A blasting safety state anomaly detection method

Abstract

The present application provides a blasting safety state anomaly detection method, which relates to the technical field of blasting safety monitoring signal processing and anomaly detection, and the specific steps include: acquiring and preprocessing the blasting safety multi-channel monitoring signal to construct a blasting safety monitoring signal dataset; introducing a dynamic rate fluctuation factor, a running channel cooperative imbalance factor and a spectrum efficiency drift factor to construct a blasting dynamic safety sparse attention mechanism and obtain a sparse attention feature sequence; under the driving of the sparse attention feature sequence, a disturbance shunt model is constructed to generate a disturbance shunt multi-scale sequence, and a safety cooperative fusion mechanism is constructed to obtain a safety cooperative fusion expression sequence; a disturbance perception decoding mechanism containing a running feedback embedding module, a disturbance driven attention allocation mechanism and a disturbance compensation prediction generation module is constructed to output a safety prediction sequence; based on the safety prediction sequence, a real-time deviation coefficient is calculated, and when it exceeds the preset threshold, a blasting safety state anomaly detection result is output.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and in particular to a method for detecting anomalies in blasting safety conditions. Background Technology

[0002] Currently, blasting operations are widely used in mining, tunneling, foundation pit excavation, and engineering demolition. Their safety status assessment relies on the simultaneous acquisition and joint analysis of multi-channel signals such as borehole pressure, near-field vibration acceleration, far-field vibration velocity, blasting sound pressure, surface displacement, dust concentration, and carbon monoxide concentration. Under the conditions of strong transient excitation and changes in channel coupling, the blasting process exhibits signal characteristics such as scattered abnormal evidence, difficulty in locating key moments, and difficulty in stably characterizing state evolution. This places higher demands on the temporal consistency and cross-channel correlation modeling capabilities of anomaly detection methods.

[0003] Publication No. CN119989245A discloses an anomaly analysis method and system based on blasting data. This method divides the blasting data of monitoring nodes into regions and calculates the anomaly density within a preset historical period to set regional weights for the training samples of an autoencoder. The weighted autoencoder is then used to reconstruct the current round of data and calculate node residuals to filter activation path information. Furthermore, a graph structure model between monitoring nodes is constructed, and anomaly propagation paths are calculated based on node correlations. Path matching is then performed to determine the anomaly status of the target node. Publication No. CN118194203A discloses an intelligent blasting control method based on collaborative management. This method analyzes the operating status of target equipment through the credibility of sub-regional parameters and constructs a fault tree to analyze the causes of operational anomalies. Quantitative analysis of the accident incidence rate is performed, and the degree of deviation is judged to obtain a fault risk assessment standard. Combined with on-site detection results and the fault risk assessment standard, control instructions are generated, and faulty equipment is located to determine the cause of deviation, thereby improving the efficiency and safety of blasting operations.

[0004] However, existing technologies tend to focus on the statistical anomaly measurement and management decision-making chain construction at the monitoring node level. They lack a unified sequence-level correlation characterization and key moment identification constraints for the characteristics of anomaly evidence in multi-channel monitoring signals of blasting operations under strong transient excitation, such as time dispersion, rapid changes in cross-channel coupling strength, and drift of spectral structure with the switching of operation stages. This results in insufficient stability of anomaly detection results under different working conditions and time segments, and affects the continuity and consistency of safety status assessment. Summary of the Invention

[0005] This invention proposes a method for detecting anomalies in blasting safety status. Addressing the problems of scattered abnormal evidence, difficulty in focusing at critical moments, and unstable safety status prediction in multi-channel monitoring signals during blasting operations under strong transient excitation and channel coupling changes, this invention proposes a multi-stage anomaly detection mechanism integrating sparse attention processing, disturbance diversion multi-scale modeling, and disturbance perception decoding prediction. The mechanism includes: acquiring and preprocessing multi-channel blasting safety monitoring signals to construct a blasting safety monitoring signal dataset; introducing dynamic rate fluctuation factors, operational channel collaborative imbalance factors, and spectral efficiency drift factors to construct a dynamic safety sparse attention mechanism and obtain a sparse attention feature sequence; constructing a disturbance diversion model under the drive of the sparse attention feature sequence to generate a disturbance diversion multi-scale sequence; constructing a safety collaborative fusion mechanism based on the disturbance diversion multi-scale sequence to obtain a safety collaborative fusion expression sequence; constructing a disturbance perception decoding mechanism consisting of an operational feedback embedding module, a disturbance-driven attention allocation mechanism, and a disturbance compensation prediction generation module to output a safety prediction sequence; calculating a real-time deviation coefficient based on the safety prediction sequence; and outputting the blasting safety status anomaly detection result when the real-time deviation coefficient exceeds a preset threshold, thereby achieving blasting safety status anomaly detection.

[0006] A method for detecting abnormal blasting safety conditions, the specific method is as follows: S1. Acquire and preprocess the multi-channel blasting safety monitoring signals to obtain the blasting safety monitoring signal dataset; S2. A sparse attention mechanism for blasting dynamic safety is constructed by introducing dynamic rate fluctuation factor, operational channel coordination imbalance factor and spectral efficiency drift factor. The sparse attention mechanism for blasting dynamic safety is then used to perform sparse attention processing on the blasting safety monitoring signal dataset to obtain a sparse attention feature sequence. S3. Construct a perturbation splitting model driven by sparse attention feature sequences, generate perturbation splitting multi-scale sequences based on the perturbation splitting model, construct a secure collaborative fusion mechanism based on the perturbation splitting multi-scale sequences, and obtain a secure collaborative fusion expression sequence using the secure collaborative fusion mechanism. S4. Construct a disturbance-aware decoding mechanism. Input the secure collaborative fusion expression sequence into the disturbance-aware decoding mechanism to obtain a secure prediction sequence. The disturbance-aware decoding mechanism includes a runtime feedback embedding module, a disturbance-driven attention allocation mechanism, and a disturbance compensation prediction generation module. S5. Calculate the real-time deviation coefficient based on the safety prediction sequence. When the real-time deviation coefficient exceeds the preset threshold, output the abnormal detection result of the blasting safety status.

[0007] Preferably, for the construction of the blasting safety monitoring signal dataset, firstly, multi-channel monitoring signals generated during blasting operations under conventional initiation conditions, delay deviation conditions, charge deviation conditions, packing deviation conditions, near-field disturbance enhancement conditions, and ventilation condition changes are collected. The multi-channel monitoring signals include borehole pressure signals, near-field vibration acceleration signals, far-field vibration velocity signals, blasting sound pressure signals, surface displacement signals, dust concentration signals, and carbon monoxide concentration signals. During the acquisition process, a unified sampling configuration is performed on all channels, and time alignment processing is completed to ensure the consistency of multi-channel records under the same time index. Subsequently, format verification and missing data repair are performed on the original signals, and amplitude limiting and smoothing processing is performed on abnormal abrupt segments to improve temporal continuity. Then, based on preset sliding window parameters, the signals of each channel are divided into windows, and local standardization processing is performed on the data within the window to unify the amplitude scale. Next, one-dimensional convolutional feature mapping is performed on the standardized windows, and position embedding and time embedding are introduced for fusion and organization. Finally, each window segment is bound to its corresponding time index, channel number, and working condition record to construct a blasting safety monitoring signal dataset for subsequent sparse attention modeling.

[0008] Preferably, the process of obtaining the blasting safety monitoring signal dataset includes: S11. Acquire multi-channel monitoring signals generated by blasting operations under different working conditions, and construct an original signal matrix composed of multi-channel monitoring signals. The multi-channel monitoring signals include borehole pressure signal, near-field vibration acceleration signal, far-field vibration velocity signal, blasting sound pressure signal, ground surface displacement signal, dust concentration signal, and carbon monoxide concentration signal. S12. Divide the original signal matrix into windows according to the preset sliding window length to obtain a window segmentation matrix composed of continuous window segments; S13. Perform local normalization on each channel of the window segmentation matrix to construct a normalized window matrix; S14. Perform one-dimensional convolutional feature mapping on the standardized window matrix to obtain the convolutional feature matrix. Introduce positional embedding and temporal embedding and fuse them with the convolutional feature matrix to construct a blasting safety monitoring signal dataset.

[0009] Preferably, in step S2, the construction process of the dynamic rate fluctuation factor, the operating channel coordination imbalance factor, and the spectral performance drift factor includes: S21. Perform rate of change analysis on each channel of the standardized window matrix along the time direction and aggregate them along the channel dimension to obtain the dynamic rate fluctuation factor. S22. Construct a multi-channel operating condition vector based on the standardized window matrix, and calculate the cross-channel energy coupling relationship of the multi-channel operating condition vector at each time position to obtain the energy coupling matrix. Statistically analyze the deviation of each element in the energy coupling matrix from the overall average level to obtain the operating channel coordination imbalance factor. S23. Perform spectral analysis on local segments of the standardized window matrix to obtain the spectral energy distribution of each channel. Calculate the spectral centroid of each channel based on the spectral energy distribution. Statistically analyze the dispersion of the spectral centroid between channels to obtain the spectral performance drift factor.

[0010] Preferably, the construction process of the dynamic safety sparse attention mechanism for blasting includes: S24. By integrating the dynamic rate fluctuation factor, the operational channel coordination imbalance factor and the spectral efficiency drift factor, the sparsity score sequence of the blasting unit is obtained. S25. Select the highest-scoring elements from the sparsity scoring sequence of explosive units. A sparse set of locations of interest is constructed. Based on the sparse set of locations of interest, query features of corresponding time locations in the blasting safety monitoring signal dataset are extracted. Suppression processing is performed on time locations not included in the sparse set of locations of interest to obtain a sparse query representation set. S26. Construct an attention key representation based on the blasting safety monitoring signal dataset and the sparse query representation set, and calculate the attention scoring matrix from the attention key representation; S27. Introduce a structural bias term based on the sparsity score of explosive elements to perform structural amplification on the attention scoring matrix to obtain a structural bias attention scoring matrix. S28. Adjust the structural bias attention scoring matrix and combine it with the attention value representation generated by the blasting safety monitoring signal dataset to obtain sparse attention output results at some time locations; fill the unselected time locations with redundancy compensation according to the sparse attention location set to form a sparse attention feature sequence.

[0011] Furthermore, considering the characteristics of multi-channel monitoring signals in blasting operations—strong transient excitation dominance, rapid changes in inter-channel coupling levels with changing operating conditions, drift of the spectral energy centroid on the time axis, and sparse distribution of abnormal triggering times—this invention performs rate-of-change aggregation on the standardized window matrix along the time direction in step S2 to form a dynamic rate fluctuation factor. Based on the multi-channel operating condition vector, it calculates the cross-channel energy coupling matrix and then statistically analyzes the deviation to form a coordinating imbalance factor for the operating channels. Simultaneously, it performs spectral energy distribution statistics on local segments and uses the spectral centroid dispersion to form a spectral efficiency drift factor. Finally, it fuses these three factors to obtain a blasting unit sparsity score sequence and selects the highest-scoring units. A sparse set of attention locations is constructed. Under set constraints, query features are extracted and non-attention locations are suppressed to obtain a sparse query representation set. Then, an attention key representation is constructed using the blasting safety monitoring signal dataset and the sparse query representation set, and an attention scoring matrix is ​​calculated. A structural bias term based on sparsity scoring is introduced to perform structural amplification and adjustment on the scoring matrix and weighted combination with the attention value representation to obtain the sparse attention output result. Finally, redundancy compensation is performed on unselected locations to form a sparse attention feature sequence. This allows for focusing on key abnormal moments while maintaining the integrity of the timeline and reducing the interference of non-critical moments, thereby improving the stability of the safety status representation.

[0012] Preferably, in step S3, the process of generating the perturbation split multi-scale sequence includes: S31. Perform interval mapping processing on the sparsity scoring sequence of the blasting unit to obtain the structural disturbance intensity sequence that changes with time position, and construct a structural disturbance weight vector based on the structural disturbance intensity sequence. S32. Perform high-fidelity preservation processing, first-scale compression processing, and second-scale compression processing on the sparse attention feature sequence to obtain the multi-scale branch representation sequence. S33. Based on the structural perturbation weight vector, scale modulation processing is performed on the multi-scale branch representation sequence to construct a perturbation splitting model. The perturbation splitting model is then used to perform cross-scale recombination processing at each time position to obtain the perturbation splitting multi-scale sequence.

[0013] Preferably, the process of obtaining the secure collaborative fusion expression sequence includes: S34. Perform structural feature extraction processing on the perturbation split multi-scale sequence at each scale to obtain the structural feature sequence corresponding to each scale. S35. Utilize structural feature sequences to perform coupling analysis on structural relationships between different scales and generate cross-scale structural similarity maps. S36. Perform structural correlation regulation on the perturbation-splitting multi-scale sequences, and construct a safe collaborative fusion mechanism by combining cross-scale structural similarity maps; S37. Based on the secure collaborative fusion mechanism, feature recombination processing is performed on the regulated multi-scale representation sequence to obtain the secure collaborative fusion expression sequence.

[0014] Furthermore, considering that the blasting safety monitoring signal simultaneously contains a short-term spike response caused by impact detonation and a gradual trend caused by ventilation dust diffusion within the same operation cycle, exhibiting a multi-scale structural superposition, in step S3 of this invention, interval mapping is first performed on the sparsity scoring sequence of the blasting unit to generate a structural disturbance intensity sequence that varies with time position, and a structural disturbance weight vector is constructed accordingly. Then, high-fidelity preservation processing, first-scale compression processing, and second-scale compression processing are performed on the sparse attention feature sequence to form a multi-scale branch representation sequence. Under the drive of the structural disturbance weight vector, scale modulation and cross-scale recombination are performed on each scale branch to obtain a disturbance splitting multi-scale sequence. Subsequently, structural feature extraction processing is performed on the perturbation split multi-scale sequence at each scale to obtain structural feature sequences. Based on the structural feature sequences, coupling analysis is performed on the structural relationships between different scales to generate cross-scale structural similarity maps. Then, structural association regulation is performed on the multi-scale representation sequences, and a safe collaborative fusion mechanism is constructed by combining the cross-scale structural similarity maps. Finally, based on the safe collaborative fusion mechanism, feature recombination is performed on the regulated multi-scale representation sequences to form a safe collaborative fusion expression sequence. This not only preserves key transient details but also uniformly portrays the medium- and long-term evolutionary structure and suppresses redundant fluctuations introduced by scale switching, thereby improving the overall global robustness of the fusion expression and the identification accuracy of anomalous patterns.

[0015] Preferably, in step S4, the process of constructing the perturbation-aware decoding mechanism includes: S41. Generate the operation disturbance response vector and the multi-scale control vector based on the safety collaborative fusion expression sequence, the blasting unit sparsity score sequence and the structural disturbance weight vector, respectively. Combine the safety collaborative fusion expression sequence, the operation disturbance response vector and the multi-scale control vector to construct the operation feedback embedding module. S42. Use the runtime feedback embedding module to perform state mapping processing on the secure collaborative fusion expression sequence to form decoded input features; S43. Obtain the perturbation difference adjustment term that changes with time position based on the sparsity score sequence of the blasting unit, and generate the time distance penalty term by combining the time interval. Superimpose the perturbation difference adjustment term and the time distance penalty term to form a dynamic structure mask. S44. A perturbation-driven attention allocation mechanism is constructed from the dynamic structure mask and the decoding input features. The attention weight adjustment process is then performed on the decoding input features using the perturbation-driven attention allocation mechanism to obtain the decoded representation after perturbation modulation. S45. Generate a perturbation correction amount based on the sparsity score sequence of the blasting unit, and combine the perturbation correction amount with the decoded representation after perturbation modulation to construct a perturbation compensation prediction generation module. S46, the running feedback embedding module, the disturbance-driven attention allocation mechanism, and the disturbance compensation prediction generation module together constitute the disturbance perception decoding mechanism.

[0016] Furthermore, considering the evolutionary pattern of blasting safety status exhibiting initial abrupt changes followed by a decline over time, and the dependence of abnormal responses on historical disturbance accumulation accompanied by multi-scale coupled feedback, this invention generates operational disturbance response vectors and multi-scale control vectors in step S4 based on the safety collaborative fusion expression sequence, the blasting unit sparsity score sequence, and the structural disturbance weight vector, respectively. These three vectors are then combined to construct an operational feedback embedding module, forming an embedded expression containing disturbance feedback information. The operational feedback embedding module is then used to perform state mapping processing on the safety collaborative fusion expression sequence to form decoded input features. Subsequently, a disturbance difference adjustment term varying with time position is generated based on the blasting unit sparsity score sequence, and combined with the time interval... A time distance penalty term is generated and superimposed to form a dynamic structure mask. The dynamic structure mask applies a decreasing constraint on the correlation strength as the time distance increases and applies a structural penalty to the attention score as the perturbation difference increases, thus forming a unified constraint that simultaneously reflects temporal proximity and perturbation sensitivity. Under this constraint, a perturbation-driven attention allocation mechanism is constructed to adjust the attention weights of the decoded input features to obtain the perturbation-modulated decoded representation. Finally, a perturbation correction amount is generated based on the blast unit sparsity score sequence and combined with the perturbation-modulated decoded representation to construct a perturbation compensation prediction generation module to output a safe prediction sequence. This achieves adaptive allocation of perturbation-sensitive positions and improves the continuity and stability of the prediction sequence while maintaining temporal constraints.

[0017] Preferably, in step S5, the process of outputting the abnormal detection result of the blasting safety status includes: S51. Construct a reference safety baseline value based on the safety prediction sequence, and calculate the real-time deviation coefficient based on the reference safety baseline value and the safety prediction sequence; S52. The real-time deviation coefficient is compared with the preset threshold to obtain the blasting safety status anomaly detection result. The blasting safety status anomaly detection result includes anomaly identification and anomaly level.

[0018] Furthermore, considering the characteristics of borehole pressure, near-field vibration acceleration, far-field vibration velocity, blast sound pressure, surface displacement, dust concentration, and carbon monoxide concentration during blasting operations—which exhibit baseline drift and dimensional differences over time, and whose abnormal states often superimpose normal fluctuations in the form of short-term deviations—this invention, in step S5, uses the safety prediction sequence output by the disturbance-sensing decoding mechanism as a unified representation and constructs a reference safety benchmark value. This reference safety benchmark value forms a stable reference by performing sliding aggregation on the predicted states at continuous time positions and combining it with channel consistency constraints to reduce the impact of transient noise on the discrimination boundary. Then, based on the difference between the reference safety benchmark value and the safety prediction sequence at the same time position… The real-time deviation coefficient is calculated and used as an anomaly score to achieve scale uniformity across work sections. Then, the real-time deviation coefficient is compared with a preset threshold to obtain the anomaly detection result of the blasting safety status and output the anomaly identifier and anomaly level. The preset threshold is set according to the deviation distribution range of the normal working condition historical records and corresponds one-to-one with the anomaly level boundary. In the threshold comparison stage, a continuous over-threshold judgment rule is introduced to filter isolated peak deviations. This makes the anomaly detection result real-time, traceable and hierarchically interpretable, and easy to link with the work records for tracking. At the same time, it reduces false alarm triggering, improves the reliability of on-site handling decisions, and forms a closed-loop output for blasting safety management.

[0019] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0020] 1. This invention constructs a blasting dynamic safety sparse attention mechanism driven by dynamic rate fluctuation factor, operational channel collaborative imbalance factor and spectral efficiency drift factor. It focuses on extracting key time positions from the whole sequence and amplifies the attention weight with structural bias, thereby achieving centralized expression of abnormal evidence and reducing the impact of non-critical moment interference on safety representation, thus improving the stability of sequence expression.

[0021] 2. This invention constructs a perturbation diversion model and generates a multi-scale perturbation diversion sequence. It combines a cross-scale structural similarity map with a security collaborative fusion mechanism to complete the multi-scale structural correlation regulation and feature recombination, thereby achieving unified modeling of transient impact details and medium-to-long-term evolution trends and improving the cross-scale consistency and the identification accuracy of complex anomaly patterns.

[0022] 3. This invention constructs a disturbance-aware decoding mechanism that includes runtime feedback embedding, disturbance-driven attention allocation, and disturbance compensation prediction generation, and combines it with a real-time deviation coefficient hierarchical judgment mechanism. Under the condition of maintaining time constraints, it achieves adaptive modulation and prediction compensation output of disturbance-sensitive information, thereby improving the continuity of the safe prediction sequence and realizing real-time output and hierarchical interpretable closed loop of abnormal results. Attached Figure Description

[0023] Figure 1This is a flowchart of a method for detecting abnormal blasting safety conditions provided by the present invention.

[0024] Figure 2 This is a structural diagram of the blasting dynamic safety sparse attention mechanism provided by the present invention.

[0025] Figure 3 This is a structural diagram of the construction of a secure collaborative fusion expression sequence provided by the present invention.

[0026] Figure 4 This is a structural diagram of the disturbance sensing decoding mechanism provided by the present invention.

[0027] Figure 5 This is a structural diagram of the abnormality detection results provided by the present invention.

[0028] Figure 6 This is a schematic diagram of real-time deviation coefficient anomaly determination provided by the present invention.

[0029] Figure 7 This is a detection performance curve of the comparison method provided by the present invention. Detailed Implementation

[0030] This invention proposes a method for detecting anomalies in blasting safety status. Addressing the problems of scattered abnormal evidence, difficulty in focusing at critical moments, and unstable safety status prediction in multi-channel monitoring signals during blasting operations under strong transient excitation and channel coupling changes, this invention proposes a multi-stage anomaly detection mechanism integrating sparse attention processing, disturbance diversion multi-scale modeling, and disturbance perception decoding prediction. The mechanism includes: acquiring and preprocessing multi-channel blasting safety monitoring signals to construct a blasting safety monitoring signal dataset; introducing dynamic rate fluctuation factors, operational channel collaborative imbalance factors, and spectral efficiency drift factors to construct a dynamic safety sparse attention mechanism and obtain a sparse attention feature sequence; constructing a disturbance diversion model under the drive of the sparse attention feature sequence to generate a disturbance diversion multi-scale sequence; constructing a safety collaborative fusion mechanism based on the disturbance diversion multi-scale sequence to obtain a safety collaborative fusion expression sequence; constructing a disturbance perception decoding mechanism consisting of an operational feedback embedding module, a disturbance-driven attention allocation mechanism, and a disturbance compensation prediction generation module to output a safety prediction sequence; calculating a real-time deviation coefficient based on the safety prediction sequence; and outputting the blasting safety status anomaly detection result when the real-time deviation coefficient exceeds a preset threshold, thereby achieving blasting safety status anomaly detection.

[0031] S1. Acquire and preprocess the multi-channel monitoring signals for blasting safety to obtain a dataset of blasting safety monitoring signals.

[0032] Please see Figure 1As shown, the construction process of the blasting safety monitoring signal dataset in this embodiment includes three stages: multi-channel monitoring signal acquisition, preprocessing, and structured embedding. In the signal acquisition stage, blasting operation records covering conventional detonation conditions, delay deviation conditions, charge deviation conditions, packing deviation conditions, near-field disturbance enhancement conditions, and ventilation condition changes are selected. Hole pressure signals, near-field vibration acceleration signals, far-field vibration velocity signals, blasting sound pressure signals, surface displacement signals, dust concentration signals, and carbon monoxide concentration signals are synchronously acquired and timestamped using a unified clock. The acquisition time range is set from 10 seconds before detonation to 30 seconds after detonation to cover the entire process of blasting triggering, main impact, and attenuation. The detonation time is determined by the sampling point where the hole pressure signal first abruptly changes, serving as the alignment reference. The acquisition interval is obtained by intercepting data 10 seconds before and 30 seconds after this alignment reference point. The sampling frequency is uniformly resampled to [missing information]. HZ ensures comparability of different channels under the same time index, and organizes the channel samples into a multi-channel original signal matrix according to the time index. ,in , Indicates the sampling time index. Indicates the channel index. The total number of channels. For the first Each channel at time The sampled values; in the preprocessing stage, data format verification and missing data repair are performed on the multi-channel original signal matrix, and abnormal peaks are limited and repaired. Missing sampling points are filled by linear interpolation of the nearest valid sampling points before and after the missing segment. Abnormal peaks are replaced and limited by the combination of the neighborhood median and first exponential smoothing. The mean removal and amplitude normalization processes are performed on each channel to obtain a standardized sequence. Amplitude normalization is adopted Form, in which and The first The mean and standard deviation of the channel within the range of this record. To stabilize the terms and avoid division by zero, during the windowing stage, a sliding window method is used to segment the standardized multi-channel sequence along the time axis, with the input window length set to [value missing]. The sliding step size is set to Thus, generating from the full time series. A set of window segments consisting of window fragments, where This indicates rounding down, and the first... Each window segment is organized into a standardized window matrix. ,in , For the first The passage is in the window The standardized sequence fragments within the structured embedding organization stage are used to extract local dynamic situation features. To achieve this, a one-dimensional convolutional mapping is applied along the time direction to each standardized window matrix during the structured embedding stage to obtain a fixed-length convolutional feature matrix. ,in, In this embodiment, the number of convolution output channels is set to... , For convolution kernel parameters, For bias vector The Middle A bias value This is the channel number you input. Indicates the time position index within the window. This represents the output channel index of the convolution. The time offset is the value when the kernel length is 3, corresponding to the previous sampling point, the current sampling point, and the next sampling point, respectively. Exceeding When considering the range of values, a boundary copying method is used to replace out-of-bounds values ​​with the nearest valid sampled value to ensure that the length of the convolutional output sequence remains constant. Positional and temporal embeddings are superimposed on the convolutional feature matrix to construct a unified temporal representation of the input, and the temporal position index within the window is used. Constructing sine and cosine position embedding ,in , , and with the window start timestamp Relative time index within the window Construction of time embedding ,satisfy ,in and The linear mapping parameters are used to fuse the convolutional feature matrix with the positional embedding matrix and the temporal embedding matrix to obtain the structured embedding matrix. The structured embedding matrices corresponding to all windows are uniformly collected to form a blasting safety monitoring signal dataset, which serves as the input basis for subsequent sparse attention modeling. The dataset is divided into training and test sets in a 7:3 ratio, and the training set is further divided into training subset and validation subset in a 9:1 ratio to support subsequent model training and parameter selection.

[0033] S2. A sparse attention mechanism for blasting dynamic safety is constructed by introducing a dynamic rate fluctuation factor, an operational channel coordination imbalance factor, and a spectral efficiency drift factor. The sparse attention mechanism for blasting dynamic safety is then used to process the blasting safety monitoring signal dataset with sparse attention to obtain a sparse attention feature sequence.

[0034] Furthermore, in step S2, a dynamic safety sparse attention mechanism for blasting is constructed, the process of which is as follows: Figure 2 As shown, the specific steps for constructing a dynamic safety sparse attention mechanism for blasting are as follows: First, input the normalized window matrix. With structured embedding matrix ,in, For window indexing, to reduce the number of effective query locations and keep the time axis length constant, three types of factors are constructed along the time index within each window, generating a sparse set of locations of interest; to characterize the intensity of transient acceleration changes induced by blasting impact, within the window... Internal to each time position The dynamic rate fluctuation factor sequence is obtained by aggregating the second-order difference amplitudes according to the channels. , of which The dynamic rate fluctuation factor at each time point is ,in, For standardized window matrix The Middle Channel in time position The element at that location, when or When the boundary is exceeded, the standardized sampled value at the out-of-bounds location is replaced with the nearest valid boundary sampled value, that is, let , To ensure that the dynamic rate fluctuation factor sequence length is The dynamic rate fluctuation factor sequence is represented as Then, open the window. The multi-channel standardized values ​​at each time point are organized into a cross-channel operating condition vector. Based on this, a cross-channel energy coupling matrix is ​​constructed. Further calculate the overall average coupling level And statistically analyze each coupling element relative to The degree of deviation is used to obtain the sequence of operational channel coordination imbalance factors. , of which The operational channel coordination imbalance factor at each time point is ,in Indicates channel With channel The energy coupling strength at a given time location, and the sequence of operational channel coordination imbalance factors, are expressed as follows: Then, using the preset spectrum analysis length... , to the window Each time position within Extracting local sequences from each channel ,when At that time, The item "China crosses the border" has been uniformly replaced with To maintain consistent sequence length, a discrete Fourier transform is performed to obtain the spectral energy distribution. ,in, The imaginary unit, For discrete frequency indexing, For the first A normalized frequency point, The sampling point offset index within the spectrum window is used to calculate the channel based on the spectrum energy distribution. In time location Spectral centroid at With the mean of the cross-channel spectrum centroid ,in, To maintain consistency with the stabilization term in step S1, the denominator is kept at zero or close to zero when the local spectral energy is too low; and the spectral efficiency drift factor sequence is obtained by statistically analyzing the dispersion between channels. , of which The spectral efficiency drift factor at each time location is In this embodiment, it is set To ensure a balance between the transient spectral resolution and computational cost of the blasting, the spectral performance drift factor sequence is expressed as follows: The dynamic rate fluctuation factor, the operational channel coordination imbalance factor, and the spectral performance drift factor are linearly fused under the same time index to obtain the sparsity score sequence of the blasting unit. ,in, , , , For learnable weights and satisfying In this embodiment , , All initialized to The training process involves synchronous updates via end-to-end backpropagation; a sparse scoring sequence for burst units is selected. The largest front A set of sparse attention locations is formed by these time locations. ,in As a preset positive integer, As a sparsity control factor, this embodiment sets , Indicates rounding up; rounds down to the nearest integer. Arrange the time positions in chronological order to obtain And subsequently, in that order from Extract the corresponding rows to form a sparse query representation set; when When, define For time position After sorting ordinal index in, and Based on sparse interest location set From structured embedding matrix Extract the query features corresponding to the time location and then identify those not included in the query. Suppression is performed on the time position in the data to form a sparse query representation set, where first the time position is... Performing linear projection yields the query representation matrix. And based on this, a set of sparse query representations is obtained. ,in For the projection parameter matrix, As the attention space dimension, this embodiment sets , Indicates a collection by index Extract corresponding rows to form a sparse set; for the structured embedding matrix Performing two sets of linear projections yields the attention key representation matrix. Attention value representation matrix ,in and For projection parameter matrix; based on sparse query representation set Attention key representation matrix Calculate the attention scoring matrix And introduce a sparsity scoring sequence based on explosive elements. The structural bias term is applied to the attention scoring matrix to perform structural amplification, resulting in the structural bias attention scoring matrix. ,in The structure is a column vector of all 1s to achieve the same structural bias applied to each query row, and the attention scoring matrix for the structural bias is used. Normalization is performed along the key dimension to obtain the attention weight matrix. and the attention value representation matrix Weighted combination yields sparse attention output results at certain time positions. Further calculate the mean vector of the sparse output. , for exist ordinal index in, when hour, ,when hour, Thus, a sparse attention feature sequence is constructed. While maintaining the full length of the timeline, the number of effective query locations is reduced, and window-level sparse attention output is obtained.

[0035] S3. Generate a structural perturbation weight vector based on the sparse attention feature sequence, construct a perturbation splitting model under the drive of the structural perturbation weight vector, generate a perturbation splitting multi-scale sequence based on the perturbation splitting model, construct a secure collaborative fusion mechanism based on the perturbation splitting multi-scale sequence, and obtain a secure collaborative fusion expression sequence using the secure collaborative fusion mechanism.

[0036] Furthermore, in step S3, the window-level sparse attention feature sequence output from step S2 is used. Sparsity scoring sequence of explosive elements As input and in each window The process involves three stages: characterizing structural disturbance intensity, constructing multi-scale splits, and performing cross-scale collaborative fusion, to obtain a secure collaborative fusion expression sequence. The workflow is as follows: Figure 3 As shown, the specific steps to obtain the safe collaborative fusion expression sequence are as follows: The sparsity scoring sequence of the explosive element... Perform interval mapping to obtain the structural disturbance intensity sequence. , of which The structural disturbance intensity at each time location is ,in , , It is a stable term and consistent with step S1; based on the structural perturbation intensity sequence Constructing three-component weights , , And form the structural perturbation weight vector ,in Used to characterize time position Perturbation splitting tendency at high-fidelity scale, first scale, and second scale; sparse attention feature sequences High-fidelity branch representation sequences are obtained by performing high-fidelity preservation processing, first-scale compression processing, and second-scale compression processing respectively. First-scale compressed branch representation sequence With second-scale compressed branch representation sequence ,in The first-scale compression branch length, and , For the second-scale compression of branch length, and , This is the time index of the first-scale compressed sequence. The time index is used for the second-scale compressed sequence; to facilitate cross-scale recombination at each time point, the first-scale compressed branch represents the sequence. With second-scale compressed branch representation sequence Interpolation aligned to length ,get and ,in , , , According to the structural perturbation weight vector High-fidelity branch representation sequence First-scale compressed branch representation sequence With second-scale compressed branch representation sequence Perform scale modulation and construct perturbation shunt multiscale sequences It should be noted that this process applies to each time position. Defined , , ,and For the perturbation splitting model at time location The cross-scale recombination output at the location; respectively for , , Perform structural feature extraction processing to obtain a three-scale structural feature sequence. , , ,in For learnable mapping matrix, For bias vectors, As a structural feature dimension, this embodiment sets ; Calculate the cross-scale structural similarity coefficient at the same time location based on three-scale structural feature sequences. , , Constructing a cross-scale structural similarity map based on cross-scale structural similarity coefficients ,in Cross-scale structural similarity maps can characterize the degree of consistency of structures at different scales in the timing of blast impacts; the cross-scale structural similarity maps are used to perform structural correlation regulation on perturbation split multi-scale sequences and generate fused intermediate representations. ,in , , As learnable weights, they are initialized to [value] in this embodiment. In subsequent model training phases, synchronous updates are performed via end-to-end backpropagation. To ensure the controllability of cross-scale contributions during the fusion process, this embodiment sets... The fusion intermediate representation is subjected to feature recombination processing to obtain a secure and collaborative fusion expression sequence. ,in, , For learnable parameters, To determine the number of output channels for the convolution, the outputs of each window are arranged into a sequence set according to their window indices. This serves as the input for step S4.

[0037] S4. Construct a disturbance-aware decoding mechanism. Input the secure collaborative fusion expression sequence into the disturbance-aware decoding mechanism to obtain a secure prediction sequence. The disturbance-aware decoding mechanism includes a runtime feedback embedding module, a disturbance-driven attention allocation mechanism, and a disturbance compensation prediction generation module.

[0038] Furthermore, in step S4, a secure collaborative fusion expression sequence is used. Explosive unit sparsity scoring sequence With structural perturbation weight vector As input, in each window The built-in runtime feedback embedding, dynamic structure masking, and perturbation compensation prediction generation form a perturbation-aware decoding mechanism, the process of which is as follows: Figure 4 As shown, the specific steps for constructing the perturbation-aware decoding mechanism are as follows: To characterize the perturbation excitation intensity of the sparse score on the decoding input, based on the blasting unit sparse score sequence... Constructing the runtime disturbance response vector The running disturbance response vector satisfies ,in It is a learnable scale vector. To enable learnable bias and to achieve bounded excitation intensity of perturbation, It is a Sigmoid function and its output range is To characterize the impact of the splitting weights on the scaling of the decoder input, a structural perturbation weight vector is used. Constructing multi-scale control vectors Multi-scale control vector satisfies ,in For learnable scale basis vectors, , , They are respectively In time location The three-component weights are used; during the construction phase of the runtime feedback embedding module, the secure collaborative fusion expression sequence is incorporated. , Operational disturbance response vector With scale control vector Combined processing is performed to construct decoded input features ,in, This represents element-wise multiplication; to characterize the difference in influence between future positions and current perturbation intensity, it is based on the sparsity scoring sequence of explosive elements. Calculate the disturbance difference adjustment term as it changes with time and location. The disturbance difference adjustment term satisfies the condition when hour, ,when hour, ,in, Indicates the current time position Another time location index for comparison, As learnable weights, they can be used to adjust the strength of the effect of perturbation differences in the structural mask; to characterize the attention decay as time intervals increase, a time distance penalty term is constructed based on the time index difference. The time distance penalty item satisfies the condition. hour, ,when hour, ,in, The learnable weights can be used to adjust the slope intensity of distance decay; the dynamic structure mask is obtained by superimposing the perturbation difference adjustment term and the time distance penalty term. To maintain temporal order constraints and introduce dynamic decay of future attention, a dynamic structure mask is used as a structural constraint term in the perturbation-driven attention allocation mechanism. During the construction phase of the perturbation attention allocation mechanism, based on the decoded input features... Construct decoding queries respectively Decoding key Decoded value ,in For the projection parameter matrix, As the attention space dimension, this embodiment sets Combined with dynamic structure mask Calculate the attention scoring matrix The attention weights are obtained by normalizing the attention scoring matrix. This leads to the decoded representation after perturbation modulation. To achieve compensation and adjustment during the prediction generation stage, a sparsity scoring sequence based on explosive elements is used. Generate disturbance correction amount The perturbation correction amount satisfies ,in, For learnable mapping vectors, The bias vector is a learnable bias vector; in the construction phase of the perturbation compensation prediction generation module, the prediction generation input is defined as... ,in, A learnable mapping matrix; input generated based on prediction. Constructing a gating compensation matrix With compensation candidate matrix ,in, , For learnable mapping matrix, , As a learnable bias vector, the decoder-side output sequence is obtained by jointly combining the prediction-generated input, the gating compensation matrix, and the compensation candidate matrix. The perturbation perception decoding mechanism is composed of the operation feedback embedding module, the perturbation-driven attention allocation mechanism, and the perturbation compensation prediction generation module.

[0039] S5. Calculate the real-time deviation coefficient based on the output sequence from the decoding side. When the real-time deviation coefficient exceeds the preset threshold, output the abnormal detection result of the blasting safety status.

[0040] Further, in step S5, the decoder output sequence is input, the real-time deviation coefficient is calculated, and the anomaly detection result is obtained. The process is as follows: Figure 5 As shown, the specific steps are as follows: Output sequence on the decoding side... Performing a security state mapping yields a security prediction sequence. The safe prediction sequence satisfies ,in, For learnable mapping vectors, For learnable bias, For window In time location The safety prediction value at the location; constructing a reference safety baseline sequence based on the safety prediction sequence. The reference safety benchmark value sequence satisfies ,in, The baseline window length is set in this embodiment. The real-time deviation coefficient sequence is calculated based on the reference safety baseline value sequence and the safety prediction sequence. The real-time deviation coefficient satisfies ,in, As a stable term, For window In time location The real-time deviation coefficients at the specified locations are statistically calibrated on the training set under normal operating conditions to obtain the three-level anomaly thresholds. , , And satisfy In this embodiment, we take , , ; the maximum value in the real-time deviation sequence With Level 3 Anomaly Threshold , , The results of the blasting safety condition anomaly detection were obtained through comparison. The blasting safety condition anomaly detection results include... , among which, when This indicates that the blasting safety status is normal. When, it indicates that the blasting safety status has become abnormal; further, when When the anomaly level is level one, At that time, the anomaly level was level two. At that time, the anomaly level was three; therefore, based on the comparison between the real-time deviation coefficient sequence and the preset threshold, the output window... The corresponding abnormal detection results of blasting safety status.

[0041] Furthermore, during the model parameter initialization phase, the one-dimensional convolution kernel parameters are... Use He normal initialization and bias Initialize the linear projection matrix to zero vectors using Xavier uniform initialization and initialize the corresponding biases to zero vectors. Initialize the learnable scale vector and learnable mapping vector to a zero-mean, small-variance normal distribution. During model training, divide the window segment set obtained in step S1 into training and test sets in a 7:3 ratio, and further divide the training set into training and validation subsets in a 9:1 ratio. Within each iteration batch, adjust the window matrix of the training subset. and structured embedding matrix Execute S2 sequentially to obtain and Executing S3 yields the structural perturbation weight vector. Co-fusion expression sequence with security Execute S4 to obtain the decoder output sequence. Execute S5 to output the sequence on the decoding side. Performing security state mapping yields security prediction values. Reference safety benchmark sequence With real-time deviation coefficient sequence Accordingly, the Adam optimizer is used on the normal operating condition training subset with a learning rate of [missing information]. With a batch size of 32 and a maximum number of rounds of 100, perform end-to-end backpropagation to update all training parameters, save the final model parameters for test set inference, and output anomaly flags. And the corresponding anomaly level.

[0042] During the experimental environment configuration phase, this embodiment was implemented and trained under a Windows 10 64-bit operating system. The hardware configuration consisted of an Intel Core i7 processor, 32GB of RAM, an NVIDIA RTX 3060 12GB graphics card with CUDA 11.8 and cuDNN 8.9 acceleration enabled. The software environment included Python 3.10, PyTorch 2.1, NumPy 1.26, SciPy 1.11, and Matplotlib 3.8.

[0043] Furthermore, the blasting safety multi-channel monitoring signals obtained in steps S1 to S4 are input into the anomaly detection process constructed in this invention for processing, and the real-time deviation coefficient anomaly determination result is as follows: Figure 6As shown, the upper part presents a comparison between the safety prediction sequence and the reference safety benchmark value as the time position changes. It can be seen that the two have the same overall trend within the normal window and only have a small fluctuation deviation, which is consistent with the non-ideal consistency characteristics of the real monitoring signal under noise and coupling disturbance conditions. However, near the abnormal window, the two show obvious deviation and are accompanied by local abrupt changes. Figure 6 The lower part presents the real-time deviation coefficient curve and the results of the grading threshold settings. It can be seen that within the normal window, the real-time deviation coefficient mostly remains in a low level range and fluctuates below the threshold. However, within the two abnormal windows, the real-time deviation coefficient rises rapidly and crosses the corresponding grading threshold boundaries, clearly establishing the location and grading distinction of abnormal segments. This indicates that the real-time deviation coefficient constructed based on the safety prediction sequence possesses the ability to focus on anomalies at critical moments and exhibits stable grading response characteristics. Furthermore, the detection performance curves of the comparison method are shown below. Figure 7 As shown in the figure, the horizontal axis represents recall and the vertical axis represents precision. The high recall interest area with a recall rate of not less than 0.90 is marked. It can be seen that the curve of the method of the present invention is generally above the comparison method in the whole recall range, and still maintains a higher precision level in the high recall interest area. At the same time, its AUPRC and P@R=0.90 indicators are better than the traditional threshold method, autoencoder, a type of SVM and simple RNN comparison method. Thus, under the safety monitoring constraint of emphasizing high recall, it achieves lower false negative risk and more stable anomaly judgment output. The experimental results verify the effectiveness and robustness of the present invention for the detection of blasting safety status anomalies under strong transient excitation and channel coupling change conditions.

[0044] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for detecting abnormal blasting safety conditions, characterized in that, include: S1. Acquire and preprocess the multi-channel blasting safety monitoring signals to obtain the blasting safety monitoring signal dataset; S2. A sparse attention mechanism for blasting dynamic safety is constructed by introducing a dynamic rate fluctuation factor, an operational channel coordination imbalance factor, and a spectral efficiency drift factor. The sparse attention mechanism for blasting dynamic safety is then used to perform sparse attention processing on the blasting safety monitoring signal dataset to obtain a sparse attention feature sequence. S3. Driven by the sparse attention feature sequence, a perturbation splitting model is constructed, a perturbation splitting multi-scale sequence is generated according to the perturbation splitting model, a secure collaborative fusion mechanism is constructed based on the perturbation splitting multi-scale sequence, and a secure collaborative fusion expression sequence is obtained using the secure collaborative fusion mechanism. S4. Construct a disturbance-aware decoding mechanism, input the secure collaborative fusion expression sequence into the disturbance-aware decoding mechanism to obtain a secure prediction sequence. The disturbance-aware decoding mechanism includes a runtime feedback embedding module, a disturbance-driven attention allocation mechanism, and a disturbance compensation prediction generation module. S5. Calculate the real-time deviation coefficient based on the safety prediction sequence, and output the abnormal detection result of the blasting safety status when the real-time deviation coefficient exceeds the preset threshold.

2. The method for detecting abnormal blasting safety conditions according to claim 1, characterized in that, The process of obtaining the blasting safety monitoring signal dataset includes: S11. Acquire multi-channel monitoring signals generated by blasting operations under different working conditions, and construct an original signal matrix composed of the multi-channel monitoring signals. The multi-channel monitoring signals include borehole pressure signal, near-field vibration acceleration signal, far-field vibration velocity signal, blasting sound pressure signal, ground surface displacement signal, dust concentration signal, and carbon monoxide concentration signal. S12. Divide the original signal matrix into windows according to the preset sliding window length to obtain a window segmentation matrix composed of continuous window segments; S13. Perform local normalization on each channel of the window segmentation matrix to construct a normalized window matrix; S14. Perform one-dimensional convolutional feature mapping on the standardized window matrix to obtain a convolutional feature matrix. Introduce positional embedding and temporal embedding and fuse them with the convolutional feature matrix to construct a blasting safety monitoring signal dataset.

3. The method for detecting abnormal blasting safety conditions according to claim 1, characterized in that, The construction process of the dynamic rate fluctuation factor, the operational channel coordination imbalance factor, and the spectral performance drift factor includes: S21. Perform rate of change analysis on each channel of the standardized window matrix along the time direction and aggregate them in the channel dimension to obtain the dynamic rate fluctuation factor; S22. Construct a multi-channel operating condition vector based on the standardized window matrix, and calculate the cross-channel energy coupling relationship of the multi-channel operating condition vector at each time position to obtain an energy coupling matrix. Statistically analyze the deviation of each element in the energy coupling matrix from the overall average level to obtain the operating channel collaborative imbalance factor. S23. Perform spectral analysis on local segments of the standardized window matrix to obtain the spectral energy distribution of each channel, calculate the spectral centroid of each channel based on the spectral energy distribution, and statistically analyze the dispersion of the spectral centroid among the channels to obtain the spectral performance drift factor.

4. The method for detecting abnormal blasting safety conditions according to claim 3, characterized in that, The construction process of the dynamic safety sparse attention mechanism for blasting includes: S24. By integrating the dynamic rate fluctuation factor, the operational channel coordination imbalance factor and the spectral efficiency drift factor, the sparsity score sequence of the blasting unit is obtained. S25. Select the highest-scoring element from the sparsity scoring sequence of the explosive units. A sparse set of locations of interest is constructed. Based on the sparse set of locations of interest, query features of corresponding time locations in the blasting safety monitoring signal dataset are extracted. Suppression processing is performed on time locations not included in the sparse set of locations of interest to obtain a sparse query representation set. S26. Construct an attention key representation based on the blasting safety monitoring signal dataset and the sparse query representation set, and calculate the attention scoring matrix based on the attention key representation; S27. Introduce a structural bias term based on the sparsity score of the explosive unit to perform structural amplification processing on the attention scoring matrix to obtain a structural bias attention scoring matrix. S28. The structural bias attention scoring matrix is ​​adjusted and weighted and combined with the attention value representation generated by the blasting safety monitoring signal dataset to obtain sparse attention output results at some time positions; the unselected time positions are filled with redundancy compensation according to the sparse attention position set to form a sparse attention feature sequence.

5. The method for detecting abnormal blasting safety conditions according to claim 1, characterized in that, The process of generating perturbation split multiscale sequences includes: S31. Perform interval mapping processing on the sparsity scoring sequence of the blasting unit to obtain the structural disturbance intensity sequence that changes with time position, and construct a structural disturbance weight vector based on the structural disturbance intensity sequence. S32. Perform high-fidelity preservation processing, first-scale compression processing and second-scale compression processing on the sparse attention feature sequence to obtain a multi-scale branch representation sequence. S33. The multi-scale branch representation sequence is subjected to scale modulation processing according to the structural perturbation weight vector to construct a perturbation splitting model, and cross-scale recombination processing is performed on each time position using the perturbation splitting model to obtain a perturbation splitting multi-scale sequence.

6. The method for detecting abnormal blasting safety conditions according to claim 5, characterized in that, The process of obtaining a secure collaborative fusion expression sequence includes: S34. Perform structural feature extraction processing on the perturbation split multi-scale sequence at each scale to obtain the structural feature sequence corresponding to each scale. S35. Using the structural feature sequence, perform coupling analysis on the structural relationships between different scales to generate a cross-scale structural similarity map; S36. Perform structural correlation regulation on the perturbation-splitting multi-scale sequences, and construct a secure collaborative fusion mechanism in conjunction with the cross-scale structural similarity map; S37. Perform feature recombination processing on the regulated multi-scale representation sequence according to the security collaborative fusion mechanism to obtain the security collaborative fusion expression sequence.

7. The method for detecting abnormal blasting safety conditions according to claim 1, characterized in that, The process of constructing a disturbance-aware decoding mechanism includes: S41. Generate the operation disturbance response vector and the multi-scale control vector based on the safety collaborative fusion expression sequence, the blasting unit sparsity score sequence and the structural disturbance weight vector, respectively. Combine the safety collaborative fusion expression sequence, the operation disturbance response vector and the multi-scale control vector to construct the operation feedback embedding module. S42. The running feedback embedding module is used to perform state mapping processing on the secure collaborative fusion expression sequence to form decoded input features; S43. Obtain the perturbation difference adjustment term that changes with time position according to the sparsity scoring sequence of the blasting unit, and generate a time distance penalty term by combining the time interval. Then, superimpose the perturbation difference adjustment term and the time distance penalty term to form a dynamic structure mask. S44. A perturbation-driven attention allocation mechanism is constructed from the dynamic structure mask and the decoded input features, and the attention weight adjustment processing is performed on the decoded input features using the perturbation-driven attention allocation mechanism to obtain the decoded representation after perturbation modulation. S45. Generate a perturbation correction amount based on the sparsity score sequence of the blasting unit, and combine the perturbation correction amount with the decoded representation after perturbation modulation to construct a perturbation compensation prediction generation module. S46. The running feedback embedding module, the disturbance-driven attention allocation mechanism, and the disturbance compensation prediction generation module together constitute the disturbance perception decoding mechanism.

8. The method for detecting abnormal blasting safety conditions according to claim 1, characterized in that, The process of outputting the abnormal detection results of the blasting safety status includes: S51. Construct a reference safety benchmark value based on the safety prediction sequence, and calculate the real-time deviation coefficient based on the reference safety benchmark value and the safety prediction sequence; S52. The real-time deviation coefficient is compared with a preset threshold to obtain the blasting safety status anomaly detection result, which includes anomaly identification and anomaly level.

Images