Coal mining machine bearing early degradation detection method and system adopting spiking neural network

Abstract

The invention relates to the technical field of coal cutter fault diagnosis, in particular to a coal cutter bearing early degradation detection method and system adopting a spiking neural network. The method comprises the following steps: collecting a bearing acoustic signal in real time and carrying out blind source separation and band-pass filtering preprocessing; converting the preprocessed signal into a logarithmic Mel time-frequency spectrogram, and coding the logarithmic Mel time-frequency spectrogram into a pulse event sequence through peak detection; inputting the pulse sequence into a pulse neural network model comprising a pulse convolutional layer and a pulse recurrent neural network layer, and extracting state features; and judging whether the bearing is degraded early based on the classification or deviation calculation. According to the method, by simulating biological auditory sense and a nerve processing mechanism, utilizing the characteristics of high sensitivity and low power consumption of the pulse neural network to time sequence signals and combining targeted noise reduction and coding, sensitive and accurate recognition of the weak characteristics of early degradation of the bearing in a strong noise environment is achieved, and the method is particularly suitable for intelligent edge monitoring of underground equipment.

Description

Technical Field

[0001] This invention relates to the field of coal mining machine fault diagnosis technology, specifically to a method and system for detecting early degradation of coal mining machine bearings using a pulse neural network. Background Technology

[0002] As the core equipment in fully mechanized coal mining faces, the reliability of key components of the coal mining machine (such as the bearings of the cutting and traction sections) directly affects production safety and efficiency. Bearings operating under harsh conditions for extended periods will gradually degrade due to wear, fatigue, and spalling. If not detected in time, this can lead to sudden failures, causing production stoppages or even safety accidents. Therefore, early degradation detection and predictive maintenance of coal mining machine bearings are of great significance. Currently, vibration signal analysis is the mainstream technology for bearing condition monitoring. This involves installing accelerometers to collect signals and analyzing their spectrum, envelope, and other characteristics to identify faults. However, this method faces significant challenges in coal mining machine applications: First, the working environment of coal mining machines is extremely noisy, with strong background noise such as the impact of cutting coal and rock and the vibration of other rotating parts easily drowning out the weak vibration characteristics of early bearing degradation, leading to missed detections; second, traditional vibration analysis usually relies on expert experience to set thresholds or fixed frequency bands, which is insensitive to early and atypical degradation patterns, resulting in delayed early warnings; third, although intelligent diagnostic methods based on deep learning (such as CNN and RNN) can automatically extract features, they are mostly computationally intensive "static" networks with low efficiency and high power consumption in processing continuous time-series signals, making it difficult to meet the real-time and low-power edge computing requirements of underground equipment. In recent years, spiking neural networks, as a novel computational model that mimics the information processing mechanism of biological neurons, have shown potential in fields such as speech recognition and dynamic vision due to their high sensitivity to time-series signals, event-driven nature, sparse computation, and extremely low power consumption. However, there are still no mature applications in the acoustic monitoring of industrial equipment, especially in the early warning of ultra-early degradation of mechanical components under complex noise backgrounds. Therefore, there is an urgent need to develop a new intelligent detection method that can keenly capture early signs of degradation from strong noise and is suitable for edge deployment.

[0003] Therefore, the existing technology still needs further development. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for detecting early degradation of coal mining machine bearings using a pulse neural network, so as to solve the problems existing in the prior art.

[0005] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a method for detecting early degradation of coal mining machine bearings using a pulse neural network, comprising: S1. Acquire the original acoustic signal of the operating part of the coal mining machine bearing in real time, and preprocess the original acoustic signal to obtain the preprocessed audio signal; S2. Encode the preprocessed audio signal into a pulse event sequence; S3. Input the pulse event sequence into a pre-trained spiking neural network model for processing to obtain the state feature data of the bearing; S4. Based on the state characteristic data, determine whether the coal mining machine bearing has experienced early degradation.

[0006] Specifically, the preprocessing in step S1 includes: Blind source separation is performed on the original acoustic signal to separate the source signal related to bearing vibration; the separated source signal is bandpass filtered to retain the signal components of a preset frequency band to obtain the preprocessed audio signal.

[0007] Specifically, the preset frequency band is a characteristic frequency range that is sensitive to bearing degradation, which is determined in advance based on the inherent frequency and harmonic components of the coal mining machine bearing.

[0008] Specifically, step S2 is as follows: The preprocessed audio signal is converted into a time-spectrum graph; energy peaks in the time and frequency dimensions of the time-spectrum graph are detected; each detected energy peak is encoded as a pulse event, and the pulse event sequence is generated in chronological order.

[0009] Specifically, the time spectrum is a log-Mel spectrum; the process of encoding into a pulse event is as follows: when the energy value of a frequency unit in a continuous time frame exceeds the energy value of its adjacent time frame and frequency unit and reaches a preset threshold, a pulse is generated at that time point and on the corresponding frequency channel.

[0010] Specifically, the spiking neural network model comprises a spiking convolutional layer and a spiking recurrent neural network layer connected in sequence; wherein, the spiking convolutional layer is used to extract spatial features related to bearing degradation from the input spiking event sequence; the spiking recurrent neural network layer is used to capture the temporal dependency of the spatial features over time and output the state feature data.

[0011] Specifically, the pulse recurrent neural network layer is a pulse long short-term memory network layer.

[0012] Specifically, in step S4, determining whether the coal mining machine bearing has experienced early degradation involves: The state feature data is input into a classifier to obtain a classification result indicating whether the bearing is in a healthy state or an early degradation state; or, the deviation between the state feature data and the baseline feature data of a healthy bearing is calculated, and if the deviation continues to exceed a preset threshold, it is determined that early degradation has occurred.

[0013] Specifically, the classifier is a spiking neural layer or a fully connected layer; the deviation is obtained by calculating Euclidean distance, cosine similarity, or Mahalanobis distance.

[0014] According to a second aspect of the present invention, a coal mining machine bearing early degradation detection system employing a pulse neural network is provided, comprising: The signal acquisition and preprocessing module is used to acquire the original acoustic signals of the operating parts of the coal mining machine bearing in real time, and to preprocess the original acoustic signals to obtain the preprocessed audio signals. A pulse coding module is used to encode the preprocessed audio signal into a pulse event sequence; The spiking neural network processing module integrates a pre-trained spiking neural network model to process the spiking event sequence and output the bearing's state feature data. The degradation judgment module is used to determine whether the bearing of the coal mining machine has undergone early degradation based on the state characteristic data.

[0015] Beneficial effects: Compared with existing technologies, the method and system for detecting early degradation of coal mining machine bearings using a pulse neural network provided by this invention have the following significant advantages: 1. Achieved ultra-early and highly sensitive early warning under strong noise interference: This invention actively separates and purifies bearing-related acoustic signals from complex mixed sound fields through a two-stage preprocessing process of "blind source separation - characteristic frequency band filtering," effectively suppressing environmental noise. Furthermore, by mimicking the auditory mechanism through "time-frequency peak detection coding," the sound signal is converted into a sparse sequence of pulse events. This process itself filters out stationary backgrounds and amplifies abrupt changes related to bearing dynamic events. Finally, utilizing the unique sensitivity and powerful feature extraction capabilities of spiking neural networks to spatiotemporal pulse patterns, it can identify weak patterns characterizing initial micro-damage in bearings that are difficult to detect using traditional methods from these refined pulse streams. This significantly advances the fault warning time, achieving true "early" detection and significantly improving the sensitivity and reliability of the warning.

[0016] 2. Meets the stringent requirements of high efficiency and low power consumption for underground edge deployment: The core processing unit of this invention employs a spiking neural network (SNN). Its event-driven nature means that computation is only performed when there is an input pulse (representing a valid acoustic event), resulting in extremely low power consumption when no events occur. Simultaneously, pulse coding compresses the high-data-rate continuous audio stream into a sparse pulse sequence, significantly reducing data throughput. The entire algorithm flow (preprocessing, encoding, SNN inference) is designed as lightweight streaming processing with low computational overhead. These characteristics make this invention highly suitable for deployment in embedded systems within coal mining machines or underground edge computing gateways where computing resources, storage space, and power supply are limited. It enables local real-time data processing and decision-making without uploading massive amounts of raw audio data to the cloud, reducing transmission bandwidth pressure and latency, aligning with the development trend of edge intelligence in the Industrial Internet of Things (IIoT).

[0017] 3. Strong Adaptability and Generalization Capabilities: The spiking neural network model proposed in this invention integrates spiking convolutional layers and spiking recurrent neural network layers, enabling it to automatically learn and extract degradation-related spatiotemporal features from acoustic pulse sequences without the need for complex manual feature engineering. In particular, the use of a spiking long short-term memory network endows the model with the ability to capture long-range temporal dependencies, allowing it to understand the evolution of bearing states over time and exhibiting better adaptability to degradation patterns of different degrees and rates. Furthermore, the "classification" and "deviation monitoring" strategies provided in the judgment phase can achieve high-precision classification using labeled data, and can also establish health benchmarks and monitor anomalies in an unsupervised manner when only healthy data is available, enhancing the method's practicality and generalization capabilities under different data conditions.

[0018] 4. A complete and reliable hardware and software collaborative solution is constituted: This invention not only provides an innovative detection method but also constructs a complete detection system. From signal acquisition hardware, preprocessing algorithms, and intelligent processing models to the judgment and output module, this system forms a closed loop from physical signals to decision information. This system can be integrated into the coal mining machine as an independent intelligent sensing terminal, realizing all-weather, automated online monitoring and early warning. It upgrades the traditional periodic, offline inspection mode to a continuous, online intelligent prediction mode, greatly improving the intelligence level and maintenance efficiency of equipment health management, and providing solid technical support for ensuring the safe, stable, and efficient operation of the coal mining machine. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method for detecting early degradation of coal mining machine bearings using a pulse neural network, provided in a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the system composition of the early degradation detection system for coal mining machine bearings using a pulse neural network, provided in a specific embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0021] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0022] Please see Figure 1 This invention provides a method for detecting early degradation of coal mining machine bearings using a pulse neural network, comprising: S1. Acquire the original acoustic signal of the operating part of the coal mining machine bearing in real time, and preprocess the original acoustic signal to obtain the preprocessed audio signal.

[0023] It should be further explained that this method constitutes a complete end-to-end edge intelligent detection process. In step S1, a high-fidelity, wide-bandwidth acoustic sensor (such as a microphone array or acoustic emission sensor) is installed close to the bearing housing of the coal mining machine to continuously collect the sound signals generated by the bearing during operation, forming a raw acoustic signal stream. This signal includes the vibration sound of the bearing, the meshing sound of gears, the noise of cutting coal and rock, and the background noise of other components of the coal mining machine, with an extremely low signal-to-noise ratio. The purpose of preprocessing is to initially improve the signal-to-noise ratio, creating conditions for subsequent precise information processing.

[0024] S2. Encode the preprocessed audio signal into a pulse event sequence.

[0025] It should be further explained that in step S2, the continuous audio signal is creatively converted into a sparse, event-driven pulse sequence. This encoding method mimics the signal transmission mechanism of the biological auditory nerve, generating "pulses" only at moments and feature dimensions sensitive to signal changes. This greatly compresses the amount of data and highlights the key time-frequency features in the signal, making it naturally suitable for the processing paradigm of spiking neural networks.

[0026] S3. Input the pulse event sequence into a pre-trained spiking neural network model for processing to obtain the state feature data of the bearing.

[0027] It should be further explained that in step S3, the encoded pulse event sequence is input into a spiking neural network model specifically designed and trained for bearing acoustic degradation pattern recognition. This model can efficiently process such spatiotemporal pulse patterns, learning and extracting high-dimensional features characterizing the bearing's health status from sparse pulse streams.

[0028] S4. Based on the state characteristic data, determine whether the coal mining machine bearing has experienced early degradation.

[0029] It should be further explained that in step S4, based on the high-dimensional features output by the model, an automated judgment and early warning of whether the bearing has undergone early degradation is achieved through a preset logic or classifier.

[0030] Understandably, the core advantage of this method lies in its ability to capture subtle acoustic signs that characterize the initial damage of bearings (such as micropitting and initial spalling) from a background of strong noise, which are difficult to detect by traditional methods, through the event-driven characteristics and high sensitivity to time-series signals of spiking neural networks. This enables true "early" warnings. At the same time, this method is computationally efficient and consumes little power, making it very suitable for deployment on edge computing devices at coal mining sites where computing power and energy are limited.

[0031] Specifically, the preprocessing in step S1 includes: performing blind source separation on the original acoustic signal to separate the source signal related to bearing vibration; performing bandpass filtering on the separated source signal to retain the signal components of a preset frequency band, thereby obtaining the preprocessed audio signal.

[0032] It should be further explained that the preprocessing process in step S1 adopts a two-stage optimization strategy of "separation first, then focusing" to cope with the harsh acoustic environment of the coal mining machine. The first stage of processing is blind source separation. The acoustic signal at the coal mining machine's working site is a mixture of multiple sound sources (bearings, gearboxes, motors, cutting heads, hydraulic systems, coal and rock crushing, etc.). Directly processing the mixed signal would easily mask the weak fault characteristics of the bearings. This invention preferably uses an independent component analysis algorithm for blind source separation. Assume that M sensors collect M mixed signal vectors. The goal of ICA is to find a demixing matrix. This makes the output signal Each component The components are statistically independent as much as possible. By maximizing the non-Gaussianity of each component of the output signal (e.g., using negative entropy as a metric), the algorithm can estimate the inverse of the mixing matrix, thereby separating the individual source signal estimates, including the source signal component of interest related to bearing vibration. Denotes the M-dimensional mixed observation signal vector at time t. It is the M×M dimensional unmixing matrix to be solved. This is the estimated vector of the separated source signal. After separation, operators can manually select or identify the target bearing source signal based on prior knowledge (such as the bearing's characteristic frequency range) or the signal's energy distribution, or through simple automatic filtering rules (such as selecting the component with the highest energy proportion within the bearing's characteristic frequency band). The second stage of processing involves bandpass filtering. Even after separating the bearing source signal, its spectrum remains very wide. Abnormal vibration energy caused by early degradation is typically concentrated near the bearing's natural frequencies (such as the fault characteristic frequencies of the inner ring, outer ring, rolling elements, and cage calculated from bearing geometry parameters) and their harmonics. Energy in other frequency bands is mostly irrelevant interference or noise. Therefore, this invention focuses on the separated bearing source signal... Digital bandpass filtering is performed to retain only signals within a preset characteristic frequency range that is sensitive to bearing degradation. For example, for a typical rolling bearing in the cutting section of a coal mining machine, its outer ring fault characteristic frequency... It is likely around 85Hz. Considering the broadband excitation effect of modulation sidebands and early faults, the preset frequency band can be set to... That is, approximately the range of [60Hz, 255Hz]. The signal obtained after filtering. This is the pre-processed audio signal, whose signal-to-noise ratio is significantly improved compared to the original signal, and the signal energy is concentrated in the frequency band most likely to contain fault information, laying a solid foundation for subsequent accurate encoding and recognition.

[0033] Understandably, this two-stage preprocessing strategy has extremely strong environmental noise robustness, enabling it to "extract" the sound of the target bearing from an extremely complex sound field and "purify" the components of key frequency bands, effectively solving the core problem of early fault characteristics being submerged under strong background noise.

[0034] Specifically, the preset frequency band is a characteristic frequency range that is sensitive to bearing degradation, which is determined in advance based on the inherent frequency and harmonic components of the coal mining machine bearing.

[0035] It should be further explained that determining the preset frequency band is a process based on a combination of bearing physical models and historical experience data, ensuring the scientific nature and effectiveness of the filtering operation. First, based on the specific model and installation location of the bearing to be monitored on the coal mining machine, its geometric parameters need to be obtained: bearing pitch diameter... , rolling element diameter Contact angle Number of rolling elements and spindle speed (Revolutions / second). Calculate the inner ring failure characteristic frequency using the standard bearing failure characteristic frequency calculation formula. Outer ring fault characteristic frequency characteristic frequency of rolling element failure and cage failure characteristic frequency .in, These are the rotational frequencies of the shaft containing the bearing. These frequencies are the impact characteristic frequencies that periodically appear in the vibration / acoustic signal when local damage occurs to various components of the bearing. However, in the early degradation stage, the damage is very small, and the resulting impact may not strictly occur at these calculated frequencies. The spectrum will exhibit a certain broadband characteristic, with sidebands appearing near the characteristic frequencies. Therefore, the preset frequency band should not be a narrow band around a single characteristic frequency, but rather a wider band covering the fundamental frequency and several of its harmonics. A preferred setting method is: the lower limit of the preset frequency band... Pick upper limit Pick For example, if the calculated main characteristic frequency is between 80Hz and 200Hz, the preset frequency band can be set to [40Hz, 1000Hz]. This range ensures coverage of all frequency components that may be excited by early faults, while avoiding interference from excessively high frequencies (usually dominated by noise) and excessively low frequencies (usually occupied by the fundamental frequencies of other large rotating components). This frequency band can be pre-configured in the system and can be directly applied to bearings of the same model of coal mining machine, thus standardizing the detection process.

[0036] Understandably, this physics-based frequency band determination method makes the preprocessing highly targeted, enabling it to adaptively focus on the spectral region where the specific physical phenomenon of "bearing degradation" may produce an acoustic response, thereby maximizing the extraction of effective information, filtering out irrelevant noise, and improving the signal-to-noise ratio and feature correlation in subsequent processing.

[0037] Specifically, step S2 involves: converting the preprocessed audio signal into a time-frequency spectrogram; detecting energy peaks in the time and frequency dimensions of the time-frequency spectrogram; encoding each detected energy peak as a pulse event; and generating a sequence of pulse events in chronological order.

[0038] It should be further explained that step S2 implements an efficient, bio-inspired signal-to-pulse conversion mechanism. First, the one-dimensional time-domain audio signal... Converted to a two-dimensional time-spectrum diagram A commonly used method is the short-time Fourier transform. A window function (such as the Hanning window) and a fixed window length (e.g., corresponding to 1024 sampling points) and overlap rate (e.g., 50%) are set, and the window is slid along the time axis to calculate the signal spectrum within each time window. This results in a two-dimensional matrix where the horizontal axis represents time and the vertical axis represents frequency, with each point representing the signal energy (amplitude spectrum or power spectrum) at a specific time and frequency. The time-spectrum plot visually displays the time-varying frequency domain characteristics of the signal. Next, in the time-spectrum plot... Peak detection is performed on the spectrum. This mimics the "feature-triggered" response of neurons in the auditory system to sounds of specific frequencies. For each point in the spectrum... to its energy value It is compared with its neighbors in time and frequency. A specific peak detection algorithm is: if Simultaneously satisfying a frequency greater than its own A moment before and the next moment The energy value, and greater than its energy value at the same time. Upper adjacent frequency and The energy value is considered to be It is a local energy peak point. This condition can be expressed as: , , ,and Finally, the encoding process. Each detected energy peak point... , representing in Time, frequency channel Significant signal activity was observed. We encode this event as a pulse event. After the entire processing is complete, we will obtain a series of timestamped data. The ordered pulse events form a sparse pulse event sequence. Compared to encoding the energy of each moment and each frequency channel, this peak encoding method has extremely high sparsity. It only retains the most significant feature points in the signal that are most likely triggered by bearing dynamic events (such as the collision between the rolling element and the defect), while suppressing smooth background noise. This is very much in line with the characteristics of spiking neural networks in processing sparse, event-driven data.

[0039] Understandably, this encoding scheme achieves efficient data compression, significantly reducing the amount of data input to the spiking neural network. At the same time, by extracting the "significant" features of the signal in the time-frequency domain, it plays a role in feature enhancement, making it easier for the subsequent neural network to learn key patterns related to changes in bearing state.

[0040] Specifically, the time spectrum is a log-Mel spectrum; the process of encoding into a pulse event is as follows: when the energy value of a frequency unit in a continuous time frame exceeds the energy value of its adjacent time frame and frequency unit and reaches a preset threshold, a pulse is generated at that time point and on the corresponding frequency channel.

[0041] It should be further noted that this invention preferably uses a logarithmic Mel-scale spectrum rather than a standard linear frequency scale spectrum. This is because the human ear's perception of sound frequencies is not linear; it has high resolution in the low-frequency range and low resolution in the high-frequency range. The Mel-scale simulates this non-linear perception. (The text then repeats the previous sentence about linear frequencies.) (Hz) converted to Mel frequency The formula is: .in, It is a linear frequency measured in Hertz. This is the converted Mel frequency. After calculating the linear spectrum, the spectrum is smoothed and its dimensionality reduced using a set of triangularly overlapping Mel filters. Then, the logarithm of the energy for each Mel channel is taken to obtain the final logarithmic Mel spectrum. The advantages of using logarithmic Mel spectrograms are: Firstly, it is more in line with auditory characteristics and can enhance the ability to represent sound features; Secondly, the dimensionality reduction effect of the Mel filter bank reduces the amount of data in the frequency dimension, thereby improving computational efficiency; Third, logarithmic operations compress the dynamic range, allowing the energy of weak signals (potentially corresponding to early faults) to be highlighted. Regarding the specific implementation of the encoding process, the aforementioned peak detection condition can be further strengthened by adding an absolute energy threshold condition to filter out peaks with excessively weak energy that may be caused by random noise. A preferred encoding rule is: in time... and Mel frequency channel Above, if and only if its logarithmic energy value A pulse event will only be generated at this time point and on this frequency channel if all three of the following conditions are met: (1) and ; (2) and ; (3) ,in This is a preset energy threshold. The choice of [the appropriate setting] is crucial, and a preferred approach is to adaptively determine it based on the background noise level. For example, an audio segment acquired under known device health conditions could be selected, and the global average of its log-Mel spectrum could be calculated. and standard deviation Then set .in, It is an adjustable parameter, with an optimal value between 2 and 3. When This means that only peak values ​​with energy values ​​exceeding 2.5 standard deviations above the average level of healthy background noise will be encoded as pulses, effectively suppressing random noise interference. This encoding process ultimately outputs a pulse... The represented list of impulse events can be directly mapped to the input layer of a spiking neural network. Time, Number One input neuron fired a pulse.

[0042] Understandably, peak detection coding using log-Mel spectrograms combined with adaptive thresholds not only inherits the advantage of Mel scale conforming to auditory perception, but also further enhances the robustness of the coding through threshold filtering, ensuring that the pulse sequence input to the spiking neural network mainly contains acoustic events significantly related to the bearing state, greatly improving the accuracy and reliability of subsequent feature extraction.

[0043] Specifically, the spiking neural network model comprises a spiking convolutional layer and a spiking recurrent neural network layer connected in sequence; wherein, the spiking convolutional layer is used to extract spatial features related to bearing degradation from the input spiking event sequence; the spiking recurrent neural network layer is used to capture the temporal dependency of the spatial features over time and output the state feature data.

[0044] It should be further explained that the spiking neural network model designed in this invention is a deep spatiotemporal spiking neural network, whose structure is specifically optimized for the spatiotemporal characteristics of bearing acoustic pulse event sequences. The model input is an encoded pulse event sequence, which can be viewed as a two-dimensional matrix (time × input channel), where each channel corresponds to a Mel frequency channel. The first layer is a spiking convolutional layer. This layer contains multiple three-dimensional spiking convolutional kernels that slide along the time and spatial (frequency) dimensions. Each convolutional kernel learns to detect specific local spatiotemporal patterns in the input pulse stream. For example, a convolutional kernel might learn to detect a pattern of "pulses appearing successively on several adjacent frequency channels within several consecutive time steps," which might correspond to the spectral characteristics of short-term impacts generated by periodic impacts at bearing damage points. The neurons in the spiking convolutional layer employ a leak-in firing model. Each neuron has a membrane potential. Its dynamics are determined by the input pulse. And the decision on the leaked items: .in, It is the membrane time constant, which controls the rate of potential decay. It is the weighted pulse current from the convolution input. When Exceeding a distribution threshold At that time, the neuron fires an output pulse, and simultaneously Reset to a reset potential The bearing enters a brief refractory period. Through this mechanism, the spiked convolutional layer extracts higher-level, space-frequency features related to bearing degradation from the raw, sparse spike events, outputting another set of feature spike sequences. The second layer is a spiked recurrent neural network layer. Bearing degradation is a continuous dynamic process with long-range temporal dependencies in its features. For example, the weak impacts caused by early failures may be sparse and irregular at the beginning, but their periodicity gradually increases over time. To capture this temporal evolution, a spiked recurrent neural network layer, such as a spiked long short-term memory network, is introduced after the convolutional layer. The Spiking LSTM unit contains input gates, forget gates, output gates, and cell states, but all operations are driven by spike events. It can remember long-term contextual information and decide when to pass the memory to the output. This layer receives the feature pulse sequence output by the pulsed convolutional layer. At each time step, its internal state is updated based on the current input pulse and the hidden state of the previous time step, ultimately outputting a high-dimensional, dense state feature vector that comprehensively represents all historical information from the beginning of the sequence to the current time step. This vector is the "state feature data," which is a highly abstract and condensed representation of the bearing's health status from the start of operation to the current time step.

[0045] Understandably, this network architecture combines the powerful local spatiotemporal pattern extraction capability of the pulse convolutional layer with the long-range temporal dependency modeling capability of the pulse recursive layer. It can learn deep spatiotemporal dynamic features that characterize the evolution of bearings from health to early degradation from complex and sparse acoustic pulse event streams, providing strong feature support for making accurate degradation judgments.

[0046] Specifically, the pulse recurrent neural network layer is a pulse long short-term memory network layer.

[0047] It should be further noted that, preferably, the spiking recurrent neural network layer employs a spiking long short-term memory network layer. Spiking LSTM is a variant of traditional LSTM in the spiking neural network domain, where both neurons and gating signals are transmitted in pulse form. A basic Spiking LSTM unit at time... The calculation involves the following core steps (taking the pulse rate coding approximation as an example): 1. Input Modulation: Modulates the current input pulse vector And the hidden state impulse vector of the previous time step (After transformation through a fully connected layer) the pulse rate of the candidate cell state is calculated. .

[0048] 2. Gating signal generation: also based on and The input gate pulse rate is calculated using an independent sigmoid activation function (the output value is considered as the pulse firing rate). Forget gate pulse rate and output gate pulse rate .

[0049] 3. Cell state update: Cell state The update of (an analog, non-pulse) combines the preservation of long-term memory with the addition of new information: .in This represents element-wise multiplication. Forget gate. Control the previous state The degree of retention, the input gate rate Control candidate state The degree of inclusion.

[0050] 4. Hidden State Output: The final hidden state (pulse firing rate) is determined by the current cell state through a non-linear transformation (such as tanh) and controlled by the output gate. . This can be further interpreted as the pulse firing rate, used to drive the next layer or as a final feature. During training, due to the high complexity of the BPTT algorithm that directly processes pulses, a surrogate gradient method is often used. This involves defining a differentiable surrogate function around the pulse firing moment to approximate the gradient of the non-differentiable pulse firing function. During the inference phase, the trained gating weights and state update rules are directly used to process the input pulse sequence. The advantage of using Spiking LSTM in this invention is that its gating mechanism (especially the forget gate) allows it to adaptively decide which historical information to remember or forget, which is crucial for processing long-term continuous signals such as bearing acoustic signals. For example, it can learn to "remember" the periodic patterns of anomalous pulse events in the past few minutes while "forgetting" earlier information that may be irrelevant to the current state. This selective memory ability allows the model to accurately capture subtle temporal pattern changes that indicate the onset of degradation.

[0051] Understandably, the use of Spiking LSTM layers enables the entire model to have excellent long-term temporal correlation modeling capabilities, and can identify periodic and trending degradation signs from acoustic event streams over long periods of time. This is of key significance for providing early warnings hours or even days before a failure occurs, and is a core component for improving the "early detection" capability.

[0052] Specifically, in step S4, determining whether the coal mining machine bearing has experienced early degradation involves: inputting the state feature data into a classifier to obtain a classification result indicating whether the bearing is in a healthy state or an early degradation state; or, calculating the deviation between the state feature data and the baseline feature data of a healthy bearing, and if the deviation continuously exceeds a preset threshold, determining that early degradation has occurred.

[0053] It should be further explained that step S4 provides two flexible and complementary degradation judgment strategies to adapt to different application scenarios and data conditions, as detailed below: The first approach is classifier-based. This is suitable for scenarios with a certain number of labeled "healthy" and "early degradation" samples. During model training, a simple classifier is connected to the end of the spiking neural network (such as Spiking CNN-LSTM), after the state feature vector output by the recursive layer. This classifier can be a fully connected layer followed by a softmax activation function, outputting the probabilities of "healthy" and "early degradation" for the two neurons, respectively. During training, the entire network (including the feature extraction network and the classifier) ​​is trained end-to-end using data labeled with health / degradation. During the inference (detection) phase, the state feature vector of the final time step obtained from real-time processing is input into this classifier, and the class with the higher probability value is used as the judgment result. This method is direct, explicit, and the output is easy to understand.

[0054] The second approach is based on deviation (or health index). This is suitable for scenarios where it's difficult to obtain a large number of clearly defined "early degradation" labels, and it focuses more on unsupervised or self-supervised learning. In this approach, a "health baseline" needs to be established first. During the initial operational phase when the bearing is known to be healthy, sufficient duration of normal sound data is collected. Through steps S1-S3 described above, its state feature data is extracted (e.g., taking the statistics of the feature vector output by the Spiking LSTM layer at each time step, such as the mean vector). Covariance Matrix During the real-time monitoring phase, for newly acquired data with a length of... The time window data is also used to extract the sequence of its state feature vectors. Then, the deviation between the sequence characteristics and the health baseline is calculated. A commonly used and effective method is the Mahalanobis distance: .in, It is the mean vector of the feature vectors within the current window. Mahalanobis distance considers the correlation and variance between the various dimensions of the features, making it more reasonable than Euclidean distance. A threshold is set. If the calculated deviation continuous Each (e.g., N=5) time window exceeds This triggers an early degradation alarm. Threshold The setting can be based on the distribution of deviations in health data, for example, set to three times the standard deviation of the average deviation. This method does not require a "early degradation" label; it only requires health data to establish a benchmark, making it more universal, and the magnitude of the deviation can intuitively reflect the degree of degradation.

[0055] Understandably, this invention provides two judgment mechanisms: the classifier method offers high accuracy and direct judgment when labeled data is plentiful; the deviation method remains effective even when labels are scarce and can quantify the degree of anomaly. Both can be based on the same high-performance spiking neural network feature extractor, providing users with flexible options and enhancing the method's practicality and applicability.

[0056] Specifically, the classifier is a spiking neural layer or a fully connected layer; the deviation is obtained by calculating Euclidean distance, cosine similarity, or Mahalanobis distance.

[0057] It's worth noting that two implementations compatible with spiking neural networks are provided for the classifier. One is a spiking neural layer, such as one or two fully connected spiking neural network layers. Its input is the spiking firing rate or spiking events output from the previous layer (SpikingLSTM), and the output layer typically consists of two neurons, corresponding to two classes. The network transmits class information through time encoding or rate encoding. This approach maintains the entirely event-driven nature of the entire model, theoretically has the lowest power consumption, and is particularly suitable for deployment on edge devices where energy efficiency is paramount. The other is a traditional fully connected layer (simulated neurons). The state feature vector (a set of simulated values, such as membrane potential or spiking firing rate) output from the last time step of the Spiking LSTM layer is used as input, passed through one or more fully connected layers, and finally output as class probabilities using a softmax function. This approach is more stable in training, easier to integrate with mainstream deep learning frameworks, and is a more common choice in engineering. For the calculation of deviation, various metrics are provided to suit different needs. Euclidean distance is the simplest metric. It calculates the straight-line distance between the current feature mean and the healthy baseline mean in the feature space. It has low computational cost but is sensitive to feature scale and correlation. Cosine similarity focuses on the direction of the vector rather than its length. The deviation can be defined as It is not sensitive to the absolute magnitude of features, but focuses more on the relative distribution changes of feature patterns. Mahalanobis distance, as mentioned earlier, is statistically superior because it considers the variance and correlation of each dimension of the feature, effectively "whitening" the feature space and making the distance metric more reasonable. In practical applications, the choice can be made based on the distribution characteristics of the feature data. Generally, if the features are well normalized and the correlation between dimensions is weak, Euclidean distance is a simple and effective choice. If the physical meanings and dimensions of the feature vectors are different and there is a strong correlation, Mahalanobis distance is a better choice because it automatically adjusts the importance of different dimensions.

[0058] It is understood that this invention refines the specific implementation and measurement methods of the judgment layer, providing technicians with clear and selectable implementation schemes, ensuring the operability and flexibility of the method. Different choices can balance computational complexity, accuracy requirements, and deployment environment constraints, enabling this invention to adapt to different application scenarios, from high-precision cloud analysis to low-power embedded terminals.

[0059] Please see Figure 2 The present invention provides another embodiment, which provides a coal mining machine bearing early degradation detection system using a pulse neural network. The coal mining machine bearing early degradation detection system using a pulse neural network includes: The signal acquisition and preprocessing module 100 is used to acquire the original acoustic signal of the operating part of the coal mining machine bearing in real time, and preprocess the original acoustic signal to obtain the preprocessed audio signal. Pulse coding module 200 is used to encode the preprocessed audio signal into a pulse event sequence; The spiking neural network processing module 300 integrates a pre-trained spiking neural network model for processing the spiking event sequence and outputting the bearing's state feature data. The degradation judgment module 400 is used to determine whether the coal mining machine bearing has experienced early degradation based on the state characteristic data.

[0060] It should be further explained that this detection system is a hardware-software co-processing system integrating hardware interfaces and intelligent algorithms, deployed in the edge computing unit of the coal mining equipment or a nearby industrial gateway. The signal acquisition and preprocessing module 100 includes hardware and software components. The hardware component includes high-sensitivity, waterproof and shockproof acoustic sensors (such as microphones or acoustic emission probes) and their signal conditioning circuits (such as amplification and anti-aliasing filtering), responsible for converting the vibration sound of the bearing into an analog electrical signal, and then converting it into a digital audio stream through a high-precision analog-to-digital converter. The software component runs on the processor of the edge computing unit, loaded with blind source separation algorithms and digital filter programs, preprocessing the acquired multi-channel or single-channel raw acoustic signals in real time, and outputting the audio signal after noise reduction and frequency band focusing. The pulse coding module 200 is a dedicated signal processing program that receives the preprocessed audio signal and executes algorithm steps including short-time Fourier transform, log-Mel transform, peak detection, and pulse generation, converting the continuous audio stream into a sparse pulse event stream in real time. The spiking neural network processing module 300 is the core intelligent unit of the system, internally containing or loading the parameters of the fully trained network model of the aforementioned structure. This module can be an optimized inference engine running on a general-purpose processor (CPU / GPU) or deployed on a dedicated neuromorphic computing chip, processing the pulse stream from the encoding module in real time with extremely low power consumption and outputting a high-dimensional state feature vector. The degradation judgment module 400, according to the system configuration, loads a classifier model or stores health baseline features and thresholds. It receives the state feature data output by the neural network module, performs classification or deviation calculation and comparison logic, and ultimately generates a "healthy" or "early degradation" judgment result. This result and warning information are then uploaded to the central control center or displayed locally via a communication interface (such as CAN bus or industrial Ethernet). The entire system operates in a pipeline manner, achieving fully automated, real-time processing from sound acquisition to intelligent warning.

[0061] Understandably, this system engineered and productized advanced spiking neural network algorithms, forming an independent, embeddable intelligent detection terminal. It fully leverages the low power consumption and high time-sensitivity advantages of SNNs, combined with targeted preprocessing and encoding, to achieve unmanned, intelligent, and early-stage online monitoring of the health status of coal mining machine bearings in the complex and harsh underground coal mine environment, significantly improving the level and reliability of predictive maintenance for equipment.

[0062] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, implement the method for detecting early degradation of coal mining machine bearings using a spiking neural network. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0063] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0064] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0065] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0066] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting early degradation of coal mining machine bearings using a pulse neural network, characterized in that, Includes the following steps: S1. Acquire the original acoustic signal of the operating part of the coal mining machine bearing in real time, and preprocess the original acoustic signal to obtain the preprocessed audio signal; S2. Encode the preprocessed audio signal into a pulse event sequence; S3. Input the pulse event sequence into a pre-trained spiking neural network model for processing to obtain the state feature data of the bearing; S4. Based on the state characteristic data, determine whether the coal mining machine bearing has experienced early degradation.

2. The method for detecting early degradation of coal mining machine bearings according to claim 1, characterized in that, The preprocessing in step S1 specifically includes: Blind source separation is performed on the original acoustic signal to separate the source signal related to bearing vibration; the separated source signal is bandpass filtered to retain the signal components of a preset frequency band to obtain the preprocessed audio signal.

3. The method for detecting early degradation of coal mining machine bearings according to claim 2, characterized in that, The preset frequency band is a characteristic frequency range that is sensitive to bearing degradation, based on the inherent frequency and harmonic components of the coal mining machine bearing.

4. The method for detecting early degradation of coal mining machine bearings according to claim 1, characterized in that, Step S2 specifically involves: The preprocessed audio signal is converted into a time-spectrum graph; energy peaks in the time and frequency dimensions of the time-spectrum graph are detected; each detected energy peak is encoded as a pulse event, and the pulse event sequence is generated in chronological order.

5. The method for detecting early degradation of coal mining machine bearings according to claim 4, characterized in that, The time-frequency spectrum is a log-Mel spectrum; the process of encoding into a pulse event is as follows: when the energy value of a frequency unit in a continuous time frame exceeds the energy value of its adjacent time frame and frequency unit and reaches a preset threshold, a pulse is generated at that time point and on the corresponding frequency channel.

6. The method for detecting early degradation of coal mining machine bearings according to claim 1, characterized in that, The spiking neural network model comprises a spiking convolutional layer and a spiking recurrent neural network layer connected in sequence; wherein, the spiking convolutional layer is used to extract spatial features related to bearing degradation from the input spiking event sequence; the spiking recurrent neural network layer is used to capture the temporal dependency of the spatial features over time and output the state feature data.

7. The method for detecting early degradation of coal mining machine bearings according to claim 6, characterized in that, The pulse recurrent neural network layer is a pulse long short-term memory network layer.

8. The method for detecting early degradation of coal mining machine bearings according to claim 1, characterized in that, In step S4, determining whether the coal mining machine bearing has experienced early degradation specifically involves: The state feature data is input into a classifier to obtain a classification result indicating whether the bearing is in a healthy state or an early degradation state; or, the deviation between the state feature data and the baseline feature data of a healthy bearing is calculated, and if the deviation continues to exceed a preset threshold, it is determined that early degradation has occurred.

9. The method for detecting early degradation of coal mining machine bearings according to claim 8, characterized in that, The classifier is a spiking neural layer or a fully connected layer; the deviation is obtained by calculating Euclidean distance, cosine similarity, or Mahalanobis distance.

10. A coal mining machine bearing early degradation detection system employing a pulse neural network, used to implement the detection method as described in any one of claims 1-9, characterized in that, include: The signal acquisition and preprocessing module is used to acquire the original acoustic signals of the operating parts of the coal mining machine bearing in real time, and to preprocess the original acoustic signals to obtain the preprocessed audio signals. A pulse coding module is used to encode the preprocessed audio signal into a pulse event sequence; The spiking neural network processing module integrates a pre-trained spiking neural network model to process the spiking event sequence and output the bearing's state feature data. The degradation judgment module is used to determine whether the bearing of the coal mining machine has undergone early degradation based on the state characteristic data.

Images