A device management and control method for mine safety

Abstract

The present application relates to the technical field of mine safety, and more particularly to a device management and control method for mine safety. The method comprises: collecting environmental parameters of a plurality of monitoring points in a mine and device parameters of ventilation equipment, and constructing an environmental vector and a device vector respectively; determining the stability of the environmental parameters of each monitoring point when the ventilation equipment is normal based on the variation amplitude of the above vectors within a preset time period; constructing and training an autoencoder model, and updating the model parameters using an adaptive learning rate; splicing the environmental vector to obtain an overall environmental vector and inputting the overall environmental vector into the model to obtain a reconstructed overall environmental vector and a hidden layer feature vector, and calculating an anomaly score based on the reconstruction loss and the deviation of the feature vector from a preset standard feature vector; and when the anomaly score triggers a preset management and control condition, performing a management and control operation on the ventilation equipment. The present application mines the correlation of multi-dimensional parameters, strengthens the learning of key area features through differentiated learning, and improves the early fault recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of mine safety technology, and in particular to a method for equipment control in mine safety. Background Technology

[0002] In mine working spaces, toxic gases such as methane and carbon monoxide can easily accumulate, threatening the lives of underground workers. Therefore, mines are generally equipped with ventilation systems that dilute harmful gases by supplying fresh air; these systems are crucial for ensuring operational safety. If ventilation equipment malfunctions, the concentration of harmful gases may surge. Therefore, real-time monitoring of the underground environment, accurate assessment of the ventilation equipment's status, and timely control are core requirements for mine safety management.

[0003] Current monitoring of the underground working environment relies heavily on threshold values ​​for environmental parameters such as temperature, humidity, methane concentration, and coal dust concentration. When a single parameter exceeds a preset threshold, an alarm or equipment control is triggered. However, this approach only focuses on instantaneous parameter values ​​and lacks in-depth analysis of the patterns of environmental parameter changes caused by equipment malfunctions. It struggles to distinguish between subtle changes caused by equipment anomalies and normal fluctuations, resulting in high false alarm and false negative rates.

[0004] Related technologies introduce autoencoder models, which learn the distribution of environmental parameters under normal operating conditions and use reconstruction errors to identify anomalies. However, these autoencoder models employ a uniform optimization strategy, failing to consider the differences in environmental stability across different monitoring areas: parameters in the core downhole region fluctuate significantly, requiring the autoencoder model to focus its learning; while auxiliary downhole regions are relatively stable. The uniform optimization strategy leads to insufficient learning in key areas or overfitting in stable areas, affecting the accuracy of early identification of ventilation equipment faults. Summary of the Invention

[0005] To address the technical problem of insufficient early identification accuracy of ventilation equipment faults caused by the aforementioned autoencoder model's use of a uniform optimization strategy and failure to consider the differences in environmental stability across different monitoring areas, this invention provides an equipment management method for mine safety, comprising the following steps:

[0006] Environmental parameters from several monitoring points within the mine are collected to construct environmental vectors for these monitoring points, and equipment parameters from ventilation equipment at these monitoring points are also collected to construct equipment vectors for these monitoring points. Under normal operating conditions of the ventilation equipment, the stability of the environmental parameters at each monitoring point is determined based on the variation amplitude of the environmental and equipment vectors within a preset time period. An autoencoder model containing at least one hidden layer is constructed and trained, and the autoencoder model updates its parameters using an adaptive learning rate. The adaptive learning rate is determined such that the learning rate allocated to the model parameters corresponding to each monitoring point is negatively correlated with the stability of its environmental parameters. Taking any collection time as the target time, the environmental vectors from several monitoring points at the target time are concatenated to obtain an overall environmental vector. This overall environmental vector is then input into the trained autoencoder model to obtain a reconstructed overall environmental vector and a feature vector output by the hidden layer of the autoencoder model. Based on the reconstruction loss between the overall environmental vector and the reconstructed overall environmental vector, and the deviation between the feature vector output by the hidden layer and a preset standard feature vector, an anomaly score for the mine's environmental state at the target time is calculated. When the anomaly score triggers a preset control condition, control operations are performed on the ventilation equipment.

[0007] This invention constructs vectors by collecting environmental and equipment parameters, and combines this with the stability of environmental parameters at monitoring points to achieve adaptive learning rate training of the autoencoder model. Finally, it calculates anomaly scores based on reconstruction loss and feature bias to manage mine ventilation equipment. Its effects are: first, it overcomes the limitations of threshold judgment, utilizing the autoencoder to mine the nonlinear correlation of multi-dimensional environmental parameters, effectively capturing subtle collaborative changes caused by equipment failures, and significantly reducing false alarm and false negative rates; second, it achieves differentiated learning based on stability, allocating a larger learning rate to monitoring points with large environmental fluctuations and strong correlation with equipment, strengthening feature learning in key areas, and improving the accuracy of early fault identification; third, it comprehensively calculates anomaly scores based on reconstruction loss and feature bias, assessing environmental conditions from multiple dimensions, making anomaly judgment more comprehensive, providing accurate basis for timely management of ventilation equipment, and effectively ensuring mine operation safety.

[0008] Preferably, the adaptive learning rate satisfies the following relationship: ;in, It is the first autoencoder model In the first round of training, The learning rate of the model parameters corresponding to each monitoring point; It is the preset base learning rate; It is the first The stability of environmental parameters at each monitoring point; This is the preset number of training rounds; It represents the current round number during the training process.

[0009] This invention combines the stability of environmental parameters at monitoring points with dynamic adjustment of the learning rate based on the training rounds. This ensures that a larger learning rate is allocated to monitoring points with low stability to focus on learning their complex features, while also achieving rapid convergence in the early stages of training and fine-tuning in the later stages. This avoids overfitting and further improves the training efficiency of the autoencoder model and the accuracy of learning mine environmental features.

[0010] Preferably, the variation range of the environmental vector and device vector of the plurality of monitoring points within a preset time period is determined by calculating the Euclidean distance between the environmental vector or the device vector within the preset time period.

[0011] This invention quantifies the variation of environmental and equipment vectors within a preset time period using Euclidean distance. This objectively reflects the fluctuation of environmental and equipment parameters, providing a precise quantitative basis for subsequent calculations of the stability of environmental parameters at monitoring points. It ensures that the stability index truly reflects the differences in environmental characteristics at different monitoring points.

[0012] Preferably, the stability of the environmental parameters satisfies the following relationship: ;in, It is the first The stability of environmental parameters at each monitoring point; Within a preset time period, the first The maximum Euclidean distance between the environmental vectors of any two monitoring points at any two times; The calculation method is as follows: for the first For each ventilation device at each monitoring point, calculate the maximum Euclidean distance of its device vector within the preset time period, and average the calculated maximum Euclidean distances. It is the standard normalized function.

[0013] This invention integrates the inherent volatility of environmental parameters and their correlation with ventilation equipment, and constrains the results within a reasonable range through normalization. This allows the stability of environmental parameters to comprehensively and quantitatively characterize the stability of environmental parameters at monitoring points, providing a reliable basis for the adaptive learning rate adjustment of the autoencoder model and ensuring the model's targeted learning of key areas.

[0014] Preferably, the autoencoder model includes an input layer, at least one hidden layer, and an output layer; wherein the input layer is used to receive the overall environment vector, and the output layer is used to output the reconstructed overall environment vector.

[0015] Preferably, the reconstruction loss of the overall environment vector and the reconstructed overall environment vector is determined by the Euclidean distance between the overall environment vector and the reconstructed overall environment vector; the deviation between the feature vector output by the hidden layer and the preset standard feature vector is determined by the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector.

[0016] Preferably, the anomaly score of the mine environment state at the target time. Satisfying the relation: ;in, This is the standard loss value under normal mining conditions; It is the reconstruction loss between the overall environment vector at the target time and the reconstructed overall environment vector; It is the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector; yes Type activation function, It is a preset microvalue used to prevent the denominator from being 0.

[0017] The sigmoid function of this invention maps the relative deviation of the reconstruction loss to a reasonable range, and combines the negative correlation normalization result of cosine similarity to integrate the abnormal information of the two dimensions, so that the output abnormal score can accurately and quantitatively reflect the degree of abnormality of the mine environment, providing a clear judgment basis for the control and operation of ventilation equipment.

[0018] Preferably, the standard feature vector is the mean of the feature vectors output by the hidden layer in the final stage of training the autoencoder model, after a preset training epoch.

[0019] Preferably, the control operation on the ventilation equipment specifically includes: calculating the abnormal scores of the mine environment at multiple collection times under normal operating conditions, and calculating their average value, which is recorded as the standard abnormal score; when the abnormal score of the mine environment at the target time is greater than a preset multiple of the standard abnormal score, it is determined that the ventilation equipment has an abnormal operating state and is included in the control state.

[0020] Preferably, the environmental parameters include at least one of temperature, humidity, gas concentration, and coal dust concentration; the equipment parameters include at least one of motor current and wind pressure difference of the ventilation equipment.

[0021] The beneficial effects of this invention are as follows: In autoencoder training, the stability of each monitoring point is used as the learning rate adjustment factor. A larger learning rate and increased resampling are applied to unstable areas with large environmental fluctuations and strong correlation with equipment, focusing on learning their complex features. For stable areas, the learning rate is reduced to avoid overtraining on easily fitted features, achieving differentiated learning, strengthening the capture of key area features, and significantly improving the model's accuracy in identifying early faults. By measuring the mine environment state through anomaly scores, and comprehensively considering reconstruction loss and feature bias, the environmental state is evaluated from multiple dimensions, greatly enhancing the ability to distinguish between minor and overall anomalies, enabling earlier detection of abnormal environmental changes. Simultaneously, it breaks through the limitations of traditional threshold judgment, utilizing the autoencoder to mine the nonlinear correlation of multi-dimensional parameters, effectively capturing subtle collaborative changes caused by equipment faults, significantly reducing false alarm and false negative rates, providing accurate basis for timely control of ventilation equipment, and effectively ensuring mine operation safety. Attached Figure Description

[0022] Figure 1 A flowchart of an equipment management method for mine safety provided in an embodiment of the present invention. Detailed Implementation

[0023] This invention provides a method for equipment management and control in mine safety, such as... Figure 1 As shown, the method includes steps S100-S500:

[0024] Step S100: Collect environmental parameters from several monitoring points in the mine to construct environmental vectors for several monitoring points, and collect equipment parameters from ventilation equipment at several monitoring points to construct equipment vectors for several monitoring points.

[0025] It should be noted that this step is the data foundation for the entire equipment management method. Obtaining comprehensive and accurate downhole environment and equipment operation data is a prerequisite for subsequent condition assessment and adaptive training of the autoencoder model, and the quality of this data directly determines the effectiveness of the final management decisions.

[0026] Specifically, various types of sensors are deployed at key locations in the mine, such as working faces, transport roadways, and pump rooms. These sensors are used to collect two types of core data in real time: one is environmental parameters, such as temperature, humidity, gas concentration, and coal dust concentration; the other is equipment operating parameters directly related to environmental conditions, especially the motor current and air pressure difference of ventilation equipment.

[0027] To facilitate subsequent mathematical modeling and calculations, the collected raw data needs to be normalized, and the normalized values ​​are then constructed into a vector. As an example, this embodiment uses the min-max normalization method to map all collected parameter values ​​to... Within the interval. The min-max normalization method is existing technology and will not be elaborated upon here.

[0028] Regarding vector construction, the locations of the deployed sensors are denoted as monitoring points, and the time of any data acquisition is taken as the target time. Then, the... The target time of the first The environmental vector of each monitoring point can be denoted as: ,in, , , , The first The target time of the first The normalized values ​​of temperature, humidity, gas concentration, and coal dust concentration at each monitoring point. The first time of the target time Of the monitoring points, the first The equipment vector of each ventilation device is denoted as: ,in, , The first The target time of the first The first monitoring point The normalized values ​​of the motor current and the normalized values ​​of the air pressure difference of each ventilation device.

[0029] It should be noted that the ventilation equipment at the monitoring point here refers to all ventilation equipment that can directly affect the monitoring point, and the ventilation equipment at the monitoring points below also has this meaning.

[0030] It should be noted that the above environmental vector and equipment vector are only exemplary expressions. Implementers can add or remove feature dimensions in the vectors according to the actual situation of the mine and the management and control requirements.

[0031] At this point, data acquisition and vector construction were completed, and environmental and equipment vectors for multiple times and several monitoring points within the mine were obtained.

[0032] Step S200: Under normal operating conditions of the ventilation equipment, determine the stability of the environmental parameters at each monitoring point based on the variation range of the environmental vector and equipment vector at several monitoring points within a preset time period.

[0033] It should be noted that this step is the core technology for adaptive training, used to quantitatively assess the environmental parameter variation characteristics of each monitoring point in the mine. Because different monitoring points have different structural characteristics, uses, and relationships with ventilation equipment, their environmental parameters also differ in their sensitivity to equipment status changes and their own fluctuation characteristics. If data from all locations are treated equally during autoencoder model training, it may mask key, subtle changes caused by equipment failures. Therefore, this step establishes a stability level, providing a basis for the differentiated learning of the subsequent autoencoder model.

[0034] Specifically, first, a target time period needs to be determined, that is, a preset time period, for example, selecting any one hour when the ventilation equipment is in normal operating condition as the analysis window. Environmental vectors from all monitoring points within this time period are then collected. and the equipment vector of all ventilation equipment .

[0035] Secondly, for each monitoring point The fluctuation of environmental parameters within the target time period is calculated. This can be achieved by calculating the maximum Euclidean distance between environmental vectors at that location within the target time period.

[0036] Then, for each monitoring point The correlation between the monitoring point and changes in related ventilation equipment is calculated. This can be achieved by calculating the maximum Euclidean distance between the equipment vectors of the ventilation equipment at that monitoring point within the target time period and then taking the average of these maximum Euclidean distances.

[0037] Based on the above logic, the first Stability of environmental parameters at each monitoring point Satisfying the relation:

[0038] ;

[0039] in, It is the first The stability of environmental parameters at each monitoring point; Within the target time period, the first The maximum Euclidean distance between the environmental vectors of any two monitoring points at any two times; The calculation method is as follows: for the first For each ventilation device at each monitoring point, calculate the maximum Euclidean distance of its device vector within a preset time period, and average the calculated maximum Euclidean distances. It is a standard normalization function used to quantize calculation results to... For intervals, specific methods such as min-max normalization and Z-score standardization can be used. Normalization methods and Euclidean distance calculations are existing technologies and will not be elaborated upon here.

[0040] In this formula, The larger the value, the higher the value of the ventilation equipment during normal operation. The greater the change in environmental parameters at each monitoring point, the more likely the [number]th [point] will change. The less stable the environmental parameters at each monitoring point; The smaller the value, the lower the value when the ventilation equipment is operating normally. The smaller the change in environmental parameters at each monitoring point, the better. The greater the stability of the environmental parameters at each monitoring point. The larger the value, the more likely it is to be the first. The environmental parameters at each monitoring point are more likely to change due to variations in ventilation equipment, therefore The larger, the more The less stable the environmental parameters at each monitoring point; The smaller the value, the better. The less likely the environmental parameters at a monitoring point are to change due to variations in ventilation equipment, therefore... The smaller, the first The greater the stability of the environmental parameters at each monitoring point.

[0041] Thus, the stability of the environmental parameter variation characteristics of each monitoring point was obtained.

[0042] Step S300: Construct and train an autoencoder model containing at least one hidden layer. The autoencoder model updates its parameters using an adaptive learning rate. The adaptive learning rate is determined such that the learning rate allocated to the model parameters corresponding to each monitoring point is negatively correlated with the stability of its environmental parameters.

[0043] It should be noted that there is a complex and nonlinear cooperative relationship between environmental parameters at various locations in the mine and the operating status of all ventilation equipment. Using an autoencoder model, an unsupervised learning model, can effectively learn and capture the inherent distribution characteristics of such high-dimensional data under normal operating conditions. However, traditional autoencoder models treat all input features equally during training, which is insufficient for mine safety monitoring scenarios. The core innovation of this step lies in using the stability calculated in step S200... By introducing a training process and dynamically adjusting the learning rate of the model parameters, differentiated attention can be paid to the regional characteristics corresponding to different monitoring points.

[0044] Specifically, an autoencoder model is first constructed. Structurally, the autoencoder model includes an input layer, at least one hidden layer, and an output layer. Functionally, it consists of an encoder and a decoder. The input layer receives the overall environment vector, and the output layer outputs the reconstructed overall environment vector. The mapping from the input layer to the hidden layer constitutes the encoder, and the mapping from the hidden layer to the output layer constitutes the decoder. The number of hidden layers can be determined by the implementers based on the actual situation.

[0045] It should be noted that the overall environment vector is obtained by splicing the environment vectors of several monitoring points at the target time.

[0046] In a preferred embodiment, both the encoder and decoder can be composed of multiple stacked fully connected neural network layers. The input layer dimension of the encoder is the same as the dimension of the overall environment vector. After data input, the network dimension is progressively reduced through at least one hidden layer, ultimately outputting a hidden layer feature vector with the same dimension as the overall device vector. For ease of representation, this hidden layer feature vector is denoted as the hidden layer output feature vector. Non-linear activation functions such as ReLU can be used between layers to enhance the expressive power of the autoencoder model. The input to the decoder is the feature vector output from the hidden layers. Through at least one hidden layer, the network dimension is progressively increased, ultimately restoring the output layer dimension to the same dimension as the original overall environment vector, thus completing the reconstruction of the original input. The detailed construction of the encoder and decoder is prior art and will not be elaborated upon here.

[0047] During training, this invention employs an adaptive strategy for adjusting the learning rate. The autoencoder model's first... In the first round of training, The learning rate of the model parameters corresponding to each monitoring point Satisfying the relation:

[0048] ;

[0049] in, It is the first autoencoder model In the first round of training, The learning rate of the model parameters corresponding to each monitoring point; It is the preset base learning rate, usually set to... In this embodiment The implementers can choose according to the actual situation. The value; It is the first The stability of environmental parameters at each monitoring point; This is the preset number of training rounds, which is also the condition for ending training. In this embodiment... The implementers can choose according to the actual situation. The value; It represents the current round number during the training process.

[0050] In this formula, The smaller the value, the better. Environmental parameters at each monitoring point can fluctuate significantly under normal conditions, potentially exhibiting complex changes. These should be closely monitored. For each monitoring point, the environmental parameters correspond to the autoencoder model parameters, and a larger learning rate is used. The larger the value, the more likely it is to be the first. The environmental parameters of each monitoring point change steadily under normal conditions. The features of these parameters are easy to fit, so a smaller learning rate should be used to maintain training stability. The larger the value, the more the autoencoder model is in the initial stage of training, and the higher the learning rate should be for parameter training to quickly learn data features. The smaller the value, the closer the autoencoder model is to the termination stage of training. Therefore, the learning rate for parameter training should be kept as small as possible to achieve fine-tuning of the autoencoder model and avoid overfitting.

[0051] It should be added that there are various strategies in the existing technology regarding the conditions for ending training, which will not be elaborated here. Implementers can adjust the ending conditions according to the actual situation. The value of .

[0052] Thus, an autoencoder model containing at least one hidden layer was constructed and trained, and an adaptive learning rate for updating the model parameters was obtained.

[0053] Step S400: Taking any acquisition time as the target time, splice the environmental vectors of several monitoring points at the target time to obtain the overall environmental vector; and input the overall environmental vector into the trained autoencoder model to obtain the reconstructed overall environmental vector and the feature vector output by the hidden layer of the autoencoder model; based on the reconstruction loss between the overall environmental vector and the reconstructed overall environmental vector and the deviation between the feature vector output by the hidden layer and the preset standard feature vector, calculate the anomaly score of the mine environment state at the target time.

[0054] It should be noted that during the training process of the autoencoder model, it has already learned the characteristics of the overall environment vector under normal conditions. When the overall environment vector obtained under normal conditions is input into the autoencoder model, the output reconstructed overall environment vector should be very close to the input overall environment vector. When the environmental parameters of each monitoring point change abnormally, when the overall environment vector is input into the autoencoder model, the output reconstructed overall environment vector will differ significantly from the input overall environment vector. At the same time, the feature representation in its hidden layer, i.e., the encoding, will also deviate significantly from the encoding under normal conditions. This step aims to integrate the degree of deviation in these two dimensions to construct an anomaly score that can reflect the degree of anomaly in the mine environment.

[0055] Specifically, firstly, after the autoencoder model has been trained, a baseline for the normal state needs to be established. As a preferred implementation, the average loss value from the last 30 rounds of training can be used as the standard loss value. The average of the hidden layer vectors output from these 30 rounds is taken as the preset standard feature vector.

[0056] Then, the overall environment vector at the target time is input into the trained autoencoder model, yielding two key outputs: one is the reconstruction loss, calculated by the Euclidean distance between the overall environment vector and the reconstructed overall environment vector; the other is the feature bias, calculated by the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector.

[0057] Based on the above logic, the anomaly score of the mine environment at the target time satisfies the following relationship:

[0058] ;

[0059] in, It is the anomaly score of the mine environment at the target time. This is the standard loss value under normal mining conditions; It is the reconstruction loss between the overall environment vector at the target time and the reconstructed overall environment vector; It is the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector; It is standard Type activation function, It is a preset micro value used to prevent the denominator from being 0. It can be set to 0.01, and implementers can set it according to their needs.

[0060] In this formula, This represents the difference between the loss value obtained during the monitoring process and the standard loss value. Since the autoencoder model is trained using environmental parameters under normal conditions, therefore... The larger the value, the more the environmental parameters corresponding to the overall input environmental vector deviate from the normal environmental parameters, and the larger the anomaly score of the mine environment. The smaller the value, the more closely the environmental parameters corresponding to the overall input environmental vector fit the normal environmental parameters, and the smaller the anomaly score of the mine environment.

[0061] Since the autoencoder model has the function of feature extraction, the feature vector output by the hidden layer represents the features of the mine environment after feature extraction. The larger the value, the closer the characteristics of the mine environment at the target time are to those of the mine environment under normal conditions. This means that the mine environment at the target time is more likely to be under normal conditions, and the smaller the abnormality score of the mine environment at the target time. The smaller the value, the greater the difference between the mine environment characteristics at the target time and those under normal conditions. This indicates a higher likelihood that the mine environment at the target time is in an abnormal state, resulting in a larger anomaly score. In order to Normalization yields an indicator that is positively correlated with the degree of anomaly.

[0062] At this point, the anomaly score of the mine's environmental condition was obtained.

[0063] Step S500: When the abnormal score triggers the preset control conditions, control operations are performed on the ventilation equipment.

[0064] It should be noted that this step is the decision output and closed-loop execution link of the control method. Step S400 has calculated a quantified anomaly score for the current operating condition. The core task of this step is to make a decision based on this score and trigger the corresponding control operation.

[0065] Specifically, the first step is to establish a standard anomaly score. As an example, all the overall environment vectors used during the training of the autoencoder model can be re-input into the autoencoder model, the anomaly score corresponding to each vector can be calculated, and their average can be taken as the standard anomaly score.

[0066] During real-time monitoring, when the mine environment anomaly score at a certain target time is calculated and it is found to be greater than the preset multiple of the standard anomaly score, it is determined that the ventilation equipment is in an abnormal operating state. The preset multiple can be set to 1.1, and the implementers can adjust it according to actual needs.

[0067] Upon detecting an anomaly, the system will automatically execute the control procedures: adding the ventilation equipment to the key control list; pushing alarm information to the control center and the terminals of personnel on duty underground, including anomaly scores and relevant monitoring point data to quickly locate the problem; and automatically adjusting the operating parameters of the relevant ventilation equipment according to the preset emergency plan, or prompting operators to intervene manually to promptly troubleshoot the fault and ensure the safety of underground operations.

[0068] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for equipment control in mine safety, characterized in that, Including the following steps: Environmental parameters from several monitoring points within the mine are collected to construct environmental vectors for those monitoring points, and equipment parameters from ventilation equipment at several monitoring points are collected to construct equipment vectors for those monitoring points. Under normal operating conditions of the ventilation equipment, based on the variation range of the environmental vector and equipment vector at the aforementioned monitoring points within a preset time period, the stability of the environmental parameters at each monitoring point is determined, satisfying the following relationship: , It is the first The stability of environmental parameters at each monitoring point; Within a preset time period, the first The maximum Euclidean distance between the environmental vectors of any two monitoring points at any two times; The calculation method is as follows: for the first For each ventilation device at each monitoring point, calculate the maximum Euclidean distance of its device vector within the preset time period, and average the calculated maximum Euclidean distances. It is a standard normalized function; Construct and train an autoencoder model containing at least one hidden layer. The autoencoder model updates its parameters using an adaptive learning rate. The adaptive learning rate is determined such that the learning rate allocated to the model parameters corresponding to each monitoring point is negatively correlated with the stability of its environmental parameters. The adaptive learning rate satisfies the following relationship: ; It is the first autoencoder model In the first round of training, The learning rate of the model parameters corresponding to each monitoring point; It is the preset base learning rate. It is the preset number of training rounds. This is the current round number in the training process; Taking any acquisition time as the target time, the environmental vectors of several monitoring points at the target time are concatenated to obtain the overall environmental vector; the overall environmental vector is then input into the trained autoencoder model to obtain the reconstructed overall environmental vector and the feature vector output by the hidden layer of the autoencoder model; based on the reconstruction loss between the overall environmental vector and the reconstructed overall environmental vector and the deviation between the feature vector output by the hidden layer and the preset standard feature vector, the anomaly score of the mine environment at the target time is calculated; When the abnormal score triggers a preset control condition, a control operation is performed on the ventilation equipment.

2. The equipment control method for mine safety according to claim 1, characterized in that, The variation range of the environmental vector and device vector of the plurality of monitoring points within a preset time period is determined by calculating the Euclidean distance between the environmental vector or the device vector within the preset time period.

3. The equipment control method for mine safety according to claim 1, characterized in that, The autoencoder model includes an input layer, at least one hidden layer, and an output layer; The input layer is used to receive the overall environment vector, and the output layer is used to output the reconstructed overall environment vector.

4. The equipment control method for mine safety according to claim 1, characterized in that, The reconstruction loss of the overall environment vector and the reconstructed overall environment vector is determined by the Euclidean distance between the overall environment vector and the reconstructed overall environment vector. The deviation between the feature vector output by the hidden layer and the preset standard feature vector is determined by the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector.

5. The equipment control method for mine safety according to claim 4, characterized in that, The anomaly score of the mine environment state at the target time. Satisfying the relation: ; in, This is the standard loss value under normal mining conditions; It is the reconstruction loss between the overall environment vector at the target time and the reconstructed overall environment vector; It is the cosine similarity between the feature vector output by the hidden layer and the preset standard feature vector; yes Type activation function, It is a preset microvalue used to prevent the denominator from being 0.

6. The equipment control method for mine safety according to claim 1, characterized in that, The standard feature vector is the mean of the feature vectors output by the hidden layer in the final stage of training the autoencoder model, after a preset number of training rounds.

7. The equipment control method for mine safety according to claim 1, characterized in that, The control operation performed on the ventilation equipment specifically includes: Calculate the anomaly score of the mine environment at multiple acquisition times under normal operating conditions, and calculate its mean, which is recorded as the standard anomaly score; When the mine environment anomaly score at the target time is greater than a preset multiple of the standard anomaly score, the ventilation equipment is determined to be in an abnormal operating state and is placed under control.

8. The equipment control method for mine safety according to claim 1, characterized in that, The environmental parameters include at least one of temperature, humidity, gas concentration, and coal dust concentration; the equipment parameters include at least one of motor current and wind pressure difference of the ventilation equipment.

Images