A mine shaft safety real-time monitoring method based on deep learning

Abstract

The application discloses a kind of based on deep learning's mine vertical shaft safety real-time monitoring method, comprising: step one: collection vertical shaft operation monitoring data and hoisting control signal and carry out uniform sampling period alignment, generate multivariate time series;Step two: identify hoisting cycle boundary and construct continuous phase variable mapping into ring space angle variable;Step three: rearrange data according to ring space angle and construct ring space sequence tensor;Step four: calculate phase symmetry amplitude difference vector and form closed consistency matrix;Step five: tensor input improved TranAD model and introduce ring space closed phase flow conservation attention mechanism to generate reconstruction error;Step six: execute self-conditioned modulation reconstruction to form abnormal score;Step seven: threshold is calculated based on historical distribution and output alarm result.The application improves the identification ability of early anomaly under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of deep learning and anomaly detection technology, and in particular to a real-time safety monitoring method for mine shafts based on deep learning. Background Technology

[0002] With the advancement of intelligent mining construction and the continuous improvement of automation levels in vertical shaft hoisting systems, real-time monitoring and anomaly early warning of vertical shaft operation status have become crucial technical aspects for ensuring safe mine production. Regarding operational safety issues during vertical shaft hoisting, existing technologies primarily rely on traditional threshold monitoring methods or statistical model-based fault diagnosis methods. These methods perform threshold judgments or trend analyses on single-variable or multi-variable signals such as motor current, hoisting speed, wire rope tension, and braking status. However, under complex operating conditions, the following problems commonly exist: traditional methods based on fixed thresholds or empirical rules struggle to adapt to the dynamic changes in acceleration, constant speed, and deceleration phases during the hoisting cycle, easily leading to false alarms and missed alarms; time series modeling methods based on sliding windows typically treat vertical shaft operating data as a linear time flow, ignoring the periodic closure characteristics of the hoisting cycle and failing to effectively characterize the structural relationships between the beginning and end of the cycle; existing deep learning anomaly detection methods mostly employ general-purpose autoencoders or... While recurrent neural networks (RNNs) reconstruct time series data as a whole, improving their ability to express complex patterns, they lack a topological modeling mechanism for industrial cycle processes and do not explicitly introduce cycle closure consistency constraints, resulting in insufficient sensitivity of the model to abnormal changes at cycle boundaries. Furthermore, in multi-head self-attention structures, existing methods mainly use position encoding or distance bias to enhance time-dependent modeling capabilities, but do not consider the continuity and closure conservation characteristics of phase flow during cycle operation, limiting their ability to identify phase flow breaks or local structural damage. In addition, existing anomaly score thresholds are mostly determined using fixed proportions or static statistical methods, lacking a dynamic adaptive threshold update mechanism based on improving the cycle's historical distribution, making it difficult to maintain stable alarm performance at different operational stages.

[0003] Therefore, how to provide a real-time safety monitoring method for mine shafts based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a real-time safety monitoring method for mine vertical shafts based on deep learning. This invention maps vertical shaft operation data to a ring space structure, constructs a phase-symmetric closed consistency matrix, and introduces a ring space closed phase flow conservation mechanism into the attention scoring path of the improved TranAD model to achieve structural perception that enhances the cyclic periodic closure characteristics and phase flow continuity. Combined with self-conditional modulation and dynamic quantile threshold determination strategies, it improves the accuracy of anomaly identification and alarm stability.

[0005] A real-time safety monitoring method for mine vertical shafts based on deep learning according to an embodiment of the present invention includes the following steps:

[0006] Step 1: Collect vertical shaft operation monitoring data and hoisting control signals, align the sampling periods, and output a multivariate time series;

[0007] Step 2: Based on the lifting control signal, identify the lifting cycle boundary, construct continuous phase variables within each lifting cycle, and map them as ring space angle variables;

[0008] Step 3: Rearrange the multivariate time series according to the ring space angle variable, and construct the ring space sequence tensor corresponding to each lifting cycle;

[0009] Step 4: Calculate the symmetric magnitude difference vector of each lifting cycle in the ring space based on the ring space sequence tensor, forming a closed consistency matrix of phase symmetric mapping;

[0010] Step 5: Concatenate the ring space sequence tensor with the closure consistency matrix to form the model input tensor, and input the model input tensor into the improved TranAD model. Introduce the ring space closed phase flow conservation attention mechanism in the attention scoring calculation path of the first stage reconstruction to generate the first stage reconstruction error matrix.

[0011] Step 6: Construct a self-conditional modulation matrix based on the first-stage reconstruction error matrix and the closed consistency matrix, perform the second-stage reconstruction and calculate the weighted reconstruction error to form anomaly scores;

[0012] Step 7: Establish a historical distribution sequence based on the abnormal scores, calculate the corresponding threshold, compare the current abnormal score with the threshold, generate an alarm result, and output the corresponding boosting cycle number and abnormal time interval.

[0013] Optionally, step one is as follows:

[0014] Acquire the set of operation monitoring channels and control channels within the vertical shaft hoisting system and shaft;

[0015] For each channel in the operation monitoring channel set and the control channel set, a timestamp and the original sampled value are recorded respectively. For adjacent sampling points of any channel, the difference between adjacent timestamps is calculated and a sampling interval sequence is formed. The median of the sampling interval sequence is selected as the channel sampling period of the channel.

[0016] The minimum channel sampling period is selected from all channel sampling periods as the unified sampling period. A unified time axis is established with the unified sampling period as the interval, and each time point of the unified time axis is used as the alignment time.

[0017] Align the original sampled values ​​of each channel to the unified time axis. For numerical channels, use linear interpolation to calculate the alignment sampled value at the alignment time between two adjacent original sampled timestamps. For switch and command channels, use forward hold to determine the alignment sampled value at the alignment time between two adjacent original sampled timestamps. Generate missing markers for alignment times where the original sampled values ​​have not been obtained.

[0018] Following the order of the unified time axis, the aligned sampled values ​​of the operation monitoring channel set and the aligned sampled values ​​of the control channel set are concatenated to form a multivariate time series.

[0019] Optionally, step two is as follows:

[0020] Extract the boost control signal sequence from the multivariate time series. The boost control signal sequence includes the aligned sampled values ​​of the speed setpoint channel, speed feedback channel, braking command channel and arrival switch channel.

[0021] A sequence of absolute speed values ​​is constructed based on the aligned sampled values ​​of the speed feedback channel. A static threshold and a running threshold are preset. When the absolute speed value is continuously greater than the running threshold and has been continuously less than the static threshold before, the corresponding time point is determined as the candidate start position of the boosting cycle. When the absolute speed value is continuously less than the static threshold and has been continuously greater than the running threshold before, the corresponding time point is determined as the candidate end position of the boosting cycle. The candidate boosting cycle interval is formed based on the start position and the end position.

[0022] Boundary checks are performed on each candidate lifting cycle interval. At the start position of the candidate lifting cycle, the braking command status is detected as released and the arrival switch is in an untriggered state. At the end position of the candidate lifting cycle, the braking command status is detected as triggered and the arrival switch is in a triggered state. Only candidate lifting cycle intervals that meet the above conditions are retained as lifting cycles.

[0023] For each lift cycle, calculate the time span, which is the timestamp of the lift cycle ending position minus the timestamp of the lift cycle starting position. Then, number the sampling points within the time span consecutively according to the time order of a unified time axis.

[0024] Within each lifting cycle interval, continuous phase variables are linearly allocated according to the proportion of the sampling point number to the total number of sampling points;

[0025] Based on the range of values ​​of the continuous phase variables, the continuous phase variables are mapped to the ring space angle variables, and a one-to-one correspondence between the ring space angle variables and the corresponding sampling points is established.

[0026] Optional, step three is as follows:

[0027] Within each improvement cycle, the operation monitoring data and corresponding loop space angle variables in the multivariate time series are extracted according to a unified time axis order, and then rearranged in ascending order of the loop space angle variables.

[0028] A three-dimensional data structure is established for the rearranged set of sampling points. The first dimension represents the time index, the second dimension represents the monitoring channel index, and the third dimension represents the channel value, forming a ring space sequence tensor. The length of the time index is equal to the number of sampling points in the lifting loop, and the length of the channel index is equal to the number of monitoring channels running.

[0029] To address the issue of inconsistent cyclic sampling points, a maximum sampling length N is preset. When the number of sampling points is less than N, zeros are padded at the end of the time dimension, and a padding mask matrix is ​​generated to identify the padding position. When the number of sampling points is greater than N, the first N sampling points are truncated in ascending order of the ring space angle.

[0030] The ring space sequence tensor is segmented into segments in the time dimension using a sliding window. The window length is set to L and the sliding step size is set to S. Sub-tensors of length L are extracted sequentially in the direction of increasing time index to form a set of ring space subsequence tensors.

[0031] Optional, step four is as follows:

[0032] In the ring space sequence tensor corresponding to each lifting cycle, a phase-symmetric mapping relationship is constructed based on the ring space angle variable. A one-to-one correspondence is established between the time index with a ring space angle of θ and the time index with a ring space angle of 360 degrees minus θ, forming a set of phase-symmetric index pairs.

[0033] For each phase symmetric index pair, the corresponding channel value difference is calculated in each monitoring channel dimension to form a symmetric difference vector, and the absolute value of the symmetric difference vector is processed to obtain a symmetric amplitude difference vector.

[0034] Based on the time index dimension, the symmetric magnitude difference vector corresponding to each phase symmetric index pair is written into the corresponding time index position to construct a closed consistency matrix consistent with the time dimension and channel dimension of the ring space sequence tensor.

[0035] When there is no perfectly symmetrical index for the ring space angle variable, the index with the smallest ring space angle difference is selected as the matching index pair for difference calculation.

[0036] Optional, step five is as follows:

[0037] The improved TranAD model includes an input splicing module, a closed-phase flow conservation attention mechanism in the ring space, a first-stage reconstruction module, and a second-stage self-conditional reconstruction module.

[0038] The input splicing module splices the ring space sequence tensor and the closure consistency matrix in the channel dimension to form the model input tensor. The time dimension of the model input tensor is the same as that of the ring space sequence tensor.

[0039] The closed-phase flow conservation attention mechanism in the ring space includes a phase flow gradient calculation unit, a closure mapping unit, and a conservation matrix construction unit. The phase flow gradient calculation unit calculates the difference between adjacent time indices for each monitoring channel in the time dimension of the ring space sequence tensor, forming a phase flow gradient tensor. The closure mapping unit establishes a head-to-tail closure correspondence based on the ring space angle variables, creating a closure index pair between the start and end indices of the time dimension. The conservation matrix construction unit calculates the cumulative difference between the forward gradient, backward gradient, and closure gradient at each time index position based on the phase flow gradient tensor and the closure index pair, forming a closed-phase flow conservation matrix in the ring space with the same length as the time dimension.

[0040] The first-stage reconstruction module includes a Transformer encoder and a Transformer decoder. When performing multi-head self-attention calculation, the Transformer encoder first calculates the scoring matrix of the query vector and the key vector, then maps the closed phase flow conservation matrix of the ring space to the corresponding scoring position according to the time index, expands the conservation matrix according to the time index, and superimposes it with the corresponding position of the scoring matrix to generate the encoded feature representation. The Transformer decoder performs decoding and reconstruction calculation based on the encoded feature representation to generate the first-stage reconstruction result, and calculates the first-stage reconstruction error matrix based on the model input tensor and the first-stage reconstruction result.

[0041] The second-stage self-conditional reconstruction module is connected in series with the first-stage reconstruction module and receives the encoded feature representation output by the first-stage reconstruction module.

[0042] Optional, step six specifically includes:

[0043] The first-stage reconstruction error matrix and the closure consistency matrix are aligned element by element according to the time index and the channel index. The linear weighted sum of the first-stage reconstruction error value and the closure consistency value is calculated at each corresponding position to generate the self-conditional modulation matrix.

[0044] The self-conditional modulation matrix is ​​input into the second-stage self-conditional reconstruction module of the improved TranAD model. The self-conditional modulation matrix is ​​converted into a gated weight matrix of the same dimension as the encoded feature representation through linear mapping. The encoded feature representation is then multiplied element-wise by time index and channel index to obtain the modulated feature representation.

[0045] The difference between the reconstruction result of the second stage and the model input tensor is calculated according to the time index and the channel index respectively to form the reconstruction error matrix of the second stage.

[0046] The second-stage reconstruction error matrix is ​​summed in the channel dimension and divided by the number of channels to obtain the time dimension error sequence. Then, the time dimension error sequence is summed in the time dimension and divided by the time length to obtain the anomaly score corresponding to a single boosting loop.

[0047] Optionally, step seven is as follows:

[0048] Construct a sequence of historical outlier scores according to the order in which the promotion cycles occurred, and assign a unique cycle number to each promotion cycle;

[0049] Select a sliding window sequence of length M from the historical abnormal score sequence, sort the abnormal scores in the sliding window according to their numerical values, and determine the abnormal score that is located at the preset quantile P position after sorting as the current threshold.

[0050] The abnormal score of the current boosting loop is compared with the current threshold. If the abnormal score is greater than or equal to the current threshold, it is determined to be an abnormal loop. If the abnormal score is less than the current threshold, it is determined to be a normal loop.

[0051] For a boosting loop that is determined to be an abnormal loop, identify the time intervals in the time dimension error sequence of the corresponding boosting loop where the difference between adjacent time indices is 1 and the error value is continuously greater than the average value of the time dimension error sequence multiplied by a preset coefficient K. Determine the continuous time intervals that meet the conditions as abnormal time intervals and output the corresponding loop number and abnormal time interval.

[0052] The beneficial effects of this invention are:

[0053] This invention constructs a ring space structure modeling framework based on lifting cycles, reconstructing linear time series into ring space sequence tensors with periodic closure characteristics. It then combines this with a phase-symmetric mapping mechanism to construct a closure consistency matrix. Addressing the issues of weak correlation of periodic boundaries and difficulty in characterizing the beginning and end structures during shaft hoisting, as well as the neglect of periodic topological relationships in traditional time modeling, this invention proposes a structure enhancement strategy based on ring space rearrangement and symmetric index matching, significantly improving the ability to identify periodic closure breaks and structural anomalies. In terms of model structure design, an improved TranAD model is introduced, embedding a ring space closure phase flow conservation mechanism into the multi-head self-attention scoring path reconstructed in the first stage. By constructing a phase flow gradient tensor and conservation matrix and participating in the attention weight normalization calculation, the model... In the feature extraction stage, the phase flow continuity and closure conservation constraints of the hoisting cycle are explicitly perceived, effectively enhancing the sensitivity to local phase flow breaks and abnormal fluctuations. In the second-stage reconstruction, a self-conditional modulation mechanism based on the fusion of reconstruction error and closure consistency is constructed. The encoded features are modulated element-wise through gating mapping, strengthening the response of abnormal regions in the reconstruction process and avoiding the masking of subtle anomalies under complex working conditions by simple numerical reconstruction. Finally, combined with a dynamic quantile threshold calculation strategy based on the historical distribution of the hoisting cycle, the adaptive judgment of anomaly scores and the location of anomaly time intervals are realized, and the corresponding cycle number and anomaly time period information are output, thereby achieving safe real-time monitoring of the mine shaft hoisting system with sensitivity to the periodic structure, phase flow conservation constraints, and dynamic threshold adaptation. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a schematic diagram of the overall process of a real-time safety monitoring method for mine vertical shafts based on deep learning proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the improved TranAD model structure in a deep learning-based real-time monitoring method for safety in mine shafts proposed in this invention.

[0057] Figure 3 This is a schematic diagram illustrating the role of the closed-phase flow conservation attention mechanism in the first stage of reconstruction proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figures 1-3 A real-time safety monitoring method for mine vertical shafts based on deep learning includes the following steps:

[0060] Step 1: Collect vertical shaft operation monitoring data and hoisting control signals, align the sampling periods, and output a multivariate time series;

[0061] Step 2: Identify the lifting cycle boundary based on the lifting control signal, construct continuous phase variables within each lifting cycle, and map them to ring space angle variables;

[0062] Step 3: Rearrange the multivariate time series based on the ring space angle variable, and construct the ring space sequence tensor corresponding to each lifting cycle;

[0063] Step 4: Calculate the symmetric magnitude difference vector of each lifting cycle in the ring space based on the ring space sequence tensor, forming a closed consistency matrix of phase symmetric mapping;

[0064] Step 5: Concatenate the ring space sequence tensor with the closure consistency matrix to form the model input tensor. Input the model input tensor into the improved TranAD model. Introduce the ring space closed phase flow conservation attention mechanism in the attention scoring calculation path of the first stage reconstruction to generate the first stage reconstruction error matrix.

[0065] Step 6: Construct a self-conditional modulation matrix based on the reconstruction error matrix and the closure consistency matrix of the first stage, perform the second stage reconstruction and calculate the weighted reconstruction error to form anomaly scores;

[0066] Step 7: Establish a historical distribution sequence based on the abnormal scores, calculate the corresponding threshold, compare the current abnormal score with the threshold, generate alarm results, and output the corresponding boosting cycle number and abnormal time interval.

[0067] In this embodiment, step one specifically includes:

[0068] Acquire the set of operation monitoring channels and control channels within the vertical shaft hoisting system and shaft;

[0069] The operation monitoring channel set includes the hoisting motor current channel, drum speed channel, wire rope tension channel, cage acceleration channel, braking pressure channel, shaft wind speed channel, and gas concentration channel. The control channel set includes the speed setting channel, speed feedback channel, braking command channel, arrival switch channel, hoisting direction channel, and load level channel. The signal output by the control channel set is defined as the hoisting control signal.

[0070] For each channel in the operation monitoring channel set and the control channel set, record the timestamp and the original sampled value. For adjacent sampling points of any channel, calculate the difference between adjacent timestamps and form a sampling interval sequence. Select the median of the sampling interval sequence as the channel sampling period.

[0071] The minimum channel sampling period is selected from all channel sampling periods as the unified sampling period. A unified time axis with the unified sampling period as the interval is established, and each time point of the unified time axis is used as the alignment time.

[0072] Align the original sampled values ​​of each channel to a unified time axis. For numerical channels, use linear interpolation to calculate the aligned sampled value at the alignment time between two adjacent original sampled timestamps. For switch and command channels, use forward hold to determine the aligned sampled value at the alignment time between two adjacent original sampled timestamps. Generate missing markers for alignment times where the original sampled values ​​have not been obtained.

[0073] Following a unified timeline, the aligned sampled values ​​of the operation monitoring channel set and the aligned sampled values ​​of the control channel set are concatenated to form a multivariate time series.

[0074] In this embodiment, step two specifically involves:

[0075] Extract the boost control signal sequence from the multivariate time series. The boost control signal sequence includes the aligned sampled values ​​of the speed setpoint channel, speed feedback channel, braking command channel and arrival switch channel.

[0076] A sequence of absolute speed values ​​is constructed based on the aligned sampled values ​​of the speed feedback channel. A static threshold and a running threshold are preset. When the absolute speed value is continuously greater than the running threshold and has been continuously less than the static threshold before, the corresponding time point is determined as the candidate start position of the boosting cycle. When the absolute speed value is continuously less than the static threshold and has been continuously greater than the running threshold before, the corresponding time point is determined as the candidate end position of the boosting cycle. The candidate boosting cycle interval is formed based on the start position and the end position.

[0077] Boundary checks are performed on each candidate lifting cycle interval. At the start position of the candidate lifting cycle, the braking command status is detected as released and the arrival switch is in an untriggered state. At the end position of the candidate lifting cycle, the braking command status is detected as triggered and the arrival switch is in a triggered state. Only candidate lifting cycle intervals that meet the above conditions are retained as lifting cycles.

[0078] For each lift cycle, calculate the time span, which is the timestamp of the lift cycle ending position minus the timestamp of the lift cycle starting position. Then, number the sampling points within the time span consecutively according to the time order of a unified time axis.

[0079] Within each lifting cycle interval, continuous phase variables are linearly allocated according to the proportion of the sampling point number to the total number of sampling points;

[0080] Based on the range of values ​​of the continuous phase variables, the continuous phase variables are mapped to the ring space angle variables, and a one-to-one correspondence between the ring space angle variables and the corresponding sampling points is established.

[0081] In this implementation, the stationary threshold is set to 0.10 m / s, the running threshold is set to 0.30 m / s, the continuous judgment length is set to 5 sampling points, and the uniform sampling period is 0.02 s. A stationary state is determined when the absolute velocity value of 5 consecutive sampling points is less than 0.10 m / s, and a running state is determined when the absolute velocity value of 5 consecutive sampling points is greater than 0.30 m / s. The candidate lift cycle starts at the time point when the stationary state is satisfied and the running state is entered for the first time, and ends at the time point when the running state is satisfied and the stationary state is entered for the first time. The phase variable is linearly distributed within the lift cycle according to the ratio of the sampling point number to the total number of sampling points, with the starting position corresponding to 0 and the ending position corresponding to 1, and is proportionally mapped to the 0 to 360 degree ring space angle range.

[0082] In this embodiment, step three specifically includes:

[0083] Within each improvement cycle, the operation monitoring data and corresponding loop space angle variables in the multivariate time series are extracted according to a unified time axis order, and then rearranged in ascending order of the loop space angle variables.

[0084] A three-dimensional data structure is established for the rearranged set of sampling points. The first dimension represents the time index, the second dimension represents the monitoring channel index, and the third dimension represents the channel value, forming a ring space sequence tensor. The length of the time index is equal to the number of sampling points in the lifting loop, and the length of the channel index is equal to the number of monitoring channels running.

[0085] To address the issue of inconsistent cyclic sampling points, a maximum sampling length N is preset. When the number of sampling points is less than N, zeros are padded at the end of the time dimension, and a padding mask matrix is ​​generated to identify the padding position. When the number of sampling points is greater than N, the first N sampling points are truncated in ascending order of the ring space angle.

[0086] The ring space sequence tensor is segmented into segments in the time dimension using a sliding window. The window length is set to L and the sliding step size is set to S. Sub-tensors of length L are extracted sequentially in the direction of increasing time index to form a set of ring space subsequence tensors.

[0087] In this implementation, the maximum sampling length N is set to 1024. When the number of sampling points in the lifting loop is less than 1024, the end of the time dimension is filled with the value 0, and the filling mask matrix is ​​marked as 1 at the corresponding position, while the original sampling point position is marked as 0. When the number of sampling points is greater than 1024, the first 1024 sampling points are truncated in ascending order of the ring space angle. The sliding window length L is set to 128, and the sliding step size S is set to 32. Subtensors are truncated sequentially in the time dimension in groups of 128 sampling points, with 96 sampling points overlapping between adjacent subtensors. Each subtensor maintains the original channel order and the correspondence between the ring space angles.

[0088] In this embodiment, step four specifically includes:

[0089] In the ring space sequence tensor corresponding to each lifting cycle, a phase-symmetric mapping relationship is constructed based on the ring space angle variable. A one-to-one correspondence is established between the time index with a ring space angle of θ and the time index with a ring space angle of 360 degrees minus θ, forming a set of phase-symmetric index pairs.

[0090] For each phase symmetric index pair, the corresponding channel value difference is calculated in each monitoring channel dimension to form a symmetric difference vector. The absolute value of the symmetric difference vector is then processed to obtain a symmetric amplitude difference vector.

[0091] Based on the time index dimension, the symmetric magnitude difference vector corresponding to each phase symmetric index pair is written into the corresponding time index position to construct a closed consistency matrix consistent with the time dimension and channel dimension of the ring space sequence tensor.

[0092] When there is no perfectly symmetrical index for the ring space angle variable, the index with the smallest ring space angle difference is selected as the matching index pair for difference calculation.

[0093] In this implementation, the phase symmetry mapping relationship is achieved by discretizing and sorting the ring space angle variables. The ring space angles within each lifting cycle are numbered in ascending order, and corresponding index pairs for the first and second halves are established according to their number positions. When the number of sampling points is odd, the intermediate indices do not participate in symmetry matching. When the ring space angle distribution is uneven, the minimum angle difference matching rule is used to determine the corresponding index pairs. Symmetry difference calculation is performed independently in each monitoring channel, maintaining consistency in channel order. The closure consistency matrix is ​​aligned with the ring space sequence tensor in both the time and channel dimensions.

[0094] In this embodiment, step five specifically includes:

[0095] The improved TranAD model includes an input splicing module, a closed-phase flow conservation attention mechanism in the ring space, a first-stage reconstruction module, and a second-stage self-conditional reconstruction module.

[0096] The input splicing module splices the ring space sequence tensor and the closure consistency matrix in the channel dimension to form the model input tensor. The time dimension of the model input tensor is the same as that of the ring space sequence tensor.

[0097] The closed-phase flow conservation attention mechanism in the ring space includes a phase flow gradient calculation unit, a closure mapping unit, and a conservation matrix construction unit. The phase flow gradient calculation unit calculates the difference between adjacent time indices for each monitoring channel in the time dimension of the ring space sequence tensor, forming a phase flow gradient tensor. The closure mapping unit establishes a head-to-tail closure correspondence based on the ring space angle variables, creating a closure index pair between the start and end indices of the time dimension. The conservation matrix construction unit calculates the cumulative difference between the forward gradient, backward gradient, and closure gradient at each time index position based on the phase flow gradient tensor and the closure index pair, forming a closed-phase flow conservation matrix in the ring space with the same length as the time dimension.

[0098] The first-stage reconstruction module includes a Transformer encoder and a Transformer decoder. When performing multi-head self-attention calculation, the Transformer encoder first calculates the scoring matrix of the query vector and the key vector, then maps the closed phase flow conservation matrix of the ring space to the corresponding scoring position according to the time index, expands the conservation matrix according to the time index, and superimposes it with the corresponding position of the scoring matrix to generate the encoded feature representation. The Transformer decoder performs decoding and reconstruction calculation based on the encoded feature representation to generate the first-stage reconstruction result, and calculates the first-stage reconstruction error matrix based on the model input tensor and the first-stage reconstruction result.

[0099] The second-stage self-conditional reconstruction module is connected in series with the first-stage reconstruction module and receives the encoded feature representation output by the first-stage reconstruction module.

[0100] The improved TranAD model inherits the two-stage reconstruction framework of the TranAD model in its overall structure, including a first-stage reconstruction module based on a Transformer encoder and decoder, and a second-stage self-conditional reconstruction module set in series with the first stage. The model input is a multivariate time series tensor of uniform length. The time dimension feature representation is extracted through a multi-head self-attention mechanism, and anomaly detection is performed based on the reconstruction error, maintaining the consistency of the original model's end-to-end training and inference process.

[0101] The improved TranAD model introduces a closed-loop phase flow conservation mechanism in the first-stage reconstructed attention scoring path. Specifically, it calculates the difference between adjacent sampling points in the time dimension of the loop space sequence tensor to form a phase flow gradient tensor, establishes the first and last closed index pairs and calculates the closed gradient, and constructs a phase flow conservation matrix at each time index. The conservation matrix is ​​then extended according to the time index to the same dimension as the attention scoring matrix, and then superimposed element by element before normalization calculation is performed.

[0102] By introducing the closed phase flow conservation constraint of the loop space during the attention scoring stage, the model can explicitly perceive and improve the periodic closure characteristics and phase flow continuity of the loop during feature extraction. This results in higher sensitivity to closure breaks and abnormal phase flow regions, improving the stability and anti-interference ability of anomaly identification and enhancing the model's accuracy in identifying periodic industrial process anomalies.

[0103] In this embodiment, step six specifically includes:

[0104] The first-stage reconstruction error matrix and the closure consistency matrix are aligned element by element according to the time index and the channel index. The linear weighted sum of the first-stage reconstruction error value and the closure consistency value is calculated at each corresponding position to generate the self-conditional modulation matrix.

[0105] The self-conditional modulation matrix is ​​input into the second-stage self-conditional reconstruction module of the improved TranAD model. The self-conditional modulation matrix is ​​converted into a gated weight matrix of the same dimension as the encoded feature representation through linear mapping. The encoded feature representation is then multiplied element-wise by time index and channel index to obtain the modulated feature representation.

[0106] The difference between the reconstruction result of the second stage and the model input tensor is calculated according to the time index and the channel index respectively to form the reconstruction error matrix of the second stage.

[0107] The second-stage reconstruction error matrix is ​​summed in the channel dimension and divided by the number of channels to obtain the time dimension error sequence. Then, the time dimension error sequence is summed in the time dimension and divided by the time length to obtain the anomaly score corresponding to a single boosting loop.

[0108] In this implementation, the first-stage reconstruction error weight coefficient in the linear weighted sum is set to 0.70, and the closure consistency weight coefficient is set to 0.30; the linear mapping is implemented using a fully connected layer, with the input dimension consistent with the self-conditional modulation matrix and the output dimension consistent with the encoded feature representation dimension; the absolute value operation adopts the element-wise absolute value method; the number of channels is set to 32, and the time length is set to 128; the anomaly score is obtained by first averaging the 32 channels and then averaging the 128 time steps.

[0109] In this embodiment, step seven specifically includes:

[0110] Construct a sequence of historical outlier scores according to the order in which the promotion cycles occurred, and assign a unique cycle number to each promotion cycle;

[0111] Select a sliding window sequence of length M from the historical abnormal score sequence, sort the abnormal scores in the sliding window according to their numerical values, and determine the abnormal score that is located at the preset quantile P position after sorting as the current threshold.

[0112] The abnormal score of the current boosting loop is compared with the current threshold. If the abnormal score is greater than or equal to the current threshold, it is determined to be an abnormal loop. If the abnormal score is less than the current threshold, it is determined to be a normal loop.

[0113] For a boosting loop that is determined to be an abnormal loop, identify the time intervals in the time dimension error sequence of the corresponding boosting loop where the difference between adjacent time indices is 1 and the error value is continuously greater than the average value of the time dimension error sequence multiplied by a preset coefficient K. Determine the continuous time intervals that meet the conditions as abnormal time intervals and output the corresponding loop number and abnormal time interval.

[0114] In this implementation, the sliding window length M is set to 50 lifting cycles, and the preset quantile ratio P is set to 0.95, meaning that the abnormal score located at the ⌈0.95×50⌉th position after sorting is selected as the current threshold; the preset coefficient K is set to 1.20. The time dimension error sequence length is 128. When the difference between consecutive time indices is 1 and the error value is continuously greater than the mean of the time dimension error sequence multiplied by 1.20, it is determined to be an abnormal time interval; when the number of time points that continuously meet the conditions is not less than 5 sampling points, the corresponding abnormal time interval and cycle number are output.

[0115] Example 1: To verify the feasibility of this invention in practice, it was applied to the hoisting system of a main shaft in a large metal mine in North China. The shaft is 860m deep and uses a double-drum hoist with a rated hoisting speed of 12.5m / s. It operates on average approximately 420 hoisting cycles per day, and the operating environment is characterized by complex conditions such as dust, high humidity, and load fluctuations. During long-term operation, the mine has repeatedly experienced problems such as delayed braking response, abnormal fluctuations in wire rope tension, and short-term overloads. Traditional monitoring methods based on fixed thresholds and human experience are insufficient to identify early anomalies in a timely manner, resulting in high false alarm rates and difficulty in identifying anomalies at cycle boundaries. Particularly during the transition between acceleration and deceleration phases, the data jumps significantly at the beginning and end, making it difficult for conventional time series models to accurately distinguish between normal operating condition fluctuations and genuine anomalies.

[0116] In this embodiment, the hoisting motor current, drum speed, wire rope tension, cage acceleration, braking pressure, shaft wind speed, and gas concentration are used as operational monitoring data. Simultaneously, speed command signals, speed feedback signals, braking command signals, and arrival switch signals are collected as hoisting control signals. First, all channels are aligned with a unified sampling period of 0.02s. After interpolation and forward hold processing, a multivariate time series is formed. Then, hoisting cycles are identified based on the speed feedback signals. A stationary threshold of 0.10m / s and a running threshold of 0.30m / s are set, and the start and end positions of the cycle are determined under the condition of five consecutive sampling points. By mapping each hoisting cycle to a ring space angle variable from 0 to 360 degrees, topological rearrangement of the cycle data is achieved, and a phase-symmetric mapping relationship is constructed to generate a closed consistency matrix.

[0117] In the modeling phase, the ring space sequence tensor and the closure consistency matrix are concatenated and input into the improved TranAD model. A ring space closed phase flow conservation mechanism is introduced into the multi-head self-attention scoring during the first-stage reconstruction. A phase flow gradient tensor is constructed by calculating the difference between adjacent sampling points in the time dimension, and a conservation matrix is ​​generated by combining the beginning and end closure mappings. This matrix participates in the scoring modulation before attention normalization, enabling the model to perceive the periodic closed structure and the continuity of the phase flow. In the second stage, a self-conditional modulation mechanism is employed. The error from the first stage is linearly fused with the closure consistency matrix and mapped to gating weights, which modulate the encoded features element-wise, enhancing the reconstruction differences in anomalous regions. Finally, a 95th percentile threshold is calculated based on a sliding window of anomaly scores from 50 historical boosting cycles, achieving dynamic anomaly detection.

[0118] To verify the beneficial effects of this invention, 12,000 boosting cycles were randomly selected from data collected over 30 consecutive days, including 286 manually identified abnormal cycles. The detection accuracy, false alarm rate, and false negative rate were compared with those of the traditional fixed threshold method and the standard TranAD model without the introduction of a loop space conservation mechanism. The experimental results are shown in Table 1.

[0119] Table 1. Comparison of Anomaly Detection Results of Main Shaft Hoisting System in a Metal Mine

[0120] Comparison indicators Fixed threshold method Standard TranAD Method of the present invention Sample size 12000 12000 12000 Number of abnormalities detected 412 318 294 Actual number of outliers 286 286 286 Accuracy (%) 86.42 92.15 97.58 False alarm rate (%) 18.75 10.06 3.42 Missed Reporting Rate (%) 27.62 11.89 4.19 Average response time (s) 4.8 3.1 2.4 Periodic boundary anomaly detection rate (%) 52.3 71.6 93.8

[0121] As shown in Table 1, the overall accuracy of the method of this invention reaches 97.58%, which is 11.16 percentage points higher than the fixed threshold method and 5.43 percentage points higher than the standard TranAD. The false alarm rate decreased from 18.75% to 3.42%, and the false negative rate decreased from 27.62% to 4.19%. Regarding the periodic boundary anomaly identification rate, the method of this invention reaches 93.8%, significantly better than the comparative methods, indicating that the closed-phase flow conservation mechanism of the ring space can effectively identify structural fracture phenomena at periodic boundaries. The average response time is shortened to 2.4s, meeting the requirements for real-time monitoring.

[0122] This embodiment demonstrates that the present invention effectively solves the problem of neglecting the periodic closure characteristics of traditional linear-time modeling by using ring space structure modeling and phase flow conservation attention mechanism, thereby improving the ability to identify early anomalies under complex working conditions. At the same time, combined with dynamic quantile threshold strategy, it enhances the stability and adaptability of the system in long-term operation, and has good prospects for engineering applications.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A real-time safety monitoring method for mine vertical shafts based on deep learning, characterized in that, Includes the following steps: Step 1: Collect vertical shaft operation monitoring data and hoisting control signals, align the sampling periods, and output a multivariate time series; Step 2: Based on the lifting control signal, identify the lifting cycle boundary, construct continuous phase variables within each lifting cycle, and map them as ring space angle variables; Step 3: Rearrange the multivariate time series according to the ring space angle variable, and construct the ring space sequence tensor corresponding to each lifting cycle; Step 4: Calculate the symmetric magnitude difference vector of each lifting cycle in the ring space based on the ring space sequence tensor, forming a closed consistency matrix of phase symmetric mapping; Step 5: Concatenate the ring space sequence tensor with the closure consistency matrix to form the model input tensor, and input the model input tensor into the improved TranAD model. Introduce the ring space closed phase flow conservation attention mechanism in the attention scoring calculation path of the first stage reconstruction to generate the first stage reconstruction error matrix. Step 6: Construct a self-conditional modulation matrix based on the first-stage reconstruction error matrix and the closed consistency matrix, perform the second-stage reconstruction and calculate the weighted reconstruction error to form anomaly scores; Step 7: Establish a historical distribution sequence based on the abnormal scores, calculate the corresponding threshold, compare the current abnormal score with the threshold, generate an alarm result, and output the corresponding boosting cycle number and abnormal time interval.

2. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step one specifically involves: Acquire the set of operation monitoring channels and control channels within the vertical shaft hoisting system and shaft; For each channel in the operation monitoring channel set and the control channel set, a timestamp and the original sampled value are recorded respectively. For adjacent sampling points of any channel, the difference between adjacent timestamps is calculated and a sampling interval sequence is formed. The median of the sampling interval sequence is selected as the channel sampling period of the channel. The minimum channel sampling period is selected from all channel sampling periods as the unified sampling period. A unified time axis is established with the unified sampling period as the interval, and each time point of the unified time axis is used as the alignment time. Align the original sampled values ​​of each channel to the unified time axis. For numerical channels, use linear interpolation to calculate the alignment sampled value at the alignment time between two adjacent original sampled timestamps. For switch and command channels, use forward hold to determine the alignment sampled value at the alignment time between two adjacent original sampled timestamps. Generate missing markers for alignment times where the original sampled values ​​have not been obtained. Following the order of the unified time axis, the aligned sampled values ​​of the operation monitoring channel set and the aligned sampled values ​​of the control channel set are concatenated to form a multivariate time series.

3. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step two specifically involves: Extract the boost control signal sequence from the multivariate time series. The boost control signal sequence includes the aligned sampled values ​​of the speed setpoint channel, speed feedback channel, braking command channel and arrival switch channel. A sequence of absolute speed values ​​is constructed based on the aligned sampled values ​​of the speed feedback channel. A static threshold and a running threshold are preset. When the absolute speed value is continuously greater than the running threshold and has been continuously less than the static threshold before, the corresponding time point is determined as the candidate start position of the boosting cycle. When the absolute speed value is continuously less than the static threshold and has been continuously greater than the running threshold before, the corresponding time point is determined as the candidate end position of the boosting cycle. The candidate boosting cycle interval is formed based on the start position and the end position. Boundary checks are performed on each candidate lifting cycle interval. At the start position of the candidate lifting cycle, the braking command status is detected as released and the arrival switch is in an untriggered state. At the end position of the candidate lifting cycle, the braking command status is detected as triggered and the arrival switch is in a triggered state. Only candidate lifting cycle intervals that meet the above conditions are retained as lifting cycles. For each lift cycle, calculate the time span, which is the timestamp of the lift cycle ending position minus the timestamp of the lift cycle starting position. Then, number the sampling points within the time span consecutively according to the time order of a unified time axis. Within each lifting cycle interval, continuous phase variables are linearly allocated according to the proportion of the sampling point number to the total number of sampling points; Based on the range of values ​​of the continuous phase variables, the continuous phase variables are mapped to the ring space angle variables, and a one-to-one correspondence between the ring space angle variables and the corresponding sampling points is established.

4. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step three specifically involves: Within each improvement cycle, the operation monitoring data and corresponding loop space angle variables in the multivariate time series are extracted according to a unified time axis order, and then rearranged in ascending order of the loop space angle variables. A three-dimensional data structure is established for the rearranged set of sampling points. The first dimension represents the time index, the second dimension represents the monitoring channel index, and the third dimension represents the channel value, forming a ring space sequence tensor. The length of the time index is equal to the number of sampling points in the lifting loop, and the length of the channel index is equal to the number of monitoring channels running. To address the issue of inconsistent cyclic sampling points, a maximum sampling length N is preset. When the number of sampling points is less than N, zeros are padded at the end of the time dimension, and a padding mask matrix is ​​generated to identify the padding position. When the number of sampling points is greater than N, the first N sampling points are truncated in ascending order of the ring space angle. The ring space sequence tensor is segmented into segments in the time dimension using a sliding window. The window length is set to L and the sliding step size is set to S. Sub-tensors of length L are extracted sequentially in the direction of increasing time index to form a set of ring space subsequence tensors.

5. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step four specifically involves: In the ring space sequence tensor corresponding to each lifting cycle, a phase-symmetric mapping relationship is constructed based on the ring space angle variable. A one-to-one correspondence is established between the time index with a ring space angle of θ and the time index with a ring space angle of 360 degrees minus θ, forming a set of phase-symmetric index pairs. For each phase symmetric index pair, the corresponding channel value difference is calculated in each monitoring channel dimension to form a symmetric difference vector, and the absolute value of the symmetric difference vector is processed to obtain a symmetric amplitude difference vector. Based on the time index dimension, the symmetric magnitude difference vector corresponding to each phase symmetric index pair is written into the corresponding time index position to construct a closed consistency matrix consistent with the time dimension and channel dimension of the ring space sequence tensor. When there is no perfectly symmetric index for the ring space angle variable, the index with the smallest ring space angle difference is selected as the matching index pair for difference calculation.

6. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step five specifically involves: The improved TranAD model includes an input splicing module, a closed-phase flow conservation attention mechanism in the ring space, a first-stage reconstruction module, and a second-stage self-conditional reconstruction module. The input splicing module splices the ring space sequence tensor and the closure consistency matrix in the channel dimension to form the model input tensor. The time dimension of the model input tensor is the same as that of the ring space sequence tensor. The closed-phase flow conservation attention mechanism in the ring space includes a phase flow gradient calculation unit, a closed mapping unit, and a conservation matrix construction unit. The phase flow gradient calculation unit calculates the difference between adjacent time indices for each monitoring channel in the time dimension of the ring space sequence tensor, forming the phase flow gradient tensor; the closure mapping unit establishes a head-to-tail closure correspondence based on the ring space angle variable, and establishes a closure index pair between the start index and the end index of the time dimension; The conservation matrix construction unit is based on the phase flow gradient tensor and the closed index pair. At each time index position, the cumulative difference between the forward gradient, the backward gradient and the closed gradient is calculated to form a closed phase flow conservation matrix in the ring space with the same length as the time dimension. The first-stage reconstruction module includes a Transformer encoder and a Transformer decoder. When performing multi-head self-attention calculation, the Transformer encoder first calculates the scoring matrix of the query vector and the key vector, then maps the closed phase flow conservation matrix of the ring space to the corresponding scoring position according to the time index, expands the conservation matrix according to the time index and superimposes it with the corresponding position of the scoring matrix, performs normalization processing on the superposition result and performs weighted summation with the value vector to generate the encoded feature representation. The Transformer decoder performs decoding and reconstruction calculations based on the encoded feature representations, generates the first-stage reconstruction result, and calculates the first-stage reconstruction error matrix based on the model input tensor and the first-stage reconstruction result. The second-stage self-conditional reconstruction module is connected in series with the first-stage reconstruction module and receives the encoded feature representation output by the first-stage reconstruction module.

7. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step six specifically involves: The first-stage reconstruction error matrix and the closure consistency matrix are aligned element by element according to the time index and the channel index. The linear weighted sum of the first-stage reconstruction error value and the closure consistency value is calculated at each corresponding position to generate the self-conditional modulation matrix. The self-conditional modulation matrix is ​​input into the second-stage self-conditional reconstruction module of the improved TranAD model. The self-conditional modulation matrix is ​​converted into a gated weight matrix of the same dimension as the encoded feature representation through linear mapping. The encoded feature representation is then multiplied element-wise by time index and channel index to obtain the modulated feature representation. The difference between the reconstruction result of the second stage and the model input tensor is calculated according to the time index and the channel index respectively to form the reconstruction error matrix of the second stage. The second-stage reconstruction error matrix is ​​summed in the channel dimension and divided by the number of channels to obtain the time dimension error sequence. Then, the time dimension error sequence is summed in the time dimension and divided by the time length to obtain the anomaly score corresponding to a single boosting loop.

8. The method for real-time safety monitoring of mine vertical shafts based on deep learning according to claim 1, characterized in that, Step seven specifically involves: Construct a sequence of historical outlier scores according to the order in which the promotion cycles occurred, and assign a unique cycle number to each promotion cycle; Select a sliding window sequence of length M from the historical abnormal score sequence, sort the abnormal scores in the sliding window according to their numerical values, and determine the abnormal score that is located at the preset quantile P position after sorting as the current threshold. The abnormal score of the current boosting loop is compared with the current threshold. If the abnormal score is greater than or equal to the current threshold, it is determined to be an abnormal loop. If the abnormal score is less than the current threshold, it is determined to be a normal loop. For a boosting loop that is determined to be an abnormal loop, identify the time intervals in the time dimension error sequence of the corresponding boosting loop where the difference between adjacent time indices is 1 and the error value is continuously greater than the average value of the time dimension error sequence multiplied by a preset coefficient K. Determine the continuous time intervals that meet the conditions as abnormal time intervals and output the corresponding loop number and abnormal time interval.

Images