Automatic equipment control method and system based on multi-source data

Abstract

The invention belongs to the technical field of automatic control, and discloses an automatic equipment control method and system based on multi-source data. Comprising the steps of collecting multi-source data of an industrial equipment group in a mining area, integrating the multi-source data into multi-equipment operation data, and preprocessing the multi-equipment operation data to obtain basic equipment operation data; constructing a multi-modal hybrid model based on the basic equipment operation data, updating the multi-modal hybrid model, and performing equipment health state evaluation on each industrial equipment by using the updated multi-modal hybrid model to obtain comprehensive risk data; constructing a deep learning framework, generating an equipment control strategy priority based on the comprehensive risk data and the deep learning framework, and generating an instruction sequence based on the equipment control strategy priority; performing dynamic parameter adjustment on each industrial device in the industrial device group based on the instruction sequence to obtain an optimal operation parameter combination; sending the optimal operation parameter combination to a preset mining area equipment parameter database; and automatic control of each device is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of automatic control, and more specifically, to a method and system for automatic control of equipment based on multi-source data. Background Art

[0002] A patent with the application publication number CN112415961A discloses an environmental protection device for automatic control in an industrial production environment, including a cooling device. The right end of the cooling device is fixed with a housing, the inner surface of the housing is fixed with a connecting pipe, the right end of the connecting pipe is fixed with a dust removal device, the right end of the dust removal device is fixed with a heat exchange device, and the right end of the housing is fixed with an air outlet pipe. The cooling device can well cool the incoming air through the cooling water between the closed box and the cooling box, and at the same time remove the dust in the air through the water curtain, which well reduces the temperature of the air and makes the subsequent steps more stable. At the same time, the cooling water isolates the blower from the outside, reducing the noise of the blower and ensuring the comfort of the production environment.

[0003] In the field of automatic control of mining area industrial equipment, multi-source data fusion and equipment health status assessment are the keys to realizing intelligent management and control. However, the mining area environment is complex, and there are often problems such as noise interference and environmental factors resulting in low-quality multi-source data collected by sensors. General equipment health status assessment models are mostly constructed based on single-modal data, lacking the ability to fuse multi-source heterogeneous data, resulting in one-sided or misjudged assessment results. The working conditions of mining area industrial equipment are often dynamically changing, especially in the scenario of multi-device collaboration. It is difficult to use simple models to generate control strategies to adapt to the working condition changes, and it is also difficult to adjust the optimal parameters of each device, resulting in a decline in work efficiency.

[0004] In view of this, the present invention proposes a method and system for automatic control of equipment based on multi-source data to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention provides the following technical solution: A method for automatic control of equipment based on multi-source data, including:

[0006] S1. Collect multi-source data of industrial equipment groups in the mining area and integrate them into multi-device operation data, and preprocess the multi-device operation data to obtain basic equipment operation data;

[0007] S2. Build a multi-modal hybrid model based on the basic equipment operation data and update the multi-modal hybrid model, and use the updated multi-modal hybrid model to evaluate the equipment health status of each industrial equipment to obtain comprehensive risk data;

[0008] S3. Build a deep learning framework, generate the priority of device control strategies based on the comprehensive risk data and the deep learning framework, and generate an instruction sequence based on the priority of device control strategies;

[0009] S4. Dynamically adjust the parameters of each industrial device in the industrial device group based on the instruction sequence to obtain the optimal operating parameter combination; send the optimal operating parameter combination to the preset mining area device parameter database.

[0010] Furthermore, the method for preprocessing the multi-device operation data includes:

[0011] The multi-device operation data includes environmental monitoring data, industrial device data, and geological monitoring data. Among them, the environmental monitoring data includes gas concentration data, dust concentration data, temperature data, and humidity data; the industrial device data includes the vibration data of mining and excavation equipment and the conveyor belt operation data; the geological monitoring data includes the tunnel displacement data and the water level monitoring data;

[0012] Perform data cleaning on the environmental monitoring data, build a temperature and humidity non-linear coupling model to eliminate the sensor errors of the cleaned environmental monitoring data, and perform normalization processing on the environmental monitoring data after error elimination to obtain the environmental monitoring standard data; for the vibration data of the acquisition equipment in the industrial device data, use wavelet packet decomposition to extract the multi-band energy characteristics and remove the noise to obtain the denoised mining and excavation equipment data; based on the conveyor belt operation data, build a three-dimensional state matrix to quantify the conveyor belt operation state and perform feature extraction to obtain a mixed matrix; perform standardization processing on the mixed matrix to obtain the standard conveyor belt operation data, and perform spatio-temporal alignment on the denoised mining and excavation equipment data and the standard conveyor belt operation data to obtain the device operation feature data; perform noise suppression and fill in the missing data for the geological monitoring data to obtain the geological safety data; combine the environmental monitoring standard data, the device operation feature data, and the geological safety data into the basic device operation data.

[0013] Furthermore, the method for performing data cleaning on the environmental monitoring data includes:

[0014] Construct a dynamic window and set the initial window length based on the sampling frequencies of gas concentration data and dust concentration data; when the dynamic window is sliding, add natural boundary conditions to the environmental monitoring data points at both ends of the dynamic window, and simultaneously calculate the gradient change rate of adjacent environmental data points in real time. When the change amplitude of the gradient change rate is greater than or equal to the preset gradient fluctuation threshold, trigger the window adjustment mechanism; use the cubic spline interpolation optimization algorithm to fill in the missing values for all environmental monitoring data points within the sliding window at any given moment to obtain complete environmental monitoring data; design a multi-physical parameter joint distribution anomaly detection mechanism. When the gas concentration change rate within the sliding window at any given moment is greater than the preset concentration change threshold, if the temperature change rate exceeds the preset temperature change threshold, mark all the complete environmental monitoring data points within the corresponding sliding window; conversely, if the temperature does not change synchronously, enable the combustion interference elimination algorithm; when the humidity within the sliding window at any given moment is greater than the preset relative humidity threshold, if the fluctuation of the dust concentration change rate within this sliding window exceeds the theoretical sedimentation rate, mark all the complete environmental monitoring data points within this sliding window; calculate the outliers among all the marked points based on the Mahalanobis distance method. The outliers are the abnormal points. Delete the abnormal points to obtain the cleaned environmental monitoring data;

[0015] The methods for eliminating sensor errors from the cleaned environmental monitoring data include:

[0016] Take the multi-modal compensation network as the basic structure of the temperature-humidity non-linear coupling model. Decouple the temperature-humidity coupling effect based on Laguerre orthogonal polynomials, construct a two-channel joint compensation function and eliminate sensor errors based on the two-channel joint compensation function. At the same time, construct an adaptive compensation function based on the aging attenuation coefficient and the temperature change rate to dynamically adjust the compensation period to obtain the environmental monitoring data after error elimination.

[0017] Furthermore, the method for extracting multi-band energy features and removing noise from the vibration data of the acquisition device in the industrial equipment data includes:

[0018] Perform N-layer wavelet packet decomposition on the vibration data of the acquisition device through a preset wavelet basis function to generate 2 NFrequency band nodes; calculate the node energy of each frequency band node and calculate the energy proportion of each frequency band node based on the node energy of each frequency band node; sort the frequency band nodes in descending order based on the energy proportions of all frequency band nodes to obtain a frequency band node sequence; set a threshold for the energy proportion of frequency band nodes, determine the frequency band nodes with energy proportions greater than the threshold of the energy proportion of frequency band nodes in the frequency band node sequence as high-energy nodes and save the original coefficients, and determine the remaining frequency band nodes as low-energy nodes; construct an adaptive threshold function to perform dynamic threshold processing on all low-energy nodes to obtain a denoised energy node set; merge all high-energy nodes and the denoised energy node set for signal reconstruction and calculate the signal-to-noise ratio index at this time. If the signal-to-noise ratio index is less than the expected value, perform multiple rounds of processing on all frequency band nodes after signal reconstruction using wavelet packet decomposition until the signal-to-noise ratio index is greater than or equal to the expected value to obtain denoised mining equipment data;

[0019] The method for constructing a three-dimensional state matrix based on the conveyor belt operation data to quantify the conveyor belt operation state and perform feature extraction includes:

[0020] Construct a three-dimensional coordinate system, extract the timestamp of each data point in the conveyor belt operation data as the time axis, construct a time window on the time axis and adjust the size of the time window based on the length of the time axis to ensure that it can cover the full cycle state of the conveyor belt operation; integrate the physical space distributions of sensors deployed at different positions on the conveyor belt and use it as the sensor axis; extract the physical parameters in the conveyor belt operation data as the parameter axis; couple the time axis, the sensor axis, and the parameter axis to construct a three-dimensional state matrix; expand the three-dimensional state matrix within each time window into a two-dimensional matrix and calculate the covariance matrix of the two-dimensional matrix, and perform eigen decomposition on each covariance matrix to obtain tensor state features.

[0021] Furthermore, the method for suppressing noise and filling missing data in the geological monitoring data includes:

[0022] Regard the mine roadway displacement data and water level monitoring data in the geological monitoring data as multi-dimensional vector data and map them to a vector space; use the principal component analysis algorithm to reduce the dimension of the multi-dimensional vector data and perform spatio-temporal alignment; construct a three-dimensional vector field, map the multi-dimensional vector data to the three-dimensional vector field, and any vector in the three-dimensional vector field is a part of the multi-dimensional vector data; determine the geological structure trend in the mining area by querying the mine roadway displacement data or water level monitoring data in the preset mining area equipment parameter database, construct a filtering window based on the geological structure trend and adjust its shape; calculate the two-norm distance from each vector in the filtering window to any other vector; introduce a direction-sensitive weight and use this weight to weight each vector to obtain a weighted vector, and at the same time adjust the direction-sensitive weights of the vectors in different directions according to the geological structure trend; sum up the two-norm distances of each weighted vector, that is, the sum of the total norm distances from each weighted vector to other weighted vectors, to obtain the cumulative norm distance of each weighted vector; take the weighted vector with the minimum cumulative norm distance as the output vector; perform inverse mapping on the output vector to obtain the filtered geological data;

[0023] Construct a missing data filling model, and use the pre-trained LSTM model as the basic structure of the missing data filling model; introduce physical constraints into the loss function of the missing data filling model and reconstruct the constraint loss function; use the filtered geological data as the input data of the input layer of the missing data filling model, and dynamically fill the input data through the missing data filling model to obtain geological safety data.

[0024] Furthermore, the method for constructing a multi-modal hybrid model based on the basic equipment operation data and updating the multi-modal hybrid model includes:

[0025] Classify the basic equipment operation data into numerical type data and matrix type data based on the data type; construct a dual-channel hybrid structure and use this structure as the core structure of the multi-modal hybrid model; the dual-channel hybrid structure includes a temporal convolutional channel and a self-attention channel for processing numerical type data and matrix type data respectively; adopt a depthwise separable temporal convolutional network as the basic structure of the temporal convolutional channel; shuffle the numerical type data into disordered numerical data and input this data into the causal convolutional layer of the temporal convolutional channel, and use the multi-scale convolutional kernels in the causal convolutional layer to extract the local temporal patterns of the disordered numerical data; at the same time, introduce a dilated convolutional layer, construct a dilation factor and use this factor to gradually expand the temporal coverage range of the temporal convolutional channel; add a channel attention mechanism, extract the statistical features of the disordered numerical data through global average pooling and max pooling, and generate a sensor weight vector through the fully connected layer of the temporal convolutional channel; construct a spatio-temporal two-dimensional attention network as the basic structure of the self-attention channel, perform a linear transformation on the matrix type data to generate a triple; at the same time, input the triple into the self-attention channel and embed the device physical topology constraint matrix in the spatial dimension; calculate the attention score based on the device physical topology constraint matrix and the triple; normalize the attention score to generate a probability weight, and perform weighted aggregation on all elements in the triple based on the probability weight to obtain a spatio-temporal dependence relationship output matrix; perform adaptive weighted fusion on the outputs of the two channels through the dynamic gating fusion layer of the multi-modal hybrid model to generate a joint feature representation;

[0026] Pre-train three groups of structurally heterogeneous expert models and deploy them in the dynamic gating fusion layer, namely the spatial feature expert model, the temporal feature expert model, and the cross-modal expert model; input the joint feature representation into the dynamic gating fusion layer and calculate the expert activation probability; construct an elastic selection mechanism based on the reinforcement learning algorithm and dynamically adjust the number of activated expert models based on this mechanism; encode the joint feature representation as a state vector; construct a reward function, and update the parameters of the dynamic gating fusion layer using the policy gradient algorithm based on the state vector, the number of activated expert models, and the reward function;

[0027] The methods for evaluating the health status of each industrial device include:

[0028] Process the joint feature representation in the output layer of the multi-modal hybrid data and compare it with the fault categories in the preset mine equipment parameter database to obtain the fault mode probability distribution; at the same time, extract the device features in the joint feature representation, construct a graph attention network and calculate the risk propagation coefficient between each node in the network; construct a scoring function based on the fault mode probability distribution and the risk propagation coefficient, and output the device health status evaluation score based on the scoring function; combine the fault mode probability distribution and the device health status evaluation score into a comprehensive risk code, that is, comprehensive risk data.

[0029] Further, the method for constructing the deep learning framework includes:

[0030] The deep learning framework includes two sub-architectures, namely a risk decoding architecture and a policy generation architecture; taking the dynamic routing capsule network as the core structure of the risk decoding architecture, constructing the underlying capsule unit and the high-level capsule unit as sub-structures in the core structure; inputting the comprehensive risk data into the underlying capsule unit to obtain the underlying capsule input vector corresponding to each piece of comprehensive risk data; initializing the coupling coefficient between the underlying capsule unit and the high-level capsule unit to obtain the initialized coupling coefficient; initializing the capsule weight matrix to obtain the initialized capsule weight matrix; using the initialized capsule weight matrix to convert the underlying capsule input vector into a capsule prediction vector; performing weighted summation on all capsule prediction vectors to obtain the high-level capsule input vector; constructing an anti-interference function based on the high-level capsule input vector, calculating the anti-interference function and outputting the result to obtain the high-level capsule output vector; updating the coupling coefficient based on the initialized coupling coefficient, the capsule prediction vector and the high-level capsule output vector; performing the above process iteratively multiple times until the coupling coefficient converges, at this time the high-level capsule output vector is the output of the risk decoding architecture, and at the same time using the backpropagation algorithm to update the initialized capsule weight matrix; taking the deep Q network as the core structure of the policy generation architecture, and separating the core structure of the policy generation architecture into two parallel branches, namely the state value branch and the action advantage branch; taking the high-level capsule output vector as the input data of the policy generation architecture, calculating the state value of the input high-level capsule output vector in the state value branch, and calculating the action advantage of the input high-level capsule output vector in the action advantage branch; constructing a fusion function, and using the fusion function to calculate the fusion function value of the state value and the action advantage of any input high-level capsule output vector; performing physical interpretability mapping on the high-level capsule output vector with the largest fusion function value in each calculation to obtain an executable policy; integrating all executable policies to obtain a policy set.

[0031] Further, the method for generating the priority of the device control policy based on the comprehensive risk data and the deep learning framework includes:

[0032] Using the deep learning framework to process the comprehensive risk data to obtain a policy set; classifying the urgency and grading the fault intensity of each executable policy in the policy set; determining the success rate of the executable policy in the historical situation by querying the preset mine equipment parameter database to obtain the historical execution effect record; constructing a priority objective function, and calculating the function value of the priority objective function based on the urgency, fault intensity and historical execution effect record of each executable policy; sorting all executable policies in descending order based on the function value of the priority objective function to obtain an instruction sequence.

[0033] Further, the method for dynamically adjusting parameters of each industrial device in the industrial device group based on the instruction sequence includes:

[0034] Query the preset mine area equipment parameter database to obtain the adjustable parameter list of each industrial device, and determine the number of parameters adjusted by each industrial device each time; construct a parameter dimension space, map the executable policies belonging to any one device in the instruction sequence to the parameter dimension space, and mark the parameters of the device corresponding to each executable policy; group the marked parameters based on the number of parameters adjusted by each industrial device each time, and attach a priority weight to each group of grouped parameters based on the priority of the device control policy of the device corresponding to the executable policy in the instruction sequence; regard each group as an individual and use an optimization algorithm to optimize the population composed of all individuals to obtain the optimal individual, which is the optimal operating parameter combination of any one device.

[0035] A device automatic control system based on multi-source data, which is used to implement a device automatic control method based on multi-source data, includes:

[0036] A data acquisition module, which acquires multi-source data of the industrial device group in the mine area and integrates it into multi-device operation data, preprocesses the multi-device operation data to obtain basic device operation data;

[0037] A device health status evaluation module, which constructs a multi-modal hybrid model based on the basic device operation data and updates the multi-modal hybrid model, and uses the updated multi-modal hybrid model to evaluate the device health status of each industrial device to obtain comprehensive risk data;

[0038] A policy priority generation module, which constructs a deep learning framework, generates device control policy priorities based on the comprehensive risk data and the deep learning framework, and generates an instruction sequence based on the device control policy priorities;

[0039] A parameter adjustment module, which dynamically adjusts the parameters of each industrial device in the industrial device group based on the instruction sequence to obtain the optimal operating parameter combination; sends the optimal operating parameter combination to the preset mine area equipment parameter database; the various modules are connected by wired and / or wireless means.

[0040] The technical effects and advantages of the device automatic control method and system based on multi-source data of the present invention:

[0041] By collecting multi-source data of various devices in the industrial equipment group in the mining area and performing preprocessing, more accurate basic equipment operation data is obtained; a model is constructed to evaluate the equipment health status of the basic equipment operation data, determine the status of each device, and at the same time obtain comprehensive risk data; a deep learning framework is constructed to process the comprehensive risk data, generate the equipment control strategy priorities of each industrial device and sort them to obtain an instruction sequence, and finally based on various executable strategies in the instruction sequence, parameter adjustment is performed on each industrial device to obtain the optimal operation parameter combination of each device and send the parameter combination to the database, realizing the automatic control of each device in the mining area; compared with existing experience, more accurate preprocessing methods are used, and the most suitable method is used for preprocessing each type of data; by constructing an updatable model to evaluate the health status of the equipment, the model can adapt to the dynamic changes of the working conditions in the mining area; a deep learning framework is constructed to generate executable strategies for each device, so that the efficiency is higher and more in line with the actual working conditions in the subsequent dynamic parameter adjustment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 FIG. is a schematic diagram of an automatic equipment control method based on multi-source data of the present invention;

[0043] Figure 2 FIG. is a schematic diagram of an automatic equipment control system based on multi-source data of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0045] Embodiment 1

[0046] Please refer to Figure 1 As shown, an automatic equipment control method based on multi-source data in this embodiment includes:

[0047] S1. Collect multi-source data of the industrial equipment group in the mining area and integrate it into multi-device operation data, and perform preprocessing on the multi-device operation data to obtain basic equipment operation data;

[0048] S2. Construct a multi-modal hybrid model based on the basic equipment operation data and update the multi-modal hybrid model, and use the updated multi-modal hybrid model to evaluate the equipment health status of each industrial device to obtain comprehensive risk data;

[0049] S3. Build a deep learning framework, generate the priority of device control strategies based on the comprehensive risk data and the deep learning framework, and generate an instruction sequence based on the priority of the device control strategies;

[0050] S4. Dynamically adjust the parameters of each industrial device in the industrial device group based on the instruction sequence to obtain the optimal combination of operating parameters; send the optimal combination of operating parameters to the preset mine area device parameter database.

[0051] Sensors are deployed on each industrial device in the industrial device group in the mining area to collect multi-source data. The multi-source data includes gas concentration data, dust concentration data, temperature data, humidity data, vibration data of mining equipment, conveyor belt operation data, mine roadway displacement data, and water level monitoring data (where the gas concentration data represents the concentration data of hazardous gases in the mining area, such as methane, carbon monoxide, and hydrogen sulfide; the dust concentration data represents the concentration data of inhalable particulate matter in the mining operation area, such as PM2.5 or PM10 concentration; the gas concentration data and the dust concentration data are collected by sensors installed on the ventilation system and are used to judge the ventilation efficiency and the health risk of workers; the temperature data and the humidity data are collected by sensors placed at the working face or in the mine roadway and are used to warn of fire and equipment overheating risks; the vibration data of mining equipment represents the vibration frequency of key equipment components in the drill or roadheader and is collected by acceleration sensors installed on the mining equipment to identify mechanical failures; the conveyor belt operation data includes data such as belt tension data and material movement trajectories and is used to monitor the operation status of the conveyor belt to prevent material spillage or equipment jamming; the mine roadway displacement data represents the rock deformation of the inner wall of the mine roadway to ensure the stability of the support structure; the water level monitoring data represents the water level change and permeability data caused by rainfall, etc. in the mine roadway and is used to judge the operation status of the drainage system and prevent water inrush accidents).

[0052] During the data collection process of the sensors, the components of the sensors may age due to environmental interference, so the collected data will have errors; in the spectrum type data, in order to make the required frequency bands more obvious and accurate, noise processing needs to be performed on it. Therefore, a method for preprocessing multi-device and multi-type data is proposed.

[0053] The methods for preprocessing the multi-device operation data include:

[0054] The multi-device operation data includes environmental monitoring data, industrial device data, and geological monitoring data, where the environmental monitoring data includes gas concentration data, dust concentration data, temperature data, and humidity data; the industrial device data includes vibration data of mining equipment and conveyor belt operation data; the geological monitoring data includes mine roadway displacement data and water level monitoring data.

[0055] Clean the environmental monitoring data and construct a nonlinear coupling model of temperature and humidity to eliminate sensor errors in the cleaned environmental monitoring data, and perform normalization processing on the environmental monitoring data after error elimination to obtain standard environmental monitoring data (in a high-temperature or high-humidity environment, the sensors may have certain errors in the collected data due to material reasons or component aging, so it is necessary to correct the errors to improve the accuracy of the collected data); for the vibration data of the acquisition equipment in the industrial equipment data, use wavelet packet decomposition to extract multi-band energy features and remove noise to obtain denoised mining equipment data (the vibration data of the acquisition equipment is a vibration spectrum data. Using wavelet packet decomposition to process it can enhance the energy of the effective frequency band signals while removing noise, indirectly improving the accuracy of the subsequent construction of the multi-modal hybrid model for equipment health status assessment); construct a three-dimensional state matrix based on the conveyor belt operation data to quantify the conveyor belt operation status and perform feature extraction to obtain a hybrid matrix; perform standardization processing on the hybrid matrix to obtain standard conveyor belt operation data (constructing a three-dimensional state matrix quantifies the conveyor belt operation status and simultaneously completes data dimensionality reduction, improving the training efficiency of the subsequent multi-modal hybrid model), and perform spatio-temporal alignment on the denoised mining equipment data and the standard conveyor belt operation data to obtain equipment operation feature data (synchronize the timestamps based on the NTP protocol and map the equipment coordinates to the same spatial coordinate system to achieve spatio-temporal alignment; spatio-temporal alignment can ensure the timeliness of emergency control instructions, and in the formed multi-dimensional parameter matrix, the features of each dimension have consistent spatio-temporal labels, facilitating the training of the subsequent constructed multi-modal hybrid model); suppress the noise of the geological monitoring data and fill in the missing data to obtain geological safety data; combine the standard environmental monitoring data, the equipment operation feature data, and the geological safety data into basic equipment operation data.

[0056] The methods for data cleaning of environmental monitoring data include:

[0057] Construct a dynamic window, and set the initial window length based on the sampling frequencies of gas concentration data and dust concentration data (the initial window length W0 = max(3×TI, 10s); where TI represents the sensor sampling period); when the dynamic window is sliding, add natural boundary conditions (i.e., the second derivative is zero, which is used to prevent edge oscillations) to the environmental monitoring data points at both ends of the dynamic window, and simultaneously calculate the gradient change rate of adjacent environmental data points in real time ( represents the gas concentration gradient change rate, denoted by ΔG). When the change amplitude of the gradient change rate is greater than or equal to the preset gradient fluctuation threshold (in this embodiment, the gradient fluctuation threshold is set to 15%, that is, when the change amplitude of the gradient change rate is greater than or equal to ±15%, the window adjustment mechanism is triggered), trigger the window adjustment mechanism (the window adjustment formula of the window adjustment mechanism is: Where, represents the window length of the adjusted dynamic window; σW (which represents the standard deviation of all data points within the dynamic window at any given moment); use the cubic spline interpolation optimization algorithm to fill in the missing values for all environmental monitoring data points within the sliding window at any given moment (the cubic spline interpolation optimization algorithm can effectively fit the changing trend of the data and is quite suitable for data with characteristics of continuity and smooth change such as the environmental monitoring data in this embodiment; in the technical background of this embodiment, the startup and working condition changes of various equipment in the mining area will cause short-term changes in the data, and the cubic spline interpolation optimization algorithm is more sensitive to local data fluctuations, resulting in improved efficiency of filling in missing values), obtaining complete environmental monitoring data; design a multi-physical parameter joint distribution anomaly detection mechanism. When the gas concentration change rate within the sliding window at any given moment is greater than the preset concentration change threshold (in this embodiment, the concentration change threshold is set to 1%), and if the temperature change rate exceeds the preset temperature change threshold (in this embodiment, the temperature change threshold is set to 5°C), then mark all the complete environmental monitoring data points within the corresponding sliding window; conversely, if the temperature does not change synchronously, then enable the combustion interference elimination algorithm (the combustion interference elimination algorithm is used to eliminate the interference signals generated during the combustion process and distinguish real combustion events from abnormal gas concentrations caused by equipment failures or environmental interferences); when the humidity within the sliding window at any given moment is greater than the preset relative humidity threshold, and if the fluctuation of the dust concentration change rate within this sliding window exceeds the theoretical sedimentation rate (the formula for calculating the theoretical sedimentation rate is: where VG represents the theoretical sedimentation velocity of the dust; r represents the particle radius of the dust; ρ g represents the density of the dust particles; ρ s represents the air density; g represents the acceleration due to gravity; μ0 represents the air viscous drag coefficient, which is determined according to specific circumstances. For example, when the temperature is 20°C, this value is 1.8×10 -5 ; λ0 represents the correction coefficient, which is related to the humidity and takes a value of 0.05 in this embodiment), then mark all the complete environmental monitoring data points within this sliding window; calculate the outliers among all the marked points based on the Mahalanobis distance method (the Mahalanobis distance method eliminates the correlation interference between variables through the covariance matrix, and the calculation formula is: where D M represents the Mahalanobis distance from the marked point x to the distribution center, and the distribution center refers to the data mean; μ represents the marked point mean; ∑ represents the covariance matrix; when the Mahalanobis distance of a certain marked point is greater than the preset Mahalanobis distance threshold, then determine this marked point as an outlier), the outliers are the abnormal points, and delete the abnormal points to obtain the cleaned environmental monitoring data.

[0058] Since the environmental monitoring data after cleaning may still have errors that cannot be completely eliminated by data cleaning because the sensor is vulnerable to temperature and humidity, a multi-modal compensation network is selected to compensate the environmental monitoring data after cleaning to minimize the errors in the data.

[0059] The methods for eliminating sensor errors from the environmental monitoring data after cleaning include:

[0060] Taking the multi-modal compensation network as the basic structure of the temperature-humidity non-linear coupling model, decoupling the temperature-humidity coupling effect based on Laguerre orthogonal polynomials (the basis functions of Laguerre orthogonal polynomials make the temperature and humidity compensation terms not interfere with each other after decoupling, avoiding the parameter coupling problem caused by the correlation of the traditional polynomial basis; moreover, the computational complexity of Laguerre orthogonal polynomials is lower and the decoupling efficiency is higher), constructing a two-channel joint compensation function and eliminating sensor errors based on the two-channel joint compensation function (the calculation formula of the two-channel joint compensation function is: where, C after represents the environmental monitoring data after error elimination; C before represents the environmental monitoring data after cleaning; a i represents the temperature compensation coefficient, a i ∈[-0.5, 0.5]; b i1 represents the humidity compensation coefficient, b i1 ∈[-0.3, 0.3]; L i (T) represents the i-th order Laguerre orthogonal polynomial temperature basis function; L i1 (H) represents the i1-th order Laguerre orthogonal polynomial humidity basis function; n0 and n1 represent the orders of the orthogonal polynomials, n0 ∈ [5, 9], n1 ∈ [7, 11], adjusted based on the sensor sensitivity), and at the same time constructing an adaptive compensation function based on the aging attenuation coefficient and the temperature change rate to dynamically adjust the compensation period (the calculation formula of the adaptive compensation function is: τ = τ0 × (1 + k0 × ΔT) × e -ωt ; where, τ represents the adaptive compensation period; τ0 represents the preset initial compensation period, which takes the value of 100s in this embodiment; k0 represents the thermal inertia correction factor, where, C0 represents the heat capacity of the device; m represents the mass of the device; v0 represents the heat dissipation coefficient, determined based on the device material; the above-mentioned device refers to the industrial devices in the industrial device group in the mining area; ΔT represents the temperature change rate; ω represents the aging attenuation coefficient, where, represents the half-life of the sensor; t represents time, and the unit time is one second), to obtain the environmental monitoring data after error elimination.

[0061] Processing the vibration data of the acquisition device using wavelet packet decomposition can not only remove the noise in the data, but also extract the energy characteristics of the high-energy frequency bands therein to enhance the signals in the high-energy frequency bands and avoid feature loss caused by traditional filtering.

[0062] The method of using wavelet packet decomposition to extract multi-band energy characteristics and remove noise from the vibration data of the acquisition device in industrial equipment data includes:

[0063] Perform N-layer wavelet packet decomposition on the vibration data of the acquisition device through a preset wavelet basis function (such as the Daubechies4 wavelet basis function) (the number of wavelet packet decomposition layers is set according to the specific situation. In this embodiment, the decomposition layer N = 3), generating 2 N frequency band nodes (calculate the number of frequency band nodes based on the decomposition layer, and each node corresponds to a specific frequency range. For example, the second N frequency band node covers the high-frequency band of 1 - 2 kHz); calculate the node energy of each frequency band node and calculate the energy proportion of each frequency band node based on the node energy of each frequency band node (the calculation formula for the node energy of each frequency band node is: where, E k,j represents the node energy of the j-th frequency band node in the k-th layer; CI k,j represents the wavelet coefficient of the j-th frequency band node in the k-th layer; the calculation formula for the energy proportion of each frequency band node is: where, P k,j represents the node energy proportion of the j-th frequency band node in the k-th layer); sort the frequency band nodes in descending order based on the energy proportion of all frequency band nodes to obtain a frequency band node sequence; set a threshold for the energy proportion of the frequency band nodes (the value in this embodiment is 20%), determine the frequency band nodes with an energy proportion greater than the threshold of the energy proportion of the frequency band nodes as high-energy nodes and save the original coefficients (separate the first 20% of the frequency band nodes in the frequency band node sequence as high-energy nodes. Since the occurrence of equipment failures is often accompanied by an increase in energy, the coefficients of each high-energy node are saved to protect the frequency bands that can reflect the fault characteristics, which is convenient for subsequent construction of a multi-modal hybrid model to evaluate the health status of the equipment), and determine the remaining frequency band nodes as low-energy nodes; construct an adaptive threshold function to perform dynamic threshold processing on all low-energy nodes (the calculation formula for the adaptive threshold function is: where, XI represents the adaptive threshold; OR represents the length of the original signal, that is, the frequency band length in the vibration data of the acquisition device; σI represents the estimated noise level of the low-energy node, which is obtained by calculating the median absolute deviation of the wavelet coefficients of the low-energy node; make σI change dynamically by performing non-linear shrinkage on the wavelet coefficients of the low-energy node, and the calculation formula for performing non-linear shrinkage on the wavelet coefficients of the low-energy node is: Among them, represents the wavelet coefficient of the low-energy node after non-linear contraction; CI represents the wavelet coefficient of the low-energy node; by constructing an adaptive threshold function to perform dynamic threshold processing on the low-energy node and dynamically adjust the parameters in the function, the noise in the low-energy node can be greatly suppressed, and at the same time, the high-frequency feature distortion can be reduced), obtaining a denoised energy node set; merging all high-energy nodes and the denoised energy node set for signal reconstruction (i.e., inverse wavelet packet transform) and calculating the signal-to-noise ratio index at this time. If the signal-to-noise ratio index is less than the expected value, then all frequency band nodes after signal reconstruction are processed in multiple rounds using wavelet packet decomposition until the signal-to-noise ratio index is greater than or equal to the expected value, obtaining the denoised data of the mining equipment.

[0064] Converting the conveyor belt operation data into a three-dimensional state matrix not only completes the quantification of the conveyor belt operation state, but also compresses the data dimension, improves the data processing efficiency, and at the same time solves the problem of inconsistent data time and space for each device in the case of multiple devices, providing high-quality data input for the subsequent constructed multi-modal hybrid model.

[0065] The methods for constructing a three-dimensional state matrix based on the conveyor belt operation data to quantify the conveyor belt operation state and extract features include:

[0066] Construct a three-dimensional coordinate system, extract the timestamp of each data point in the conveyor belt operation data as the time axis (each unit length on this axis represents one second), construct a time window on the time axis and adjust the size of the time window based on the length of the time axis to ensure that it can cover the entire cycle state of the conveyor belt operation (in this embodiment, data is continuously collected with a 5-second time window and the window length is 1 second); integrate the physical space distribution of sensors deployed at different positions on the conveyor belt and use it as the sensor axis (for example, three groups of sensors deployed at the head, middle, and tail of the conveyor belt); extract the physical parameters in the conveyor belt operation data as the parameter axis (the physical parameters such as belt tension, running speed, and conveyor belt lateral offset, etc.); couple the time axis, sensor axis, and parameter axis to construct a three-dimensional state matrix (for example, there is a sensor-physical parameter matrix within each time window, that is where, F 头 represents the belt tension collected by the head sensor, VC 头 represents the running speed collected by the head sensor, D 头 represents the conveyor belt lateral offset collected by the head sensor; F 中 represents the belt tension collected by the middle sensor, VC 中 represents the running speed collected by the middle sensor, D 中 represents the conveyor belt lateral offset collected by the middle sensor; F 尾 represents the belt tension collected by the tail sensor, VC 尾Represents the running speed collected by the tail sensor, D 尾 Represents the lateral offset of the conveyor belt collected by the tail sensor; match each sensor - physical parameter matrix with the time axis to obtain a three - dimensional state matrix); expand the three - dimensional state matrix within each time window into a two - dimensional matrix and calculate the covariance matrix of this two - dimensional matrix (the calculation formula of the covariance matrix is: Among them, CT(a,q) represents the covariance matrix of physical parameter a and physical parameter q within any time window; S represents the number of sensors; num represents the number of sample groups within any time window; MT(num,a) represents the two - dimensional matrix expanded based on physical parameter a within any time window; μa represents the mean value of parameter a; MT(num,q) represents the two - dimensional matrix expanded based on physical parameter q within any time window; μq represents the mean value of parameter q; use the covariance matrix to quantify the correlation between different parameters and capture the co - variation across parameters), perform eigen - decomposition on each covariance matrix (calculate the eigenvalues of each covariance matrix using mathematical methods, take the eigen - vector corresponding to the largest eigenvalue, and project the physical parameters to the direction of this eigen - vector using the principal component analysis algorithm), to obtain the tensor state feature (the tensor state feature is a vector. If the proportion of any physical parameter component in the vector suddenly changes, it indicates that the change of this physical parameter may be the cause of the fault).

[0067] In order to more efficiently suppress noise and fill in missing data for geological detection data, map the geological monitoring data to a vector space for processing, and at the same time combine the actual geological structure trend to make the results of noise suppression and missing data filling more in line with the actual situation.

[0068] The methods for suppressing noise and filling in missing data for geological monitoring data include:

[0069] Regard the mine roadway displacement data and water level monitoring data in the geological monitoring data as multi - dimensional vector data and map them to the vector space (for example, the mine roadway displacement data XT at a certain moment = (X0 T ,Y0 T ,Z0 T ); where X0 T ,Y0 T and Z0 TCorresponding to the displacement amounts in three directions respectively); using the principal component analysis algorithm to reduce the dimension of the multi-dimensional vector data and perform spatio-temporal alignment (ensuring that the timestamps of the multi-dimensional vector data are consistent and the spatial coordinates are unified in the same spatial grid); constructing a three-dimensional vector field, mapping the multi-dimensional vector data to the three-dimensional vector field, and any vector in the three-dimensional vector field is a part of the multi-dimensional vector data (for example, the multi-dimensional vector data includes vector data such as displacement vectors, stress tensors, and seepage vectors, and after mapping the multi-dimensional vector data to the three-dimensional vector field, each vector data in the multi-dimensional vector data is separated for easy observation and calculation); determining the geological structure trend in the mining area by querying the mine roadway displacement data or water level monitoring data in the preset mining area equipment parameter database (for example, the deformation trend of the rock wall in the area corresponding to the mine roadway displacement data or the water flow trend in the area corresponding to the water level monitoring data), constructing a filtering window based on the geological structure trend and adjusting the shape (for example, using an elliptical filtering window for the linear fault area because the linear fault belongs to a geological structure trend with strong directionality, and the long axis of the elliptical filtering window can extend along the fault trend, and the short axis is perpendicular to the fault direction, so the data points in the window mainly come from the area on the same side of the fault, reducing cross-boundary interference); calculating the two-norm distance from each vector in the filtering window to any other vector (the two-norm distance is the Euclidean distance between two vectors, and since the calculation is performed in the vector space, in this embodiment, the two-norm distance related to vector calculation is used to replace the Euclidean distance); introducing a direction-sensitive weight and using this weight to weight each vector to obtain a weighted vector, and at the same time adjusting the direction-sensitive weights of vectors in different directions according to the geological structure trend (for example, in the area where the fault trend is obvious, increasing the weight of the displacement vector along the fault direction); summing the two-norm distances of each weighted vector, that is, the sum of the total norm distances from each weighted vector to other weighted vectors, to obtain the cumulative norm distance of each weighted vector( where, SP di represents the cumulative norm distance of the $d_i$-th weighted vector, $d_i\in(1,total)$; total represents the number of weighted vectors; SI di the $d_i$-th weighted vector; SI to represents the $t_o$-th weighted vector); taking the weighted vector with the minimum cumulative norm distance as the output vector; performing inverse mapping on the output vector to obtain the filtered geological data.

[0070] Build a missing data filling model, using the pre-trained LSTM model as the basic structure of the missing data filling model (the LSTM model has excellent ability to capture complex data patterns, especially can effectively identify the dependencies in time series data, and based on this advantage, the LSTM model can make relatively accurate predictions for missing data); introduce physical constraints into the loss function of the missing data filling model and reconstruct the constraint loss function (since the filtered geological data are all physical parameter type data, in order to make the missing data filling model better identify the dependencies between data points in the filtered geological data, physical constraints are added to make the loss function satisfy physical laws); use the filtered geological data as the input data of the input layer of the missing data filling model, and dynamically fill the input data through the missing data filling model to obtain geological safety data (the missing data filling model learns the spatio-temporal evolution law of the filtered geological data and outputs predicted values. After each output, the KL divergence is calculated. If the KL divergence is greater than the preset KL divergence threshold, incremental learning is triggered to update the parameters of the missing data filling model. In this embodiment, the KL divergence threshold is set to 0.1).

[0071] Build two channels in the model to extract the time-dependent features and space-dependent features of the input data respectively, reducing the calculation delay and the interference of non-physical connection sensors, and obtaining a more accurate feature representation; in order to improve the timeliness of the multi-modal hybrid model, the parameters of the dynamic gating fusion layer of the multi-modal hybrid model are updated by reinforcement learning to flexibly select an expert model, so that the multi-modal hybrid model can adapt to the real-time requirements under different working conditions.

[0072] The methods for building a multi-modal hybrid model based on the basic equipment operation data and updating the multi-modal hybrid model include:

[0073] Classify the basic equipment operation data into numerical type data and matrix type data based on the data type (the environmental monitoring standard data and geological safety data in the basic equipment operation data belong to the numerical type data, and the equipment operation characteristic data belongs to the matrix type data); construct a dual-channel hybrid structure and use this structure as the core structure of the multi-modal hybrid model; the dual-channel hybrid structure includes a temporal convolutional channel and a self-attention channel (the temporal convolutional channel is used to capture the temporal features in the numerical type data, and the self-attention channel is used to extract the spatio-temporal dependency features in the matrix type data. Using the joint processing of the dual channels can better capture the joint feature representation of the location and time of the equipment and provide a unified feature expression for subsequent health state assessment) for processing numerical type data and matrix type data respectively; adopt a depthwise separable temporal convolutional network as the basic structure of the temporal convolutional channel (the depthwise separable temporal convolutional network is a network structure that combines depthwise separable convolution and temporal convolutional network, which can efficiently extract local features and capture the long-term and short-term dependencies in the data); shuffle the numerical type data into disordered numerical data (enhancing randomness and used to improve the generalization ability of the model) and input this data into the causal convolutional layer of the temporal convolutional channel, and use the multi-scale convolutional kernels in the causal convolutional layer to extract the local temporal patterns of the disordered numerical data (perform convolutional operations on the disordered numerical data through the multi-scale convolutional kernels, extract the temporal features in the disordered numerical data, and the obtained local temporal patterns represent short-term temporal features); at the same time, introduce a dilated convolutional layer, construct a dilation factor and use this factor to gradually expand the time coverage of the temporal convolutional channel layer by layer (the dilation factor is an exponentially growing dilation factor. Use this factor to gradually expand the receptive field of the temporal convolutional channel, so that the receptive field of the top layer obtains the maximum time coverage and is used to capture the long-term temporal features of equipment operation; for example, when the convolutional kernel size is 5, the dilated convolutional layer can perform four-layer dilation. If the dilation factors in the four-layer dilation are 2 1 ,2 2 ,2 4 ,2 8, the top-level receptive field can cover the data of 128 time steps); adding a channel attention mechanism to extract the statistical features of the disordered numerical data through global average pooling and max pooling, and generating a sensor weight vector through the fully connected layer of the time convolutional channel (global average pooling is used to reflect the overall energy level of the sensor signal, and max pooling is used to capture the abnormal peaks of the signal. Using the channel attention mechanism can more clearly identify the sensors with high fault correlation, and then discover the fault-related devices; for example, if the weight of the sensor collecting the bearing vibration signal is 40% higher than the weight of the sensor collecting the temperature signal, it indicates that the vibration signal is much more sensitive to faults than the temperature signal); constructing a spatio-temporal two-dimensional attention network as the basic structure of the self-attention channel, and performing a linear transformation on the matrix-type data to generate a triple (generating the Q / K / V triple through linear transformation can reflect the information in three dimensions: entity recognition, relationship matching, and feature transfer; for example, the Q matrix focuses on the sensor state (such as vibration amplitude), the K matrix matches the historical fault patterns based on the sensor state through the preset mine equipment parameter database, and the V matrix transfers the fault path (such as bearing - gearbox - motor)); at the same time, inputting the triple into the self-attention channel and embedding the device physical topology constraint matrix in the spatial dimension (the device physical topology constraint matrix is used to represent the position distribution of each device or sensor in the physical space, and is used to suppress the correlation calculation of the data collected by the sensors with non-physical connections); calculating the attention score based on the device physical topology constraint matrix and the triple (using the dot product calculation of the device physical topology matrix and the triple to obtain the attention score, which is used to observe the correlation between elements); normalizing the attention score to generate a probability weight, and performing weighted aggregation on all elements in the triple based on the probability weight to obtain a spatio-temporal dependence relationship output matrix (each element in the spatio-temporal dependence relationship output matrix is the weighted aggregation result of the global information, and this step integrates the dependence relationship of each element in the time dimension and the spatial dimension); through the dynamic gating fusion layer of the multimodal hybrid model, adaptively weighted fusion is performed on the outputs of the two channels to generate a joint feature representation (the dynamic gating fusion layer is used to fuse features, and the update of the multimodal hybrid model is also carried out on this layer).

[0074] Pre-train three groups of structurally heterogeneous expert models and deploy them in the dynamic gating fusion layer, namely the spatial feature expert model, the temporal feature expert model, and the cross-modal expert model (the expert model refers to the model specifically trained to solve specific sub-problems. The spatial feature expert model is used to capture the spatial associations of each device in the joint feature representation, the temporal feature expert model is used to analyze the time-frequency features in the joint feature representation, and the cross-modal expert model is used to couple the multi-source features in the joint feature representation); input the joint feature representation into the dynamic gating fusion layer and calculate the expert activation probability (the calculation formula of the expert activation probability is: where, AI l(tz) represents the expert activation probability of the l-th expert model; represents the learnable weight matrix of the l-th expert model, and the dimension of this matrix is the dimension of the input joint feature representation; ZS represents the number of expert models; tz represents any joint feature representation, and each joint feature representation is a kind of matrix; represents the bias term of the l-th expert model, which is used to adjust the activation threshold); construct an elastic selection mechanism based on the reinforcement learning algorithm (this mechanism uses the Top-k algorithm for selective activation) and dynamically adjust the number of activated expert models based on this mechanism (select the most appropriate corresponding expert model for activation based on the current working condition); encode the joint feature representation into a state vector (encoding into a vector is convenient for subsequent updating of parameters using the policy gradient update algorithm); construct a reward function, and update the parameters of the dynamic gating fusion layer using the policy gradient algorithm based on the state vector, the number of activated expert models, and the reward function (the calculation formula of the reward function is: RE = 0.7×ACC + 0.2×FLOP + 0.1×ZJ; where, RE represents the function value of the reward function; ACC represents the recognition accuracy rate. Since each update is after completing a device health status assessment, the recognition accuracy rate of the previous assessment needs to be used as a variable; FLOP represents the computational latency; ZJ represents the expert model weight; use the policy gradient algorithm to maximize the reward function, and construct an update function based on the maximized reward function, the number of activated expert models, and the state vector, and use this function to update the parameters of the dynamic gating fusion layer).

[0075] The methods for evaluating the device health status of each industrial device include:

[0076] Process the joint feature representation in the output layer of the multi-modal mixed data and compare it with the fault categories in the preset mine equipment parameter database to obtain the fault mode probability distribution (process the joint feature representation through multi-class Softmax, and the fault mode probability distribution is Pro f ∈GZ SL , where Pro f represents the fault mode probability of the f-th joint feature representation, and GZ SL represents the fault set containing SL types of faults); at the same time, extract the device features in the joint feature representation, construct a graph attention network and calculate the risk propagation coefficient between each node in the network (each node in the graph attention network represents an industrial device, and the calculation formula of the risk propagation coefficient is: XS Q1,Q2 = LeakyReLU(β·[W3·F Q1 ||W3·F Q2 )]; where, XS Q1,Q2Denote the risk propagation coefficient from node Q1 to node Q2; LeakyReLU represents the activation function; β represents the vector of learnable parameters that determines the interaction relationship of weighted nodes; W3 represents the learnable weight matrix of the graph attention network; F Q1 Denote the device features of node Q1, F Q2 Denote the device features of node Q2, such as vibration features, timing features, and spatial features, etc.; [... ∥...] represents the vector concatenation operation); construct a scoring function based on the fault mode probability distribution and the risk propagation coefficient, and output the device health status evaluation score based on the scoring function (the calculation formula of the scoring function is: SCORE = γ1 × Pro + γ2 × XS; where, SCORE represents the function value of the scoring function, that is, the device health status evaluation score; γ1 and γ2 are the weights of the fault mode probability distribution and the risk propagation coefficient respectively. In this embodiment, γ1 = 0.6, γ2 = 0.4; use the device health status evaluation score to determine the device health status, and set three score thresholds, namely the safe score threshold, the general health status threshold, and the poor health status threshold; when the device health status evaluation score is less than or equal to the safe score threshold, it indicates that the device health status is good; when the device health status evaluation score is greater than the safe score threshold and less than or equal to the general health status threshold, it indicates that the device health status is general; when the device health status evaluation score is greater than the general health status threshold and less than or equal to the poor health status threshold, it indicates that the device health status is poor; when the device health status evaluation score is greater than the poor health status threshold, it indicates that the device urgently needs repair and a serious fault has occurred); combine the fault mode probability distribution and the device health status evaluation score into a comprehensive risk code (such as a code of the type of fault mode probability distribution - device health status evaluation score), that is, comprehensive risk data.

[0077] Construct two sub - architectures under the deep learning framework using the dynamic routing capsule network and the deep Q - network respectively. For the risk decoding architecture in the sub - architecture, if only using traditional neural networks, it is easy to have the situation of fuzzy feature representation of multi - modal data; using the deep Q - network makes the generated executable strategy more specific and shortens the response time.

[0078] The ways to construct the deep learning framework include:

[0079] The deep learning framework includes two sub-architectures, namely, a risk decoding architecture and a policy generation architecture (the risk decoding architecture and the policy generation architecture are in a progressive relationship, and the policy generation architecture further processes the data generated by the risk decoding architecture); the dynamic routing capsule network is used as the core structure of the risk decoding architecture, and the underlying capsule units and the high-level capsule units are constructed as sub-structures in the core structure (the number of underlying capsule units is set based on the comprehensive risk data, and each underlying capsule unit corresponds to a comprehensive risk code; the high-level capsule units are adjusted based on the number of underlying capsule units); the comprehensive risk data is input into the underlying capsule units to obtain the underlying capsule input vectors corresponding to each piece of comprehensive risk data; the coupling coefficient between the underlying capsule units and the high-level capsule units is initialized to obtain the initialized coupling coefficient (the coupling coefficient is initialized to 0); the capsule weight matrix is initialized to obtain the initialized capsule weight matrix (the initialized capsule weight matrix is obtained by random generation); the underlying capsule input vectors are converted into capsule prediction vectors using the initialized capsule weight matrix (the underlying capsule input vectors are dot-product operated with the initialized capsule weight matrix to obtain the capsule prediction vectors); the weighted sum of all the capsule prediction vectors is calculated to obtain the high-level capsule input vectors; an anti-interference function is constructed based on the high-level capsule input vectors, the anti-interference function is calculated and the result is output to obtain the high-level capsule output vectors (the anti-interference function is the spuash function, and the calculation formula is: where, represents the hi-th high-level capsule output vector; represents the hi-th high-level capsule input vector; each high-level capsule output vector is a semantic vector, and the semantic content included therein has the fault category, the fault occurrence probability, and the fault degree); the coupling coefficient is updated based on the initialized coupling coefficient, the capsule prediction vectors, and the high-level capsule output vectors (the calculation formula for updating the coupling coefficient is: where, represents the updated coupling coefficient between the di-th underlying capsule unit and the hi-th high-level capsule unit; YU diindicating the predicted vector of the \(d_i\)-th capsule); The above process is iterated multiple times until the coupling coefficient converges. At this time, the output vector of the high-level capsule is the output of the risk decoding architecture. At the same time, the initial capsule weight matrix is updated using the backpropagation algorithm (the capsule weight matrix is updated to gradually learn the correlation between the features in the low-level capsule units and the features in the high-level capsule units); The deep Q-network is used as the core structure of the policy generation architecture (the deep Q-network is a reinforcement learning method for evaluating the goodness or badness of an action in a given state), and the core structure of the policy generation architecture is separated into two parallel branches, namely the state value branch and the action advantage branch (the state value branch evaluates the state value of the device state. For example, if the vibration frequency of a certain device suddenly changes, the state value branch evaluates this state; the action advantage branch evaluates the action being performed by the device. For example, if a certain device adjusts the voltage by 50 degrees, the action advantage branch evaluates this action); The high-level capsule output vector is used as the input data of the policy generation architecture (the high-level capsule output vector is used as the shared input data for the two parallel branches). The state value of the input high-level capsule output vector is calculated in the state value branch (in the fully connected layer of the state value branch, the high-level capsule output vector is mapped to a constant, which is the state value of this vector). The action advantage of the input high-level capsule output vector is calculated in the action advantage branch (in the fully connected layer of the action advantage branch, the high-level capsule output vector is mapped to multiple constants, including positive and negative numbers, and the sum is obtained as the action advantage score of this vector); A fusion function is constructed, and the fusion function value of the state value and action advantage of any input high-level capsule output vector is calculated using the fusion function (the calculation formula of the fusion function is: where, represents the fusion function value of any high-level capsule output vector; represents the state value of any high-level capsule output vector; represents the action advantage score of any high-level capsule output vector in the Stu state. The corresponding action and score can be obtained by querying the preset mining area equipment parameter database; For example, when a certain industrial device is in good condition, the state value is 90. At this time, the action advantage score for performing a shutdown operation is -20, and the action advantage score for reducing the load is 10. Then \(Q_H(\text{shutdown}) = 70\), \(Q_H(\text{reduce load}) = 100\)); The high-level capsule output vector with the largest fusion function value in each calculation is subjected to physical interpretability mapping to obtain an executable policy (the form of the executable policy is: (Device A0 power abnormal, temperature mutation) → (Cool down device A0)); All executable policies are integrated to obtain a policy set.

[0080] The methods for generating the priority of device control policies based on comprehensive risk data and deep learning frameworks include:

[0081] Process the comprehensive risk data using a deep learning framework to obtain a set of strategies; classify the urgency and grade the fault intensity for each executable strategy in the set of strategies (query the historical fault records in the preset mine equipment parameter database to judge the urgency and fault intensity of the situation at that time, and assign different weights to different urgencies and fault intensities; for example, grade the fault intensity, including safety intensity, minor fault intensity, major fault intensity, and critical fault intensity; classify the urgency into four levels: normal, general, urgent, and extremely urgent; when a certain equipment is short-circuited due to water seepage, at this time the local function of the equipment cannot operate but it is not completely ineffective, then judge the fault intensity as the major fault intensity, and since the equipment needs to be processed quickly but a short delay is allowed, then judge the urgency as the urgent level; in this embodiment, the weights of safety intensity, minor fault intensity, major fault intensity, and critical fault intensity are respectively taken as 0.1, 0.2, 0.3, and 0.4; the weights of normal, general, urgent, and extremely urgent levels are respectively taken as 0.1, 0.2, 0.3, and 0.4); determine the success rate of the executable strategy in the historical situation by querying the preset mine equipment parameter database to obtain the historical execution effect record (attach a weight to the executable strategy corresponding to the historical execution effect record based on the success rate of the executable strategy in the historical situation, in this embodiment when the success rate is 0%-40%, the weight is taken as 0; when the success rate is 40%-80%, the weight is taken as 0.3; when the success rate is 80%-100%, the weight is taken as 0.7); construct a priority objective function, and calculate the function value of the priority objective function based on the urgency, fault intensity, and historical execution effect record of each executable strategy (the calculation formula of the priority objective function is: YX(CL h0 ) = CL h0 (ω1 + ω2 + ω3); where, YX(CL h0 ) represents the function value of the priority objective function of the h0th executable strategy, and CL h0 (ω1 + ω2 + ω3) represents the sum of the urgency weight, fault intensity weight, and historical execution effect record weight corresponding to the h0th executable strategy); sort all the executable strategies in descending order based on the function value of the priority objective function to obtain an instruction sequence.

[0082] Using an optimization algorithm to adjust the parameters of each industrial equipment replaces manual empirical parameter adjustment, improves the equipment operation efficiency, and at the same time stores the optimal parameter combination in the preset database after each parameter adjustment. When the equipment adjusts parameters next time, the historical optimal parameters can be directly read, greatly reducing the equipment startup preparation time.

[0083] The method of dynamically adjusting the parameters of each industrial equipment in the industrial equipment group based on the instruction sequence includes:

[0084] Query the preset database of mining area equipment parameters to obtain the list of adjustable parameters for each industrial equipment, and determine the number of parameters to be adjusted for each industrial equipment each time (for example, the number of parameters that a certain equipment can adjust is 5); construct a parameter dimension space, map the executable strategies belonging to any one equipment in the instruction sequence to the parameter dimension space, and mark the parameters of the equipment corresponding to each executable strategy (map the specific steps of the executable strategy to parameter changes, for example, reducing the voltage of a certain equipment by 10 volts is mapped to -10 for the voltage of this equipment; marking the parameters is equivalent to separating the parameters corresponding to the same executable strategy of the same equipment from the parameter dimension space for subsequent optimization processing); group the marked parameters based on the number of parameters to be adjusted for each industrial equipment each time, and attach a priority weight to each group of parameters after grouping based on the equipment control strategy priority of the corresponding executable strategy of this equipment in the instruction sequence (for example, the number of marked parameters is 100, but the number of parameters that can be adjusted by this equipment each time is 5, so it is divided into 20 groups, and a priority weight is attached based on the equipment control strategy priority for each adjustment. If the equipment control strategy priority is larger, a larger weight is attached to the parameters in this group); regard each group as an individual and use an optimization algorithm to optimize the population composed of all individuals to obtain the optimal individual, which is the optimal operating parameter combination for any one equipment (perform crossover and mutation operations on the population using a genetic algorithm, calculate the function value of the fitness function of the genetic algorithm, construct a Pareto front based on the function values of the fitness functions of each individual until the maximum number of iterations is reached, obtain the constructed Pareto front, and screen the optimal individual in the Pareto front, which is the optimal operating parameter combination).

[0085] In this embodiment, multi-source data of various equipment in the industrial equipment group in the mining area is collected and preprocessed to obtain more accurate basic equipment operation data; a model is constructed to evaluate the equipment health status of the basic equipment operation data to determine the status of each equipment, and at the same time, comprehensive risk data is obtained; a deep learning framework is constructed to process the comprehensive risk data, generate the equipment control strategy priority for each industrial equipment and sort it to obtain an instruction sequence, and finally, based on various executable strategies in the instruction sequence, parameter adjustment is performed on each industrial equipment to obtain the optimal operating parameter combination for each equipment and send this parameter combination to the database, realizing the automatic control of each equipment in the mining area; compared with existing experience, more accurate preprocessing methods are used, and the most suitable method is used for preprocessing each type of data; the health status of the equipment is evaluated by constructing an updatable model, enabling the model to adapt to the dynamic changes in the working conditions in the mining area; a deep learning framework is constructed to generate the executable strategy for each equipment, resulting in higher efficiency and better compliance with the actual working conditions in the subsequent dynamic parameter adjustment process.

[0086] Embodiment 2

[0087] Please refer to Figure 2 As shown, for the parts not described in detail in this embodiment, refer to the description of Embodiment 1. A device automatic control system based on multi-source data is provided, including:

[0088] A data acquisition module that acquires multi-source data of an industrial equipment group in a mining area, integrates it into multi-device operation data, and preprocesses the multi-device operation data to obtain basic equipment operation data;

[0089] A device health status evaluation module that constructs a multi-modal hybrid model based on the basic equipment operation data, updates the multi-modal hybrid model, and uses the updated multi-modal hybrid model to evaluate the device health status of each industrial equipment to obtain comprehensive risk data;

[0090] A policy priority generation module that constructs a deep learning framework, generates device control policy priorities based on the comprehensive risk data and the deep learning framework, and generates an instruction sequence based on the device control policy priorities;

[0091] A parameter adjustment module that dynamically adjusts the parameters of each industrial equipment in the industrial equipment group based on the instruction sequence to obtain an optimal operation parameter combination; sends the optimal operation parameter combination to a preset mining area equipment parameter database; and each module is connected by wired and / or wireless means.

[0092] Embodiment 3

[0093] This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the above-provided device automatic control method based on multi-source data.

[0094] Since the electronic device introduced in this embodiment is the electronic device used to implement a device automatic control method based on multi-source data in an embodiment of the present application, based on the device automatic control method based on multi-source data introduced in an embodiment of the present application, those skilled in the art can understand the specific implementation manner and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in an embodiment of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device used to implement a device automatic control method based on multi-source data in an embodiment of the present application, it falls within the scope of protection of the present application.

[0095] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters and threshold selection in the formulas are set by those skilled in the art according to the actual situation.

[0096] The above are only the preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the concept of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. An automatic control method for a device based on multi-source data, characterized in that, Including: S1. Collect multi-source data of the industrial equipment group in the mining area and integrate it into multi-device operation data, preprocess the multi-device operation data to obtain basic equipment operation data; S2. Build a multi-modal hybrid model based on the basic equipment operation data and update the multi-modal hybrid model, use the updated multi-modal hybrid model to evaluate the equipment health status of each industrial equipment to obtain comprehensive risk data; S3. Build a deep learning framework, generate equipment control strategy priorities based on the comprehensive risk data and the deep learning framework, and generate an instruction sequence based on the equipment control strategy priorities; S4. Dynamically adjust the parameters of each industrial equipment in the industrial equipment group based on the instruction sequence to obtain an optimal operation parameter combination; send the optimal operation parameter combination to the preset mining area equipment parameter database.

2. The automatic control method for a device based on multi-source data according to claim 1, characterized in that The method for preprocessing the multi-device operation data includes: The multi-device operation data includes environmental monitoring data, industrial equipment data, and geological monitoring data, where the environmental monitoring data includes gas concentration data, dust concentration data, temperature data, and humidity data; the industrial equipment data includes excavation equipment vibration data and conveyor belt operation data; the geological monitoring data includes mine roadway displacement data and water level monitoring data; Perform data cleaning on the environmental monitoring data, build a temperature and humidity non-linear coupling model to eliminate sensor errors in the cleaned environmental monitoring data, and perform normalization processing on the environmental monitoring data after error elimination to obtain environmental monitoring standard data; for the acquisition equipment vibration data in the industrial equipment data, use wavelet packet decomposition to extract multi-band energy features and remove noise to obtain denoised excavation equipment data; build a three-dimensional state matrix based on the conveyor belt operation data to quantify the conveyor belt operation state and perform feature extraction to obtain a mixed matrix; perform standardization processing on the mixed matrix to obtain standard conveyor belt operation data, and perform spatio-temporal alignment on the denoised excavation equipment data and the standard conveyor belt operation data to obtain equipment operation feature data; perform noise suppression and fill in missing data on the geological monitoring data to obtain geological safety data; combine the environmental monitoring standard data, equipment operation feature data, and geological safety data into basic equipment operation data.

3. The automatic control method for a device based on multi-source data according to claim 2, wherein, The method for performing data cleaning on the environmental monitoring data includes: Construct a dynamic window and set the initial window length based on the sampling frequencies of gas concentration data and dust concentration data; when the dynamic window is sliding, add natural boundary conditions to the environmental monitoring data points at both ends of the dynamic window, and simultaneously calculate the gradient change rate of adjacent environmental data points in real time. When the change amplitude of the gradient change rate is greater than or equal to the preset gradient fluctuation threshold, trigger the window adjustment mechanism; use the cubic spline interpolation optimization algorithm to fill in the missing values of all environmental monitoring data points within the sliding window at any given moment to obtain complete environmental monitoring data; design a multi-physical parameter joint distribution anomaly detection mechanism. When the gas concentration change rate within the sliding window at any given moment is greater than the preset concentration change threshold, if the temperature change rate exceeds the preset temperature change threshold, mark all the complete environmental monitoring data points within the corresponding sliding window; conversely, if the temperature does not change synchronously, then enable the combustion interference elimination algorithm; when the humidity within the sliding window at any given moment is greater than the preset relative humidity threshold, if the fluctuation of the dust concentration change rate within the sliding window exceeds the theoretical sedimentation rate, mark all the complete environmental monitoring data points within the sliding window; calculate the outliers among all the marked points based on the Mahalanobis distance method. The outliers are the abnormal points, and delete the abnormal points to obtain the cleaned environmental monitoring data; The methods for eliminating sensor errors from the cleaned environmental monitoring data include: Take the multi-modal compensation network as the basic structure of the temperature-humidity non-linear coupling model, decouple the temperature-humidity coupling effect based on Laguerre orthogonal polynomials, construct a two-channel joint compensation function and eliminate sensor errors based on the two-channel joint compensation function. At the same time, construct an adaptive compensation function based on the aging attenuation coefficient and the temperature change rate to dynamically adjust the compensation period to obtain the environmental monitoring data after error elimination.

4. The automatic control method of a device based on multi-source data according to claim 3, characterized in that, The method for extracting multi-band energy features and removing noise from the vibration data of the acquisition device in the industrial equipment data includes: Perform N-layer wavelet packet decomposition on the vibration data of the acquisition device through a preset wavelet basis function to generate 2^N frequency band nodes; calculate the node energy of each frequency band node and calculate the energy proportion of each frequency band node based on the node energy of each frequency band node; sort the frequency band nodes in descending order based on the energy proportions of all frequency band nodes to obtain a frequency band node sequence; set a frequency band node energy proportion threshold, determine the frequency band nodes with energy proportions greater than the frequency band node energy proportion threshold in the frequency band node sequence as high-energy nodes and save the original coefficients, and determine the remaining frequency band nodes as low-energy nodes; construct an adaptive threshold function to perform dynamic threshold processing on all low-energy nodes to obtain a denoised energy node set; merge all high-energy nodes and the denoised energy node set for signal reconstruction and calculate the signal-to-noise ratio index at this time. If the signal-to-noise ratio index is less than the expected value, perform multiple rounds of processing on all frequency band nodes after signal reconstruction using wavelet packet decomposition until the signal-to-noise ratio index is greater than or equal to the expected value to obtain the denoised mining equipment data; The method for constructing a three-dimensional state matrix to quantify the running state of the conveyor belt based on the conveyor belt running data and extracting features includes: Construct a three-dimensional coordinate system, extract the timestamps of each data point in the conveyor belt operation data as the time axis, construct a time window on the time axis and adjust the size of the time window based on the length of the time axis to ensure that it can cover the full-cycle state of the conveyor belt operation; integrate the physical space distributions of sensors deployed at different positions on the conveyor belt and use it as the sensor axis; extract the physical parameters in the conveyor belt operation data as the parameter axis; couple the time axis, the sensor axis and the parameter axis to construct a three-dimensional state matrix; expand the three-dimensional state matrix within each time window into a two-dimensional matrix and calculate the covariance matrix of this two-dimensional matrix, and perform eigen decomposition on each covariance matrix to obtain the tensor state features.

5. The automatic control method of a device based on multi-source data according to claim 4, characterized in that, The method for suppressing noise and filling missing data in the geological monitoring data includes: Regard the mine roadway displacement data and water level monitoring data in the geological monitoring data as multi-dimensional vector data and map them to the vector space; use the principal component analysis algorithm to reduce the dimension of the multi-dimensional vector data and perform spatio-temporal alignment; construct a three-dimensional vector field, map the multi-dimensional vector data to the three-dimensional vector field, and any vector in the three-dimensional vector field is a part of the multi-dimensional vector data; determine the geological structure trend in this mining area by querying the mine roadway displacement data or water level monitoring data in the preset mining area equipment parameter database, construct a filtering window based on the geological structure trend and adjust its shape; calculate the two-norm distance between each vector in the filtering window and any other vector; introduce a direction-sensitive weight and use this weight to weight each vector to obtain a weighted vector, and at the same time adjust the direction-sensitive weights of vectors in different directions according to the geological structure trend; sum up the two-norm distances of each weighted vector, that is, the sum of the total norm distances from each weighted vector to other weighted vectors, to obtain the cumulative norm distance of each weighted vector; take the weighted vector with the minimum cumulative norm distance as the output vector; perform inverse mapping on the output vector to obtain the filtered geological data; Construct a missing data filling model, and use the pre-trained LSTM model as the basic structure of the missing data filling model; introduce physical constraints into the loss function of the missing data filling model and reconstruct the constraint loss function; use the filtered geological data as the input data of the input layer of the missing data filling model, and dynamically fill the input data through the missing data filling model to obtain the geological safety data.

6. The automatic control method for a device based on multi-source data according to claim 5, characterized in that, The method for constructing a multi-modal hybrid model based on the basic equipment operation data and updating the multi-modal hybrid model includes: Classify the basic device operation data into numerical type data and matrix type data based on the data type; construct a dual-channel hybrid structure and use this structure as the core structure of the multi-modal hybrid model; the dual-channel hybrid structure includes a temporal convolutional channel and a self-attention channel for processing numerical type data and matrix type data respectively; adopt a depthwise separable temporal convolutional network as the basic structure of the temporal convolutional channel; shuffle the numerical type data into disordered numerical data and input this data into the causal convolutional layer of the temporal convolutional channel, and use the multi-scale convolutional kernels in the causal convolutional layer to extract the local temporal patterns of the disordered numerical data; at the same time, introduce a dilated convolutional layer, construct a dilation factor and use this factor to gradually expand the time coverage of the temporal convolutional channel; add a channel attention mechanism, extract the statistical features of the disordered numerical data through global average pooling and max pooling, and generate a sensor weight vector through the fully connected layer of the temporal convolutional channel; construct a spatio-temporal two-dimensional attention network as the basic structure of the self-attention channel, perform a linear transformation on the matrix type data to generate a triple; at the same time, input the triple into the self-attention channel and embed the device physical topology constraint matrix in the spatial dimension; calculate the attention score based on the device physical topology constraint matrix and the triple; normalize the attention score to generate a probability weight, and perform weighted aggregation on all elements in the triple based on the probability weight to obtain a spatio-temporal dependence relationship output matrix; perform adaptive weighted fusion on the outputs of the two channels through the dynamic gating fusion layer of the multi-modal hybrid model to generate a joint feature representation; Pre-train three groups of structurally heterogeneous expert models and deploy them in the dynamic gating fusion layer, namely a spatial feature expert model, a temporal feature expert model, and a cross-modal expert model; input the joint feature representation into the dynamic gating fusion layer and calculate the expert activation probability; construct an elastic selection mechanism based on the reinforcement learning algorithm and dynamically adjust the number of activated expert models based on this mechanism; encode the joint feature representation into a state vector; construct a reward function, and update the parameters of the dynamic gating fusion layer using the policy gradient algorithm based on the state vector, the number of activated expert models, and the reward function; The methods for evaluating the device health status of each industrial device include: Process the joint feature representation in the output layer of the multi-modal hybrid data and compare it with the fault categories in the preset mine area device parameter database to obtain the fault mode probability distribution; at the same time, extract the device features in the joint feature representation, construct a graph attention network and calculate the risk propagation coefficient between each node in the network; construct a scoring function based on the fault mode probability distribution and the risk propagation coefficient, and output the device health status evaluation score based on the scoring function; combine the fault mode probability distribution and the device health status evaluation score into a comprehensive risk code, that is, comprehensive risk data.

7. An automatic control method for a device based on multi-source data according to claim 6, characterized in that, The methods for constructing the deep learning framework include: The deep learning framework includes two sub-architectures, namely a risk decoding architecture and a policy generation architecture. The dynamic routing capsule network is used as the core structure of the risk decoding architecture, and the underlying capsule unit and the high-level capsule unit are constructed as sub-structures in the core structure. The comprehensive risk data is input into the underlying capsule unit to obtain the underlying capsule input vector corresponding to each piece of comprehensive risk data. The coupling coefficient between the underlying capsule unit and the high-level capsule unit is initialized to obtain the initialized coupling coefficient. The capsule weight matrix is initialized to obtain the initialized capsule weight matrix. The initialized capsule weight matrix is used to convert the underlying capsule input vector into a capsule prediction vector. The weighted sum of all capsule prediction vectors is calculated to obtain the high-level capsule input vector. An anti-interference function is constructed based on the high-level capsule input vector, the anti-interference function is calculated and the result is output to obtain the high-level capsule output vector. The coupling coefficient is updated based on the initialized coupling coefficient, the capsule prediction vector and the high-level capsule output vector. The above process is iterated multiple times until the coupling coefficient converges. At this time, the high-level capsule output vector is the output of the risk decoding architecture, and the initialized capsule weight matrix is updated using the backpropagation algorithm. The deep Q network is used as the core structure of the policy generation architecture, and the core structure of the policy generation architecture is separated into two parallel branches, namely the state value branch and the action advantage branch. The high-level capsule output vector is used as the input data of the policy generation architecture. The state value of the input high-level capsule output vector is calculated in the state value branch, and the action advantage of the input high-level capsule output vector is calculated in the action advantage branch. A fusion function is constructed, and the fusion function value of the state value and the action advantage of any input high-level capsule output vector is calculated using the fusion function. The high-level capsule output vector with the largest fusion function value in each calculation is subjected to physical interpretability mapping to obtain an executable policy. All executable policies are integrated to obtain a policy set.

8. An automatic control method for a device based on multi-source data according to claim 7, characterized in that, The method for generating the priority of the device control policy based on the comprehensive risk data and the deep learning framework includes: Using the deep learning framework to process the comprehensive risk data to obtain a policy set; classifying the emergency level and grading the fault intensity of each executable policy in the policy set; determining the success rate of the executable policy in the historical situation by querying the preset mine equipment parameter database to obtain the historical execution effect record; constructing a priority objective function, and calculating the function value of the priority objective function based on the emergency level, fault intensity and historical execution effect record of each executable policy; sorting all executable policies in descending order based on the function value of the priority objective function to obtain an instruction sequence.

9. The device automatic control method based on multi-source data according to claim 8, characterized in that The method for dynamically adjusting the parameters of each industrial device in the industrial device group based on the instruction sequence includes: Query the preset database of mining area equipment parameters to obtain the adjustable parameter list of each industrial equipment, and determine the number of parameters adjusted each time for each industrial equipment; construct a parameter dimension space, map the executable strategies belonging to any one equipment in the instruction sequence to the parameter dimension space, and mark the parameters of the equipment corresponding to each executable strategy; group the marked parameters based on the number of parameters adjusted each time for each industrial equipment, and attach a priority weight to each group of grouped parameters based on the priority of the equipment control strategy of the corresponding executable strategy of the equipment in the instruction sequence; regard each group as an individual and use an optimization algorithm to optimize the population composed of all individuals to obtain the optimal individual, which is the optimal operating parameter combination of any one equipment.

10. An automatic control system for devices based on multi-source data, which is used to implement an automatic control method for devices based on multi-source data according to any one of claims 1 to 9, characterized in that, Including: A data acquisition module that collects multi-source data of the industrial equipment group in the mining area and integrates it into multi-equipment operation data, and preprocesses the multi-equipment operation data to obtain basic equipment operation data; An equipment health status evaluation module that constructs a multi-modal hybrid model based on the basic equipment operation data and updates the multi-modal hybrid model, and uses the updated multi-modal hybrid model to evaluate the equipment health status of each industrial equipment to obtain comprehensive risk data; A strategy priority generation module that constructs a deep learning framework, generates equipment control strategy priorities based on the comprehensive risk data and the deep learning framework, and generates an instruction sequence based on the equipment control strategy priorities; A parameter adjustment module that dynamically adjusts the parameters of each industrial equipment in the industrial equipment group based on the instruction sequence to obtain an optimal operating parameter combination; sends the optimal operating parameter combination to the preset mining area equipment parameter database; each module is connected by wired and / or wireless means.

Images