Intelligent industrial equipment state monitoring device and method

Abstract

The invention relates to an intelligent industrial equipment state monitoring device and method, and the method comprises the steps: installing a vibration sensor, an acoustic sensor and an image sensor at key parts of industrial equipment, and comprehensively collecting the operation data of the equipment. In the data preprocessing stage, vibration and acoustic signals are filtered and denoised, and image data are denoised and enhanced. Extracting multi-dimensional characteristic parameters from three aspects of vibration, acoustics and images; a random forest algorithm is adopted for model training, and the method comprises the steps of data preparation, segmentation, decision tree construction, voting mechanism, majority voting, model verification and optimization, testing and the like. In the model identification stage, the real-time features are input into the trained model, and the equipment state is determined through decision tree prediction and voting summarization. And when an abnormal state is detected, setting a multi-level alarm response according to the probability, and generating a maintenance suggestion. According to the method, the whole-process intelligent monitoring of the industrial equipment is realized, and the fault identification accuracy and the maintenance efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring, and in particular to an intelligent industrial equipment status monitoring device and method thereof. Background Art

[0002] In modern industrial production, it is crucial to ensure the stable operation of industrial equipment. With the continuous development of industrial technology, the complexity and intelligence of industrial equipment are increasing, and traditional equipment status monitoring methods have gradually revealed many shortcomings. Traditional equipment status monitoring often relies on a single type of sensor, such as using only vibration sensors to judge the operating status of equipment. However, the information that a single sensor can provide is limited, and it is difficult to fully and accurately reflect the overall condition of the equipment. For example, some potential faults caused by acoustic anomalies or surface physical changes may not be discovered in time by vibration sensors alone. At the same time, traditional monitoring methods are relatively simple in data processing and lack effective processing of noise in complex environments. In actual industrial environments, background noise and environmental interference are prevalent. These interferences will seriously affect the quality of data collected by sensors, making it difficult to accurately judge the status of equipment. For example, without effective filtering and noise reduction processing, the noise in the vibration signal may mask the weak characteristics of early equipment failures, resulting in the failure to be discovered in time.

[0003] The patent with publication number KR20240091364A discloses a machine fault diagnosis and prediction system based on deep learning, but it still has certain limitations in practical applications: in terms of feature extraction, it only focuses on a few simple feature parameters, and it is difficult to capture subtle changes and complex patterns in the equipment status. For example, only extracting a single feature such as vibration frequency cannot fully reflect the mechanical state and potential fault types of the equipment. In addition, equipment status monitoring also has limitations in model building and decision-making. Using only a single classification model is prone to overfitting or underfitting problems, and the recognition accuracy of different fault types is not high. Moreover, there is a lack of effective verification and optimization mechanisms, making it difficult to adapt to the ever-changing equipment operating environment and failure modes.

[0004] In summary, in order to meet the requirements of modern industry for high reliability and efficient operation of equipment, there is an urgent need for a more intelligent, comprehensive and accurate industrial equipment status monitoring method that can comprehensively utilize data collected by multiple sensors to perform effective data preprocessing, feature extraction and model training to achieve accurate monitoring of equipment status and fault diagnosis. Summary of the invention

[0005] 1. Technical issues to be resolved

[0006] In view of the shortcomings of the prior art, the present invention aims to provide an intelligent industrial equipment status monitoring device and method thereof, which solves the problems existing in the prior art. In terms of data acquisition, the vibration sensor, acoustic sensor and image sensor are integrated to comprehensively collect the status information of the equipment in different dimensions, breaking the limitations of the traditional single sensor. In the data preprocessing stage, the vibration signal is accurately filtered and denoised by an adaptive algorithm combined with the LMS algorithm, the background noise is effectively removed by spectrum subtraction of the acoustic signal, and the image data is median filtered and enhanced, which greatly improves the data quality. Feature extraction covers multiple key parameters in the three fields of vibration, acoustics and images, and comprehensively reflects the equipment status and potential faults. The model training adopts the random forest algorithm, and multiple decision trees are constructed by Bootstrap sampling. The output is determined by the majority voting mechanism, and the performance is ensured to be stable by combining the verification set evaluation and parameter optimization. In the abnormality detection and decision support stage, it can accurately mark the abnormal state, identify the fault type, set a multi-level alarm response, trigger the alarm and generate maintenance suggestions, and continuously monitor the feedback, which effectively guarantees the stable operation and efficient maintenance of industrial equipment.

[0007] (II) Technical solution

[0008] To achieve the above object, the present invention provides the following technical solution: a method for monitoring the state of intelligent industrial equipment, comprising the following steps:

[0009] Data collection: Install sensor modules at key locations of industrial equipment to continuously monitor the operating status of the equipment, collect vibration, acoustics, and surface image data in real time, and transmit them to the computer system module;

[0010] Data preprocessing: The computer system module filters and de-noises the collected original vibration signals and acoustic signals, eliminates background noise and environmental interference, improves signal clarity and usability, performs median filtering and denoising on the collected image data, enhances the image through histogram equalization and contrast stretching, and improves the accuracy of image feature recognition;

[0011] Feature extraction: The computer system module extracts features from the preprocessed data, including vibration features, acoustic features, and image features;

[0012] Model training: The computer system module combines vibration, acoustic and image features into a comprehensive feature vector, which is input into the random forest model. The Bootstrap sampling method is used to randomly extract subsamples from the training set to construct a decision tree. Each decision tree independently makes predictions and outputs classification results and probability values. The final output is then determined by the voting aggregation of all trees. The model performance indicators are evaluated through the validation set. The model parameters are optimized using cross-validation and grid search methods. Finally, the test set is input into the trained model to evaluate its performance on unseen data and analyze the confusion matrix.

[0013] Model identification: The computer system module inputs the vibration, acoustic and image features extracted in real time into the trained random forest model for model prediction; the decision tree processes the input feature vector, starting from the root node and passing it layer by layer to the leaf node to generate the prediction result; the prediction results of all decision trees are collected, and the majority decision mechanism is used to determine the final prediction output and output the equipment state label and the prediction probability value of each state;

[0014] Anomaly detection: The computer system module performs anomaly detection. When the model outputs an abnormal state and the predicted probability of the state is higher than the set alarm threshold, the system will mark the state as an abnormal state that requires attention and identify the specific fault type. The probability value is used for decision-making: different levels of alarm response are set according to the predicted probability value;

[0015] Decision support and further actions: If the model identifies a fault state and the predicted probability exceeds the alarm threshold, the system will trigger an alarm and notify the operator. Based on the specific fault type identified, the system will generate corresponding maintenance recommendations.

[0016] Preferably, in the data preprocessing step, the filtering and noise reduction of the vibration signal include: dynamically adjusting the filtering parameters according to the changes in the background noise through an adaptive algorithm to achieve more accurate noise suppression, performing Fourier transform on the vibration signal, designing an FIR filter, applying an adaptive algorithm to estimate the noise characteristics and adjust the filter weights, extracting the effective signal and evaluating the noise suppression effect.

[0017] Preferably, in the data preprocessing step, filtering and denoising the acoustic signal includes:

[0018] Signal acquisition and preprocessing: obtain the original acoustic signal s(t) from the acoustic sensor;

[0019] Convert the time domain signal to the frequency domain: Convert the time domain signal s(t) to the frequency domain signal S(f) through fast Fourier transform;

[0020] Noise spectrum estimation: During the operation of the equipment, background noise signals are collected and their spectrum is obtained through FFT conversion. A representative noise spectrum N(f) is obtained through multiple sampling and averaging calculations.

[0021] Spectral subtraction: subtract the estimated noise spectrum N(f) from the spectrum S(f) of the original signal, and introduce the noise reduction coefficient α to obtain the denoised frequency domain signal;

[0022] Inverse Fourier transform: The processed frequency domain signal is converted back to the time domain through the inverse fast Fourier transform (IFFT) to obtain the denoised time domain signal.

[0023] Preferably, in the feature extraction step, vibration feature extraction includes: extracting characteristic parameters from the preprocessed vibration signal, including time domain features such as amplitude, root mean square value, peak-to-peak value and kurtosis, frequency domain features such as main frequency, frequency component, spectral energy distribution and harmonic components, and time-frequency domain features such as short-time Fourier transform and wavelet transform.

[0024] Preferably, in the feature extraction step, acoustic feature extraction includes: analyzing the preprocessed acoustic signal, extracting frequency domain features such as acoustic frequency distribution, sound pressure level, and spectrum center, as well as time domain features such as envelope analysis, autocorrelation function, and pitch analysis.

[0025] Preferably, in the feature extraction step, image feature extraction includes: extracting physical features that can reflect the surface condition of the device from the preprocessed image data, including texture features, edge features, shape features and color features.

[0026] Preferably, the step of edge feature extraction includes:

[0027] Gaussian filtering: Use a Gaussian filter to perform a convolution operation on the image. The Gaussian filter performs weighted averaging through a Gaussian function to smooth the pixel values ​​in the image.

[0028] Calculate the gradient strength and direction of the image: Use the Sobel operator to calculate the horizontal gradient and vertical gradient respectively, calculate the gradient strength by the square root of the sum of the squares of the horizontal and vertical gradients, and calculate the gradient direction by the inverse tangent function to indicate the direction of the edge;

[0029] Non-maximum suppression: Check the gradient strength of each pixel and compare the pixel with the neighboring pixels along the gradient direction according to the gradient direction, and only keep the position where the pixel is a local maximum;

[0030] Determine strong edges and weak edges: pixels with gradient strength greater than a high threshold are marked as strong edges, pixels with gradient strength between a high threshold and a low threshold are marked as weak edges, and pixels with gradient strength lower than a low threshold are suppressed as non-edges;

[0031] Generate edge image: Start from the strong edge and check its adjacent weak edges. If a weak edge is connected to any strong edge, keep it as part of the edge, otherwise suppress it. Finally, generate a continuous edge image and output it.

[0032] Preferably, in the model training step, the comprehensive feature vector = [vibration feature | acoustic feature | image feature].

[0033] Preferably, the model prediction comprises the following steps:

[0034] Input feature vector: The standardized feature vector is input to each decision tree in the random forest, and these trees process the input features independently;

[0035] Layer-by-layer processing: The decision tree starts from the root node and transfers the input features layer by layer according to the predefined decision rules;

[0036] Node judgment and splitting: At each node, the tree makes a decision based on the feature threshold and passes the data to the corresponding child node;

[0037] Arrival at a leaf node: The feature vector reaches a leaf node, which represents a specific device status category;

[0038] Generate prediction results: Each decision tree generates a prediction result based on the leaf node reached by the feature vector.

[0039] An intelligent industrial equipment status monitoring device, comprising:

[0040] Sensor module: including a vibration sensor for detecting vibration signals of the device; an acoustic sensor for collecting sound signals generated when the device is running; and an image sensor for photographing the surface of the device;

[0041] Computer system module: used to receive and process the data collected by the sensor module;

[0042] Communication and alarm module: It is used to issue alarms and has communication functions to transmit monitoring data and alarm information to the computer system module.

[0043] (III) Beneficial effects

[0044] The purpose of the present invention is to provide an intelligent industrial equipment status monitoring device and method thereof, which has significant beneficial effects. First, the data acquisition method of multi-sensor fusion greatly enriches the monitoring information. The comprehensive use of vibration sensors, acoustic sensors and image sensors can comprehensively reflect the operating status of the equipment from multiple angles such as mechanical vibration, sound characteristics and surface physical conditions, overcoming the problem of limited information of traditional single sensors, and providing a more comprehensive basis for accurately judging the status of the equipment.

[0045] Secondly, sophisticated data preprocessing significantly improves data quality. The filtering and noise reduction of vibration and acoustic signals effectively removes background noise and interference in complex industrial environments, making the effective information in the signal clearer and more discernible. The denoising and enhancement operations in image preprocessing improve the accuracy of image feature recognition and provide a more reliable data basis for subsequent fault diagnosis.

[0046] Furthermore, comprehensive feature extraction can provide in-depth insights into the status of equipment. Multi-parameter extraction of vibration features in the time domain, frequency domain, and time-frequency domain, comprehensive analysis of acoustic features in the frequency domain and time domain, and extraction of image features in terms of texture, edge, shape, and color enable potential equipment failures to be discovered more promptly and accurately, helping to intervene in the early stages of failures and reduce equipment damage and production losses.

[0047] Then, the advanced random forest model training method improves the fault identification accuracy. Multiple decision trees are constructed through Bootstrap sampling, and the final output is determined by the majority voting mechanism, which avoids the overfitting or underfitting problem of a single classification model, can better capture the complex patterns of the data, and has a higher identification accuracy for different types of faults.

[0048] Finally, the perfect anomaly detection and decision support mechanism provides timely and effective response measures. When an abnormal state is detected, the fault type can be accurately marked and a multi-level alarm response can be set according to the predicted probability. When the alarm is triggered, corresponding maintenance suggestions are generated, and the equipment status is continuously monitored and fed back, providing strong support for equipment maintenance and management, ensuring that the equipment can be quickly handled when problems occur, and ensuring the stable progress of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0050] Figure 1 It is a flowchart of an intelligent industrial equipment status monitoring device and method thereof in an embodiment of the present application.

[0051] Figure 2 This is a flow chart of data preprocessing in an intelligent industrial equipment status monitoring device and method in an embodiment of the present application.

[0052] Figure 3 This is a flowchart of feature extraction in an intelligent industrial equipment status monitoring device and method in an embodiment of the present application.

[0053] Figure 4 This is a flowchart of model training in an intelligent industrial equipment status monitoring device and method in an embodiment of the present application.

[0054] Figure 5 This is a flowchart of model identification in an intelligent industrial equipment status monitoring device and method in an embodiment of the present application. DETAILED DESCRIPTION

[0055] The following will be combined with the attached examples of the present invention Figure 1 - Figure 5 A clear and complete description is given to the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] A method for monitoring the state of intelligent industrial equipment comprises the following steps:

[0057] 1. Data collection:

[0058] 1) Sensor installation: Before monitoring, vibration sensors, acoustic sensors and image sensors are installed at key locations of industrial equipment and calibrated to ensure that the accuracy and sensitivity of the sensors meet the requirements. Equipment status baseline collection: When the equipment is operating normally, the initial status data is collected as the baseline data for subsequent testing and comparison. The baseline data includes the vibration frequency, acoustic mode and surface image features of the normal operation of the equipment.

[0059] 2) Real-time data collection: The sensor module continuously monitors the operating status of the equipment and collects vibration, acoustic and surface image data in real time. The collection frequency and resolution are dynamically adjusted according to the working environment, importance and monitoring requirements of the equipment to ensure the comprehensiveness and accuracy of the data.

[0060] 2. Data preprocessing:

[0061] 1) Signal filtering and noise reduction: Filter and reduce the noise of the collected original vibration signals and acoustic signals to eliminate background noise and environmental interference and improve the clarity and usability of the signals. The specific steps include:

[0062] 11) Filtering and noise reduction of vibration signals: Through adaptive algorithms, filter parameters are dynamically adjusted according to changes in background noise to achieve more accurate noise suppression. This method is particularly suitable for complex industrial environments and can adjust the filter response in real time to maximize the retention of effective signal information.

[0063] 111) Signal preprocessing

[0064] Preliminary signal processing: In the collected vibration signal, there are high-frequency background noise and low-frequency vibration signals under normal working conditions of the equipment. The system collects vibration signals in real time and performs Fourier transform to convert the signal from time domain to frequency domain. The frequency domain signal at this time can be expressed as:

[0065] X(f)=S(f)+N(f)

[0066] Where X(f) is the spectrum of the total signal, S(f) is the spectrum of the vibration signal of the equipment, and N(f) is the noise spectrum.

[0067] 112) Adaptive filter design

[0068] FIR filter structure, the output of the FIR filter can be expressed as:

[0069]

[0070] Where: y(n) is the output of the filter; M is the order of the filter (i.e. the length of the filter). k (n) is the kth tap coefficient (weight), which is adaptively adjusted as time n changes. x(nk) is the delayed sample of the input signal.

[0071] 113) Application of adaptive algorithms

[0072] Noise estimation: The adaptive algorithm first estimates the noise characteristics in the current environment. The system will continuously monitor the spectral characteristics of the background noise through the LMS algorithm to determine its changing trend. It includes the following specific steps:

[0073] Step 1: Noise model initialization

[0074] When starting noise estimation, we first establish an initial noise model through preliminary analysis of the ambient noise, and input the preprocessed frequency domain signal. This model contains the initial frequency distribution and amplitude information of the noise.

[0075] Step 2: Noise Power Spectral Density Estimation

[0076] The system estimates the power spectral density (PSD) of the noise by averaging the signal over multiple time periods. The PSD estimation formula is:

[0077]

[0078] Among them, P N (f) is the noise power spectral density at frequency f, N is the number of time periods used for averaging, and X n (f) is the signal spectrum in the nth time period.

[0079] Step 4: Filter weight adjustment

[0080] Combined with the LMS algorithm, the system adjusts the filter weight w(n) in real time according to the PSD of the noise. The real-time update formula of the LMS algorithm is:

[0081] w(n+1)=w(n)+μe(n)x(n)

[0082] Where w(n) is the weight vector of the filter at the current time n. μ is the step size factor, which determines the speed of weight update. e(n) = d(n) - y(n) is the current error signal, d(n) is the expected signal, and y(n) is the output signal of the filter. x(n) is the signal vector input at the current time

[0083] In frequency ranges with greater noise intensity, the bandwidth of the filter is increased, thereby more effectively suppressing the noise in these frequency bands.

[0084] Step 5: Dynamic Update of Noise Estimate

[0085] Over time, the characteristics of the ambient noise may change. The system continuously monitors the noise spectrum, updates the noise model in real time, and adjusts the filter response based on the latest noise estimate. In this way, the filter can adapt to dynamically changing noise environments.

[0086] 114) Effective signal extraction

[0087] Signal reconstruction: After adaptive filtering, the system reconstructs the vibration signal and extracts the effective information. This includes retaining key vibration characteristics of equipment operation, such as specific frequency, amplitude, acceleration and other parameters. These parameters are crucial for equipment status monitoring and fault diagnosis.

[0088] Noise suppression effect evaluation: The system evaluates the noise suppression effect by comparing the signal quality before and after filtering. If there is still significant noise in the filtered signal, the system will record this and make corresponding adjustments. At the same time, the system performs statistical analysis on the filtered signal to confirm the validity and integrity of the signal.

[0089] 12) Filtering and noise reduction of acoustic signals: Reduce background noise by calculating the noise spectrum and subtracting it from the signal spectrum. This method can effectively remove continuous background noise while retaining useful acoustic information about the operation of the equipment. It includes the following specific steps:

[0090] 121) Signal acquisition and preprocessing

[0091] First, the original acoustic signal s(t) is obtained from the acoustic sensor. This signal contains the useful sound information generated by the operation of the equipment and the background noise in the environment.

[0092] 122) Convert time domain signal to frequency domain

[0093] In order to analyze and subtract the noise spectrum, the time domain signal needs to be converted into a frequency domain signal. Through the fast Fourier transform (FFT), the time domain signal s(t) can be converted into a frequency domain signal S(f), where S(f) is the spectral component of the signal at frequency f:

[0094] S(f)=FFT{s(t)}

[0095] 123) Noise spectrum estimation

[0096] In the frequency domain, the spectrum of the background noise is estimated. The noise spectrum is obtained as follows:

[0097] Noise collection: During the operation of the equipment, the background noise signal n(t) is collected and its spectrum is obtained through FFT conversion. A representative noise spectrum N(f) is obtained through multiple sampling and averaging.

[0098] 124) Spectral Subtraction

[0099] Subtract the estimated noise spectrum N(f) from the spectrum S(f) of the original signal, and introduce the noise reduction coefficient α to obtain the denoised frequency domain signal

[0100]

[0101] The purpose of introducing the noise reduction coefficient α is to remove noise more gently to ensure that the useful acoustic information of the equipment is preserved.

[0102] 125) Inverse Fourier Transform

[0103] The processed frequency domain signal The denoised time domain signal is converted back to the time domain through the inverse fast Fourier transform (IFFT)

[0104]

[0105] This time domain signal is the signal after noise suppression processing, which contains useful acoustic information about the operation of the equipment, while the background noise is significantly weakened.

[0106] 126) Subsequent processing and analysis

[0107] Denoised acoustic signal Further analysis can be performed, such as extracting acoustic features for equipment condition monitoring and fault detection.

[0108] The frequency domain noise reduction method can effectively remove background noise from the acoustic signal of equipment operation. The key to this process is to accurately estimate the noise spectrum and appropriately subtract it to retain the useful acoustic information of equipment operation while significantly reducing interference. This method is particularly suitable for dealing with continuous background noise and can improve the quality of acoustic signals in complex industrial environments.

[0109] 2) Image preprocessing: Preprocess the collected image data, including denoising and image enhancement operations, to improve the accuracy of image feature recognition.

[0110] 21) Denoising

[0111] The median filter is used to remove noise from the image, including: replacing the value of the central pixel with the median value of the pixel neighborhood, thereby effectively removing noise. Formula definition: Let f(x,y) be the original image, the filter window size is n×n, and the median filtered image g(x,y) is calculated as follows:

[0112] g(x,y)=median{f(i,j)},(i,j)∈neighborhood(x,y)

[0113] 22) Image Enhancement

[0114] 221) Histogram Equalization

[0115] Histogram equalization enhances the overall contrast of an image by adjusting the grayscale histogram distribution of the image. This method can stretch the grayscale range of the image, making the bright and dark areas of the image more distinct, and highlighting details such as wear and cracks. It includes the following steps:

[0116] Step 1: Calculate the grayscale histogram h(i) of the image.

[0117] Step 2: Calculate the cumulative distribution function CDF(i)

[0118]

[0119] 3. Use CDF(i) to map the grayscale of the original image to the equalized grayscale.

[0120] After histogram equalization, the visual effect of low-contrast images is significantly improved, and the details are clearer.

[0121] 222) Contrast stretch

[0122] Contrast stretching is to increase the contrast between bright and dark areas by expanding the dynamic range of the image. This method can enhance the subtle features of the device surface in the image, making wear or cracks easier to identify. The formula for contrast stretching is:

[0123]

[0124] where f mi n and f ma x is the minimum and maximum grayscale value of the original image, g is min and g maxis the grayscale range of the target image. By contrast stretching, the features on the device surface are made more prominent, which is helpful for subsequent fault detection.

[0125] 3. Feature extraction: The computer system module extracts features from the preprocessed data, including vibration features, acoustic features, and image features.

[0126] 1) Vibration feature extraction

[0127] Vibration feature extraction is to obtain important parameters that can reflect the mechanical state of the equipment by analyzing the pre-processed vibration signal. It includes the following specific steps:

[0128] 11) Time domain feature extraction

[0129] Amplitude: Amplitude is a basic feature of vibration signals, indicating the maximum amplitude of the signal change. Find the maximum and minimum values ​​from the time series data, and then calculate the difference between them, i.e., the amplitude. The amplitude can directly reflect the operating status of the equipment. For example, abnormal large vibrations may indicate mechanical failure or imbalance of the equipment.

[0130] Root mean square value: The RMS value is the energy characteristic of the vibration signal and represents the effective value of the signal. It can comprehensively reflect the overall strength of the vibration signal and is of great significance in equipment health status monitoring. The calculation formula is:

[0131]

[0132] where x i is the sample value of the signal, and N is the total number of samples.

[0133] Peak-to-peak value: The peak-to-peak value indicates the maximum difference between the positive peak and the negative peak of the vibration signal, and is used to evaluate the severity of the vibration signal. A larger peak-to-peak value indicates that the equipment has abnormal impact or collision.

[0134] Kurtosis: Kurtosis is an indicator that measures the shape characteristics of the vibration signal and can reflect the degree of impulse components in the signal. High kurtosis values ​​are usually related to early failure of equipment or impact events. The formula for calculating kurtosis is:

[0135]

[0136] Where μ is the mean of the signal and σ is the standard deviation.

[0137] 12) Frequency domain feature extraction

[0138] Main frequency: The main frequency is the frequency with the most concentrated energy in the vibration signal spectrum. By converting the time domain signal into the frequency domain through Fourier transform, the main vibration frequency of the equipment can be identified to help locate possible fault sources.

[0139] Frequency components: Extract multiple frequency components from the vibration signal and analyze their amplitude distribution. These frequency components can help identify different types of mechanical faults, such as gear wear or bearing damage.

[0140] Spectral Energy Distribution: Calculates the energy distribution of frequency domain signals in different frequency bands to analyze the spectral characteristics of vibration signals. Increased energy in a specific frequency band may indicate the presence of a specific mechanical fault.

[0141] Harmonic content: By analyzing the harmonic content in the signal, it is possible to detect resonance phenomena or periodic fault characteristics of the equipment, such as shaft misalignment or gear meshing problems.

[0142] 13) Time-frequency domain feature extraction

[0143] Short-time Fourier transform: STFT divides the time domain signal into segments and performs Fourier transform on each segment to generate a time-frequency spectrogram, which reflects the change of the frequency of the vibration signal over time. This method is particularly suitable for analyzing non-stationary signals, such as transient failures of equipment. The extraction method is: divide the signal into multiple short windows (such as Hanning window, Hamming window), perform FFT on each window, and obtain the time-frequency spectrogram. By sliding the time window, the spectral characteristics of the signal changing over time are obtained.

[0144] Wavelet transform: Wavelet transform can decompose vibration signals into frequency components of different scales and analyze the instantaneous frequency changes of signals. Wavelet coefficients can reflect sudden events and detail changes in signals and are an effective tool for detecting early faults. The extraction method is to select a suitable wavelet basis (such as Daubechies wavelet), perform continuous or discrete wavelet transform on the signal, and extract wavelet coefficients of each scale.

[0145] 2) Acoustic feature extraction

[0146] Acoustic feature extraction extracts the acoustic features generated by the equipment during operation by analyzing the preprocessed acoustic signals. These features help identify abnormal acoustic patterns of the equipment. The following are the main acoustic feature extraction methods:

[0147] 21) Frequency domain feature extraction

[0148] Frequency distribution: The frequency distribution of the acoustic signal is analyzed by Fourier transform. The normal operation of the equipment usually has a specific frequency distribution, and abnormal changes in the spectrum may indicate the occurrence of a fault.

[0149] Sound Pressure Level: SPL is a measure of the strength of an acoustic signal, indicating the energy of the sound. By monitoring changes in SPL, it is possible to determine whether the equipment is generating abnormal noise, such as excessive friction or collision.

[0150]

[0151] Where p is the sound pressure and p0 is the reference sound pressure (set to 20 μPa here).

[0152] Spectrum Center: The spectrum center is a weighted average of frequencies and reflects the center of gravity of the signal spectrum. Movement of the spectrum center indicates changes in the sound characteristics of the equipment, such as material wear or loose components. The formula for calculating the spectrum center is:

[0153]

[0154] Among them, f i is the frequency, A i is the amplitude corresponding to the frequency.

[0155] 22) Time domain feature extraction

[0156] Envelope analysis: Envelope analysis is used to extract the amplitude envelope of acoustic signals, which can identify periodic impact or friction sounds in equipment. The frequency components of the envelope signal are often used to diagnose bearing or gear failures. Extraction method: The envelope of the signal is calculated through the Hilbert transform. The envelope signal is analyzed, such as frequency component extraction, to diagnose bearing or gear failures. If there is a clear periodic signal in the envelope, it may correspond to a fault in a rotating or reciprocating component in the equipment.

[0157] Autocorrelation function: Autocorrelation analysis extracts the periodic characteristics of a signal by calculating the similarity between the signal and itself after delay. This method is particularly suitable for identifying periodic mechanical vibrations or acoustic phenomena.

[0158]

[0159] Where xi is the signal sample value and τ is the delay time.

[0160] Tone analysis: Tone analysis is used to detect specific tones produced by the equipment. Changes in the tone may indicate loose or worn parts in the equipment. Extraction method: Perform FFT on the signal and analyze the amplitude of each frequency in the spectrum. Identify and track changes in specific tones, such as the fundamental frequency or its multiples when the equipment is operating normally. Tone shifts or the appearance of new tones may indicate loose or damaged parts.

[0161] 3) Image feature extraction

[0162] Image feature extraction is to extract important physical features that can reflect the surface condition of the equipment, such as wear and cracks, from the preprocessed image data. These features provide intuitive evidence for the physical failure of the equipment. The following are the main image feature extraction methods:

[0163] 31) Texture feature extraction

[0164] The image is grayed out, the relative positions of pixel pairs are set (e.g., distance is 1, directions are 0, 45, 90, and 135 degrees), a gray-level co-occurrence matrix is ​​generated, and texture features, including contrast, correlation, and uniformity, are extracted. These features can reveal the roughness and degree of wear on the device surface.

[0165] 32) Edge feature extraction

[0166] Edge detection: Use the Canny edge detection algorithm to extract edge features in the image. Edge features can be used to identify cracks, scratches, or wear boundaries on the surface of the device. The following steps are included:

[0167] 321) Gaussian Filtering: Performs a convolution operation on an image using a Gaussian filter. The Gaussian filter smoothes the pixel values ​​in the image by weighted averaging using a Gaussian function (normal distribution). The result is a smoothed image where bursts of noise are effectively reduced while most edge structures are still preserved.

[0168] 322) Calculate the gradient strength and direction of the image: Calculate the gradient of the image in the horizontal and vertical directions, using the Sobel operator to calculate the X direction (horizontal gradient) and Y direction (vertical gradient). The gradient strength is calculated by the square root of the sum of the squares of the horizontal and vertical gradients. The gradient direction is calculated by the inverse tangent function to indicate the direction of the edge (for example, horizontal, vertical, or diagonal).

[0169] 323) Non-maximum suppression: Check the gradient strength of each pixel and compare it with the neighboring pixels along the gradient direction according to the gradient direction. Only the position where the pixel is a local maximum is retained, that is, its gradient strength is greater than the neighboring pixels in its direction. This step will make the edge more refined and accurate.

[0170] 324) Determine strong and weak edges: Use two thresholds: a higher threshold (high threshold) is used to determine strong edges, and a lower threshold (low threshold) is used to mark weak edges. Pixels with gradient strength greater than the high threshold are marked as strong edges (deterministic edges); pixels with gradient strength between the high threshold and the low threshold are marked as weak edges (potential edges). Pixels with gradient strength lower than the low threshold are suppressed as non-edges.

[0171] 325) Generate edge image: Start with a strong edge and check its adjacent weak edges. If a weak edge is connected to any strong edge, it is retained as part of the edge, otherwise it is suppressed. In this way, all significant edges are retained while isolated weak edges are avoided, and finally a clear and continuous edge image is formed.

[0172] Canny edge detection is particularly suitable for high-precision edge detection, such as identifying tiny cracks or corrosion marks on the surface of equipment.

[0173] 33) Shape feature extraction

[0174] Shape descriptors: Shape descriptors such as Fourier descriptors and Hu moments can effectively describe the shape features of objects in images. These features can be used to identify the defect shapes on the surface of equipment, such as the length, direction, and area of ​​cracks.

[0175] Extraction method:

[0176] Fourier descriptor: Perform Fourier transform on the coordinates of the contour points to obtain a frequency domain description.

[0177] Hu moment: Calculates the seven Hu invariant moments of the image to reflect the shape characteristics of the image.

[0178] Shape comparison: Compare the extracted shape features with the standard template to identify the defect shape.

[0179] Contour analysis: By extracting and analyzing the contours of objects in the image, the geometric features of the contours, such as the perimeter and area, are calculated. Contour analysis can help identify the scope and severity of wear on the surface of the equipment.

[0180] Extraction method:

[0181] Contour detection: Use edge detection or binarization to extract the contours of objects.

[0182] Feature calculation: Calculate geometric features such as the perimeter, area, and convex hull area of ​​the contour.

[0183] Defect Analysis: Analyze shape changes of contours and identify surface defects.

[0184] 34) Color feature extraction

[0185] Color histogram: Analyzes the color distribution of an image and extracts the color features of the image through the color histogram. These features can help identify oxidation, corrosion, or other color changes on the surface of the device.

[0186] Extraction method:

[0187] Convert the image to HSV or RGB color space. Calculate the color histogram of each channel and count the number of pixels of each color. Use the color histogram to describe the overall color characteristics of the image and identify surface color changes. Use the K-means clustering algorithm to perform color segmentation on the image and extract the color characteristics of specific areas. This method can help identify the material characteristics or degree of wear of different areas on the surface of the device.

[0188] 4. Model Training

[0189] 1) Data preparation

[0190] 11) Standardization: To ensure that different types of features are processed at the same scale, all features need to be standardized before combining to make them consistent with the data distribution during training. Use Z-score standardization to adjust the feature values ​​to a distribution with a mean of 0 and a standard deviation of 1.

[0191] 12) Feature combination

[0192] Feature connection: Connect various features according to their categories. These include vibration features, acoustic features, and image features:

[0193] Vibration characteristics: [amplitude, RMS value, peak-to-peak value, kurtosis, dominant frequency, frequency component, spectral energy distribution, harmonic component, short-time Fourier transform, wavelet transform]

[0194] Acoustic characteristics: [acoustic frequency distribution, sound pressure level, spectrum center, envelope analysis, autocorrelation function, tone analysis]

[0195] Image features: [contrast, correlation, uniformity, edge features, shape features, color features]

[0196] Forming a feature vector: Connect all features together in a predetermined order to form a high-dimensional feature vector. For example, if there are 10 vibration features, 6 acoustic features, and 6 image features, then the final feature vector length is 22:

[0197] Comprehensive feature vector = [vibration feature | acoustic feature | image feature].

[0198] 13) Input of comprehensive feature vector

[0199] Input model: The standardized comprehensive feature vector is input into the random forest model for classification and recognition. Each vector represents the device status at a sampling moment, and the model performs real-time analysis and outputs the status label.

[0200] 14) Label definition: Label the status of the equipment based on historical data (such as normal, fault type A, fault type B, etc.). These labels will serve as the target output of the model.

[0201] 141) Collect historical data

[0202] Data source: Historical data comes from sensor data, acoustic data, image data, etc. collected during the operation of the equipment. These data record the characteristics of the equipment in different states.

[0203] Data wrangling: Organize and clean data to ensure data integrity and accuracy and remove outliers and noise.

[0204] 142) State identification and event classification

[0205] Status classification: Based on the working principle and operation process of the equipment, identify the possible status categories of the equipment, such as normal operation, minor fault, major fault, etc.

[0206] Event classification: Fault status is further classified into different types of faults. Mechanical equipment includes abnormal vibration, abnormal noise, abnormal temperature, etc.

[0207] 143) Define status labels

[0208] Label definition: Define clear labels for each state and fault type. These labels will be used as the target output of the model. Labels are represented by digital encoding (such as 0 for normal, 1 for fault type A, 2 for fault type B) or text encoding (such as "Normal", "FaultA", "FaultB").

[0209] 144) Label allocation

[0210] Label matching: Match the defined labels with the feature vectors in the historical data. Assign labels to each set of features in the historical data through manual labeling, expert diagnosis, or automatic labeling tools (such as rule-based systems).

[0211] 145) Label verification and calibration

[0212] Verification process: Verify the data of the assigned label to ensure the accuracy of the label. The correctness of the label can be verified through expert review or using equipment logs (such as maintenance records).

[0213] Calibration and adjustment: If the label assignment is found to be inaccurate or erroneous, the label needs to be calibrated and adjusted to ensure that the label of the training data can truly reflect the status of the device.

[0214] 2) Data segmentation

[0215] Divide the dataset: Divide the dataset into training set, validation set and test set (specifically in a ratio of 8:1:1).

[0216] Data standardization: Standardize feature data to eliminate the impact of different feature dimensions and ensure that the model can process all features evenly.

[0217] 3) Build a decision tree:

[0218] Bootstrap sampling: Random forest uses Bootstrap sampling to randomly select subsamples from the training set with replacement. The size of each subsample is the same as the training set.

[0219] Multiple decision trees: Build a decision tree for each subsample.

[0220] Tree construction: When each tree is trained based on subsamples, a recursive partitioning method is used, starting from the root node, to select a feature and a partitioning point to split the data into two parts, with the goal of making each part as pure as possible (i.e., a single state dominates). This partitioning is based on reducing the measure of impurity, such as Gini impurity or information gain.

[0221] 4) Voting mechanism

[0222] Prediction of each tree:

[0223] Independent prediction: Each decision tree makes a prediction independently and outputs a classification result (such as normal, fault type A).

[0224] Output probability: Each tree can output a probability value by counting the proportion of each state in its leaf nodes (for example, if 80% of the leaf nodes are fault type A, the probability of outputting this state is 0.8).

[0225] 5) Majority Voting:

[0226] Ensemble voting: The final output is determined by the votes of all trees. The system counts the prediction results of all trees and selects the state with the most votes as the final prediction result. This voting mechanism effectively reduces the deviation of a single tree.

[0227] Improved accuracy: By integrating the results of multiple weak classifiers (i.e., decision trees), random forests can better capture the complex patterns of data than a single decision tree, thus improving the accuracy of identifying different types of equipment faults.

[0228] 6) Model verification and optimization

[0229] Verify model performance: Evaluate the model’s accuracy, recall, F1 score and other performance indicators through the validation set.

[0230] Hyperparameter optimization: Use cross-validation and grid search methods to optimize model parameters (such as the number of trees, tree depth, minimum number of sample splits, etc.) to improve the accuracy and stability of the model.

[0231] 7) Model testing

[0232] Test set evaluation: Input the test set into the trained model to evaluate its performance on unseen data to verify the generalization ability of the model.

[0233] Confusion matrix analysis: Analyze the confusion matrix and check the recognition of various faults to ensure that the model has good detection capabilities for various states.

[0234] 5. Model Identification

[0235] 1) Feature input: During the operation of the equipment, vibration, acoustic and image features are extracted in real time, and these features are combined into a comprehensive feature vector and input into the trained random forest model. Comprehensive feature vector = [vibration feature | acoustic feature | image feature].

[0236] 2) Model prediction

[0237] 21) Decision Tree Prediction:

[0238] 211) Input feature vector: The normalized feature vector is input to each decision tree in the random forest. These trees process the input features independently.

[0239] 212) Starting from the root node:

[0240] A decision tree processes input features starting from the root node. The root node contains decision rules or conditions based on the value of a feature (whether it is above a certain threshold).

[0241] 213) Layer-by-layer transmission:

[0242] The input feature vector is passed to the child nodes one by one according to the tree's splitting condition ("Is the value of feature X greater than threshold T?"). Each node decides which child node to pass the data to based on the feature value and the decision rule.

[0243] 214)Decision Node:

[0244] At each internal node, the decision tree splits based on a certain threshold of the feature. For example, if the value of feature X is greater than T1, it enters the left subtree, otherwise it enters the right subtree. The condition of each node is determined by the optimal splitting criterion (such as information gain, Gini index) during the training process.

[0245] 215) Arriving at a leaf node:

[0246] After several layers of node judgment, the feature vector finally reaches a leaf node. The leaf node represents a final category label, such as "normal", "fault type A", "fault type B", etc.

[0247] 22) Generate prediction results

[0248] 221) Leaf node classification:

[0249] Each leaf node corresponds to a device status category. These categories are determined during the training process based on the labels of the training data. For example, one leaf node may represent "bearing wear fault" and another leaf node may represent "motor overheating fault".

[0250] 222)Decision tree prediction output:

[0251] The decision tree generates a prediction result based on the leaf node to which the feature vector reaches. For example, after the feature vector is judged by the decision tree, it may be classified as "fault type A". This result is the classification judgment of the decision tree on the input data.

[0252] Randomness of the tree:

[0253] Because each decision tree is trained with a different subset of data and features, it may generate different predictions. For example, one tree may be more inclined to judge high values ​​of feature X as "fault type B", while another tree may think that these feature values ​​represent "fault type A".

[0254] Example description:

[0255] When monitoring the device, the following feature vectors were extracted (including sound pressure level, spectrum center, texture features, etc.): sound pressure level: 85dB, spectrum center: 1500Hz, texture features: LBP value is "11001001"

[0256] Example of prediction process:

[0257] Input feature vector: Input the above feature vector into a decision tree.

[0258] Decision through the root node: The root node condition is "sound pressure level>80dB". The input feature meets this condition, so the data is passed to the left subtree.

[0259] After judging through the intermediate nodes: At an intermediate node in the left subtree, the condition is "spectrum center < 1600Hz". If the input feature meets this condition, the data will continue to be passed to the left subtree.

[0260] Arriving at a leaf node: Eventually reaching a leaf node, the leaf node corresponds to "fault type A".

[0261] Generate prediction results: This decision tree classifies the input feature vector as "Fault Type A".

[0262] 3) Voting summary:

[0263] Collect the prediction results of all decision trees: After all decision trees in the random forest complete their predictions, their prediction results are aggregated to form a prediction set. Each decision tree votes for the input data once, which is equivalent to casting a vote for the state category it "thinks" is.

[0264] Majority decision mechanism: Random forest uses a majority decision mechanism, that is, by counting the voting results of all decision trees, the category with the most votes is selected as the final prediction output. If 30 out of 50 trees predict "fault type A" and 20 predict "normal", the final prediction result is "fault type A".

[0265] 4) Interpretation of results

[0266] Status label output: Random forest finally outputs the label of the device status, indicating the current status of the device as considered by the model. These status labels can be specific categories such as "normal", "fault type A", "fault type B", etc.

[0267] Prediction probability value: In addition to outputting the state label, the model also outputs the prediction probability value for each state, which indicates the model's confidence level for each state. For example, if the prediction result is "fault type A" and the probability is 85%, it means that the model has a high confidence in this result.

[0268] Calculate the probability value: The probability value is calculated by counting the frequency of each decision tree predicting the state. For example, if there are 100 trees and 85 of them predict "fault type A", the probability of this state is 85%. This probability reflects the decision stability and consistency of the model as a whole.

[0269] 6. Anomaly Detection

[0270] Detect and mark abnormal states: When the model outputs an abnormal state (such as a fault) and the predicted probability of the state is higher than the set alarm threshold (70%), the system will mark the state as an abnormal state that requires attention.

[0271] Abnormal status type identification: The system will mark the abnormal status as a specific fault type, such as "bearing wear" or "motor overheating", which is of great reference significance for subsequent alarms and maintenance.

[0272] Probability values ​​are used for decision making: Predicted probability values ​​not only help determine the current state of the equipment, but are also used to determine whether immediate action is needed. For example, when the model predicts that the probability of "fault type B" is 95%, the system should immediately trigger an emergency alarm; if the probability is only 55%, it may indicate that the equipment is in a marginal state and can be set to an observation state instead of immediate maintenance.

[0273] Multi-level alarm mechanism: The system can set different levels of alarm responses based on probability values, for example:

[0274] Warning level: predicted probability 50% - 70%, the system only prompts the operator to pay attention.

[0275] Alarm level: Prediction probability 70% - 90%, it is recommended that operators check the equipment as soon as possible.

[0276] Emergency level: When the predicted probability is above 90%, the emergency maintenance process is triggered and the system may be shut down automatically.

[0277] VII. Decision support and further actions:

[0278] Alarm and notification: If the model identifies a fault state and the predicted probability exceeds the alarm threshold, the system will trigger an alarm. The alarm can be sent in a variety of ways, such as sound and light alarms, control room display reminders, and mobile device push messages to ensure that operators are notified of equipment abnormalities in a timely manner.

[0279] Generate maintenance suggestions: Based on the specific fault type identified, the system generates corresponding maintenance suggestions. For example, "fault type A" may prompt to check the lubrication of the equipment, while "fault type B" may suggest to check the tightness of the connector. These suggestions can be directly used to guide the actual operation of maintenance personnel.

[0280] Continuous monitoring and feedback: The system continues to monitor the status of the equipment in real time, recording all prediction results and alarm records to facilitate subsequent status analysis and model optimization. By recording and analyzing historical data, the system can also adjust alarm thresholds and response strategies to better adapt to changes in equipment operation.

[0281] Through the above detailed steps, the identification and decision-making process realizes the full process control from data input, status identification to alarm and maintenance, providing a solid guarantee for the healthy operation of the equipment. This closed-loop identification and decision-making mechanism ensures that the equipment can be discovered in time when problems occur, and appropriate maintenance measures can be taken to effectively reduce the losses caused by failures.

[0282] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring the state of intelligent industrial equipment, characterized in that: The following steps are involved: Data collection: Install sensor modules at key locations of industrial equipment to continuously monitor the operating status of the equipment, collect vibration, acoustics, and surface image data in real time, and transmit them to the computer system module; Data preprocessing: The computer system module filters and de-noises the collected original vibration signals and acoustic signals, eliminates background noise and environmental interference, improves signal clarity and usability, performs median filtering and denoising on the collected image data, enhances the image through histogram equalization and contrast stretching, and improves the accuracy of image feature recognition; Feature extraction: The computer system module extracts features from the preprocessed data, including vibration features, acoustic features, and image features; Model training: The computer system module combines vibration, acoustic and image features into a comprehensive feature vector, which is input into the random forest model. The Bootstrap sampling method is used to randomly extract subsamples from the training set to construct a decision tree. Each decision tree independently makes predictions and outputs classification results and probability values. The final output is then determined by the voting aggregation of all trees. The model performance indicators are evaluated through the validation set. The model parameters are optimized using cross-validation and grid search methods. Finally, the test set is input into the trained model to evaluate its performance on unseen data and analyze the confusion matrix. Model identification: The computer system module inputs the vibration, acoustic and image features extracted in real time into the trained random forest model for model prediction; the decision tree processes the input feature vector, starting from the root node and passing it layer by layer to the leaf node to generate the prediction result; the prediction results of all decision trees are collected, and the majority decision mechanism is used to determine the final prediction output and output the equipment state label and the prediction probability value of each state; Anomaly detection: The computer system module performs anomaly detection. When the model outputs an abnormal state and the predicted probability of the state is higher than the set alarm threshold, the system will mark the state as an abnormal state that requires attention and identify the specific fault type. The probability value is used for decision-making: different levels of alarm response are set according to the predicted probability value; Decision support and further actions: If the model identifies a fault state and the predicted probability exceeds the alarm threshold, the system will trigger an alarm and notify the operator. Based on the specific fault type identified, the system will generate corresponding maintenance recommendations.

2. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the data preprocessing step, the filtering and noise reduction of the vibration signal include: dynamically adjusting the filtering parameters according to the changes in the background noise through an adaptive algorithm to achieve more accurate noise suppression, performing Fourier transform on the vibration signal, designing an FIR filter, applying an adaptive algorithm to estimate the noise characteristics and adjust the filter weights, extracting effective signals and evaluating the noise suppression effect.

3. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the data preprocessing step, the filtering and noise reduction of the acoustic signal includes: Signal acquisition and preprocessing: obtain the original acoustic signal s(t) from the acoustic sensor; Convert the time domain signal to the frequency domain: Convert the time domain signal s(t) to the frequency domain signal S(f) through fast Fourier transform; Noise spectrum estimation: During the operation of the equipment, the background noise signal is collected and its spectrum is obtained through FFT conversion. A representative noise spectrum N(f) is obtained by averaging multiple samples; Spectral subtraction: subtract the estimated noise spectrum N(f) from the spectrum S(f) of the original signal, and introduce the noise reduction coefficient α to obtain the denoised frequency domain signal Inverse Fourier transform: transform the processed frequency domain signal The denoised time domain signal is converted back to the time domain through the inverse fast Fourier transform (IFFT) 4. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the feature extraction step, the vibration feature extraction includes: extracting characteristic parameters from the preprocessed vibration signal, including time domain features such as amplitude, root mean square value, peak-to-peak value and kurtosis, frequency domain features such as main frequency, frequency component, spectrum energy distribution and harmonic component, and time-frequency domain features such as short-time Fourier transform and wavelet transform.

5. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the feature extraction step, acoustic feature extraction includes: analyzing the preprocessed acoustic signal, extracting frequency domain features such as acoustic frequency distribution, sound pressure level, and spectrum center, as well as time domain features such as envelope analysis, autocorrelation function, and pitch analysis.

6. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the feature extraction step, image feature extraction includes: extracting physical features that can reflect the surface condition of the device from the preprocessed image data, including texture features, edge features, shape features and color features.

7. The method for monitoring the state of intelligent industrial equipment according to claim 6, characterized in that: The step of edge feature extraction comprises: Gaussian filtering: Use a Gaussian filter to perform a convolution operation on the image. The Gaussian filter performs weighted averaging through a Gaussian function to smooth the pixel values ​​in the image. Calculate the gradient strength and direction of the image: Use the Sobel operator to calculate the horizontal gradient and vertical gradient respectively, calculate the gradient strength by the square root of the sum of the squares of the horizontal and vertical gradients, and calculate the gradient direction by the inverse tangent function to indicate the direction of the edge; Non-maximum suppression: Check the gradient strength of each pixel and compare the pixel with the neighboring pixels along the gradient direction according to the gradient direction, and only keep the position where the pixel is a local maximum; Determine strong edges and weak edges: pixels with gradient strength greater than a high threshold are marked as strong edges, pixels with gradient strength between a high threshold and a low threshold are marked as weak edges, and pixels with gradient strength lower than a low threshold are suppressed as non-edges; Generate edge image: Start from the strong edge and check its adjacent weak edges. If a weak edge is connected to any strong edge, keep it as part of the edge, otherwise suppress it. Finally, generate a continuous edge image and output it.

8. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the model training step, the comprehensive feature vector = [vibration feature | acoustic feature | image feature].

9. The method for monitoring the state of intelligent industrial equipment according to claim 1, characterized in that: In the model identification step, the model predicts The following steps are involved: Input feature vector: The standardized feature vector is input to each decision tree in the random forest, and these trees process the input features independently; Layer-by-layer processing: The decision tree starts from the root node and transfers the input features layer by layer according to the predefined decision rules; Node judgment and splitting: At each node, the tree makes a decision based on the feature threshold and passes the data to the corresponding child node; Arrival at a leaf node: The feature vector reaches a leaf node, which represents a specific device status category; Generate prediction results: Each decision tree generates a prediction result based on the leaf node reached by the feature vector.

10. A device based on the intelligent industrial equipment status monitoring method according to any one of claims 1 to 9, characterized in that: include: Sensor module: including a vibration sensor, which is used to detect vibration signals of the device; An acoustic sensor is used to collect sound signals generated when the device is running; an image sensor is used to capture the surface of the device; Computer system module: used to receive and process the data collected by the sensor module; Communication and alarm module: It is used to issue alarms and has communication functions to transmit monitoring data and alarm information to the computer system module.

Images