W explosion-proof speaker is encapsulated in an explosion-proof shell, with a sound pressure level of 90dB and an explosion-proof rating of ExdIIBT4. The 5W explosion-proof speaker is used to play voice warnings.

[0062] The communication module supports dual-mode 4G and WiFi communication. The 4G module supports the CAT4 standard with an uplink rate of 50Mbps and a downlink rate of 150Mbps. The WiFi module supports the 802.11ac protocol with a transmission rate of 867Mbps. The communication module is used to transmit alarm data packets and operating status to mobile devices or PCs.

[0063] The function execution module receives control commands from the RISC-V main control chip, drives a high-brightness LED indicator and a 5W explosion-proof speaker to perform alarm output, and triggers the communication module to transmit data.

[0064] The external output module connects to the function execution module and outputs control signals to a high-brightness LED indicator, a 5W explosion-proof speaker, and a communication module.

[0065] The rechargeable lithium battery pack is composed of 18650 lithium-ion batteries connected in series and parallel, with a nominal voltage of 12V, a capacity of 20Ah, and a battery life of 72 hours.

[0066] The power module uses a step-down voltage regulator circuit to provide stable voltage for each hardware module. The power module output voltage includes 5V to power the communication module, 3.3V to power the RISC-V main control chip and storage module, and 12V to power the millimeter-wave radar module and 5W explosion-proof speaker.

[0067] The battery management system monitors the voltage, current, and temperature of the rechargeable lithium battery pack. When the voltage of the rechargeable lithium battery pack is lower than 10.5V, the battery management system triggers a low battery alarm. When the temperature of the rechargeable lithium battery pack exceeds 60℃, the battery management system triggers overheat protection. The battery management system implements overcharge protection, over-discharge protection, and short circuit protection functions.

[0068] The storage module stores instruction code. When the RISC-V main control chip runs the instruction code, it executes the above-mentioned substation personnel safety distance monitoring method to realize multi-source sensing data fusion analysis, multipath interference suppression, target positioning, safety distance judgment and active alarm functions.

[0069] The specific implementation methods of the above steps are described in detail below.

[0070] The specific implementation of step S1 involves transmitting a 77GHz linear frequency modulated continuous wave signal to the monitoring area via a millimeter-wave radar module. This signal generates a reflected echo upon encountering a target. Multiple receiving antennas of the millimeter-wave radar module are arranged at half-wavelength intervals to receive the target echo signal. Simultaneously, a four-microphone array, arranged in a square layout with a spacing of 0.1m, synchronously acquires ambient sound field signals. The sampling rate is set to 48kHz to ensure complete capture of sound signals within the 20Hz to 20kHz frequency range. The positioning module obtains the device's spatial coordinates using a BeiDou and GPS dual-mode positioning chip, achieving a positioning accuracy of 1m. The data is then stored. The module reads pre-configured preset safety distance thresholds and equipment hazardous area boundary parameters from the data storage area. The preset safety distance thresholds are determined according to the equipment voltage level: 0.7m for 10kV equipment, 1.0m for 35kV equipment, 1.5m for 110kV equipment, 3.0m for 220kV equipment, and 5.0m for 500kV equipment. The equipment hazardous area boundary parameters include the coordinates of the center point of the hazardous area and the radius of the hazardous area. The radius of the hazardous area is equal to the sum of the preset safety distance threshold and half of the physical size of the equipment. This step provides complete raw data and reference benchmarks for subsequent signal processing and safety judgment.

[0071] The specific implementation of step S2 is as follows: First, the target echo signal received by the millimeter-wave radar module is de-chewed to obtain the intermediate frequency (IF) signal. De-chewing is achieved by mixing the received signal with a delayed copy of the transmitted signal to achieve frequency down-conversion. Then, the IF signal is randomly undersampled at a rate lower than the Nyquist sampling rate to obtain the measurement vector. Undersampling reduces the amount of data but retains key signal information. Next, wavelet transform is performed on the IF signal to obtain the wavelet coefficient matrix. The wavelet transform utilizes the multi-scale characteristics of the wavelet basis functions to decompose the signal into different frequencies and time resolutions. The proportion of coefficients greater than the energy threshold in the wavelet coefficient matrix is ​​statistically verified to be less than 12%, satisfying the sparsity condition. The product of the Gaussian random observation matrix and the wavelet transform matrix is ​​constructed as the sensing matrix. The constraint constant of the sensing matrix is ​​verified to be less than 0.3 to ensure the uniqueness and stability of signal reconstruction. An optimization objective function is established to minimize the wavelet coefficient vector. The norm is the objective, and the constraint is the product of the sensing matrix and the wavelet coefficient vector minus the measurement vector. If the norm is less than the noise energy estimate, the orthogonal matching pursuit algorithm is used to iteratively solve the optimization problem. In each iteration, the wavelet basis function with the highest correlation to the residual is selected and added to the support set and the residual is updated until the residual energy is lower than the set threshold or the maximum number of iterations is reached. The sparse wavelet coefficients obtained by the solution are subjected to inverse wavelet transform to reconstruct the complete intermediate frequency signal. Finally, the reconstructed complete intermediate frequency signal is subjected to fast Fourier transform to obtain the distance spectrum. The target distance information is extracted by peak detection and the target radial velocity information is extracted by moving target detection algorithm. This step is based on compressed sensing theory to achieve high-precision signal reconstruction under undersampling conditions, which effectively reduces the data acquisition and storage costs while maintaining the target detection performance.

[0072] The specific implementation of step S3 involves performing short-time Fourier transforms on the signals from the four microphone channels to obtain time-frequency matrices. The number of rows in the time-frequency matrix corresponds to the number of frequency points, and the number of columns corresponds to the number of time frames. The amplitudes of the time-frequency matrices from the four channels are concatenated to form a joint time-frequency matrix. The dimension of the joint time-frequency matrix is ​​the number of frequency points multiplied by four times the number of time frames. The source basis matrix and activation coefficient matrix are initialized as random positive value matrices. The number of columns in the source basis matrix is ​​set to the preset number of source types. A non-negative matrix decomposition objective function is established to minimize the squared Frobenius norm between the joint time-frequency matrix and the product of the source basis matrix and activation coefficient matrix. Norm sparsity penalty term and time smoothing regularization term, The coefficient of the norm sparsity penalty term is set to 0.01. The time smoothing regularization coefficient was set to 0.05. The source basis matrix and activation coefficient matrix were iteratively optimized using a multiplicative update rule, with 200 iterations. The columns corresponding to footsteps and speech were identified from the decomposed source basis matrix. The row vectors of the corresponding activation coefficient matrices were extracted as time-domain activation sequences. The peak interval, zero-crossing rate, and energy envelope were extracted as footstep feature vectors from the footstep time-domain activation sequence. The first 13 dimensions of the Mel-frequency cepstral coefficients were extracted as speech feature vectors from the speech time-domain activation sequence. The arrival time difference of the signals from the four microphone channels was calculated using a generalized cross-correlation time delay estimation algorithm. The azimuth of the sound source was calculated using a triangulation formula with an accuracy of 5 degrees. This step used a non-negative matrix factorization algorithm to achieve multi-source separation and feature extraction, providing acoustic auxiliary information for subsequent multipath interference suppression.

[0073] The specific implementation of step S4 involves concatenating the target distance information, target radial velocity information, echo amplitude, sound source azimuth angle, and the first 5 dimensions of the footstep sound feature vector into a model input feature vector with a total dimension of 9. This input is then used to perform inference on the multipath interference suppression model. The model first uses a first convolutional layer with 32 convolutional kernels to perform one-dimensional convolution on the input feature vector to extract local patterns. The first-level classifier performs global average pooling on the output of the first convolutional layer and then connects it to a hierarchical soft-maximum classifier. A coarse-grained classifier classifies the target into three coarse categories: real targets, metal-reflecting false targets, and ground-reflecting false targets, outputting a 3-dimensional coarse category probability vector. The second convolutional layer uses 64 convolutional kernels to... The output of the first convolutional layer is further used for feature extraction. The second-level classifier performs global average pooling on the output of the second convolutional layer and then connects it to a fine-grained classifier. The fine-grained classifier further subdivides the coarse class probability vector into 6 subclasses. The third convolutional layer uses 128 convolutional kernels to extract high-level semantic features. The output layer performs global average pooling on the output of the third convolutional layer and then connects it to a fully connected layer. The fully connected layer outputs the two-dimensional corrected offset and confidence score of the target's true position coordinates. The model outputs the target's true position coordinates and a set of false multipath echo markers. This step effectively suppresses multipath interference in the complex electromagnetic environment of the substation by fusing radar and acoustic multimodal information through a deep learning model.

[0074] The specific implementation of step S5 involves calculating the shortest distance between the target and the equipment based on the target's true location coordinates and the boundary parameters of the equipment's hazardous area. First, the Euclidean distance between the target's true location coordinates and the center point coordinates of the hazardous area is calculated. The shortest distance value between the target and the equipment is equal to the Euclidean distance minus the radius of the hazardous area. When the shortest distance value between the target and the equipment is less than a preset safe distance threshold, an alarm is triggered. The function execution module controls the flashing mode and color of the high-brightness LED indicator light according to the alarm level. When the alarm level is 1, it flashes red light continuously at a frequency of 5Hz; when the alarm level is 2, it flashes orange light intermittently at a frequency of 2Hz; and when the alarm level is 3, it flashes yellow light slowly at a frequency of 1Hz. At the same time, the corresponding voice warning content is played through a 5W explosion-proof speaker. This step realizes real-time safe distance judgment and on-site active alarm, providing timely safety protection for substation personnel.

[0075] Step S6 is an optional step. In its specific implementation, the communication module encapsulates the target's true location coordinates, the target device's shortest distance value, the sound source azimuth angle, and the timestamp into an alarm data packet. The data packet format includes a 4-byte data packet header, an 8-byte Unix timestamp, a 16-byte device identifier, an 8-byte target's true location coordinates, a 4-byte target device's shortest distance value, a 4-byte sound source azimuth angle, a 1-byte alarm level, and a 4-byte cyclic redundancy check code. The communication module supports 4G and WiFi dual-mode communication and transmits the alarm data packet to a mobile terminal or PC through the selected communication protocol to realize remote monitoring and data recording functions.

[0076] It should be noted that this invention also solves the following technical problems: Traditional radar signal processing requires high-speed sampling to satisfy the Nyquist sampling theorem, resulting in high hardware costs and huge data volumes. This invention uses compressed sensing theory to establish a sparse representation in the wavelet domain, and reconstructs the complete signal through random undersampling and orthogonal matching pursuit algorithms. Even with a sampling rate reduced to 30% of traditional methods, it can still accurately extract target distance and velocity information, significantly reducing the rate requirements of analog-to-digital converters and the burden of data storage. Traditional sound source localization methods suffer from accuracy degradation in substation reverberant environments. This invention uses non-negative matrix factorization to decompose the mixed sound field signal into independent sound source components, suppresses reverberation interference through sparsity constraints and time smoothing regularization terms, extracts footstep feature vectors and speech feature vectors, and utilizes a generalized cross-correlation time delay estimation algorithm combined with phase transform weighting to enhance noise and reverberation suppression capabilities, achieving a sound source azimuth angle calculation with an accuracy of 5 degrees.

[0077] Specifically, the principle of this invention is as follows: This invention establishes a correlation constraint between radar echoes and acoustic features through multimodal signal fusion. It utilizes the physical characteristic that real targets have consistent spatial positions while false multipath echoes lack a corresponding sound source, effectively suppressing multipath interference. Compressed sensing reconstruction recovers the complete radar signal under undersampling conditions, reducing the data acquisition burden while maintaining the accuracy of distance and velocity extraction. Non-negative matrix factorization decomposes the mixed sound field signal into independent sound source components, extracting the temporal features and spatial orientation information of footsteps. The coarse-grained classification of the hierarchical soft maximum classifier uses shallow features to quickly classify target types, providing clear boundaries for fine-grained classification. Fine-grained classification learns sub-category discrimination features under the constraints of coarse classification, avoiding overfitting and improving generalization ability. The temperature parameter adaptively adjusts the decision smoothness based on echo amplitude, angle deviation, and footstep energy, maintaining caution to avoid misjudgment when the signal is reliable and making decisive decisions to avoid hesitation when the signal is poor.

[0078] The following provides a specific embodiment 1 of the present invention. The specific implementation methods of steps S1, S5 and S6 in this embodiment 1 are the same as those described above, and will not be repeated in detail here. The specific implementation methods of other steps are described in detail below.

[0079] In step S2, the specific implementation of the undersampling reconstruction processing based on compressed sensing is as follows: The target echo signal received by the millimeter-wave radar module is processed by deskewing to obtain the intermediate frequency signal. ,in This is a time variable, with the unit being seconds. For intermediate frequency signals... Random undersampling is performed at a rate lower than the Nyquist sampling rate, with a sampling rate of 30% of the Nyquist sampling rate, to obtain the measurement vector. Measurement vector The dimension is ,in The number of sampling points is 512 (empirical value), and the measurement vector is... The element unit is volt. For intermediate frequency signals... Wavelet transform is performed to obtain the wavelet coefficient matrix. wavelet coefficient matrix The dimension is ,in The number of wavelet basis functions, typically 1024, represents the wavelet coefficient matrix. The elements are dimensionless parameters. Construct a Gaussian random observation matrix. Its dimensions are Observation matrix The elements in the set have a mean of 0 and a variance of 0. Gaussian distribution, observation matrix The elements are dimensionless parameters. Perception matrix. The statement is as follows:

[0080] ;

[0081] In the formula, Here is the wavelet transform matrix, and its dimension is... wavelet transform matrix The elemental unit is the volt; The dimension is Perception Matrix The element unit is volt. Verify the perception matrix. Limitation equidistant constant From the perception matrix 500 different column combinations are randomly selected, with each group containing a preset number of columns with a sparsity of [split length]. sparsity The empirical value is an integer of 32. The submatrix formed by each column combination. Calculate singular values ​​and submatrices The dimension is The maximum singular value is denoted as The minimum singular value is denoted as Singular values ​​are dimensionless parameters. Condition number The statement is as follows:

[0082] ;

[0083] In the formula, For dimensionless parameters, and All were obtained using the singular value decomposition algorithm. The condition number was then counted. maximum value , For dimensionless parameters, restrict equidistant constants The approximate value is expressed as follows:

[0084] ;

[0085] In the formula, the limiting equidistant constant is... The parameter is dimensionless, ranging from 0 to 1. An optimization objective function is established, expressed as follows:

[0086] The constraints are ;

[0087] In the formula, Wavelet coefficient matrix of Norm, calculated as follows ,in Wavelet coefficient matrix The One element; For the residual vector Norm, the unit is volt, and it is calculated as follows: ,in matrix multiplication The One element, the unit is volt. For measurement vector The One element, the unit is volt; The reference voltage is set to 1 volt. This is an estimate of the noise energy, in volts. The calculation method is as follows: when the millimeter-wave radar module is not transmitting a linear frequency modulated continuous wave signal, the receiver background noise is collected. Background noise The unit is volts, which is used to measure background noise. Undersampling is performed to obtain the noise measurement vector Noise measurement vector The dimension is The unit is volts. Noise energy estimate. The statement is as follows:

[0088] ;

[0089] In the formula, the coefficient 1.5 is a safety margin coefficient, a dimensionless parameter used to compensate for noise power fluctuations; Noise measurement vector of Norm, the unit is volt, and it is calculated as follows: ,in Noise measurement vector The The elements are in volts. An orthogonal matching pursuit algorithm is used iteratively to solve the objective function, and the residual vector is initialized. residual vector The dimension is The unit is volts, and the initial support set is... Set to an empty set and initialize the number of iterations. , It is a dimensionless integer. In the next iteration, the perception matrix is ​​calculated. Each column With residual vector The inner product, with the unit of the inner product being volts squared, selects the column index with the largest absolute value of the inner product. The statement is as follows:

[0090] ;

[0091] In the formula, This is the dot product operator. ,in For the perception matrix No. Liede The elements of the row, in units of volts. For residual vectors The One element, the unit is volt; For the perception matrix The Columns, dimensions ; The value range is 1 to ; Integer index. Column index. Add support set Extracting the perception matrix Central support set The corresponding columns form a submatrix submatrix The dimension is ,in To support the collection The number of elements in the submatrix Least square fitting is performed to estimate the wavelet coefficients. The statement is as follows:

[0092] ;

[0093] In the formula, superscript This represents the matrix transpose operation; Invert a matrix; The dimension is The elements are dimensionless parameters. Update the residual vector. The statement is as follows:

[0094] ;

[0095] In the formula, The dimension is The unit is volts. Calculate the residual vector. norm The unit is volts. or number of iterations The iteration terminates when the maximum number of iterations (200) is reached; otherwise... And continue iterating. Output support set. Wavelet coefficient estimates at corresponding positions The wavelet coefficients at the remaining positions are set to 0 to obtain the complete sparse wavelet coefficient vector. , The dimension is The elements are dimensionless parameters. For the sparse wavelet coefficient vector... Perform inverse wavelet transform to reconstruct the complete intermediate frequency signal. The unit is volts. For reconstructing the complete intermediate frequency signal... The distance spectrum is obtained by performing a fast Fourier transform. ,in Distance frequency, measured in Hertz, distance spectrum The unit is volts. Target distance information is extracted through peak detection. Target distance information The statement is as follows:

[0096] ;

[0097] In the formula, The speed of light, with a value of meters per second; Distance spectrum The distance frequency corresponding to the mid-peak value, in Hertz; The bandwidth of a linear frequency modulated continuous wave signal is expressed in Hertz (Hz), and is typically taken as a value of 1000. hertz; The unit is meters. For reconstructing the complete intermediate frequency signal... The Doppler spectrum is obtained by performing a Fast Fourier Transform along the slow time dimension. ,in Doppler frequency, measured in Hertz (Hz), Doppler spectrum The unit is volt. Calculate the Doppler spectrum. The energy distribution, measured in volts squared, is used to extract frequency points where the energy exceeds three times the background noise energy as the target Doppler frequency. The unit is Hertz, where the background noise energy is obtained through Doppler spectrum statistics when there is no target. Target radial velocity information. The statement is as follows:

[0098] ;

[0099] In the formula, The radar carrier frequency, measured in Hertz, is typically taken as a value of [value missing]. hertz; The unit is meters per second. When A target is identified as a moving target when the speed is meters per second. A target measured in meters per second is considered stationary. Echo amplitude Distance spectrum The amplitude value corresponding to the mid-peak value, in volts.

[0100] In step S3, the specific implementation of the nonnegative matrix decomposition process is as follows. For the signals from the four microphone channels... Perform short-time Fourier transforms on each, where Number the microphone channel. The variable is time, and the unit is seconds; The unit is Pascal. The time-frequency matrix for each channel. The dimension is ,in This refers to the number of frequency points, with an empirical value of 256. This refers to the number of time frames, with an empirical value of 1000. The element unit is Pascal. The amplitude values ​​of the time-frequency matrices from the four channels are concatenated to form a joint time-frequency matrix. Joint time-frequency matrix The dimension is , The element unit is Pascal. Initialize the source basis matrix. and activation coefficient matrix The source basis matrix is ​​a random positive matrix. The dimension is Activation coefficient matrix The dimension is ,in The preset number of sound source types is 6 (based on experience). The elemental unit is Pascal. The elements are dimensionless parameters. The objective function for nonnegative matrix factorization. The statement is as follows:

[0101] ;

[0102] In the formula, It is the Frobenius norm. ,in Joint time-frequency matrix No. Line number The elements of the column, in Pascals. matrix multiplication No. Line number The elements of the column, in Pascals; For reference sound pressure level, the value is [value to be filled in]. Pascal; Activation coefficient matrix of Norm, as a sparsity penalty term, ; for The coefficient of the norm sparsity penalty term, with a value of 0.01, is a dimensionless parameter; The coefficient for the time smoothing regularization term is 0.05, a dimensionless parameter. Activation coefficient matrix No. Line number The elements of the column, dimensionless parameters; The objective function value is dimensionless. Iterative optimization using a multiplicative update rule is employed, and the source basis matrix is... The update formula is expressed as follows:

[0103] ;

[0104] In the formula, This is the element-wise multiplication operator; To prevent tiny positive numbers from being divided by zero, the value is set to... The unit is Pascal; The dimension is The unit of element is Pascal; The dimension is The element unit is Pascal. Activation coefficient matrix. The update formula is expressed as follows:

[0105] ;

[0106] In the formula, To and A matrix of all 1s of the same dimension, with elements being dimensionless parameters; for The matrix shifted one column to the left is padded with zeros in the leftmost column, and the elements are dimensionless parameters. for The matrix shifted one column to the right is padded with zeros in the rightmost column, and the elements are dimensionless parameters. The dimension is The unit of element is Pascal square; The dimension is The element unit is Pascal squared. The number of iterations is set to 200. The source basis matrix obtained from the decomposition... Chinese footstep sound recognition column Corresponding columns for voice The identification method is through manual annotation or automatic classification algorithm. and All dimensions are Extract the corresponding activation coefficient matrix The row vectors are used as time-domain activation sequences and The temporal activation sequence is a dimensionless parameter with dimension . Footstep temporal activation sequence Extracting peak intervals Zero crossing rate and energy envelope As a feature vector of footsteps Footstep feature vector The dimension is 5, including the peak interval. mean Standard deviation Zero crossing rate mean Energy envelope maximum value and average Peak interval The unit is seconds, zero crossing rate For dimensionless parameters, energy envelope These are dimensionless parameters. For speech temporal activation sequences... Extract the first 13 dimensions of the Mel frequency cepstral coefficients as the speech feature vector. Speech feature vector The parameter is dimensionless. The arrival time difference of the signals from the four microphone channels is calculated using the generalized cross-correlation time delay estimation algorithm. ,in Number the microphone channel. and , The unit is seconds. Cross-correlation is calculated for signals from adjacent microphone channels; the cross-correlation function is... The statement is as follows:

[0107] ;

[0108] In the formula, This is a time delay, measured in seconds. The unit is Pascal squared second. For the cross-correlation function... Phase transformation weighting, weighting function The statement is as follows:

[0109] ;

[0110] In the formula, Cross-correlation function The frequency domain representation of is obtained through Fourier transform, with units of Pascals squared. For magnitude operators, The unit is Pascal squared; The variable is a frequency, measured in Hertz. This is a dimensionless weighted function. The weighted cross-correlation function... The generalized cross-correlation function is obtained by performing an inverse Fourier transform. The unit is Pascal squared, and the generalized cross-correlation function is extracted. The time delay corresponding to the peak position The unit is seconds. Based on 6 sets of time delays. Based on the geometric position of the microphone array, the three-dimensional coordinates of the sound source are solved using the least squares method. The unit is meters. Sound source azimuth. The statement is as follows:

[0111] ;

[0112] In the formula, This is the arctangent function, and the output unit is radians. and The coordinates of the sound source projected onto the horizontal plane are in meters. The unit is radians; when converting to degrees, multiply by [missing value]. ,in The value is 3.14159, with a precision of 5 degrees. Footstep sound feature vector energy. The statement is as follows:

[0113] ;

[0114] In the formula, This represents the average of the peak intervals, in seconds. , where is the standard deviation of the peak interval, in seconds; For reference time, the value is 1 second; The mean of the zero-crossing rates is a dimensionless parameter. The maximum value of the energy envelope is a dimensionless parameter. The average value of the energy envelope is a dimensionless parameter. It is a dimensionless parameter.

[0115] In step S4, the specific implementation of the multipath interference suppression model is as follows. The model input feature vector... For target distance information Target radial velocity information echo amplitude azimuth of the sound source Footstep feature vector The target distance information is composed of the first five dimensions, and each dimension needs to be normalized before stitching. The statement is as follows:

[0116] ;

[0117] In the formula, The maximum monitoring distance is expressed in meters, with an empirical value of 50 meters. This is a dimensionless parameter. Normalized value of target radial velocity information. The statement is as follows:

[0118] ;

[0119] In the formula, The maximum radial velocity is expressed in meters per second, with an empirical value of 5 meters per second. This is a dimensionless parameter. The normalized value of the echo amplitude. The statement is as follows:

[0120] ;

[0121] In the formula, The maximum echo amplitude is measured in volts and is obtained through statistical analysis of the training set; the empirical value is 1 volt. This is a dimensionless parameter. The normalized value of the sound source azimuth angle. The statement is as follows:

[0122] ;

[0123] In the formula, The unit is radians; The value is 3.14159; These are dimensionless parameters. The first 5 elements of the footstep feature vector are denoted as... , respectively , , , , All of these are dimensionless parameters. The model input feature vector... The statement is as follows:

[0124] ;

[0125] In the formula, superscript Indicates vector transpose; The total dimension is 9, and it is a dimensionless column vector. Radar measurement angle. The calculation method is as follows: The millimeter-wave radar module contains multiple receiving antennas, and the spacing between the receiving antennas is half a wavelength. ,in Wavelength, in meters. The phase difference between the target echo signals received by adjacent receiving antennas. Obtained through signal processing, measured in radians. Radar measurement angle. The statement is as follows:

[0126] ;

[0127] In the formula, The distance between receiving antennas, in meters. ; The value is 3.14159; The unit is radians. Angular deviation. The statement is as follows:

[0128] ;

[0129] In the formula, The unit is radians; when converting to degrees, multiply by [missing value]. Temperature parameters of the hierarchical soft maximum classification head. The calculation formula is expressed as follows:

[0130] ;

[0131] In the formula, The base temperature value is 1.0, and it is a dimensionless parameter. This is the amplitude adjustment factor, a dimensionless parameter. Angle adjustment factor, a dimensionless parameter; The energy regulation factor is a dimensionless parameter. This is a dimensionless parameter. Amplitude adjustment factor. The statement is as follows:

[0132] ;

[0133] In the formula, The standard echo amplitude was obtained from the median of the training set, and the unit is volts. This is a limiting function that sets the input value... Limited to arrive Within the range, when Time output ,when Time output Otherwise output Both input and output parameters are dimensionless. This is a dimensionless parameter. Angle adjustment factor. The statement is as follows:

[0134] ;

[0135] In the formula, The unit is radians. Convert the radians to normalized values ​​in units of 90 degrees, which are dimensionless parameters; This is a dimensionless parameter. Energy regulation factor. The statement is as follows:

[0136] ;

[0137] In the formula, The standard energy value is obtained from the median of the training set statistics and is a dimensionless parameter. These are dimensionless parameters. The multipath interference suppression model outputs the true target position coordinates. Two-dimensional correction offset And a set of spurious multipath echo markers. Two-dimensional corrected offset. The unit is meters, and the normalized offset is obtained through the fully connected layer of the model output layer. The normalized offset is a dimensionless parameter, multiplied by the maximum monitoring distance. Restore to the actual offset. Target's true position coordinates. The statement is as follows:

[0138] ;

[0139] ;

[0140] In the formula, The target's position coordinates are measured directly by radar, in meters, based on target distance information. and radar measurement angle Calculations show that , ; The unit is meters; The unit is meters; The unit is meters.

[0141] Among them, the shortest distance value of the target device The calculation formula is expressed as follows:

[0142] ;

[0143] In the formula, These are the coordinates of the center point of the hazardous area, in meters, and are the same as the spatial coordinates of the equipment. The radius of the danger zone is in meters. The unit is meters. Danger zone radius. The statement is as follows:

[0144] ;

[0145] In the formula, The preset safe distance threshold, in meters, is determined based on the equipment voltage level; The physical dimensions of the equipment are in meters and are obtained through laser scanning or manual measurement.

[0146] Among them, alarm levels The division rules are described as follows:

[0147] when hour, This indicates a serious danger;

[0148] when hour, This indicates a moderate risk.

[0149] when hour, This indicates a slight danger;

[0150] when hour, This indicates a safe state;

[0151] In the formula, It is a dimensionless parameter; It is an integer, and its value ranges from 0 to 3.

[0152] To better understand and implement this invention, the following is an embodiment 2 of a specific application scenario: A technical team deployed a personnel safety distance monitoring device in a 220kV substation. The main transformer area of ​​this substation has complex metal equipment cabinets and pipe structures, and the ground is made of concrete, resulting in significant multipath interference. The monitoring device was installed 5m east of the main transformer, at a height of 2.8m above the ground. Figure 2 and Figure 6 As shown, the device is powered by a rechargeable lithium battery pack with a nominal voltage of 12V and a capacity of 20Ah. Actual testing showed a battery life of 71 hours. The positioning module uses a dual-mode BeiDou and GPS positioning chip to acquire the device's spatial coordinates and store them in the data storage area. Simultaneously, it reads the preset safe distance threshold of 3.0m and the boundary parameters of the device's hazardous area. The center point coordinates of the hazardous area represent the device's installation location, and the radius of the hazardous area is 3.5m.

[0153] The millimeter-wave radar module operates at a frequency of 77 GHz, with a bandwidth of 4 GHz, a range resolution of 0.04 m, and a maximum detection range of 50 m. The technical team conducted 80 hours of data acquisition on-site. During the acquisition, personnel moved along a pre-set trajectory within the monitoring area, covering a distance range of 0.5 m to 15 m and an angle range covering a 180-degree fan-shaped area in front. Movement included standing still, slow walking, and rapid passage. The millimeter-wave radar module transmits a linear frequency modulated continuous wave signal and receives the target echo signal. The echo signal is de-skewed to obtain the intermediate frequency (IF) signal. The IF signal is then randomly undersampled at a rate lower than 60% of the Nyquist sampling rate to obtain the measurement vector. Wavelet transform is performed on the IF signal to obtain the wavelet coefficient matrix. Statistical analysis revealed that the proportion of coefficients greater than the energy threshold was 8.7%, satisfying the sparsity condition. The product of the Gaussian random observation matrix and the wavelet transform matrix is ​​constructed as the sensing matrix. 500 different column combinations are randomly selected from the sensing matrix for verification, with each group containing 120 columns. The calculated constraint equidistant constant is 0.27, which meets the requirement of being less than 0.3, thus ensuring the uniqueness and stability of the signal reconstruction.

[0154] The technical team employed an orthogonal matching pursuit algorithm for iterative solution. The residual vector was initialized as the measurement vector, and the support set was initialized as an empty set. The noise energy estimate, calculated from the background noise of the acquisition receiver, was 0.032. Multiplying this by a safety margin factor of 1.5 yielded a termination threshold of 0.048. During the iteration process, the wavelet basis function with the highest correlation to the residual was selected and added to the support set each time. After 87 iterations, the residual vector... When the norm drops to 0.041, below the termination threshold, the algorithm terminates and outputs the sparse wavelet coefficient vector. An inverse wavelet transform is performed on the sparse wavelet coefficients to reconstruct the complete intermediate frequency signal. A fast Fourier transform is then performed on the reconstructed signal to obtain the distance spectrum, as shown below. Figure 3 As shown, the target distance information extracted through peak detection is 6.8m, and the echo amplitude is 0.73. A Fast Fourier Transform is performed on the reconstructed signal along the slow time dimension to obtain the Doppler spectrum. The frequency point with energy exceeding three times the background noise energy is extracted as 267Hz. The calculated radial velocity information of the target is 0.52m / s, thus identifying it as a moving target.

[0155] The 4-microphone array uses a square layout with a microphone spacing of 0.1m and a sampling rate of 48kHz to synchronously acquire ambient sound field signals. Short-time Fourier transforms are performed on the signals from each of the four microphone channels to obtain a time-frequency matrix with 1024 frequency points in each row and 300 time frames in each column. The amplitudes of the time-frequency matrices from the four channels are concatenated to form a joint time-frequency matrix with dimensions of 1024×1200. The source basis matrix and activation coefficient matrix are initialized as random positive matrices. The number of columns in the source basis matrix is ​​set to 5, corresponding to five preset sound source types: footsteps, speech, device noise, ambient noise, and wind noise. A multiplicative update rule is used for 200 iterations of optimization. The coefficient of the norm sparsity penalty term is 0.01, and the coefficient of the time smoothing regularization term is 0.05, ultimately converging the Frobenius norm to 12% of its initial value. Footsteps are identified in the second column of the decomposed source basis matrix, and speech is identified in the first column. The row vectors of the corresponding activation coefficient matrices are extracted as the temporal activation sequence. The footstep temporal activation sequence is extracted with a peak interval of 0.68s, a zero-crossing rate of 0.34, and an energy envelope peak value of 0.91, forming the first three dimensions of the footstep feature vector. The remaining two dimensions are 0.27 and 0.19, respectively. The energy of the footstep feature vector is calculated as the sum of squares of all elements, which is 1.28. The first 13 dimensions of the Mel-frequency cepstral coefficients are extracted from the speech temporal activation sequence as the speech feature vector. The arrival time difference of signals from four microphone channels was calculated using a generalized cross-correlation time delay estimation algorithm. Six cross-correlation functions were weighted by phase transformation, and the weighted cross-correlation functions were subjected to inverse Fourier transform to obtain the generalized cross-correlation function. The time delays corresponding to the peak positions were extracted to be 0.00029s, 0.00021s, 0.00015s, 0.00008s, 0.00012s, and 0.00019s, respectively. Based on the time delays and the geometric position of the microphone array, the three-dimensional coordinates of the sound source were solved using the least squares method. Projecting these coordinates onto the horizontal plane yielded a sound source azimuth angle of 32 degrees, with an accuracy of 5 degrees.

[0156] The millimeter-wave radar module contains four receiving antennas spaced 1.95 mm apart (half a wavelength). The phase difference between the target echo signals received by adjacent antennas is calculated to be 0.58 radians, and the radar measurement angle is calculated to be 28 degrees. The angular deviation is defined as the absolute value of the difference between the sound source azimuth angle of 32 degrees and the radar measurement angle of 28 degrees, which is 4 degrees. The target distance information (6.8 m), target radial velocity information (0.52 m / s), echo amplitude (0.73), sound source azimuth angle of 32 degrees, and the first five dimensions of the footstep sound feature vector are concatenated into a 9-dimensional model input feature vector, which is then input into the multipath interference suppression model. The multipath interference suppression model employs a hierarchical soft-maximum classification header structure, such as... Figure 4 As shown.

[0157] The shortest distance to the target equipment is calculated based on the target's actual location coordinates and the boundary parameters of the hazardous area. The Euclidean distance between the target's actual location coordinates and the center point of the hazardous area is calculated to be 6.71m. The shortest distance to the target equipment is equal to the Euclidean distance of 6.71m minus the radius of the hazardous area of ​​3.5m, which is 3.21m. Since the shortest distance to the target equipment (3.21m) is greater than the preset safe distance threshold of 3.0m, the alarm level is set to level 0, indicating a safe state. The function execution module does not trigger the flashing of the high-brightness LED indicator or the voice warning from the 5W explosion-proof speaker. The communication module encapsulates the target's true location coordinates, the target device's shortest distance (3.21m), the sound source azimuth angle (32 degrees), and a timestamp into an alarm data packet. The data packet header uses a 4-byte identifier (0x5A5AA5A5), the timestamp uses the Unix timestamp format (1703084725) occupying 8 bytes, the device identifier is STA220KV001 occupying 16 bytes, the target's true location coordinates are represented by two floating-point numbers occupying 8 bytes, the alarm level (0) occupies 1 byte, and the data packet checksum is generated using a cyclic redundancy check algorithm (0x3C2F1A8B) occupying 4 bytes, for a total encapsulation of 41 bytes. The communication module transmits the data packet to the mobile and PC terminals via a 4G network at a CAT4 standard uplink rate of 50Mbps, with a transmission latency of 0.23s. Figure 5 As shown, after receiving the data packet, the PC parses and displays the target location, distance information, and alarm level.

[0158] The technical team simulated a scenario where personnel gradually approached the danger zone. When personnel moved to a position 3.8m east and 1.5m south of the device, after the same signal processing and model inference process, the shortest distance to the target device was calculated to be 0.58m, which is less than the preset safe distance threshold of 3.0m. The alarm level was determined based on the ratio of the shortest distance to the target device (0.58m) to the preset safe distance threshold (3.0m), which was 0.19, classified as Level 1 indicating severe danger. The function execution module triggered a high-brightness LED indicator to flash continuously at a frequency of 5Hz and display red, with a visibility distance of 100m. Simultaneously, a 5W explosion-proof speaker issued a voice warning of severe danger, urging immediate evacuation from the danger zone, with a sound pressure level of 90dB. On-site personnel immediately responded and evacuated to a safe area. During continuous monitoring, the device completed a full signal acquisition, processing, inference, and decision-making process every 0.1s, meeting the real-time requirements for safety monitoring. During the 80-hour data collection period, a total of 120,000 training samples were recorded. The samples were divided into a training set of 96,000 and a validation set of 24,000 in an 8:2 ratio, as shown in Table 1.

[0159] Table 1. Statistics on the distribution of training dataset samples

[0160]

[0161] The multipath interference suppression model training process employs a three-stage strategy. The first stage trains a coarse-grained classifier for 50 epochs, improving the accuracy from an initial 58% to 92%. The second stage trains a fine-grained classifier for 80 epochs, improving the accuracy from an initial 43% to 86%. The third stage involves joint training for 120 epochs, stabilizing the coarse-grained classification accuracy at 94% and improving the fine-grained classification accuracy to 89%, while reducing the mean absolute error of coordinate regression to 0.11m. The learning rate is dynamically adjusted using a cosine annealing strategy, starting from an initial value of 0.001, decaying to 0.0001 at the end of the first stage, 0.00001 at the end of the second stage, and 0.000001 at the end of the third stage. The model's overall performance on the validation set reaches its optimal level at epoch 186. The model weights from this epoch are selected and saved as the final model, with a file size of 87MB stored in the program storage area.

[0162] It should be noted that the variables involved in this invention are explained in detail in Tables 2, 3, and 4.

[0163] Table 2. Variable Explanation Table (Part 1)

[0164]

[0165] Table 3. Variable Explanation Table (Part Two)

[0166]

[0167] Table 4. Variable Explanation Table (Part 3)

[0168]

[0169] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring personnel safety distance in a substation, characterized in that, The system includes a millimeter-wave radar module that transmits linear frequency modulated continuous wave signals and receives target echo signals; a four-microphone array that synchronously acquires ambient sound field signals; a positioning module that obtains the device's spatial coordinates; undersampling reconstruction processing of the target echo signal based on compressed sensing; establishing a sparse representation in the wavelet domain; solving the L1 norm minimization problem using the orthogonal matching pursuit algorithm to recover the complete echo signal; extracting target distance and radial velocity information; performing non-negative matrix decomposition processing on the ambient sound field signal, decomposing the time-frequency matrix into a source basis matrix and an activation coefficient matrix; extracting footstep sound feature vectors and speech feature vectors; calculating the source azimuth angle; and determining the target distance. The multipath interference suppression model inputs the target radial velocity information, echo amplitude, sound source azimuth angle, and footstep sound feature vector. It outputs the target's true position coordinates and a set of false multipath echo markers. Based on the target's true position coordinates and the boundary parameters of the equipment's hazardous area, it calculates the shortest distance value between the target and the equipment. When the target's shortest distance value is less than the preset safe distance threshold, the function execution module triggers a high-brightness LED indicator to flash and issues a voice warning through a 5W explosion-proof speaker. The communication module encapsulates the target's true position coordinates, the target's shortest distance value, the sound source azimuth angle, and the timestamp into an alarm data packet and transmits it to the mobile terminal or PC terminal.

2. The method according to claim 1, characterized in that, The compressed sensing-based undersampling reconstruction process specifically involves deskewing the received target echo signal to obtain an intermediate frequency (IF) signal, randomly undersampling the IF signal at a rate lower than the Nyquist sampling rate to obtain a measurement vector, performing wavelet transform on the IF signal to obtain a wavelet coefficient matrix, and constructing the product of the Gaussian random observation matrix and the wavelet transform matrix as the sensing matrix.

3. The method according to claim 2, characterized in that, The orthogonal matching pursuit algorithm is solved iteratively. The residual vector is initialized as the measurement vector, the support set is initialized as an empty set, the inner product of each column of the perception matrix and the residual vector is calculated, the column index with the largest absolute value of the inner product is selected and added to the support set, and the columns corresponding to the support set in the perception matrix are extracted to form a submatrix for least squares fitting.

4. The method according to claim 3, characterized in that, Before the undersampling reconstruction process based on compressed sensing, the following steps are also included: verifying that the equidistant constant of the sensing matrix is ​​less than 0.3; randomly selecting 500 different column combinations from the sensing matrix; calculating the singular values ​​of the submatrix formed by each column combination; and calculating the ratio of the maximum singular value to the minimum singular value of the submatrix as the condition number.

5. The method according to claim 2, characterized in that, The wavelet coefficient matrix satisfies the sparsity condition by having less than 12% of the coefficients in the wavelet coefficient matrix that are greater than the energy threshold. The optimization objective function is to minimize the L1 norm of the wavelet coefficient vector, with the constraint that the product of the sensing matrix and the wavelet coefficient vector minus the L2 norm of the measurement vector is less than the noise energy estimate.

6. The method according to claim 1, characterized in that, The non-negative matrix decomposition process specifically involves performing short-time Fourier transforms on the signals from the four microphone channels to obtain time-frequency matrices, concatenating the amplitudes of the time-frequency matrices of the four channels into a joint time-frequency matrix, initializing the source basis matrix and activation coefficient matrix as random positive matrices, and establishing a non-negative matrix decomposition objective function using a multiplicative update rule for iterative optimization.

7. The method according to claim 6, characterized in that, The objective function of the nonnegative matrix decomposition is expressed as minimizing the squared Frobenius norm of the product of the joint time-frequency matrix, the source basis matrix, and the activation coefficient matrix, plus an L1 norm sparsity penalty term and a time smoothing regularization term, where the coefficient of the L1 norm sparsity penalty term is 0.01 and the coefficient of the time smoothing regularization term is 0.

05.

8. The method according to claim 6, characterized in that, The extraction of the footstep sound feature vector is achieved by identifying the corresponding column of the footstep sound from the decomposed sound source basis matrix, extracting the row vector of the corresponding activation coefficient matrix as the time-domain activation sequence, and extracting the peak interval, zero-crossing rate and energy envelope of the footstep sound time-domain activation sequence as the footstep sound feature vector.

9. A substation personnel safety distance monitoring device, characterized in that, The system includes a millimeter-wave radar module, a microphone array, an external input module, a main control chip, a storage module, a positioning module, an LED indicator, an explosion-proof speaker, a communication module, a function execution module, an external output module, a rechargeable lithium battery pack, a power supply module, and a battery management system. The main control chip is electrically connected to the millimeter-wave radar module, the microphone array, the external input module, the storage module, the positioning module, the LED indicator, the explosion-proof speaker, the communication module, the function execution module, the external output module, the rechargeable lithium battery pack, the power supply module, and the battery management system. The millimeter-wave radar module is used to transmit linear frequency modulated continuous wave signals and receive target echo signals. The microphone array... The system is used to acquire ambient sound field signals and locate sound sources. The external input module receives target echo signals and ambient sound field signals and performs analog-to-digital conversion and signal preprocessing. The storage module stores instruction codes, model parameters, running data, and alarm records. The positioning module acquires the spatial coordinates of the device and provides location information. The high-brightness LED indicator outputs light warnings in different colors and flashing modes. The explosion-proof speaker plays voice warnings. The communication module transmits alarm data packets and operating status to a mobile terminal or PC. The function execution module drives the LED indicator and explosion-proof speaker to execute alarm outputs and triggers the communication module to transmit data. The power supply module provides stable voltage to each hardware module.

10. The apparatus according to claim 9, characterized in that, The storage module stores instruction code, and the main control chip executes the instruction code to perform the method described in any one of claims 1 to 8.

Images