Coal mill degradation degree evaluation method based on deep learning of sound-vibration-temperature multi-modal fusion

Abstract

The application discloses a sound-vibration-temperature multi-modal fusion coal mill degradation degree evaluation method based on deep learning, and relates to the technical field of sound-vibration-temperature multi-modal fusion coal mills. A sound-vibration-temperature multi-modal time series data set is constructed; high-dimensional time series dynamic characteristic sequences of each mode are obtained; splicing and reorganization are performed in a characteristic channel dimension to generate a sound-vibration-temperature multi-modal initial fusion characteristic tensor; an angle lizard optimization algorithm model with a characteristic saliency degree measurement as a target function is constructed, and an optimal mode characteristic weighting matrix is calculated; a multi-modal degradation sensitive representation set resistant to background noise interference is synthesized; and a final coal mill degradation degree evaluation index is output. The application ensures that a global receptive field of historical data is maintained under extremely low computing overhead, and significantly improves the efficiency and accuracy of long-period time series modeling.

Description

Technical Field

[0001] This invention relates to the field of acoustic, vibration, and temperature multimodal fusion coal mill technology, and in particular to a deep learning-based method for assessing the deterioration of an acoustic, vibration, and temperature multimodal fusion coal mill. Background Technology

[0002] With the rapid development of modern large-scale thermal power units and industrial automation technology, the coal mill, as the core hub equipment of the pulverizing system, directly affects the safety, stability, and combustion efficiency of the entire unit. Therefore, real-time and accurate assessment of the coal mill's degradation level has become a standard requirement in the field of industrial intelligent operation and maintenance. In real-world operating environments characterized by high noise, strong vibration, and complex heat exchange, monitoring equipment degradation faces numerous challenges.

[0003] Traditional coal mill monitoring methods typically rely on threshold alarms for single-mode signals or simple independent parallel processing of acoustic, vibration, and temperature signals. In industrial settings involving complex, multi-source, heterogeneous data, this often leads to one-sided feature extraction. Traditional time-domain alignment methods fail to consider the thermal inertial hysteresis effect of temperature sensors and the high-frequency transients and asynchronous delays in the propagation of acoustic and vibration signals, resulting in significant spatiotemporal misalignment errors introduced in the early fusion stage.

[0004] Existing deep learning time series prediction models generally suffer from long-range memory loss or quadratic explosion of computational complexity when dealing with ultra-long-cycle equipment operation sequences. They are unable to effectively capture the long-span underlying dynamic evolution of coal mills from health to slight deterioration and then to serious failure under limited computing power.

[0005] Traditional multimodal feature fusion mechanisms typically employ simple rigid splicing or result-based voting, failing to bridge the significant high-dimensional heterogeneity gap between different physical signals. When faced with complex background noise that drowns out weak degradation signals, existing feature weight allocation methods often rely on manual experience or conventional heuristic optimization algorithms. Heuristic optimization algorithms are prone to getting stuck in local optima when dealing with high-dimensional and non-convex fusion feature spaces, failing to accurately remove redundant features and adaptively amplify sensitive degradation characteristics. Furthermore, traditional regression evaluation models lack resilience to transient abnormal conditions, easily leading to drastic jumps and artifact perturbations in the final output health assessment index, failing to provide field maintenance personnel with a smooth, stable, and forward-looking degradation assessment benchmark. Summary of the Invention

[0006] One objective of this invention is to propose a deep learning-based method for evaluating the deterioration of coal mills using a multimodal fusion of acoustic, vibration, and temperature data. This invention ensures the maintenance of a global receptive field for historical data with extremely low computational overhead, significantly improving the efficiency and accuracy of long-cycle time-series modeling.

[0007] According to an embodiment of the present invention, a deep learning-based multimodal fusion method for assessing the degradation degree of a coal mill based on acoustic, vibration, and temperature modulation includes:

[0008] The system synchronously acquires continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences during the operation cycle of the target coal mill, and performs spatiotemporal synchronization and resampling alignment operations on multi-source heterogeneous signals to construct a multi-modal time-series data set of acoustic, vibration, and temperature.

[0009] The acoustic, vibration and temperature multimodal time series data set is input into the pre-constructed Mamba-2 parallel state space model. The structured state space duality mechanism is used to perform long-range context dependency modeling on acoustic, vibration and temperature signals respectively, and obtain high-dimensional time series dynamic feature sequences that are independent of each mode.

[0010] Perform cross-modal cross-mapping and joint alignment on each modally independent high-dimensional temporal dynamic feature sequence, and splice and recombine them in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration and temperature multimodal modes.

[0011] An optimization algorithm model for horned lizards was constructed with feature salience measurement as the objective function. The initial fusion feature tensor of acoustic, vibration and temperature multimodal features was used as the optimization search space. Global and local iterative optimization was performed by executing the adaptive camouflage mechanism and covert predation and wandering strategy of the horned lizard population. The optimal modal feature weighting matrix was calculated.

[0012] The optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic, vibration and temperature multimodal modes are subjected to Hadamard product modulation and spatial dimension reconstruction to synthesize a multimodal degradation sensitive characterization set that is resistant to background noise interference.

[0013] The multimodal degradation sensitive characterization set is input into the nonlinear evaluation mapper for feature decoding and regression calculation, and the final coal mill degradation assessment index is output.

[0014] Optionally, the synchronous acquisition of continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences during the target coal mill's operating cycle, and the execution of spatiotemporal synchronization and resampling alignment operations for multi-source heterogeneous signals, constructs a multi-modal time-series data set of acoustic, vibration, and temperature, including:

[0015] By synchronously collecting raw sound, vibration and temperature signals in the environmental background through sensor arrays deployed in key parts of the coal mill, an uncleaned multi-source heterogeneous data stream is formed.

[0016] Fractional-order calculus filtering is applied to the original acoustic sequence in the uncleaned multi-source heterogeneous data stream to enhance the abrupt high-frequency features and denoise the steady state, resulting in a purified acoustic sequence.

[0017] Hilbert envelope demodulation and polynomial baseline elimination are performed on the original vibration sequence in the uncleaned multi-source heterogeneous data stream to obtain the cleaned shell vibration sequence.

[0018] A first-order inertial inverse model is applied to the original temperature sequence in the uncleaned multi-source heterogeneous data stream for dynamic thermal hysteresis compensation, and the temperature sequence of the key purification node is obtained.

[0019] Based on the vibration sequence of the purification shell, the cross-correlation function is calculated using a phase-locked loop to compensate for the time delay between the acoustic sound generation sequence of the purification and the temperature sequence of the key purification nodes, thereby generating a time-domain aligned multimodal sequence.

[0020] A cubic spline interpolation algorithm is applied to time-domain aligned multimodal sequences for uniform step size resampling and alignment, and then stacked in the channel dimension to construct an acoustic-vibration-temperature multimodal time series dataset.

[0021] Optionally, the feature is that the step of inputting the multimodal time-series data set of acoustic, vibration, and temperature into a pre-constructed Mamba-2 parallel state-space model, and using a structured state-space dual mechanism to perform long-range context-dependent modeling of acoustic, vibration, and temperature signals respectively, to obtain high-dimensional time-series dynamic feature sequences that are independent of each mode, including:

[0022] The set of acoustic, vibration and temperature multimodal time series data is input into the linear projection layer to perform spatial dimension upsizing and form the initial hidden state tensor.

[0023] Initialize the continuous evolution parameters of the state-space model and apply a zero-order hold for discretization transformation to construct a discretized state-space parameter group;

[0024] A causal one-dimensional depthwise convolution operation is performed on the initial hidden state tensor to extract local transient fluctuations, resulting in a local temporal-aware feature sequence.

[0025] Based on the discretized state-space parameter group, a semi-separable matrix and prefix sum parallel scan are applied to the local time-series sensing feature sequence to perform structured state-space dual computation and output a global long-range sensing state sequence.

[0026] The global long-range sensing state sequence is modulated by element-wise multiplication with the nonlinear gated branch and then linear dimensionality reduction is performed to obtain a high-dimensional temporal dynamic feature sequence that is independent of each mode.

[0027] Optionally, the step of performing cross-modal cross-mapping and joint alignment on the modally independent high-dimensional temporal dynamic feature sequences, and splicing and recombining them in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration, and temperature multimodal modes, includes:

[0028] By using radial basis kernel functions, the modally independent high-dimensional temporal dynamic feature sequences are mapped to a shared high-dimensional manifold topological space, resulting in a manifold projection aligned feature set.

[0029] Calculate the dynamic time warping distance of the temporal evolution trajectory of each mode in the manifold projection alignment feature set, and construct the cross-modal geodesic distance correlation matrix;

[0030] Based on the threshold set by the cross-modal geodesic distance correlation matrix, a sparse cross-modal dynamic graph topology is constructed to form a sparse adjacency weight matrix.

[0031] Based on the sparse adjacency weight matrix, weighted aggregation and nonlinear propagation of node features are performed on the graph structure to obtain the enhanced feature representations of each modality after cross-modal graph neural alignment;

[0032] The enhanced feature representations of each modality after cross-modal graph neural alignment are stacked sequentially by channel and then compressed by dimensionality reduction through a fully connected bottleneck layer to generate the initial fusion feature tensor of acoustic, vibration and temperature multimodal modes.

[0033] Optionally, the method is characterized in that the construction of the horned lizard optimization algorithm model with feature salience measurement as the objective function, using the initial fused feature tensor of acoustic, vibrational, and thermal multimodal features as the optimization search space, and performing global and local iterative optimization by executing the adaptive camouflage mechanism and covert predation movement strategy of the horned lizard population, calculates the optimal modal feature weighting matrix, including:

[0034] A feature saliency objective function is constructed by combining the information entropy reward term and the inter-class variance penalty term to establish a fitness evaluation benchmark;

[0035] Using the channel dimension of the initial fusion feature tensor of acoustic-vibration-temperature multimodal modes as the search boundary, an initial horned lizard population position matrix containing multiple sets of candidate feature weights is randomly generated.

[0036] Based on the fitness evaluation benchmark, a local optimum is selected, and adaptive camouflage position update is performed by smoothly approaching it to obtain the position matrix of the first candidate population.

[0037] The individuals trapped in the iteration stagnation in the first candidate population position matrix are injected with a logistic chaotic perturbation sequence to perform escape transitions, thereby generating the second candidate population position matrix.

[0038] The second candidate population position matrix is ​​guided to perform a Levy flight walk with a heavy-tailed distribution towards the globally optimal individual. After iterative optimization, the combination of the globally optimal horned lizard individual position variables is extracted.

[0039] Perform absolute value normalization and tensor structure transformation on the global optimal combination of individual horned lizard position variables to output the optimal modal feature weighting matrix.

[0040] Optionally, the step of performing Hadamard product modulation and spatial dimension reconstruction on the optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic-vibration-temperature multimodal modes to synthesize a multimodal degradation-sensitive characterization set resistant to background noise interference includes:

[0041] The optimal modal feature weighting matrix is ​​broadcast and extended in the spatial dimension to construct a dimension-extended weighting tensor with the same shape as the initial fusion feature tensor of the acoustic, vibration and temperature multimodal modes.

[0042] The weighted feature tensor is obtained by performing an element-wise Hadamard product operation on the dimension-expanded weighted tensor and the initial fusion feature tensor of the acoustic-vibration-temperature multimodal modes.

[0043] The local mean and maximum values ​​are extracted by applying a multi-scale sliding window in parallel to the weighted modulation fusion feature tensor and then performing depth concatenation to output a multi-scale spatial feature convergence tensor.

[0044] Perform subchannel group decomposition, transpose, and flattening interleaving operations on the multi-scale spatial feature convergence tensor to generate a channel interleaving reconstructed feature map;

[0045] A soft-threshold shrinkage nonlinear activation function is applied to the channel interleaving reconstructed feature map to remove minor background noise, and the output is a multimodal degradation sensitive characterization set that is resistant to background noise interference.

[0046] Optionally, the step of inputting the multimodal degradation-sensitive characterization set into a nonlinear evaluation mapper for feature decoding and regression calculation, and outputting the final coal mill degradation assessment index, includes:

[0047] The set of multimodal degradation sensitive representations that resist background noise interference is spatially flattened along the channel priority order to form a one-dimensional global representation vector.

[0048] The one-dimensional global representation vector is input into a deep network with Gaussian error linear unit activation function for progressive abstraction and nonlinear mapping, and the output is nonlinear decoded high-order features;

[0049] In the forward computation of the hidden layer, a random deactivation regularization mechanism is applied to the nonlinear decoding high-order features to suppress overfitting and output robust high-order features.

[0050] The robust high-order features are passed to the linear regression layer, and the continuous degradation assessment scalar prediction value is output by calculating the algebraic sum of the inner product and the global bias.

[0051] The scalar prediction value of continuous degradation assessment output from continuous time slices is subjected to historical trend fusion and high-frequency anti-jitter algorithm by applying the exponential moving weighted average filtering algorithm to output the final coal mill degradation assessment index.

[0052] The beneficial effects of this invention are:

[0053] (1) This invention introduces a first-order inertial link inverse model and fractional-order calculus filtering, combined with a semi-separable matrix structure and a mechanism of prefix sum parallel forward scan computation, which enables the low-level physical reality restoration and long-range context-dependent modeling of acoustic, vibration and temperature signals with linear complexity. This improves the integrity of feature extraction under ultra-long operating cycles and can accurately extract deep dynamic evolution laws even when early weak wear features are submerged by the background noise of massive normal operating equipment. This invention uses the structured state-space duality mechanism of Mamba-2 to ensure that the global receptive field of historical data is maintained with extremely low computational overhead, which significantly improves the efficiency and accuracy of long-cycle time series modeling.

[0054] (2) This invention constructs a sparse adjacency matrix based on cross-modal geodesic distance and introduces a horned lizard population iteration mechanism using the logistic chaotic escape operator and the Lévy flight random walk strategy. This enables the model to accurately and adaptively find the optimal modal weighting matrix that maximizes information entropy and inter-class variance in an extremely complex high-dimensional non-convex feature space. By introducing an improved horned lizard algorithm based on chaotic escape, this invention can instantly escape local extremum traps and dynamically adjust the multi-source feature fusion weights. This allows the network to adaptively amplify high-value channels containing specific degradation mechanisms in complex acoustic, vibration, and temperature heterogeneous data, resulting in a significant improvement in the accuracy of anti-interference fusion when processing industrial field data with extremely low signal-to-noise ratios.

[0055] (3) This invention combines a multi-scale spatial feature interweaving reconstruction mechanism with a soft-threshold shrinkage nonlinear activation function and a robust regression decoder with exponential moving average weighted filtering to perform high-order denoising and trend smoothing prediction on the extracted multi-modal degradation characteristics at the microscopic physical dimension. The soft-threshold shrinkage completely eliminates numerical jitter caused by residual weak background noise, and the EMA filtering algorithm deeply integrates long-term historical stable trends to resist transient spikes, accurately outputting a stable and reliable coal mill degradation assessment index. This invention superimposes multi-scale translation invariance extraction and historical trend weighting at the network end to ensure that the model can both keenly track transient real degradation and effectively filter out high-frequency noise. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a deep learning-based multimodal fusion method for evaluating the deterioration of a coal mill. Detailed Implementation

[0058] Example 1: Reference Figure 1 A deep learning-based multimodal fusion method for assessing the deterioration of coal mills based on acoustic, vibration, and temperature parameters includes:

[0059] The system synchronously acquires continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences during the operation cycle of the target coal mill, and performs spatiotemporal synchronization and resampling alignment operations on multi-source heterogeneous signals to construct a multi-modal time-series data set of acoustic, vibration, and temperature.

[0060] In this embodiment, continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences are simultaneously acquired during the operating cycle of the target coal mill. Spatiotemporal synchronization and resampling alignment of multi-source heterogeneous signals are then performed to construct a multi-modal time-series data set of acoustic, vibration, and temperature signals, including:

[0061] An acoustic pickup array calibrated by frequency band, a piezoelectric accelerometer network, and an infrared thermal imaging temperature measurement node are deployed on the outside of the target coal mill cylinder and on the surface of key transmission components. The original acoustic sound emission sequence, the original shell vibration sequence, and the original temperature sequence of key nodes are acquired through a multi-channel high-frequency data acquisition front end, forming an uncleaned multi-source heterogeneous data stream.

[0062] Fractional-order calculus filtering operators are introduced into the original acoustic emission sequence in the uncleaned multi-source heterogeneous data stream for noise reduction and purification. The fractional derivative of the original acoustic signal within the sliding time window is calculated. The infinite series of the product of the combination coefficients of the gamma function and the historical observations of the original acoustic signal is calculated to enhance the transient high-frequency change components in the acoustic signal generated by the impact of steel balls inside the coal mill, and suppress the slowly changing background steady-state noise. Hamming window is used for overlapping frame processing to form a purified acoustic emission sequence.

[0063] Adaptive envelope demodulation and baseline drift elimination are performed on the original shell vibration sequence in the uncleaned multi-source heterogeneous data stream. By calculating the Hilbert transform of the original shell vibration signal, the convolution integral of the original signal and the reciprocal function of the time variable is calculated to construct the vibration analytical signal. The complex modulus of the analytical signal is extracted as the vibration amplitude envelope. Low-frequency baseline fluctuations are subtracted by polynomial fitting to obtain the cleaned shell vibration sequence.

[0064] Dynamic compensation based on the first-order inertial link inverse model is performed on the original key node temperature sequence in the uncleaned multi-source heterogeneous data stream. The discrete first-order difference value of the original temperature sequence with respect to time is multiplied by the sensor calibration time constant and then algebraically added to the original temperature sequence itself at the corresponding time point to restore the real transient temperature change trend and obtain the purification key node temperature sequence.

[0065] Using the sampling time axis of the vibration sequence of the purification shell as the reference time axis, the cross-correlation function between the purification acoustic emission sequence and the reference time axis is calculated using a preset phase-locked loop algorithm. The product integral of the two under different time sliding offsets is calculated, and the time offset corresponding to the global maximum value of the cross-correlation function is located as the time delay difference between modes. Based on the time delay difference, the purification acoustic emission sequence and the temperature sequence of the key purification node are time-shifted and compensated to complete the absolute time-domain alignment of the multi-source signals and generate a time-domain aligned multi-mode sequence.

[0066] To address the physical constraint of inconsistent sampling frequencies among modalities in time-domain aligned multimodal sequences, a cubic spline interpolation algorithm is employed for resampling alignment. A smooth curve composed of piecewise cubic polynomials is constructed between adjacent discrete sampling time points. The polynomial coefficients within each segment interval are established, ensuring that the entire spliced ​​curve satisfies the smoothing conditions of continuous function values, continuous first derivative, and continuous second derivative at the connection points. By re-evaluating the smooth curve at a unified frequency sampling point, acoustic, vibration, and temperature signals are mapped onto a unified feature grid with the same time resolution. The resampled three modal data are then stacked along the feature channel dimension to generate a tensor-form multimodal time-series data set of acoustic, vibration, and temperature.

[0067] The acoustic, vibration and temperature multimodal time series data set is input into the pre-constructed Mamba-2 parallel state space model. The structured state space duality mechanism is used to perform long-range context dependency modeling on acoustic, vibration and temperature signals respectively, and obtain high-dimensional time series dynamic feature sequences that are independent of each mode.

[0068] In this embodiment, the multimodal time-series data set of acoustic, vibration, and temperature is input into a pre-constructed Mamba-2 parallel state-space model. The structured state-space duality mechanism is used to perform long-range context-dependent modeling of the acoustic, vibration, and temperature signals, respectively, resulting in independent high-dimensional temporal dynamic feature sequences for each mode, including:

[0069] The multimodal time series data set of acoustic vibration and temperature is input into the linear projection layer along the time step direction. By calculating the matrix multiplication of each time slice data vector with the learnable initial projection weight matrix, it is mapped to a high-dimensional latent space of a preset dimension to form the initial latent state tensor of the initial representation dimension.

[0070] Based on the fundamental physical definition of the state-space model, the core parameters of the partial differential equations of the system state transition in the continuous time domain are initialized, including the continuous system evolution matrix, the continuous input control matrix, and the continuous observation matrix. The zero-order hold assumption is introduced to perform the discretization transformation of the parameters. By calculating the matrix exponent of the product of the continuous system evolution matrix and the discrete time step scalar, the infinite term matrix polynomial summation based on Taylor series expansion is calculated to obtain the discretized state transition matrix. At the same time, the inverse matrix of the difference between the discrete time step, the discrete state transition matrix and the identity matrix, and the continuous input control matrix are multiplied together to obtain the discretized input control matrix, thus constructing the discretized state-space parameter group.

[0071] A causal one-dimensional deep convolution operation is applied to the initial hidden state tensor in the time dimension. A sliding convolution kernel with a preset odd length is used to extract features using information from the current time step and historical time steps to prevent premature leakage of information from future time steps. By calculating the weighted sum of data within the local time window, the local transient fluctuation features of each modal signal at a short time scale are captured, and a local time-series-aware feature sequence is generated.

[0072] The structured state-space dual mechanism at the core of the Mamba-2 architecture is activated for the local time-series perceptual feature sequence. The feature mapping process of the input sequence is transformed into a quadratic matrix multiplication form similar to self-attention. A semi-separable matrix structure derived from the discretized state transition matrix is ​​constructed. The diagonal blocks of the semi-separable matrix contain the fast dynamic transition information of the local time steps, while the off-diagonal blocks encode the decay memory of long-range historical state information through low-rank decomposition. To break the serial computation bottleneck of traditional recurrent neural networks, the associative property of the semi-separable matrix is ​​used to perform parallel forward scan computation along the length of the entire input sequence. The hardware-aware prefix sum algorithm is used to compute the cumulative product and summation of state transfer in parallel between different data computation blocks. Thus, while maintaining linear computational complexity, the global receptive field coverage of the entire historical long sequence is achieved, and the output is a global long-range perceptual state sequence that captures the long-cycle evolution law of the coal mill deterioration trend.

[0073] The global long-range sensing state sequence is multiplied at the corresponding position level with the nonlinear gated branch after activation function processing to achieve dynamic amplification of key degradation features and suppression of redundant background features. Then, a linear projection transformation is performed to compress and restore the feature dimension, and the output is a high-dimensional time-series dynamic feature sequence that represents the deep dynamic evolution of acoustics, vibration and temperature.

[0074] Perform cross-modal cross-mapping and joint alignment on each modally independent high-dimensional temporal dynamic feature sequence, and splice and recombine them in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration and temperature multimodal modes.

[0075] In this embodiment, cross-modal cross-mapping and joint alignment are performed on the high-dimensional temporal dynamic feature sequences that are independent of each modality, and the sequences are spliced ​​and recombined in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration, and temperature multimodal modes, including:

[0076] To construct a nonlinear mapper based on the regenerative kernel Hilbert space, acoustic, vibration, and temperature features from modally independent high-dimensional temporal dynamic feature sequences are taken as independent variable inputs. Spatial dimension-up projection is performed using a Gaussian radial basis function (RBF) kernel. The Gaussian RBF kernel is calculated by: calculating the square of the Euclidean norm of the difference between any two input feature vectors, multiplying it by a negative scaling hyperparameter that controls the distribution width, and then calculating the power of the base of the natural logarithm corresponding to the independent variable. This forces the features of the three modes to be mapped to a shared, linearly separable high-dimensional manifold topological space, resulting in a manifold projection aligned feature set.

[0077] Within the shared space formed by the manifold projection aligned feature set, the similarity of different modal feature sequences on the time evolution trajectory is quantified. The dynamic time warping distance between acoustic features and vibration features, vibration features and temperature features, and temperature features and acoustic features is calculated respectively. The calculation process of the dynamic time warping distance is as follows: construct a two-dimensional distance cost matrix for two feature sequences to be compared, where each element of the matrix represents the squared error of the feature vector at the corresponding time point. From the upper left corner of the matrix to the lower right corner, a non-decreasing path with the minimum cumulative cost is found using a dynamic programming algorithm. The minimum cumulative cost is the elastic similarity measure between the two modal features, forming a cross-modal geodesic distance correlation matrix containing three distance scalar values.

[0078] A cross-modal dynamic graph topology is constructed using the cross-modal geodesic distance correlation matrix. Acoustic, vibration, and temperature features are regarded as three core nodes in the graph structure. The weight of the undirected connection edge between nodes is defined as the normalized value of the inverse of the corresponding dynamic time warp distance. When the weight is greater than the set graph connection sparsification threshold, it is retained; otherwise, the weight of the edge is forcibly set to zero, forming a sparse adjacency weight matrix representing the intensity of information interaction between modes.

[0079] Based on the sparse adjacency weight matrix, cross-modal feature information transmission and aggregation iteration are performed on the graph structure. For any central modal node, its own features and the features of all its neighboring modal nodes are summed by a weighted linear combination according to the adjacency edge weights. The summation result is input into the feedforward neural network layer for nonlinear transformation. After a specified number of graph convolution information transmission iterations, the features of each modality have deeply absorbed the complementary information of other related modalities, resulting in the enhanced feature representation of each modality after cross-modal graph neural alignment.

[0080] The modal enhancement feature representations after cross-modal graph neural alignment are strictly sequentially stacked along the feature channel dimension. The last channel of the acoustic enhancement feature is connected to the first channel of the vibration enhancement feature to form a joint high-dimensional tensor of high-order complementary information of all modes. The joint high-dimensional tensor is cross-fused and dimensionally compressed between channels using a fully connected bottleneck layer composed of multiple one-dimensional convolutional kernels to generate an initial fusion feature tensor of acoustic, vibration and temperature multimodal features for subsequent feature selection and optimization.

[0081] An optimization algorithm model for horned lizards was constructed with feature salience measurement as the objective function. The initial fusion feature tensor of acoustic, vibration and temperature multimodal features was used as the optimization search space. Global and local iterative optimization was performed by executing the adaptive camouflage mechanism and covert predation and wandering strategy of the horned lizard population. The optimal modal feature weighting matrix was calculated.

[0082] In this embodiment, an optimization algorithm model for horned lizards is constructed with feature salience measurement as the objective function. The initial fusion feature tensor of acoustic, vibrational, and thermal multimodal features is used as the optimization search space. Global and local iterative optimization is performed by executing the adaptive camouflage mechanism and covert predation movement strategy of the horned lizard population to calculate the optimal modal feature weighting matrix, including:

[0083] To address the physical challenge of easily obscuring weak signals from coal mill degradation, a feature saliency metric objective function is constructed to maximize signal-to-noise ratio and information richness. This objective function consists of two parts: the first is an information entropy reward term, calculated as the negative sum of the products of the probabilities of all feature elements and their corresponding natural logarithms; the second is an inter-class variance penalty term, calculated as the arithmetic mean of the squares of the deviations of each feature element from its mean. The objective function value is defined as the sum of the weighted values ​​of the information entropy reward term and the reciprocals of the inter-class variance penalty term, guiding the optimization algorithm to search the feature subspace with the highest information content and the most volatile features, thus establishing a fitness evaluation benchmark.

[0084] The dimension of the algorithm optimization variables is determined by the channel dimension of the initial fusion feature tensor of acoustic, vibration and temperature multimodal. In the multidimensional continuous search space composed of the upper and lower limits of the allowed feature weights, a two-dimensional floating-point data array composed of a preset number of rows and columns is randomly generated by a pseudo-random number generator with uniform distribution characteristics. Each row represents a virtual horned lizard individual, and each value inside the individual represents a candidate feature weight of a corresponding channel, thus forming the initial horned lizard population position matrix.

[0085] The fitness benchmark value corresponding to each individual horned lizard in the initial population is calculated, and the individual with the highest fitness value is selected as the current local optimal guide. An adaptive camouflage mechanism is executed to update the position to simulate the biological characteristic of horned lizards changing color to hide according to the environment: the difference vector between the current position of the horned lizard individual and the position of the local optimal guide is calculated, and the difference vector is multiplied by a camouflage control factor that decays nonlinearly with the number of iterations. The resulting product is then algebraically added to the current position of the horned lizard individual to achieve a smooth convergence to the excellent feature configuration scheme and obtain the first candidate population position matrix.

[0086] A chaotic escape operator based on the horned lizard's eye-spraying defense mechanism is introduced to improve the algorithm at the individual level: In each iteration, the number of stagnant iterations in which the fitness value of each horned lizard individual does not improve substantially is counted in real time. When the number of stagnant iterations exceeds the preset defense trigger threshold, a logistic chaotic perturbation sequence is injected into the position of the specific horned lizard individual. The specific operation is as follows: Based on the current individual position, the product of a chaotic variable between zero and one and the difference between one and the chaotic variable is calculated, and then multiplied by the chaotic expansion factor to generate a drastically changing chaotic perturbation vector. This perturbation vector is directly superimposed on the current individual position, forcing the horned lizard individual to instantly jump out of the current local extreme value trap and jump to an unknown new region in the search space, generating the second candidate population position matrix after the escape update.

[0087] In the later stages of iteration, a stealthy predatory walking strategy is implemented for high-precision local exploitation, simulating the behavior of horned lizards waiting for prey and then striking quickly: the position of the globally optimal horned lizard individual with the highest fitness among all individuals is calculated, and a random movement vector with a step size following a heavy-tailed probability distribution is generated based on the Levy flight random walk theory. This heavy-tailed distribution makes most walks have extremely small step sizes, with occasional large step jumps. The position of the globally optimal horned lizard individual is weighted and summed with the Levy flight random movement vector to guide the remaining horned lizard individuals to conduct extremely detailed micro-neighborhood exploration near the globally optimal solution. After continuous cyclic updates to reach the maximum preset number of iterations, the iteration stops and the combination of globally optimal horned lizard individual position variables corresponding to the historical maximum value of the fitness evaluation benchmark is extracted.

[0088] The extracted global optimal horned lizard individual position variable combination is re-normalized and arranged according to the channel arrangement order of the original feature tensor to ensure that the sum of the absolute values ​​of all weight coefficients is equal to one or mapped to a closed interval between zero and one. The normalized continuous floating-point sequence is transformed into a tensor structure of a specific dimension, which serves as the optimal modal feature weighting matrix reflecting the sensitivity of each feature channel of acoustic vibration temperature to the deterioration of the coal mill.

[0089] The optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic, vibration and temperature multimodal modes are subjected to Hadamard product modulation and spatial dimension reconstruction to synthesize a multimodal degradation sensitive characterization set that is resistant to background noise interference.

[0090] In this embodiment, the optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic-vibration-temperature multimodal modes are subjected to Hadamard product modulation and spatial dimension reconstruction to synthesize a multimodal degradation-sensitive characterization set that is resistant to background noise interference, including:

[0091] Extract each weight scalar element representing the importance of a channel from the optimal modal feature weighting matrix, and perform tensor copying and broadcast expansion operations in the spatial dimensions, namely height and width, to construct a three-dimensional scalar multiplier matrix with the same values ​​in the spatial dimensions but different values ​​in the channel dimensions, and whose overall three-dimensional shape is completely consistent with the initial fusion feature tensor of the acoustic-vibration-temperature multimodal model.

[0092] The dimension-expanded weighted tensor and the initial fusion feature tensor of acoustic, vibration and temperature multimodalities are subjected to element-wise Hadamard product operation. At the three-dimensional coordinate index positions that are completely corresponding to the two tensors, the stored values ​​are multiplied by scalar. Through the modulation mechanism of bitwise multiplication, the algorithm adaptively amplifies the numerical response amplitude of high-value feature channels related to wear of coal mill bearings, breakage of grinding parts or loosening and deterioration of liners. At the same time, it suppresses or clears the numerical response of low-value or redundant feature channels caused by coal powder airflow turbulence and external motor electromagnetic interference. Feature-level information filtering is completed within the tensor to obtain the weighted modulation fusion feature tensor.

[0093] To further refine the translation invariance of features, a multi-scale spatial feature convergence operation is applied to the weighted modulation fusion feature tensor. Two different sizes of local sliding receptive windows, 3x3 and 5x5, are set up, and these two windows slide across the spatial plane of the feature tensor with a set step size. Within the local receptive field region where each sliding window stays, the arithmetic mean of all pixel values ​​is calculated in parallel to preserve the global background distribution trend, and the absolute maximum value of all pixel values ​​is calculated to extract deterioration and abrupt edge information. The calculated average feature map and the maximum feature map are then combined again in the depth dimension to output a multi-scale spatial feature convergence tensor that compresses the spatial resolution but improves the semantic density.

[0094] Perform feature dimension reshaping and channel interleaving operations on the multi-scale spatial feature convergence tensor: decompose the total number of channels of the tensor into multiple sub-channel groups according to a preset ratio, and through matrix transpose and flattening operations, uniformly interleave and mix the feature vectors within different sub-channel groups, so that the output channel at any receptive field position can simultaneously come from the combined information of the three physical domains of sound, vibration and temperature, and generate a channel interleaving reconstructed feature map.

[0095] The soft-threshold shrinkage nonlinear activation function is applied to the channel interleaving reconstructed feature map. The specific mathematical calculation logic of the denoising operation is as follows: a background noise tolerance threshold parameter based on the statistical mean is preset. When the absolute value of a feature element in the feature map is strictly less than the threshold parameter, the feature element is forcibly assigned to zero. When the absolute value of a feature element in the feature map is greater than or equal to the threshold parameter, the absolute value of the feature element is subtracted from the value of the threshold parameter, and the feature element is reassigned its original positive or negative sign before processing. This completely eliminates the numerical jitter caused by small noises and outputs a highly purified multimodal degradation sensitive characterization set that is resistant to background noise interference.

[0096] The multimodal degradation sensitive characterization set is input into the nonlinear evaluation mapper for feature decoding and regression calculation, and the final coal mill degradation assessment index is output.

[0097] In this embodiment, the multimodal degradation sensitivity characterization set is input into a nonlinear evaluation mapper for feature decoding and regression calculation, outputting the final coal mill degradation assessment index, including:

[0098] The three-dimensional, anti-background noise-sensitive multimodal representation set is completely flattened in spatial dimension along the channel priority order. The original three-dimensional data array with height, width and depth is connected and combined in the memory reading order from left to right, from top to bottom and from front to back into a one-dimensional fully connected continuous floating-point vector with no spatial topological dimension and a single length dimension. This vector serves as the basic input carrier for the subsequent deep decoding network, namely the one-dimensional global representation vector.

[0099] A deep feedforward nonlinear mapping network is constructed, consisting of an input perception layer, several hidden computation layers, and a single output regression layer. A one-dimensional global representation vector is input to the first hidden computation layer. Each neuron in this layer receives a weighted sum of the output values ​​of all neurons in the previous layer. After adding an independent and learnable bias constant, the sum is input to the Gaussian error linear unit activation function for nonlinear spatial distortion transformation. The specific execution logic of the Gaussian error linear unit activation function is as follows: it calculates the product of the input value and the cumulative distribution function of its corresponding standard normal distribution at the corresponding input value. This allows the network to introduce a probability-based nonlinear smooth transition while preserving linearity. Through the step-by-step decoding and abstraction of multiple hidden computation layers, the mapping relationship of the nonlinear degradation mechanism hidden behind the multimodal signal is deeply mined, and the high-order features of nonlinear decoding are output.

[0100] To improve the generalization ability and robustness of the final evaluation model when facing unknown working conditions or drastic load changes, and to prevent the network from over-memorizing noisy data from specific training batches, a random deactivation regularization mechanism is forcibly introduced between the hidden computation layers: In each forward propagation computation cycle, using an independent and identically distributed Bernoulli random variable generator, according to a pre-set percentage probability threshold, the output activation values ​​of some neuron nodes are forcibly set to zero. The neurons that are cleared to zero will not participate in the forward information propagation and backward error gradient propagation calculation of the current round, forcing the remaining neurons in the network to learn autonomously and take on more independent feature representation responsibilities, reducing the complex cooperative adaptive dependence between neurons, and outputting robust high-order features that have been treated against overfitting.

[0101] The robust high-order features are passed to the last regression layer of the deep feedforward nonlinear mapping network. This regression layer contains a linear neuron node without any activation function. By calculating the sum of the inner products of each element in the robust high-order features and the final regression weight vector, plus a global baseline bias scalar, a dimensionless continuous floating-point value with a theoretical range of zero to one hundred is calculated. The value of zero represents that the coal mill is in a perfectly healthy new factory condition, and the value of one hundred represents that the coal mill is in an extreme deterioration state where a catastrophic failure or complete failure is imminent. This is the predicted value of the continuous deterioration assessment scalar corresponding to the current instantaneous sampling time window.

[0102] Because the predicted values ​​of a single transient time window are easily affected by sudden large coal particles crushing or voltage transients in the power grid, causing sharp fluctuations, an exponential moving weighted average filtering algorithm is applied to a series of continuous degradation assessment scalar predicted values ​​generated by continuous time slices in order to provide a smooth trend curve that conforms to the visual habits of industrial operators. The specific smoothing iteration logic of this algorithm is as follows: multiply the latest assessment scalar predicted value at the current moment by a weight parameter that determines the smoothing response speed, and add the product of the filtered historical assessment index from the previous moment and the difference between the filtered historical assessment index and the weight parameter. This ensures that the current output can not only accurately track the latest degradation state of the coal mill at the current moment, but also deeply integrate the long-term stable historical trend, effectively filtering out high-frequency jitter noise and outputting a smooth, stable, and highly valuable coal mill degradation assessment index for engineering guidance.

[0103] Example 2: During a high-load peak operation cycle, a direct-fired medium-speed roller mill of a large thermal power generating unit was operating under complex conditions of co-firing high-hardness, low-quality coal. The on-site monitoring system detected an abnormal, muffled impact sound inside the mill cylinder, but the conventional vibration amplitude alarm was not triggered.

[0104] The implementers observed that the front-end data acquisition module acquired the original acoustic sequence with a sampling rate of 44.1 kHz, the original shell vibration sequence of 10 kHz, and the critical node temperature sequence of 1 Hz within the same 5-second sliding time window. At this time, the signal-to-noise ratio of the original acoustic signal was extremely low, only -18 dB. The system introduced a fractional-order calculus filter operator of order 0.85 to process the acoustic sequence. The calculation results showed that the steady-state noise in the frequency band of 1500 Hz to 2500 Hz generated by the random impact of high-ash coal blocks was suppressed by about 60%, while the transient peak energy representing metal crushing was amplified by 3.2 times, and the signal-to-noise ratio was improved to -4 dB. For the vibration sequence, the system extracted its Hilbert envelope and subtracted the baseline drift below 0.5 Hz. The system calculation found that the temperature probe had a thermal hysteresis of about 1.2 seconds. Using a first-order inertial inverse model with a calibration time constant of 1.5, the true transient temperature peak was calculated. By calculating the cross-correlation function using a phase-locked loop, the system located a 12-millisecond air propagation time delay between the acoustic signal and the vibration signal, and performed translation compensation on the time axis. The system then used cubic spline interpolation to uniformly resample the data from the three modes into a 1000Hz feature grid, generating a multimodal time-series dataset with a dimension of 5000×3.

[0105] The system inputs a 5000×3 data tensor into the Mamba-2 model. A linear projection layer maps it to a high-dimensional initial hidden state tensor of 5000×256. The system applies a causal one-dimensional depthwise convolution with a window size of 7, effectively capturing extremely short transient fluctuations within 0.007 seconds. The discretized state-space parameter group is activated, and the system performs prefix sum parallel scan computation at the hardware level. Monitoring logs show that when processing this long sequence containing 5000 time steps, the structured dual mechanism of Mamba-2 successfully captured a periodic state decay feature with a period of 2 seconds, without gradient vanishing after the 800th time step as is common in traditional LSTM. The system outputs three independent high-dimensional temporal dynamic feature sequences, each with a dimension of 5000×128.

[0106] To overcome the physical isolation between sound, vibration, and temperature, the system utilizes a Gaussian radial basis kernel function with a scale hyperparameter set to 0.5 to map the 5000×128 feature dimension to a 512-dimensional shared manifold topological space. The dynamically time-warped distance cost matrix calculated by the system shows that the geodesic distance between acoustic and vibration nodes is 0.34, while the distance between vibration and temperature is 0.82. The system sets the graph connection sparsity threshold to 0.4, thus forcing the weights of weakly correlated edges between vibration and temperature to 0, forming a sparse adjacency weight matrix. After three iterations of graph convolution information transfer, the acoustic features successfully absorb the strongly correlated high-frequency harmonic information of vibration. The modal enhancement features are concatenated and stacked, and dimensionality is reduced through a fully connected bottleneck layer, generating a spatialized initial fusion feature tensor of size 64×64×3.

[0107] Faced with the aforementioned 64×64×3 fusion tensor, the system needs to determine which mode is most indicative under the current anomalous state. The system initializes a population of 50 virtual horned lizards, each representing a set of feature channel weights of length 3. In the initial first iteration, the fitness benchmark score is only 14.2, at which point the weight distribution is relatively even. When the iteration reaches 25, the system detects that the local optimum has failed to improve by more than 0.01 in five consecutive iterations, triggering a defense threshold. The system immediately injects a logistic chaotic perturbation sequence with an expansion factor of 3.5 into the 12 stagnant horned lizard individuals. Logs show that this perturbation causes the population to instantly leap to a new search space. Combined with the Levy flight strategy where the step size follows a heavy-tailed distribution, the fitness score soars to 45.8 in the 38th iteration and finally converges. The system extracts the globally optimal variable combination and transforms it into the optimal modal feature weighting matrix, with its channel weights precisely locked at: acoustic 0.68, vibration 0.25, and temperature 0.07. This data definitively proves that, under current operating conditions, subtle abnormal acoustic characteristics are the most crucial factor in assessing equipment degradation.

[0108] The system extracts a weight vector of [0.68, 0.25, 0.07], broadcasts and expands it spatially into a 64×64×3 three-dimensional scalar multiplier matrix, and performs element-wise Hadamard product operations with the initial fused feature tensor. Data tracking shows that the redundant high values ​​in the 3rd channel, originally caused by external environmental interference, are instantly suppressed to below 12, while the weak textures in the 1st channel containing high-frequency impact features are significantly amplified. Next, the system uses 3×3 and 5×5 dual-scale sliding windows to extract local means and maximum values, reconstructing the channel interleaving feature map. Finally, the system sets the background noise tolerance threshold parameter to 0.05 and performs soft-threshold shrinkage nonlinear activation. All residual electromagnetic background noise with absolute values ​​less than 0.05 in the matrix is ​​completely cleared, outputting an extremely pure set of multimodal degradation-sensitive representations.

[0109] The system completely flattens the purified feature tensor to form a one-dimensional global representation vector of length 4096. This vector passes through a deep nonlinear mapping network containing Gaussian error linear units. During forward propagation in the hidden layers, the system triggers a random deactivation mechanism with a 30% probability threshold, forcibly severing the connections of 1228 neurons to prevent the model from rote memorizing the current batch of specific low-quality coal noise data. In the final regression layer, the model outputs a series of continuous scalar predictions of degradation degree for each time slice. Due to the irregular impact of co-fired coal blocks, the original predictions fluctuated drastically between 74.5 and 86.2. The system immediately switched to an exponential moving average weighted average filtering algorithm with a smoothing coefficient of 0.15. Finally, the system outputs an extremely smooth and stable evaluation index curve on the main control screen—the value steadily climbs from a healthy 25 to a stable 81.4. This index exceeded the system's set severe warning threshold of 80 points. The system immediately generated a detailed deterioration diagnosis report, indicating that localized spalling of the grinding roller sleeve had occurred inside the coal mill, and sent a high-level warning signal to the on-site control room in real time. Based on this accurate warning, the operators conducted an emergency shift shutdown inspection during a break in the unit's load reduction. They discovered a 15-centimeter-long fatigue spalling pit on the surface of the No. 2 grinding roller, thus successfully preventing an unplanned shutdown accident caused by the main shaft vibration failure.

[0110] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for assessing the deterioration degree of a coal mill based on a deep learning-based multimodal fusion of acoustic, vibration, and temperature parameters, characterized in that... include: The system synchronously acquires continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences during the operation cycle of the target coal mill, and performs spatiotemporal synchronization and resampling alignment operations on multi-source heterogeneous signals to construct a multi-modal time-series data set of acoustic, vibration, and temperature. The acoustic, vibration and temperature multimodal time series data set is input into the pre-constructed Mamba-2 parallel state space model. The structured state space duality mechanism is used to perform long-range context dependency modeling on acoustic, vibration and temperature signals respectively, and obtain high-dimensional time series dynamic feature sequences that are independent of each mode. Perform cross-modal cross-mapping and joint alignment on each modally independent high-dimensional temporal dynamic feature sequence, and splice and recombine them in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration and temperature multimodal modes. An optimization algorithm model for horned lizards was constructed with feature salience measurement as the objective function. The initial fusion feature tensor of acoustic, vibration and temperature multimodal features was used as the optimization search space. Global and local iterative optimization was performed by executing the adaptive camouflage mechanism and covert predation and wandering strategy of the horned lizard population. The optimal modal feature weighting matrix was calculated. The optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic, vibration and temperature multimodal modes are subjected to Hadamard product modulation and spatial dimension reconstruction to synthesize a multimodal degradation sensitive characterization set that is resistant to background noise interference. The multimodal degradation sensitive characterization set is input into the nonlinear evaluation mapper for feature decoding and regression calculation, and the final coal mill degradation assessment index is output.

2. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The synchronous acquisition of continuous acoustic emission sequences, shell vibration sequences, and key node temperature sequences during the target coal mill's operating cycle, along with spatiotemporal synchronization and resampling alignment of multi-source heterogeneous signals, constructs a multi-modal time-series data set of acoustic, vibration, and temperature data, including: By synchronously collecting raw sound, vibration and temperature signals in the environmental background through sensor arrays deployed in key parts of the coal mill, an uncleaned multi-source heterogeneous data stream is formed. Fractional-order calculus filtering is applied to the original acoustic sequence in the uncleaned multi-source heterogeneous data stream to enhance the abrupt high-frequency features and denoise the steady state, resulting in a purified acoustic sequence. Hilbert envelope demodulation and polynomial baseline elimination are performed on the original vibration sequence in the uncleaned multi-source heterogeneous data stream to obtain the cleaned shell vibration sequence. A first-order inertial inverse model is applied to the original temperature sequence in the uncleaned multi-source heterogeneous data stream for dynamic thermal hysteresis compensation, and the temperature sequence of the key purification node is obtained. Based on the vibration sequence of the purification shell, the cross-correlation function is calculated using a phase-locked loop to compensate for the time delay between the acoustic sound generation sequence of the purification and the temperature sequence of the key purification nodes, thereby generating a time-domain aligned multimodal sequence. A cubic spline interpolation algorithm is applied to time-domain aligned multimodal sequences for uniform step size resampling and alignment, and then stacked in the channel dimension to construct an acoustic-vibration-temperature multimodal time series dataset.

3. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The process involves inputting a multimodal time-series data set of acoustic, vibration, and temperature signals into a pre-constructed Mamba-2 parallel state-space model. A structured state-space duality mechanism is then used to model the long-range context dependencies of the acoustic, vibration, and temperature signals, resulting in independent high-dimensional temporal dynamic feature sequences for each mode, including: The set of acoustic, vibration and temperature multimodal time series data is input into the linear projection layer to perform spatial dimension upsizing and form the initial hidden state tensor. Initialize the continuous evolution parameters of the state-space model and apply a zero-order hold for discretization transformation to construct a discretized state-space parameter group; A causal one-dimensional depthwise convolution operation is performed on the initial hidden state tensor to extract local transient fluctuations, resulting in a local temporal-aware feature sequence. Based on the discretized state-space parameter group, a semi-separable matrix and prefix sum parallel scan are applied to the local time-series sensing feature sequence to perform structured state-space dual computation and output a global long-range sensing state sequence. The global long-range sensing state sequence is modulated by element-wise multiplication with the nonlinear gated branch and then linear dimensionality reduction is performed to obtain a high-dimensional temporal dynamic feature sequence that is independent of each mode.

4. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The process of performing cross-modal cross-mapping and joint alignment on the modally independent high-dimensional temporal dynamic feature sequences, and splicing and recombining them in the feature channel dimension to generate an initial fusion feature tensor for acoustic, vibration, and temperature multimodal modes includes: By using radial basis kernel functions, the modally independent high-dimensional temporal dynamic feature sequences are mapped to a shared high-dimensional manifold topological space, resulting in a manifold projection aligned feature set. Calculate the dynamic time warping distance of the temporal evolution trajectory of each mode in the manifold projection alignment feature set, and construct the cross-modal geodesic distance correlation matrix; Based on the threshold set by the cross-modal geodesic distance correlation matrix, a sparse cross-modal dynamic graph topology is constructed to form a sparse adjacency weight matrix. Based on the sparse adjacency weight matrix, weighted aggregation and nonlinear propagation of node features are performed on the graph structure to obtain the enhanced feature representations of each modality after cross-modal graph neural alignment; The enhanced feature representations of each modality after cross-modal graph neural alignment are stacked sequentially by channel and then compressed by dimensionality reduction through a fully connected bottleneck layer to generate the initial fusion feature tensor of acoustic, vibration and temperature multimodal modes.

5. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The proposed optimization algorithm model for horned lizards, with feature salience measurement as the objective function, uses the initial fused feature tensor of acoustic, vibrational, and thermal multimodal features as the optimization search space. It performs global and local iterative optimization by implementing the adaptive camouflage mechanism and covert predation movement strategy of the horned lizard population, calculating the optimal modal feature weighting matrix, including: A feature saliency objective function is constructed by combining the information entropy reward term and the inter-class variance penalty term to establish a fitness evaluation benchmark; Using the channel dimension of the initial fusion feature tensor of acoustic-vibration-temperature multimodal modes as the search boundary, an initial horned lizard population position matrix containing multiple sets of candidate feature weights is randomly generated. Based on the fitness evaluation benchmark, a local optimum is selected, and adaptive camouflage position update is performed by smoothly approaching it to obtain the position matrix of the first candidate population. The individuals trapped in the iteration stagnation in the first candidate population position matrix are injected with a logistic chaotic perturbation sequence to perform escape transitions, thereby generating the second candidate population position matrix. The second candidate population position matrix is ​​guided to perform a Levy flight walk with a heavy-tailed distribution towards the globally optimal individual. After iterative optimization, the combination of the globally optimal horned lizard individual position variables is extracted. Perform absolute value normalization and tensor structure transformation on the global optimal combination of individual horned lizard position variables to output the optimal modal feature weighting matrix.

6. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The step of performing Hadamard product modulation and spatial dimension reconstruction on the optimal modal feature weighting matrix and the initial fusion feature tensor of acoustic, vibration, and temperature multimodal modes to synthesize a multimodal degradation sensitive characterization set resistant to background noise interference includes: The optimal modal feature weighting matrix is ​​broadcast and extended in the spatial dimension to construct a dimension-extended weighting tensor with the same shape as the initial fusion feature tensor of the acoustic, vibration and temperature multimodal modes. The weighted feature tensor is obtained by performing an element-wise Hadamard product operation on the dimension-expanded weighted tensor and the initial fusion feature tensor of the acoustic-vibration-temperature multimodal modes. The local mean and maximum values ​​are extracted by applying a multi-scale sliding window in parallel to the weighted modulation fusion feature tensor and then performing depth concatenation to output a multi-scale spatial feature convergence tensor. Perform subchannel group decomposition, transpose, and flattening interleaving operations on the multi-scale spatial feature convergence tensor to generate a channel interleaving reconstructed feature map; A soft-threshold shrinkage nonlinear activation function is applied to the channel interleaving reconstructed feature map to remove minor background noise, and the output is a multimodal degradation sensitive characterization set that is resistant to background noise interference.

7. The method for evaluating the deterioration degree of a coal mill based on deep learning and multimodal fusion of acoustic, vibration, and temperature as described in claim 1, is characterized in that... The process of inputting the multimodal degradation sensitivity characterization set into a nonlinear evaluation mapper for feature decoding and regression calculation, and outputting the final coal mill degradation assessment index, includes: The set of multimodal degradation sensitive representations that resist background noise interference is spatially flattened along the channel priority order to form a one-dimensional global representation vector. The one-dimensional global representation vector is input into a deep network with Gaussian error linear unit activation function for progressive abstraction and nonlinear mapping, and the output is nonlinear decoded high-order features; In the forward computation of the hidden layer, a random deactivation regularization mechanism is applied to the nonlinear decoding high-order features to suppress overfitting and output robust high-order features. The robust high-order features are passed to the linear regression layer, and the continuous degradation assessment scalar prediction value is output by calculating the algebraic sum of the inner product and the global bias. The scalar prediction value of continuous degradation assessment output by continuous time slices is subjected to historical trend fusion and high-frequency anti-jitter by applying the exponential moving weighted average filtering algorithm to output the coal mill degradation assessment index.

Images