Mining vibrating screen fault detection method based on audio frequency and domain self-adaption

Abstract

The invention discloses a mining vibrating screen fault detection method based on audio frequency and domain self-adaption, and relates to the technical field of mechanical equipment in mineral processing. Key technologies such as multi-sensor audio frequency perception, selective state space modeling, Transform sequence modeling and multi-system adversarial domain self-adaption are fused; a complete technical process from data acquisition, feature extraction and modeling to unsupervised domain adaptation and intelligent early warning is constructed, and the method has good robustness, mobility and real-time performance and is suitable for fault monitoring and early warning of the mining vibrating screen in a complex field operation environment.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment technology in mineral processing, specifically to a fault detection method for a mining vibrating screen based on audio and domain adaptation. Background Technology

[0002] As a core piece of equipment in the coal washing and processing process, the operating status of mining vibrating screens directly affects the efficiency and stability of the entire production line. During long-term, high-intensity operation, mining vibrating screens are prone to various failure modes, such as reduced spring stiffness, loose screen plates, abnormal bearings, and improper lubrication. A failure can lead to the shutdown of the entire production line, causing significant economic losses. Therefore, real-time fault detection and early warning for mining vibrating screens is of great practical significance.

[0003] Existing fault detection methods for mining vibrating screens mainly employ contact or semi-contact sensing technologies such as vibration signal monitoring, temperature monitoring, and current monitoring, which face significant limitations in practical applications. Traditional vibration sensors and temperature sensors need to be directly installed on or inside the equipment surface. In the harsh working environment of mining vibrating screens, these sensors are easily affected by dust, moisture, and strong vibrations, leading to high failure rates and high maintenance costs. Furthermore, the installation of contact sensors often requires equipment modification, increasing system complexity and safety risks.

[0004] In contrast, audio signals, as a natural representation of equipment operating status, have a significant advantage in non-contact detection. They can acquire rich fault feature information from a distance without interfering with normal equipment operation, effectively avoiding the reliability issues of sensors in harsh environments. However, existing audio-based fault detection methods have shortcomings in feature extraction and modeling, making it difficult to fully exploit fault features in audio signals, especially in recognizing weak abnormal audio patterns under complex operating conditions.

[0005] More importantly, acquiring fault data in actual industrial production faces severe challenges. First, fault events occur infrequently, resulting in an extremely limited number of natural fault samples. Second, artificially creating faults for data collection poses safety risks and is costly. Most importantly, different manufacturers and models of mining vibrating screens exhibit significant differences in acoustic characteristics and operating parameters, making it difficult to directly apply fault detection models trained on one type of equipment to other, resulting in poor cross-equipment generalization ability. Traditional supervised learning methods heavily rely on large amounts of labeled fault data, but when new equipment is put into operation or under different operating conditions, there is often a lack of sufficient fault samples for model training, severely limiting the practical deployment and application of fault detection systems.

[0006] In summary, existing technologies lack effective solutions to simultaneously address the two key challenges of in-depth mining of the audio characteristics of mining vibrating screens and the scarcity of fault data. Therefore, there is an urgent need for a fault detection technology for mining vibrating screens based on non-contact audio detection, possessing powerful feature extraction and modeling capabilities, and cross-domain adaptive capabilities, to meet the real-time monitoring and early warning needs of industrial sites under conditions of limited fault data. Summary of the Invention

[0007] To address the aforementioned shortcomings in existing technologies, this invention provides a fault detection method for mining vibrating screens based on audio and domain adaptation, which solves the problem of the lack of effective solutions in existing technologies to simultaneously address the issues of in-depth mining of audio characteristics of mining vibrating screens and the scarcity of fault data.

[0008] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A fault detection method for mining vibrating screens based on audio and domain adaptation is provided, which includes the following steps: Spatial collaborative acquisition and preprocessing of the operating audio of the target mining vibrating screen are performed to obtain the target audio signal; The target audio signal is converted into a Mel spectrogram, and an energy gradient-driven adaptive manifold topology unfolding strategy is used to serialize the Mel spectrogram to obtain the corresponding sequence. The sequence is encoded using an unsteady fractional-order state-space model to obtain a fractional-order state embedding vector that includes mechanical memory characteristics; By using a linearized frequency domain gated evolutionary deduction network to globally model the fractional-order state embedding vector, the pure frequency band energy drift characteristics of the mining vibrating screen are captured, and the corresponding probability distributions of various faults of the mining vibrating screen are output to complete the fault detection of the mining vibrating screen. The training method of the linearized frequency domain gated evolution inference network is based on unsupervised adversarial training between historical source domain fault samples and the target domain to achieve domain adaptation.

[0009] Furthermore, the specific method for spatial collaborative acquisition and preprocessing of the operating audio of the target mining vibrating screen includes the following steps: Multiple microphones are arranged in a circular array, with the center located above the center of the mining vibrating screen box. The array radius is 0.3 meters and the height from the surface of the mining vibrating screen is 1.2 meters. The raw audio data collected by each microphone is sequentially passed through a preamplifier, an anti-aliasing filter, and a dynamic range control module to achieve multi-channel spatial collaborative acquisition of the operating audio of the target mining vibrating screen. The audio signals acquired through multi-channel spatial collaborative acquisition are sequentially subjected to adaptive denoising, reverberation suppression, and signal enhancement to obtain the target audio signal.

[0010] Furthermore, the specific method for converting the target audio signal into a Mel spectrogram includes the following steps: The target audio signal is normalized to the range of [-1.00, 1.00] to obtain the normalized audio signal; The normalized audio signal is divided into sliding window segments at 60-second intervals, and each normalized audio segment is framed in 1-second units to obtain several more audio frames. For each frame of audio, a Hamming window is first used to reduce spectral leakage. Then, a short-time Fourier transform is used to extract the time-frequency structure. Finally, a Mel filter bank is introduced to compress the spectrum and generate a two-dimensional Mel spectrogram.

[0011] Furthermore, the specific method for serializing the Mel spectrogram using an energy gradient-driven adaptive manifold topology expansion strategy includes the following steps: The energy gradient field of the two-dimensional Mel spectrogram is calculated to identify energy accumulation regions and gradient change ridges. Based on the distribution of the energy gradient field, an adaptive segmentation algorithm is used to divide the two-dimensional Mel spectrogram into several locally morphologically dynamically determined local spectrogram blocks. Each local spectral block is transformed into an independent one-dimensional feature vector through feature mapping. Then, according to the distribution logic of each local spectral block in the time-frequency space, all one-dimensional feature vectors corresponding to the same two-dimensional Mel spectrogram are recombined into a one-dimensional sequence to obtain the corresponding sequence.

[0012] Furthermore, the non-steady fractional-order state-space model includes a first branch and a second branch; the first branch includes a first normalization layer, a first linear layer, a one-dimensional convolutional layer, a non-steady fractional-order state-space model, a dot product layer, a second linear layer, and a first summation layer connected in sequence; the second branch includes a third linear layer and an activation function layer connected in sequence, the input of the third linear layer is the output of the first normalization layer, and the output of the activation function layer is another input of the dot product layer; the other input of the first summation layer is the input of the first normalization layer, and the output of the first summation layer is the output of the non-steady fractional-order state-space model.

[0013] Furthermore, the calculation expression for the nonsteady-state fractional hierarchy is as follows:

[0014] in It is a fractional derivative operator; and These are the state transition matrix and the input matrix, respectively, both dynamically generated through an input-dependent hypernetwork; It is a fractional-order state embedding vector.

[0015] Furthermore, the linearized frequency-domain gated evolutionary derivation network is an improved Transformer network, comprising sequentially connected position coding layers, coding modules composed of multiple stacked linearized frequency-domain gated coding layers, and a classification head; wherein: The input to the positional coding layer is a fractional-order state embedding vector; The output of the classification head is a fault probability distribution; A single linearized frequency-domain gated coding layer includes a second normalization layer, a linearized frequency-domain gated attention module, a second summing layer, a third normalization layer, a feedforward neural network layer, and a third summing layer. The input to the linearized frequency-domain gated attention module is the output of the second normalization layer. The linearized frequency-domain gated attention module internally includes parallel query vector branches, key vector branches, and value vector branches. The query vector branch includes a Fourier transform layer, a frequency-domain gated unit, an inverse Fourier transform layer, and a first kernel mapping layer connected in sequence to filter out high-frequency environmental noise unrelated to the vibrating screen's operating frequency through a fast Fourier transform. The key vector... The branch includes a second kernel mapping layer, where the outputs of both the first and second kernel mapping layers are connected to a linear dot product layer to achieve linear reconstruction of the self-attention mechanism; the outputs of the linear dot product layer and the value vector branch are connected to a matrix multiplication layer; one input of the second summing layer is the output of the matrix multiplication layer, and the other input is the input of the second normalization layer; the output of the second summing layer is connected to the feedforward neural network layer via a third normalization layer, and the two inputs of the third summing layer are the output of the feedforward neural network layer and the input of the third normalization layer, respectively. The output of the third summing layer is the output of the linearized frequency domain gated coding layer.

[0016] Furthermore, specific methods for globally modeling fractional-order state embedding vectors using linearized frequency-domain gated evolutionary derivation networks include: Multiple fractional-order state embedding vectors are recombined into a mining vibrating screen operation evolution sequence according to the chronological order of their occurrence. This mining vibrating screen operation evolution sequence is then used as the input to a linearized frequency domain gated evolution inference network, thereby capturing the pure frequency band energy drift characteristics of the mining vibrating screen and outputting the corresponding probability distribution of various faults of the mining vibrating screen.

[0017] Furthermore, specific methods for unsupervised adversarial training based on historical source domain fault samples and the target domain include: A dual-discriminator collaborative structure is constructed. The parameters of the improved Transformer network are fixed, and two discriminators are trained separately using target domain data. At the same time, the two discriminators are trained under supervision using historical source domain fault samples. The optimization objective of training in the target domain is to maximize the Manhattan distance between the prediction distributions of the two discriminators. The parameters of the dual discriminators are fixed, and the Manhattan distance between the predicted probability distributions of the two discriminators for the same target domain data is used as the loss function constraint. The improved Transformer network is then optimized in reverse so that the improved Transformer network can extract domain-invariant features that make the prediction results of the two discriminators tend to be consistent. Then, the fault boundary is automatically discovered in the target domain through the adversarial mechanism to achieve domain adaptation. Among them, the improved Transformer network is used in unsupervised adversarial training to obtain the domain-invariant features of historical source domain fault samples and target domain data. These domain-invariant features are used as input to the discriminator for domain category prediction.

[0018] Furthermore, in unsupervised adversarial training, the weighted sum of the cross-entropy classification loss of historical source domain fault samples, the distribution difference loss after the extracted source domain features and target domain features are mapped to the same regenerating kernel Hilbert space, and the regularization loss is used as the total loss. Based on the total loss, the improved Transformer network is optimized by backpropagation to achieve fault discrimination structure transfer from the historical source domain to the target domain. This ensures that different fault types can still be accurately distinguished when transferred to the target domain, thereby improving the stability and robustness of cross-domain fault identification.

[0019] The beneficial effects of this invention are as follows: 1. This invention integrates a non-contact, high-precision sensor array and employs a spatially optimized layout strategy to comprehensively cover key parts of the vibrating screen, achieving spatial collaborative acquisition of multi-point audio signals. This adapts to the harsh working environment of mining vibrating screens characterized by high noise and strong vibration. Compared to traditional contact sensors, audio sensing technology offers advantages such as non-contact operation, strong anti-interference capabilities, and low maintenance costs. Combined with adaptive denoising algorithms and spatial prediction-based reverberation suppression technology, it effectively eliminates environmental noise interference from equipment such as crushers and conveyor belts, as well as echo effects from the plant space, significantly improving the acquisition quality and identifiability of audio signals related to vibrating screen faults in complex mining environments.

[0020] 2. This invention addresses the asymmetry in the time-frequency structure of vibrating screen fault audio signals by proposing a Mel-Chart segmentation method combining a serpentine block strategy. It utilizes a non-steady-state fractional-order state-space model (selective state-space model) to capture the local dynamic evolution characteristics of faults such as spring breakage and screen plate loosening. The encoded features are then modeled using a multi-layer Transformer structure to establish global dependencies across time periods, enhancing the characterization of subtle fault evolution processes such as friction noise caused by insufficient lubrication and abnormal vibrations due to spring fatigue. This enables accurate identification of complex abnormal patterns in the long-term operating audio of vibrating screens. This serial architecture is fundamentally different from existing technologies that use LSTM and only apply Transformer for multi-modal fusion.

[0021] 3. This invention constructs a dual-discriminator cooperative adversarial structure. Under the condition of unlabeled audio data from real mining vibrating screens, it mines and maintains the implicit discriminative structure between the source and target domains through alternating optimization of dual discriminators and generators. Combined with the source domain inter-class structure preservation mechanism and constraints suitable for the fault characteristics of vibrating screens, it achieves consistent modeling of discriminative data from the field, significantly improving the system's migration capability and online early warning stability in cross-vibrating screen and cross-mining area working conditions, and solving the domain adaptation problem from the laboratory to the field and between different fields. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a schematic diagram of the spatial arrangement method of multiple sensors in the embodiment; Figure 3 This is a flowchart of the audio preprocessing process in the embodiment; Figure 4 This is a schematic diagram of the dereverberation model in the embodiment; Figure 5 This is a schematic diagram of the audio feature extraction and sequence modeling unit in the embodiment; Figure 6 This is a schematic diagram illustrating the extraction of a single instruction (fractional-order state embedding vector) in the embodiment. Figure 7 This is a schematic diagram of the unsteady fractional-order state-space model in the embodiment; Figure 8 This is a schematic diagram of the linearized frequency domain gated evolution inference network structure in the embodiment; Figure 9 This is a schematic diagram illustrating the adaptive operation of different mining vibrating screen domains in the embodiment; Figure 10 This is a schematic diagram of the discriminator in the embodiment. Detailed Implementation

[0023] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0024] like Figure 1 As shown, the fault detection method for mining vibrating screens based on audio and domain adaptation includes the following steps: S1. Spatial collaborative acquisition and preprocessing of the operating audio of the target mining vibrating screen to obtain the target audio signal; S2. Convert the target audio signal into a Mel spectrogram, and use an energy gradient-driven adaptive manifold topology expansion strategy to serialize the Mel spectrogram to obtain the corresponding sequence; S3. Encode the sequence using an unsteady fractional-order state-space model to obtain a fractional-order state embedding vector containing mechanical memory characteristics; S4. Use a linearized frequency domain gated evolutionary deduction network to globally model the fractional-order state embedding vector, capture the pure frequency band energy drift characteristics of the mining vibrating screen, output the corresponding probability distribution of various faults of the mining vibrating screen, and complete the fault detection of the mining vibrating screen. The training method of the linearized frequency domain gated evolution inference network is based on unsupervised adversarial training between historical source domain fault samples and the target domain to achieve domain adaptation.

[0025] In some embodiments, such as Figure 2 As shown, an 8-channel high-sensitivity microphone array is arranged in a circular spatial configuration to adapt to the harsh working environment of mining vibrating screens, characterized by high noise and strong vibration. The eight microphones are distributed in a circular array, with the center located above the center of the vibrating screen box. The array radius is 0.3 meters, and the distance from the screen surface is 1.2 meters, fully covering the key structural components of the vibrating screen. The microphone frequency response range is 6.3 Hz-40 kHz, the sensitivity is 50 mV / Pa, and the dynamic range is 140 dB. The data sampling frequency is 48 kHz, with 24-bit quantization accuracy. It is equipped with a preamplifier (adjustable gain 0-20 dB), an anti-aliasing filter, and a dynamic range control module to ensure synchronous acquisition of multi-channel audio. Signal fusion is achieved through sensor-based collaborative sensing, with a synchronization error of <10 μs. This arrangement strategy is suitable for audio data acquisition from small laboratory vibrating screens and large field vibrating screens.

[0026] like Figure 3 As shown, an adaptive denoising algorithm and reverberation suppression technology are used to improve the quality of multi-channel audio signals to obtain the target audio signal. The adaptive denoising algorithm combines spectral subtraction and Wiener filtering techniques to eliminate environmental noise interference from equipment such as crushers and conveyor belts, taking into account the complex noise characteristics of the mining environment. By estimating the noise power spectral density in real time, the signal-to-noise ratio is improved by 20-25 dB. A reverberation suppression technology based on spatial prediction is also described. Figure 4 As shown, by analyzing the spatial correlation and time delay characteristics of the signals received by the multi-sensor array, a multilayer perceptron is used to learn and predict the propagation path of the echo effect in the factory space, reducing the reverberation time to below 0.15 seconds.

[0027] The target audio signal was amplitude normalized to the range [-1.00, 1.00] and converted to 32-bit floating-point PCM format. It was then structured and stored according to the acquisition timestamp and corresponding fault label in the format "date_time_fault type_channel number.wav", and an audio database index was established. Data acquisition covered normal operating conditions and various fault states, including typical operating conditions such as spring breakage, loose sieve plate, abnormal bearing noise, and abnormal lubrication, providing a high-quality audio data foundation for subsequent feature extraction and sequence modeling.

[0028] In some embodiments, such as Figure 5 and Figure 6 As shown, the normalized audio signal is divided into sliding window segments at 60-second time intervals, and each audio segment is further processed into frames of 1 second each (each frame contains 48,000 sample points). For each audio frame, a Hamming window is applied to reduce spectral leakage, and its time-frequency structure is extracted using a Short-Time Fourier Transform (STFT) (FFT points are set to 2048, frequency resolution is 23.4 Hz). Then, a 128-Melb filter bank is introduced to compress the spectrum, generating a 128×T two-dimensional Mel band energy map, where T is the number of time frames, with a value of 128. This process constructs an acoustic fingerprint containing rich time-frequency texture and treats it as a Riemannian manifold in non-Euclidean geometric space, laying the foundation for subsequent topological unfolding.

[0029] Given the non-stationary and nonlinearly damped characteristics of the time-frequency distribution of vibrating screen fault signals, this step constructs a local feature encoding architecture based on an energy gradient-driven adaptive manifold topology expansion strategy and an unsteady fractional-order state-space model. First, a block-based strategy is used to divide the two-dimensional Mel map into 16×16 local map blocks, totaling 64 blocks. Second, for each map block, instead of traditional line-by-line raster scanning, an energy gradient-driven topology expansion strategy is adopted: the energy gradient field within the block is calculated, and starting from the energy extremum point, a continuously folding traversal path (represented as an "S"-shaped trajectory in the topological structure) is dynamically generated along the ridge direction where the gradient descent is slowest, mapping the two-dimensional map into a one-dimensional sequence. This strategy ensures the physical continuity of the fault signal energy flow during the dimensionality reduction process and avoids feature truncation.

[0030] Next, the sequence is input into an unsteady fractional-order state-space model for encoding. This unsteady fractional-order state-space model introduces fractional derivative operators. To simulate the viscoelastic damping memory characteristics of a mechanical system, a dynamic equation is constructed: The state transition matrix is... and input matrix Instead of fixed parameters, the parameters are dynamically generated through an input-dependent hypernetwork. This unsteady fractional-order state-space model can adaptively distinguish between fault shocks with physical memory characteristics and random noise without memory characteristics, and finally outputs a fractional-order state embedding vector containing mechanical memory characteristics.

[0031] The supernetwork it relies on contains parallel learnable linear projection units for processing the input sequence features at the current time step. Perform feature space mapping; where the first projection unit will input features The linear mapping is a state transition matrix that evolves over time. To dynamically adjust the decay rate of fractional-order memory; the second projection unit will input features The linear mapping is an input matrix that evolves over time. This allows for the adaptive determination of the gain weight of the current input on the system state update, thereby enabling time-context-aware modeling of non-steady-state signals.

[0032] In some embodiments, such as Figure 7 As shown, the unsteady fractional-order state-space model includes a first branch and a second branch. The first branch includes a first normalized layer, a first linear layer, a one-dimensional convolutional layer, an unsteady fractional-order state-space model, a dot product layer, a second linear layer, and a first additive layer connected in sequence. The second branch includes a third linear layer and an activation function layer connected in sequence. The input of the third linear layer is the output of the first normalized layer, and the output of the activation function layer is another input of the dot product layer. The other input of the first additive layer is the input of the first normalized layer, and the output of the first additive layer is the output of the unsteady fractional-order state-space model.

[0033] This embodiment utilizes the nonlocal memory property of fractional derivative operators to establish a viscoelastic damping attenuation model for mechanical systems, thereby adaptively distinguishing fault impact signals with physical memory characteristics from random background white noise without memory characteristics. Compared to local operations based on fixed convolution kernels in Euclidean space, the unsteady fractional state-space model can adaptively track the unsteady trajectory of fault signals on the manifold topology through an evolution matrix that dynamically changes with the input signal. Compared to the quadratic computational complexity of traditional global attention mechanisms, this unsteady fractional state-space model, based on a linear solution mechanism of recursive differential equations, reduces the computational complexity to O(N), thus meeting the requirement for real-time encoding of microsecond-level high-frequency transient impact characteristics.

[0034] Then, multiple fractional-order state embedding vectors are concatenated strictly according to their temporal order in the original Mel spectrum to form a mining vibrating screen operating state sequence of length L=60. The sequence is input to a linearized frequency-domain gated evolutionary derivation network (i.e., an improved Transformer network). This network contains four linearized attention layers (8 attention heads per layer, 512 hidden dimensions), abandoning traditional exponential dot product calculations and linearizing the self-attention mechanism in the Transformer network by introducing a kernel function approximation strategy. Furthermore, frequency-domain gated units are embedded in the computation path to filter out high-frequency environmental noise unrelated to the vibrating screen's operating frequency using Fast Fourier Transform. The improved Transformer network captures microscopic damping memory through a fractional-order model and macroscopic energy drift through linearized attention, thus explicitly modeling the progressive frequency band energy drift law of the vibrating screen from normal operation to failure. Finally, the output high-order feature representation is mapped through classification heads (with 256 and 128 neurons respectively, representing the number of fault categories), outputting the final fault probability distribution.

[0035] In some embodiments, such as Figure 8 As shown, the linearized frequency-domain gated evolutionary derivation network is an improved Transformer network, comprising a position coding layer connected in sequence, a coding module composed of multiple linearized frequency-domain gated coding layers stacked together, and a classification head; wherein: The input to the positional coding layer is a fractional-order state embedding vector; The output of the classification head is a fault probability distribution; A single linearized frequency-domain gated coding layer includes a second normalization layer, a linearized frequency-domain gated attention module, a second summing layer, a third normalization layer, a feedforward neural network layer, and a third summing layer. The input to the linearized frequency-domain gated attention module is the output of the second normalization layer. The linearized frequency-domain gated attention module internally includes parallel query vector branches, key vector branches, and value vector branches. The query vector branch includes a Fourier transform layer, a frequency-domain gated unit, an inverse Fourier transform layer, and a first kernel mapping layer connected in sequence to filter out high-frequency environmental noise unrelated to the vibrating screen's operating frequency through a fast Fourier transform. The key vector... The branch includes a second kernel mapping layer, where the outputs of both the first and second kernel mapping layers are connected to a linear dot product layer to achieve linear reconstruction of the self-attention mechanism; the outputs of the linear dot product layer and the value vector branch are connected to a matrix multiplication layer; one input of the second summing layer is the output of the matrix multiplication layer, and the other input is the input of the second normalization layer; the output of the second summing layer is connected to the feedforward neural network layer via a third normalization layer, and the two inputs of the third summing layer are the output of the feedforward neural network layer and the input of the third normalization layer, respectively. The output of the third summing layer is the output of the linearized frequency domain gated coding layer.

[0036] The expression for linearizing and reconstructing the self-attention mechanism by introducing a kernel function approximation strategy is as follows:

[0037] This represents the output of the attention mechanism; Represents the kernel mapping function based on Taylor series expansion; This represents the transpose of the matrix; Q, K, and V represent the query vector, key vector, and value vector, respectively.

[0038] The above process uses a kernel mapping function based on Taylor series expansion or stochastic Fourier features to replace the traditional exponential dot product calculation, reducing the computational complexity of global dependency modeling from quadratic to linear. A frequency domain gating unit is embedded in the attention calculation path. First, a fast Fourier transform is performed on the query vector. After suppressing environmental noise interference in the non-fault frequency band through adaptive threshold filtering, an inverse transform is performed, thereby capturing the pure frequency band energy evolution trajectory of the vibrating screen from normal to fault.

[0039] The above process projects the temporal characteristics onto the regenerating kernel Hilbert space through a kernel mapping function to perform linear inner product calculation, thereby establishing a global energy coupling relationship throughout the entire life cycle. By utilizing the spectrum filtering capability of the frequency domain gating unit, it directionally enhances the specific frequency band energy related to weak fault symptoms such as friction modulation caused by lubrication failure and stiffness reduction caused by component fatigue, while suppressing broadband environmental noise in unrelated frequency bands, thus achieving high signal-to-noise ratio capture of progressive fault modes during long-term operation.

[0040] In some embodiments, such as Figure 9 As shown, a dual-discriminator collaborative structure is constructed in the unsupervised adversarial training part of the target domain adversarial training based on historical source domain fault samples (which can be obtained through a small vibrating screen test bench) and the target domain. Discriminators D1 and D2 both adopt a 2-layer convolutional neural network with a kernel size of 3×1 and 128 and 256 channels respectively. Each layer is followed by batch normalization and ReLU activation function. Considering the unlabeled nature of the audio data from real mining vibrating screens, the two discriminators use the same network architecture and different inputs to learn complementary discriminative perspectives on vibrating screen faults. In the training phase (phase 1), the feature extraction network G is first fixed, and D1 and D2 are trained using the target domain data respectively. The optimization objective is to maximize the Manhattan distance between the predicted distributions of the two discriminators. ,in The samples are from the target domain. Simultaneously, labeled data from the source domain is introduced for supervised training. Where CE is the cross-entropy loss, , The source domain samples and labels are provided. The structure of the discriminator is as follows: Figure 10 As shown.

[0041] After a certain number of iterations in Phase 1, Phase 2 begins: The parameters of the dual discriminators are fixed, and the joint objective is to minimize the Manhattan distance between the predicted probability distributions of the two discriminators for the same target domain data, while simultaneously minimizing the source domain classification loss. This process involves back-optimizing the improved Transformer network. This step aims to enable the improved Transformer network to extract domain-invariant features that possess both fault detection capabilities and make the predictions of the two discriminators more consistent, thereby automatically identifying fault boundaries in the target domain through an adversarial mechanism. The generator (improved Transformer network) total loss function includes the source domain structured discrimination loss, the domain distribution difference measurement loss, and the physical-logic regularization loss, and its expression is: ,in to The balanced weights are adaptively changed during the training process. The training process adopts an alternating optimization strategy. In each training round, the discriminator is updated in 5 steps first, and then the generator is updated in 1 step. Through multiple rounds of iteration, the model's generalization ability in different vibrating screen and mining environment is realized. The training ends when the loss is lower than the set threshold or the number of training rounds exceeds 1000, realizing domain adaptation from laboratory to field and between different fields.

[0042] The source domain structured discriminant loss in the total loss function aims to establish a clear fault discrimination boundary, including the cross-entropy classification loss of historical source domain fault samples. And a metric learning loss for explicitly optimizing the topology of the feature space. .in, Includes intra-class compactness and inter-class separation mechanisms: a center loss function is introduced to optimize intra-class compactness. ,in For sample features, The central features of the corresponding categories are used to minimize the intra-class variance; inter-class separation is achieved by maximizing the distance between the centers of different categories. By minimizing the above loss, the model forms a clear discriminative structure on the source domain data. The domain distribution difference measure in the total loss function is the loss. It is used during the feature alignment stage of model training as an explicit supplement to adversarial training. Specifically, after mapping the extracted source domain features and target domain features to the same reproducing kernel Hilbert space, their distribution differences are calculated: .in , These are the mean embeddings of features from the source and target domains, respectively. By minimizing this loss, the feature distribution centers of the source and target domains are explicitly brought closer together, helping the model to perform domain transfer more stably. The physical-logic regularization loss in the total loss function... This is introduced as a regularization term to ensure that the features learned by the model conform to the physical laws of vibrating screen faults. Specifically, it includes continuity constraints on fault severity (i.e., fault features with adjacent severity levels should remain adjacent in space) and mutual exclusion constraints on fault types (i.e., mutually exclusive faults should not coexist with a high probability at the same time). These constraints, together with classification loss and domain adaptation loss, guide the backpropagation of gradients, ensuring that when transferred to the mining site, the model can not only accurately distinguish different fault types such as reduced spring stiffness, loose screen plates, and abnormal bearings, but also maintain the physical interpretability and logical consistency of the prediction results, thereby improving the stability and robustness of cross-domain fault identification.

[0043] In one embodiment of the present invention, when the probability of a certain fault category output is greater than the confidence threshold, a fault alarm mechanism is triggered. Intelligent decision-making is performed by combining the maintenance history and fault development trend of the mining vibrating screen, generating an early warning report containing the fault location, type, and suggested handling measures to guide on-site maintenance personnel in timely handling. The early warning output includes: (1) fault type label: normal, spring broken, screen plate loose, bearing damaged, lubrication abnormal; (2) early warning timestamp and duration; (3) recommended maintenance measures. The system response delay is controlled within 100ms to meet real-time monitoring requirements.

[0044] In summary, this invention integrates key technologies such as multi-sensor audio perception, selective state-space modeling, Transformer sequence modeling, and multi-system adversarial domain adaptation to construct a complete technical process from data acquisition, feature extraction and modeling to unsupervised domain adaptation and intelligent early warning. It has good robustness, portability and real-time performance, and is suitable for fault monitoring and early warning of mining vibrating screens in complex on-site operating environments.

Claims

1. A fault detection method for a mining vibrating screen based on audio and domain adaptation, characterized in that, Includes the following steps: Spatial collaborative acquisition and preprocessing of the operating audio of the target mining vibrating screen are performed to obtain the target audio signal; The target audio signal is converted into a Mel spectrogram, and an energy gradient-driven adaptive manifold topology unfolding strategy is used to serialize the Mel spectrogram to obtain the corresponding sequence. The sequence is encoded using an unsteady fractional-order state-space model to obtain a fractional-order state embedding vector that includes mechanical memory characteristics; By using a linearized frequency domain gated evolutionary deduction network to globally model the fractional-order state embedding vector, the pure frequency band energy drift characteristics of the mining vibrating screen are captured, and the corresponding probability distributions of various faults of the mining vibrating screen are output to complete the fault detection of the mining vibrating screen. The training method of the linearized frequency domain gated evolution inference network is based on unsupervised adversarial training between historical source domain fault samples and the target domain to achieve domain adaptation.

2. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 1, characterized in that, The specific method for spatial collaborative acquisition and preprocessing of the operating audio of the target mining vibrating screen includes the following steps: Multiple microphones are arranged in a circular array, with the center located above the center of the mining vibrating screen box. The array radius is 0.3 meters and the height from the surface of the mining vibrating screen is 1.2 meters. The raw audio data collected by each microphone is sequentially passed through a preamplifier, an anti-aliasing filter, and a dynamic range control module to achieve multi-channel spatial collaborative acquisition of the operating audio of the target mining vibrating screen. The audio signals acquired through multi-channel spatial collaborative acquisition are sequentially subjected to adaptive denoising, reverberation suppression, and signal enhancement to obtain the target audio signal.

3. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 1, characterized in that, The specific method for converting a target audio signal into a Mel spectrogram includes the following steps: The target audio signal is normalized to the range of [-1.00, 1.00] to obtain the normalized audio signal; The normalized audio signal is divided into sliding window segments at 60-second intervals, and each normalized audio segment is framed in 1-second units to obtain several more audio frames. For each frame of audio, a Hamming window is first used to reduce spectral leakage. Then, a short-time Fourier transform is used to extract the time-frequency structure. Finally, a Mel filter bank is introduced to compress the spectrum and generate a two-dimensional Mel spectrogram.

4. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 1, characterized in that, The specific method for serializing Mel spectrograms using an energy gradient-driven adaptive manifold topology expansion strategy includes the following steps: The energy gradient field of the two-dimensional Mel spectrogram is calculated to identify energy accumulation regions and gradient change ridges. Based on the distribution of the energy gradient field, an adaptive segmentation algorithm is used to divide the two-dimensional Mel spectrogram into several locally morphologically dynamically determined local spectrogram blocks. Each local spectral block is transformed into an independent one-dimensional feature vector through feature mapping. Then, according to the distribution logic of each local spectral block in the time-frequency space, all one-dimensional feature vectors corresponding to the same two-dimensional Mel spectrogram are recombined into a one-dimensional sequence to obtain the corresponding sequence.

5. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 1, characterized in that, The non-steady fractional state-space model includes a first branch and a second branch; the first branch includes a first normalized layer, a first linear layer, a one-dimensional convolutional layer, a non-steady fractional state-space model, a dot product layer, a second linear layer, and a first summation layer connected in sequence; the second branch includes a third linear layer and an activation function layer connected in sequence, the input of the third linear layer is the output of the first normalized layer, and the output of the activation function layer is another input of the dot product layer; The other input to the first additive layer is the input to the first normalization layer, and the output of the first additive layer is the output of the unsteady fractional-order state-space model.

6. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 5, characterized in that, The expression for calculating the unsteady fractional hierarchy is: in It is a fractional derivative operator; and These are the state transition matrix and the input matrix, respectively, both dynamically generated through an input-dependent hypernetwork; It is a fractional-order state embedding vector.

7. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 1, characterized in that, The linearized frequency-domain gated evolutionary derivation network is an improved Transformer network, comprising sequentially connected position coding layers, coding modules composed of multiple stacked linearized frequency-domain gated coding layers, and a classification head; wherein: The input to the positional coding layer is a fractional-order state embedding vector; The output of the classification head is a fault probability distribution; A single linearized frequency-domain gated coding layer includes a second normalization layer, a linearized frequency-domain gated attention module, a second summing layer, a third normalization layer, a feedforward neural network layer, and a third summing layer. The input to the linearized frequency-domain gated attention module is the output of the second normalization layer. The linearized frequency-domain gated attention module internally includes parallel query vector branches, key vector branches, and value vector branches. The query vector branch includes a Fourier transform layer, a frequency-domain gated unit, an inverse Fourier transform layer, and a first kernel mapping layer connected in sequence to filter out high-frequency environmental noise unrelated to the vibrating screen's operating frequency through a fast Fourier transform. The key vector... The branch includes a second kernel mapping layer, where the outputs of both the first and second kernel mapping layers are connected to a linear dot product layer to achieve linear reconstruction of the self-attention mechanism; the outputs of the linear dot product layer and the value vector branch are connected to a matrix multiplication layer; one input of the second summing layer is the output of the matrix multiplication layer, and the other input is the input of the second normalization layer; the output of the second summing layer is connected to the feedforward neural network layer via a third normalization layer, and the two inputs of the third summing layer are the output of the feedforward neural network layer and the input of the third normalization layer, respectively. The output of the third summing layer is the output of the linearized frequency domain gated coding layer.

8. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 7, characterized in that, Specific methods for globally modeling fractional-order state embedding vectors using linearized frequency-domain gated evolutionary derivation networks include: Multiple fractional-order state embedding vectors are recombined into a mining vibrating screen operation evolution sequence according to the chronological order of their occurrence. This mining vibrating screen operation evolution sequence is then used as the input to a linearized frequency domain gated evolution inference network, thereby capturing the pure frequency band energy drift characteristics of the mining vibrating screen and outputting the corresponding probability distribution of various faults of the mining vibrating screen.

9. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 7, characterized in that, Specific methods for unsupervised adversarial training based on historical source domain fault samples and target domain include: A dual-discriminator collaborative structure is constructed. The parameters of the improved Transformer network are fixed, and two discriminators are trained separately using target domain data. At the same time, the two discriminators are trained under supervision using historical source domain fault samples. The optimization objective of training in the target domain is to maximize the Manhattan distance between the prediction distributions of the two discriminators, so that the two discriminators can detect the misaligned feature boundaries in the target domain. The parameters of the dual discriminators are fixed, and the Manhattan distance between the predicted distributions of the two discriminators for the same target domain data and the classification loss of historical source domain fault samples are jointly optimized to improve the Transformer network in reverse. Through the above-mentioned alternating optimization, the improved Transformer network extracts domain-invariant features that not only have fault discrimination capabilities but also make the prediction results of the two discriminators tend to be consistent, thereby automatically discovering the fault discrimination boundary in the target domain and achieving domain adaptation; wherein, the improved Transformer network is used to acquire sample features during unsupervised adversarial training, and the sample features are the input of the discriminator.

10. The fault detection method for mining vibrating screen based on audio and domain adaptation according to claim 9, characterized in that, In unsupervised adversarial training, a total loss function is constructed, comprising source domain structured discrimination loss, domain distribution difference measurement loss, and physical logic regularization loss. Based on this total loss function, the improved Transformer network is optimized through backpropagation. The source domain structured discrimination loss includes cross-entropy classification loss of historical source domain fault samples and metric learning loss to constrain intra-class compactness and inter-class separability. The domain distribution difference measurement loss is the maximum mean difference between the extracted historical source domain features and target domain features mapped to the same regeneration kernel Hilbert space, used to explicitly narrow the feature distribution centers of the historical source domain and the target domain. The physical logic regularization loss includes continuity constraints on fault severity and mutual exclusion constraints on fault types, used to ensure that the transferred feature space conforms to the physical operating logic of the mining vibrating screen.

Images