Intelligent monitoring method for strength of mine cemented filling body

Abstract

This invention relates to the field of mining, and more particularly to an intelligent monitoring method for the strength of cemented backfill in mines. First, parameters from data acquired from different sources are preprocessed to obtain preprocessed parameters. At each observation age, the preprocessed parameters are concatenated to obtain a fused feature vector, resulting in a standardized input sequence. Based on the standardized input sequence, an adaptive physically-gated multi-scale causal dilation convolution feature extraction algorithm is used to obtain the final output feature vector. Then, based on the final output feature vector, a dual-branch hybrid attention mechanism is introduced to obtain an attention-enhanced feature vector. Based on the attention-enhanced feature vector, the final predicted compressive strength value is obtained. Finally, based on the final predicted compressive strength value, a safety early warning mechanism is set to obtain the monitoring results. This method solves the technical problems of inability to adaptively adjust the multi-scale receptive field, insufficient feature capture or redundant computation at key stages, and poor generalization.

Description

Technical Field

[0001] This invention relates to the field of mining, and more particularly to an intelligent monitoring method for the strength of cemented backfill in mines. Background Technology

[0002] Cemented backfill mining has become one of the mainstream mining methods in metal mines, especially in deep, high-stress mining areas. By preparing a slurry with a certain fluidity using materials such as tailings, cement, and fly ash, and transporting it through pipelines to the goaf, a cemented backfill body with a certain strength is formed. This method effectively controls ground pressure, prevents surface subsidence, and enables comprehensive utilization of tailings and zero-waste mining, resulting in significant economic and social benefits. However, whether the actual strength of the cemented backfill body meets the design requirements directly affects its ability to support the overlying strata, the stability of the stope roof, and the safe production of the entire mine. Insufficient backfill strength can lead to major disasters such as large-scale collapse of the backfill body, rock bursts, and water inrushes, causing casualties and huge economic losses. Traditional methods for monitoring the strength of cemented backfill bodies mainly include laboratory test block compressive strength and on-site core sampling. While traditional methods for monitoring the strength of cemented backfill bodies are standardized and offer strong data comparability, they suffer from several drawbacks. Existing strength prediction methods primarily rely on laboratory standard test blocks or empirical formulas, failing to accurately reflect the strength evolution differences caused by the complex and variable actual service environment underground. These methods also suffer from low prediction accuracy and poor generalization ability. Furthermore, they have fixed receptive fields, weak adaptability to different mine geological conditions and backfilling processes, and the difficulty in directly transferring the same model to different mining areas. Therefore, research into an intelligent monitoring method for the strength of cemented backfill bodies in mines has significant theoretical importance and urgent engineering application value. Summary of the Invention

[0003] This invention provides an intelligent monitoring method for the strength of cemented backfill in mines, which solves the problems of existing methods having a fixed expansion rate, being unable to adaptively adjust the multi-scale receptive field, resulting in insufficient capture of key stage features or redundant calculations; and existing methods being unable to dynamically determine the relative importance of the short, medium and long branches based on the average state of the entire observation age data, resulting in poor generalization.

[0004] The present invention provides an intelligent monitoring method for the strength of cemented backfill in mines, comprising the following steps:

[0005] S1. Preprocess the parameters in the data from different sources to obtain preprocessed parameters; concatenate the preprocessed parameters at each observation age to obtain the fused feature vector for each observation age, and further obtain the standardized input sequence of the entire infill body observation age; based on the standardized input sequence, obtain the final output feature vector through the adaptive physical gated multi-scale causal dilation convolution feature extraction algorithm.

[0006] S2. Based on the final output feature vector, a dual-branch hybrid attention mechanism is introduced to obtain an attention-enhanced feature vector; based on the attention-enhanced feature vector, a two-layer fully connected neural network combined with residual connections is used for mapping to obtain the final compressive strength prediction value; based on the final compressive strength prediction value, a safety early warning mechanism is set to obtain the monitoring results.

[0007] Preferably, S1 specifically includes:

[0008] In the adaptive physical-gated multi-scale causal dilated convolution feature extraction algorithm, global average pooling is performed on the standardized input sequence to obtain a global average pooling vector representing the observation age of the entire infill body. This global average pooling vector is then fed into three independent fully connected layers, activated with Sigmoid, and then incremented by 1 to obtain the dilation scaling factor. This scaling factor is multiplied by the respective baseline convolutional kernel widths of 3, 5, and 7, and then rounded down to the nearest integer to obtain the dilation rate of the three branches. The calculation formula is as follows:

[0009]

[0010] in, Indicates the first The expansion rate of each branch; Indicates the first The baseline kernel width of each branch is fixed. , , ; It is an activation function; Indicates the first Each branch is used to compute the learnable weight matrix of the inflation scaling factor; Represents the global average pooling vector; Indicates the first One branch is used to calculate the learnable bias of the dilation scaling factor; This indicates rounding down to the nearest integer.

[0011] Preferably, S1 specifically includes:

[0012] Based on the dilation rate and the fused feature vector, convolution kernel weights are introduced. The first weight of the convolution kernel is always applied to the fused feature vector of the current observation age. Subsequently, the weights are applied sequentially to the observation age positions that are integer multiples of the dilation rate minus the current observation age. When the calculated past observation age is less than the initial observation age of the sequence, zeros are automatically padded to obtain the feature vector extracted by each branch at each observation age. The feature vectors extracted by the three branches are concatenated along the channel dimension to obtain a multi-scale convolution concatenated feature vector.

[0013] Preferably, S1 specifically includes:

[0014] Based on the observed age, a gated time-series fusion residual mechanism is introduced. The inputs are the current observed age of the filling body, the square of the current observed age of the filling body, and the annual periodic sine term. The output is a scalar gating coefficient between 0 and 1, as shown in the formula:

[0015]

[0016] in, Indicates the observed age The corresponding gating coefficient; This represents the Sigmoid activation function; Indicates the current observed age of the filling body; This represents the square of the current observed age of the filling body; Represents the sine term of the annual cycle; Represents the weight matrix of the gated fully connected layer; This indicates gating bias.

[0017] Preferably, S1 specifically includes:

[0018] Based on the gating coefficient, the fused feature vector, and the multi-scale convolutional concatenation of the feature vector, a weighted fusion is performed to obtain the final output feature vector. The calculation formula is as follows:

[0019]

[0020] in, Indicates the observed age The final output feature vector; Indicates the observed age Multi-scale convolutional concatenation of feature vectors; Represents the weight matrix of the gated fully connected layer; Indicates the observed age The fused feature vector.

[0021] Preferably, S2 specifically includes:

[0022] In the dual-branch hybrid attention mechanism, the channel attention mechanism is executed first, and the final output feature vector is input into the trainable fully connected layer. Softmax is used for normalization so that the sum of the weights of all channels is 1. The attention weight vector of each feature channel for each observation age is then calculated.

[0023] Preferably, S2 specifically includes:

[0024] In the dual-branch hybrid attention mechanism, based on the attention weight vector, the standard Transformer attention mechanism is applied to the final output feature vector to obtain the attention-enhanced feature vector. The calculation formula is as follows:

[0025]

[0026] in, Indicates the observation age Attention-enhanced feature vectors; This indicates element-wise multiplication; This represents the standard Transformer attention weight matrix; Indicates the observed age The final output feature vector; Indicates the observation age The attention weight vector for each feature channel.

[0027] Preferably, S2 specifically includes:

[0028] A safety early warning mechanism is set up so that when the predicted final compressive strength at the current observation age is lower than the set engineering safety lower limit, the safety early warning mechanism is triggered.

[0029] Preferably, S2 specifically includes:

[0030] The mean squared error is obtained by averaging the squared differences between the measured uniaxial compressive strength at the observed age and the predicted compressive strength in the preprocessed parameters. The gradient of all trainable parameters with respect to the mean squared error is calculated by backpropagation, and then multiplied by the learning rate. All trainable parameters in the adaptive physical-gated multi-scale causal dilatational convolution feature extraction algorithm and the dual-branch hybrid attention mechanism are updated synchronously.

[0031] The beneficial effects of the technical solution of the present invention are:

[0032] 1. An adaptive physical-gated multi-scale causal dilatational convolution feature extraction algorithm is adopted. By adaptively adjusting the dilatation rate of the short, medium and long branches through the dilatation scaling factor, the most matching receptive field can be obtained for the early rapid hydration stage, the middle stable hardening stage and the late slow degradation stage. This achieves adaptive modeling of the strength evolution law of the infill body throughout its entire lifespan, avoiding the serious bias that occurs when the fixed receptive field model is used to predict the strength of the infill body in the ultra-long age period, such as 1-5 years.

[0033] 2. By introducing a gated time-series fusion residual mechanism based on physical priors of observed age and monotonic physical constraints, the predicted curve of the final compressive strength is forced to conform to the objective mechanical law of "rapid growth in the early stage - stabilization in the middle stage - stabilization or slow decline in the later stage". This completely eliminates non-physical phenomena such as early strength oscillation and decline and unjustified continuous rise in the later stage that are common in traditional pure data-driven models, making the prediction results more interpretable and reliable in engineering.

[0034] 3. The dual-branch hybrid attention mechanism can dynamically highlight the parameter channel that contributes the most to the compressive strength, while focusing on key periods such as the rapid growth period of compressive strength, the peak inflection point, and the starting point of long-term degradation. This enables the accurate capture of the core pattern of dominant strength evolution even in downhole environments with high data noise and complex disturbances, further improving prediction accuracy. Attached Figure Description

[0035] Figure 1 This is a flowchart of an intelligent monitoring method for the strength of cemented backfill in mines, as described in this invention. Detailed Implementation

[0036] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0038] The following description, in conjunction with the accompanying drawings, details the specific scheme of the intelligent monitoring method for the strength of cemented backfill in mines provided by this invention.

[0039] See attached document Figure 1 The diagram illustrates a flowchart of an intelligent monitoring method for the strength of cemented backfill in mines, provided by an embodiment of the present invention. The method includes the following steps:

[0040] S1. Preprocess the parameters in the data from different data sources to obtain preprocessed parameters; concatenate the preprocessed parameters at each observation age to obtain the fused feature vector for each observation age, and further obtain the standardized input sequence of the entire infill body observation age; based on the standardized input sequence, obtain the final output feature vector through the adaptive physical gated multi-scale causal dilation convolution feature extraction algorithm.

[0041] Laboratory mix design data were obtained from uniaxial compressive strength tests under standard indoor curing conditions. The laboratory mix design data included: slurry concentration, cement content, aggregate gradation parameters, and observation age. Measured uniaxial compressive strength, etc.

[0042] Real-time monitoring data is obtained through sensors such as temperature and humidity sensors at the filling site. The real-time monitoring data at the filling site includes: ambient temperature, relative humidity, pore water pressure, and surrounding rock stress.

[0043] Data from existing historical engineering databases is obtained, including measured records of the entire process of mix proportion, environment, and strength of backfill bodies completed over many years in the same mining area or similar type of mine.

[0044] Laboratory mixing test data, real-time monitoring data at the filling site, and historical engineering database data are used as training datasets; the amount of laboratory mixing test data depends on the scale of the experiment; the amount of real-time monitoring data at the filling site depends on the frequency of sensor acquisition; and the amount of historical engineering database data depends on the number of historical engineering cases.

[0045] Since the number of parameters in the data from different sources is not fixed, let the total number be... This data includes laboratory mixing test data, real-time monitoring data at the filling site, and historical project database data; The number of parameters for the class data is For example, laboratory mix design data includes slurry concentration, cement content ratio, and measured uniaxial compressive strength at different observation ages; real-time monitoring data at the filling site includes surrounding rock stress, ambient temperature, and relative humidity; and historical project database data includes slurry concentration, cement content ratio, uniaxial compressive strength, surrounding rock stress, ambient temperature, and relative humidity at different observation ages in historical projects. Laboratory mix design data:

[0046] Observation age (t) (days) Pulp concentration (mass fraction, %) Cement admixture ratio (mass ratio, %) Aggregate gradation parameters (main particle size range) Measured uniaxial compressive strength (MPa) 7 76 8 0–10 mm 1.4 28 74 12 0–15 mm 3.2 56 72 15 0–20 mm 4.1 90 70 16 0–25 mm 4.6 180 70 16 0–25 mm 4.8

[0047] Among them, the measured uniaxial compressive strength is only used to calculate the loss and evaluate the prediction deviation of the model under known mix ratio conditions, and is not included in the fusion feature vector;

[0048] Real-time monitoring data at the filling site:

[0049] Observation age (t) (days) Ambient temperature (°C) Relative humidity (%) Pore ​​water pressure (MPa) Surrounding rock stress (MPa) 7 24 68 0.8 2.9 28 26 62 1.1 3.3 56 27 58 1.4 3.7 90 28 55 1.6 4.0 180 29 52 1.9 4.3

[0050] Historical engineering database data:

[0051] Observation age (t) (days) Pulp concentration (%) Cement admixture (%) Aggregate gradation (mm) UCS measured (MPa) Temperature (°C) humidity(%) Pore ​​water pressure (MPa) Surrounding rock stress (MPa) 7 77 7 0–12 1.3 23 70 0.7 2.8 28 75 11 0–18 3.0 25 65 1.0 3.2 56 73 14 0–22 3.9 26 60 1.3 3.6 90 71 15 0–25 4.4 27 57 1.5 3.9 180 71 15 0–25 4.7 28 54 1.8 4.2

[0052] To eliminate dimensional differences and ensure data integrity and synchronization, parameters from data sources obtained from different sources underwent standardized preprocessing to obtain preprocessed parameters. Standardized preprocessing included imputing missing values ​​using linear interpolation or weighted averaging of adjacent time points, performing Min-Max normalization, and mapping to... The intervals are aligned and unified through equal-interval time resampling and linear interpolation. The standardized preprocessing methods are well-known techniques to those skilled in the art and will not be elaborated upon here.

[0053] At each observation age The preprocessing parameters are concatenated into a fused feature vector in a fixed order. Further, the standardized input sequence is obtained. ,in, Indicates the length of the observed age of the filling body. An index indicating the observed age of the infill body. Indicates the observed age The fused feature vector.

[0054] Based on the standardized input sequence, the final output feature vector is obtained by processing it through an adaptive physical-gated multi-scale causal dilation convolution feature extraction algorithm.

[0055] The adaptive physics-gated multi-scale causal dilated convolutional feature extraction algorithm performs a global average pooling operation on the entire standardized input sequence to obtain a global average pooling vector representing the average state of all features across the observed age of the entire infill body. This global average pooling vector is fed into three independent fully connected layers. After sigmoid activation, the output is a scalar between 0 and 1. Adding 1 to this scalar yields a dilation scaling factor between 1 and 2, used to dynamically adjust the originally fixed base receptive field size of the three parallel branches. This demonstrates the algorithm's generalization ability to different mines and infill processes. The base receptive field of the short branch corresponds to the early rapid hydration stage, the base receptive field of the medium branch corresponds to the mid-term stable hardening stage, and the base receptive field of the long branch corresponds to the late slow degradation stage.

[0056] Multiplying the dilation scaling factor by the respective baseline convolutional kernel widths of 3, 5, and 7, and then performing standard rounding down, yields the dilation rates of the three branches. This achieves dynamic learnability of the dilation rate, expressed by the formula:

[0057]

[0058] in, Indicates the first The expansion rate of each branch; Indicates the first The baseline kernel width for each branch, the baseline kernel width without dilation, with fixed values: , , This is used to determine the size of the basic receptive field for each branch; Indicates the first The scaling factor for each branch, with a value range of [value range missing]. This is to ensure that the expansion rate is at least the size of the pronucleus and no more than twice the size of the pronucleus; As an activation function, the linear output is compressed to interval; Indicates the first The branch is used to calculate the learnable weight matrix for the inflation scaling factor. Initially, it is sampled from a random Gaussian distribution with a mean of 0 and a standard deviation of 0.1. Small-scale random initialization can avoid drastic changes in the inflation rate in the early stages of training. ， After training, the weights of the short branch are... The range of values ​​is Weights after training of the middle branch The range of values ​​is Weights after long branch training The range of values ​​is ; This represents the global average pooling vector, calculated by applying the average pooling vector to the entire input sequence. The average pooling over the time dimension is used to provide global contextual information throughout the entire observation period of the filling body; Indicates the first One branch is used to calculate the learnable bias of the dilation scaling factor, initializing... In the early stages of training, a reasonable starting point for the scaling factor is provided to ensure stable gradient propagation during training, improve generalization ability on different datasets, and after training... The range of values ​​is ; This indicates rounding down to the nearest integer.

[0059] After obtaining the dilation rates of the three branches, three parallel causal dilation convolution branches are started simultaneously, outputting the feature vector extracted by each branch at each observation age.

[0060] Specifically, the first weight of the convolution kernel is always applied to the fused feature vector of the current observation age. Subsequent weights are applied sequentially to the fused feature vectors corresponding to past observation ages obtained by subtracting an integer multiple of the inflation rate from the current observation age. When the calculated past observation age is less than the initial observation age of the sequence, it is automatically padded with zeros. The formula is expressed as follows:

[0061]

[0062] in, Indicates the first Each branch at the observation age Extracted feature vectors; Used to introduce nonlinearity; Indicates the position index within the convolution kernel; Indicates the first The first branch Each convolutional kernel weight initially follows a normal distribution with a mean of 0 and a variance of 0.02. After training, short branches... The range of values ​​is After training, the middle branch The range of values ​​is Long branch after training The range of values ​​is ; This indicates the actual sampling time point. If it is less than 1, it will be automatically filled with zeros. Indicates age fused feature vectors; Indicates the first The kernel bias of each branch is initialized. To avoid bias in output during the early stages of training, short branches should be used after training. The range of values ​​is After training, the middle branch The range of values ​​is Long branch after training The range of values ​​is .

[0063] The feature vectors extracted from the three branches are concatenated along the channel dimension to form a multi-scale convolutional concatenated feature vector. Multi-scale convolutional concatenation of feature vectors possesses both sensitivity to early rapid growth and the ability to capture later slow degradation.

[0064] A gated temporal fusion residual mechanism based on the physical prior of the observed infill age is introduced. The core of this mechanism is a lightweight multilayer perceptron, whose inputs are the current observed age of the infill, the square of the current observed age of the infill, and an annual periodic sine term. This is used to capture the subtle influence of possible seasonal temperature and humidity fluctuations on long-term compressive strength, and the output is a scalar gating coefficient between 0 and 1. This gated temporal fusion residual mechanism ensures, at the mechanism level, that the final predicted compressive strength curve conforms to the objective mechanical law of cemented infill strength: "rapid growth in the early stage—stable in the middle stage—stable or slow decline in the later stage." The double-bowl formula for calculating the gating coefficient is as follows:

[0065]

[0066] in, Indicates the observed age The corresponding gating coefficients are used to determine the concatenated feature vectors from the multi-scale convolution. With fused feature vectors The fusion ratio, with a value range of [value range missing]. ; This represents the Sigmoid activation function, which outputs a scalar gating coefficient between 0 and 1. Indicates the current observed age of the filling body; This represents the quadratic term, which is the square of the current observed age of the infill body, and is used to capture the nonlinear trend of the gradually slowing rate of intensity growth. It represents the annual cycle sine term, used to capture the subtle effects of seasonal temperature and humidity on later compressive strength; This represents the gated fully connected weight matrix, initially... Observation age gating coefficient The impact is quite significant because the strength evolution of the filling material is initially time-driven, especially in the early stages of rapid strength growth. Using these initial values ​​helps capture the impact of time progression on intensity. After training, the value range is... , Reflects Due to the influence of [the above factors], the strength increase of the filling tends to slow down over time. The impact on the gating coefficient should be negative and small. The negative weight design helps to accurately simulate the gradual slowdown of the intensity growth rate in the later stages. After training, the value range is... , This reflects the slight impact of annual seasonal variations on the strength of the infill. Although the impact of seasonal fluctuations on strength is small, it remains a factor that cannot be ignored in long-term monitoring. This ensures that the impact of seasonal fluctuations is not overly ignored; after training, the value range is... ; Indicates gated bias, initialization This ensures that the initial training process is smooth and without excessive bias. After training, the value range... .

[0067] Based on the physical prior of the observed age of the infill body, a dynamic weighted fusion of multi-scale convolutional concatenated feature vectors and fused feature vectors is achieved, expressed by the formula:

[0068]

[0069] in, Indicates the observed age The final output feature vector; Indicates the observed age Multi-scale convolutional concatenation of feature vectors; This represents the gated fully connected layer weight matrix. Initially, it is randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.02 to provide a stable starting point for training and avoid gradient vanishing or exploding. After training, its values ​​range from [value missing]. .

[0070] S2. Based on the final output feature vector, a dual-branch hybrid attention mechanism is introduced to obtain an attention-enhanced feature vector; based on the attention-enhanced feature vector, a two-layer fully connected neural network combined with residual connections is used for mapping to obtain the final compressive strength prediction value; based on the final compressive strength prediction value, a safety early warning mechanism is set to obtain the monitoring results.

[0071] A dual-branch hybrid attention mechanism is introduced to process the final output feature vector, resulting in an attention-enhanced feature vector.

[0072] The dual-branch hybrid attention mechanism consists of two sequentially executed sub-processes: first, the channel attention mechanism is executed, and then the standard Transformer attention mechanism is executed.

[0073] The goal of the channel attention mechanism is to dynamically assign different weights based on the importance of each channel. This is achieved through a linear transformation process, where the final output feature vector is input into a trainable fully connected layer and then... Normalization is performed so that the sum of the weights of all channels is 1. This highlights key channels and suppresses secondary information. The formula is expressed as:

[0074]

[0075] in, Indicates the observation age The attention weight vector for each feature channel; Indicates will The linear mapping result is normalized to a probability distribution; The learnable linear mapping weight matrix representing the channel attention mechanism is initially randomly sampled from a normal distribution with a mean of 0 and a standard deviation of 0.01 to provide a stable starting point for training and avoid gradient vanishing or exploding phenomena. After training, its values ​​range from [value missing]. ; The bias vector representing the channel attention mechanism, initially, This is used to avoid initial bias in neuron activation; after training, its value range is [value range missing]. .

[0076] The final output feature vector is weighted using a channel attention mechanism to highlight important channels; further processing is then performed using the standard Transformer attention mechanism to obtain a more expressive attention-enhanced feature vector, calculated as follows:

[0077]

[0078] in, Indicates the observation age Attention-enhanced feature vectors; This indicates element-wise multiplication; This represents the standard Transformer attention weight matrix, used to indicate the correlation between each observation age and all historical observation ages (including itself), obtained according to the standard Transformer attention mechanism.

[0079] Based on attention-enhanced feature vectors, a two-layer fully connected neural network combined with residual connections is used for mapping to obtain the final predicted compressive strength value. A safety early warning mechanism is set up so that after each prediction, it is immediately determined whether the predicted final compressive strength value for the current observation age is lower than the engineering safety lower limit, which is determined according to mine design specifications or industry standards. If the predicted final compressive strength value for the current observation age is lower than the engineering safety lower limit, the safety early warning mechanism is triggered, immediately activating an audible and visual alarm and pushing the alert to the central control room and underground terminals, achieving early warning of insufficient compressive strength.

[0080] The mean square error is used as the basic loss function. The mean square error is obtained by averaging the squared difference between the measured uniaxial compressive strength at the observed age and the predicted final compressive strength.

[0081] The gradients of all trainable parameters with respect to the mean squared error are calculated via backpropagation, then multiplied by a learning rate determined through expert experience, and all trainable parameters are updated synchronously. Trainable parameters include: the learnable weight matrix. Learnable bias Convolution kernel weights Kernel bias Gated fully connected weight matrix Gating bias The learnable linear mapping weight matrix of the channel attention mechanism Bias vector of channel attention mechanism Gated fully connected weight matrix wait.

[0082] By synchronously updating all trainable parameters using gradients and learning rates, the aim is to achieve adaptive adjustment through backpropagation. The specific steps are as follows:

[0083] Forward Propagation: Perform a forward propagation calculation, utilizing the fused feature vectors. Based on the currently trainable parameters, the final predicted compressive strength value is calculated.

[0084] Loss calculation: Based on the difference between the predicted final compressive strength and the measured uniaxial compressive strength, a loss function is calculated. A commonly used loss function is the mean squared error, i.e.:

[0085]

[0086] in, Represents the loss function; It is the observation age The final predicted compressive strength value; It is the observation age Measured uniaxial compressive strength; This represents the length of the observation period for the infill material, i.e., the total observation time. The loss function measures the deviation between the predicted final compressive strength and the measured uniaxial compressive strength.

[0087] Example of prediction results:

[0088] Age (t) (days) Actual measured UCS (MPa) Predict UCS (MPa) Relative error (%) 7 1.4 1.37 2.1 28 3.2 3.14 1.9 56 4.1 4.05 1.2 90 4.6 4.57 0.7 180 4.8 4.76 0.8

[0089] Backpropagation calculates gradients: The backpropagation algorithm is used to calculate the gradient of each trainable parameter with respect to the loss function. Gradient calculation is implemented using the chain rule, where the gradient value of each layer is passed layer by layer until it reaches the input layer. The gradient reflects the degree of influence of each trainable parameter on the deviation between the final predicted compressive strength and the measured uniaxial compressive strength.

[0090] Gradient update: Uses the calculated gradient information, combined with a preset learning rate. This updates each trainable parameter. The update formula is:

[0091]

[0092] in, Indicates the first Trainable parameters, such as or wait; Indicates the index of the trainable parameters; It is the gradient of the loss function with respect to the trainable parameters; It refers to the learning rate. The learning rate controls the step size of the trainable parameters adjustment, that is, it determines the magnitude of each trainable parameter update.

[0093] Synchronous Updates: All trainable parameters, including convolutional kernel weights and attention mechanism weights, are adjusted synchronously during each backpropagation and update process. Each trainable parameter is updated based on gradient information, thereby gradually reducing prediction error and optimizing network performance.

[0094] Adaptive adjustment effect: Through the above update process, each trainable parameter will be adaptively adjusted according to its influence on the deviation between the final predicted compressive strength value and the measured uniaxial compressive strength. The adaptive adjustment mechanism can gradually optimize the model's predictive ability during the training process.

[0095] Iterative training: Through multiple rounds of iterative training, each round including forward propagation, loss calculation, backpropagation, and parameter update, the model is gradually optimized until the loss converges. The convergence condition is that the absolute value of the difference between the loss value of the current iteration and the loss value of the previous iteration is less than a threshold set according to expert experience. .

[0096] Through the above steps, gradient updates and learning rates work together to achieve synchronous updates of all trainable parameters, ultimately achieving an adaptive adjustment effect. This enables the model to generalize well under different mines and filling processes and accurately predict compressive strength.

[0097] To prevent violations of the true mechanical laws of cemented infill bodies, an additional monotonic physical constraint is introduced: Before the observation age corresponding to the peak of the predicted final compressive strength, the predicted final compressive strength for adjacent observation ages is forced to be non-decreasing, reflecting the physical process of rapid early strength growth caused by cement hydration; after the observation age corresponding to the peak of the predicted final compressive strength, the predicted final compressive strength for adjacent observation ages is forced to be non-increasing, characterizing the stable or slow degradation of strength caused by long-term loading, drying shrinkage, and microcrack propagation. The monotonic physical constraint is added to the total loss in the form of a soft constraint to ensure that, under no circumstances will there be an unphysical assumption of early strength decline or unjustified later strength increase.

[0098] In summary, a method for intelligent monitoring of the strength of cemented backfill in mines has been developed.

[0099] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0100] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0101] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for intelligent monitoring of the strength of cemented backfill in mines, characterized in that, Includes the following steps: S1. Parameters from data sources including laboratory mix design test data, real-time monitoring data from the backfilling site, and historical project database data are preprocessed to obtain preprocessed parameters. Laboratory mix design test data includes slurry concentration, cement content ratio, and measured uniaxial compressive strength at different observation ages. Real-time monitoring data from the backfilling site includes surrounding rock stress, ambient temperature, and relative humidity. Historical project database data includes slurry concentration, cement content ratio, uniaxial compressive strength, surrounding rock stress, ambient temperature, and relative humidity at different observation ages in historical projects. The preprocessed parameters are then spliced ​​together at each observation age. The fused feature vector for each observation age is obtained, and the standardized input sequence of the entire infill body's observation ages is further obtained. Using an adaptive physical-gated multi-scale causal dilated convolutional feature extraction algorithm, global average pooling is performed on the standardized input sequence to obtain a global average pooling vector representing the entire infill body's observation ages. This global average pooling vector is fed into three independent fully connected layers, activated with Sigmoid, and then incremented by 1 to obtain the dilation scaling factor. The dilation scaling factor is multiplied by the respective baseline convolutional kernel widths of 3, 5, and 7, and then rounded down to the nearest integer to obtain the dilation rate of the three branches. The calculation formula is as follows: ； in, Indicates the first The expansion rate of each branch; Indicates the first The baseline kernel width of each branch is fixed. , , ; It is an activation function; Indicates the first Each branch is used to compute the learnable weight matrix of the inflation scaling factor; Represents the global average pooling vector; Indicates the first One branch is used to calculate the learnable bias of the dilation scaling factor; This indicates rounding down to the nearest integer. Based on the dilation rate and the fused feature vector, convolution kernel weights are introduced. The first weight of the convolution kernel is always applied to the fused feature vector of the current observation age. Subsequently, the weights are applied sequentially to the fused feature vectors corresponding to past observation ages obtained by subtracting an integer multiple of the dilation rate from the current observation age. When the calculated past observation age is less than the initial observation age of the sequence, it is automatically padded with zeros to obtain the feature vector extracted by each branch at each observation age. The formula is expressed as: ； in, This represents the feature vector extracted by the nth branch at the observed age 𝑡; Indicates the position index within the convolution kernel; Indicates the nth branch. Each convolutional kernel weight; Indicates the actual sampling time point; Indicates age fused feature vectors; This represents the kernel bias of the nth branch; the feature vectors extracted from the three branches are concatenated along the channel dimension to obtain a multi-scale convolutional concatenated feature vector; Based on the observed age, a gated time-series fusion residual mechanism is introduced. The inputs are the current observed age of the filling body, the square of the current observed age of the filling body, and the annual periodic sine term. The output is a scalar gating coefficient between 0 and 1, as shown in the formula: ； in, Indicates the observation age The corresponding gating coefficient; This represents the Sigmoid activation function; Indicates the current observed age of the filling body; This represents the square of the current observed age of the filling body; Represents the sine term of the annual cycle; Represents the weight matrix of the gated fully connected layer; Indicates gating bias; Based on the gating coefficient, the fused feature vector, and the multi-scale convolutional concatenation of the feature vector, a weighted fusion is performed to obtain the final output feature vector. The calculation formula is as follows: ； in, Indicates the observation age The final output feature vector; Indicates the observation age Multi-scale convolutional concatenation of feature vectors; Represents the weight matrix of the gated fully connected layer; Indicates the observation age fused feature vectors; S2. Based on the final output feature vector, a dual-branch hybrid attention mechanism is introduced to obtain an attention-enhanced feature vector; based on the attention-enhanced feature vector, a two-layer fully connected neural network combined with residual connections is used for mapping to obtain the final compressive strength prediction value; based on the final compressive strength prediction value, a safety early warning mechanism is set to obtain the monitoring results.

2. The intelligent monitoring method for the strength of cemented backfill in mines according to claim 1, characterized in that, S2 specifically includes: In the dual-branch hybrid attention mechanism, the channel attention mechanism is executed first, and the final output feature vector is input into the trainable fully connected layer. Softmax is used for normalization so that the sum of the weights of all channels is 1. The attention weight vector of each feature channel for each observation age is then calculated.

3. The intelligent monitoring method for the strength of cemented backfill in mines according to claim 2, characterized in that, S2 specifically includes: In the dual-branch hybrid attention mechanism, based on the attention weight vector, the standard Transformer attention mechanism is applied to the final output feature vector to obtain the attention-enhanced feature vector. The calculation formula is as follows: ； in, Indicates the observation age Attention-enhanced feature vectors; This indicates element-wise multiplication; This represents the standard Transformer attention weight matrix; Indicates the observation age The final output feature vector; Indicates the observation age The attention weight vector for each feature channel.

4. The intelligent monitoring method for the strength of cemented backfill in mines according to claim 1, characterized in that, S2 specifically includes: A safety early warning mechanism is set up so that when the predicted final compressive strength at the current observation age is lower than the set engineering safety lower limit, the safety early warning mechanism is triggered.

5. The intelligent monitoring method for the strength of cemented backfill in mines according to claim 4, characterized in that, S2 specifically includes: The mean squared error is obtained by averaging the squared differences between the measured uniaxial compressive strength at the observed age and the predicted compressive strength in the preprocessed parameters. The gradient of all trainable parameters with respect to the mean squared error is calculated by backpropagation, and then multiplied by the learning rate. All trainable parameters in the adaptive physical-gated multi-scale causal dilatational convolution feature extraction algorithm and the dual-branch hybrid attention mechanism are updated synchronously.

Images