An intelligent monitoring and fault diagnosis method and system for an industrial equipment production process

Abstract

The application relates to the technical field of intelligent manufacturing and industrial automation, and discloses an intelligent monitoring and fault diagnosis method and system for an industrial equipment production process, the method comprising the following steps: acquiring equipment multi-source signals and historical time sequences; processing the equipment multi-source signals to obtain a filtered signal set; processing the filtered signal set to obtain a frequency drift mode, and decomposing and constructing an early feature group when deviating; clustering the early feature group to obtain an initial feature cluster, performing boundary contraction to obtain an optimized clustering cluster, and matching to obtain a fault category set; calculating a stable index value according to the fault category set and the historical time sequences; and updating a clustering threshold value in a preset clustering model according to the stable index value to obtain an optimized clustering model. The method can realize self-adaptive stable diagnosis under a complex environment.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial automation technology, and in particular to an intelligent monitoring and fault diagnosis method and system for industrial equipment production processes. Background Technology

[0002] The stable and continuous operation of industrial equipment is a key foundation for modern manufacturing to ensure production efficiency and maintain product quality. With the continuous deepening of industrial automation and Internet of Things technologies, fault prediction and health management for equipment under complex operating conditions has become an important area of ​​industry research and application. Its core purpose is to promptly detect potential physical hazards in equipment through high-frequency real-time monitoring, thereby avoiding unplanned downtime and major safety accidents.

[0003] In existing technologies, signal processing techniques based on single-threshold alarms or conventional macroscopic time-frequency domain analysis are typically used to monitor equipment status. These methods primarily rely on external sensors to collect single-dimensional vibration or acoustic signals. When the overall amplitude of the acquired physical signal exceeds a pre-set safety envelope, or when a significant abnormal energy spike is identified in the basic spectrum, the system determines that the equipment has malfunctioned and outputs a qualitative diagnostic result. However, in real, complex industrial production environments, equipment is often exposed for extended periods to strong interference from multi-source environmental noise, non-stationary load conditions, and mechanically coupled oscillations of multiple components. The aforementioned existing technologies rely too heavily on significant local-dimensional anomalies for judgment, lacking physical constraints on the spatial propagation consistency of multi-source signals and the ability to deconstruct deep, multi-scale, subtle features of the signal. When equipment is in the early stages of a fault (such as micro-polarization of the rotor, early micro-pitting of gears, or slight wear of bearings), the resulting nonlinear abnormal fluctuations are extremely weak. These early anomalies are easily overwhelmed by large fluctuations in normal operating conditions and environmental background noise, rendering conventional threshold and macroscopic spectrum analysis completely ineffective.

[0004] Existing technologies cannot accurately distinguish and stably diagnose fault characteristics such as minute polarization at different evolution stages. Summary of the Invention

[0005] This invention provides an intelligent monitoring and fault diagnosis method and system for industrial equipment production processes, in order to solve the problem that existing technologies cannot accurately distinguish and stably diagnose fault characteristics such as minute polarization at different evolution stages.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent monitoring and fault diagnosis method for industrial equipment production processes, comprising: Acquire multi-source signals from the device, and preprocess the multi-source signals to obtain a purification signal set; The purified signal set is filtered by propagation feature screening according to preset propagation constraints to obtain a filtered signal set; The filtered signal set is processed using short-time Fourier transform to obtain a time-frequency distribution matrix. The time-frequency distribution matrix is ​​then differentially calculated with a preset background noise to obtain a correction spectrum column. The correction spectrum column is then subjected to shift analysis to obtain a frequency drift mode. If the frequency drift mode deviates from the preset reference value, then the filtered signal set is used to extract the modulation phase features, and the modulation phase features are used to construct an early feature group. The early feature groups are clustered using a preset clustering model to obtain an initial feature cluster. Boundary shrinkage is performed on the initial feature cluster to obtain an optimized cluster. The optimized cluster is matched with a preset mechanism feature library to obtain a fault category set. Obtain historical time series segments, compare feature drift amounts with the fault category set and the historical time series segments, and obtain stability index values; The clustering threshold in the clustering model is updated based on the stability index value to obtain an optimized clustering model.

[0007] Secondly, the present invention provides an intelligent monitoring and fault diagnosis system for industrial equipment production processes, comprising: The preprocessing module is used to acquire multi-source signals from the device and preprocess the multi-source signals to obtain a purified signal set. The signal filtering module is used to filter the purified signal set according to the propagation characteristics based on preset propagation constraints to obtain a filtered signal set. The mode determination module is used to process the filtered signal set using short-time Fourier transform to obtain a time-frequency distribution matrix, perform differential calculation between the time-frequency distribution matrix and a preset background noise to obtain a correction spectrum column, and perform shift analysis on the correction spectrum column to obtain a frequency drift mode. The feature construction module is used to extract modulation phase features from the filtered signal set if the frequency drift mode deviates from a preset reference value, and to construct an early feature group from the modulation phase features. The clustering matching module is used to cluster the early feature groups using a preset clustering model to obtain an initial feature cluster, perform boundary shrinking on the initial feature cluster to obtain an optimized cluster, and match the optimized cluster with a preset mechanism feature library to obtain a fault category set. The indicator calculation module is used to obtain historical time series segments, compare the feature drift amount with the fault category set and the historical time series segments, and obtain stable indicator values. The model optimization module is used to update the clustering threshold in the clustering model based on the stability index value, so as to obtain an optimized clustering model.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains multi-source signals from equipment, performs preprocessing and extracts abnormal fluctuation segments, calculates propagation characteristic values, verifies physical space constraints to determine the filtered signal set, and uses short-time Fourier transform and background noise difference to calculate the corrected spectrum, and then performs wander analysis to determine the frequency drift mode. By combining the propagation characteristics of multi-source signals in the time domain and physical space, random interference and workshop background noise that do not conform to the physical propagation law are effectively filtered out, and nonlinear frequency wander trajectories are tracked in a high signal-to-noise ratio characteristic space. This fundamentally overcomes the problem of strong background noise in complex industrial environments masking weak early fault signals of equipment, and achieves high-sensitivity capture and accurate identification of small, hidden anomalies when equipment is in the bud stage of a fault.

[0009] (2) This invention constructs an early feature group by extracting the modulation phase features of the sub-band spectrum set using wavelet packet decomposition algorithm when the frequency drift mode deviates from the reference value. Then, it uses the DBSCAN algorithm to perform density clustering and perform boundary shrinkage operation to obtain optimized clusters. Finally, it matches these clusters with a preset mechanism feature library to determine the fault category set. By using multi-scale decomposition to deeply deconstruct the subtle modulation and phase lag features of the signal, and combining density-based spatial clustering and boundary denoising shrinkage mechanism, it accurately removes the adhesion and fuzzy boundaries in the feature space, realizing the accurate mapping of high-dimensional abstract micro-fluctuation features to specific physical fault types. It effectively solves the problem of misjudgment caused by the superposition of multiple subtle features during fault evolution, and significantly improves the classification accuracy and diagnostic robustness of multiple concurrent faults under complex and variable working conditions.

[0010] (3) This invention calculates the deviation correction value by comparing the extracted dominant spectral features with the dynamic energy series constructed based on historical time series, and then calculates the stability index value by weighting. Finally, the PID control algorithm is used to calculate the parameter compensation amount of the stability index value to dynamically update the clustering threshold in the preset clustering model. An adaptive closed loop based on residual feedback is established between diagnostic execution and model maintenance. By utilizing the proportional, integral and derivative adjustment capabilities of the PID control algorithm, the current diagnostic deviation is transformed into real-time feedforward compensation for the future discrimination boundary, so that the clustering model can adaptively adjust the decision boundary as the equipment ages or the operating conditions change. This endows the intelligent diagnostic system with strong autonomous evolution and dynamic adaptation capabilities, completely breaking the bottleneck of response lag and generalization failure that is prone to occur in the long-term operation of traditional fixed threshold models, and making the equipment monitoring exhibit adaptive characteristics of becoming more accurate and stable with use. Attached Figure Description

[0011] Figure 1This is a schematic diagram of the intelligent monitoring and fault diagnosis method for the industrial equipment production process provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent monitoring and fault diagnosis system for the production process of industrial equipment provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides an intelligent monitoring and fault diagnosis method for industrial equipment production processes, comprising the following steps: S11, acquire multi-source signals from the device, and preprocess the multi-source signals from the device to obtain a purification signal set; S12, the purification signal set is filtered by propagation characteristics according to the preset propagation constraints to obtain the filtered signal set; S13, the filtered signal set is processed by short-time Fourier transform to obtain a time-frequency distribution matrix, the time-frequency distribution matrix is ​​differentially calculated with a preset background noise to obtain a correction spectrum column, and the correction spectrum column is subjected to wander analysis to obtain a frequency drift mode. S14, if the frequency drift mode deviates from the preset reference value, then the filtered signal set is subjected to feature extraction to obtain modulation phase features, and the modulation phase features are subjected to sequence construction to obtain an early feature group; S15, use a preset clustering model to cluster the early feature groups to obtain an initial feature cluster, perform boundary shrinkage on the initial feature cluster to obtain an optimized cluster, and match the optimized cluster with a preset mechanism feature library to obtain a fault category set; S16, Obtain historical time series segments, compare feature drift amounts with the fault category set and the historical time series segments to obtain stability index values; S17, Update the clustering threshold in the clustering model according to the stability index value to obtain an optimized clustering model.

[0014] In step S11, the device multi-source signals are acquired, and the device multi-source signals are preprocessed to obtain a purification signal set.

[0015] The process of preprocessing the multi-source signals from the device to obtain a purified signal set includes: The multi-source signals of the device are aligned using a preset sampling clock to obtain a synchronization signal segment; The complete signal sequence is obtained by interpolating the synchronization signal segment using a linear interpolation algorithm; An intermediate signal set is obtained by denoising the complete signal series using an adaptive filtering algorithm; The intermediate signal set is adjusted using a preset gain rule to obtain a purified signal set.

[0016] In one implementation, this embodiment acquires multi-source signals from the industrial equipment, including thermal inertial signals and acoustic / impact signals, through a multi-sensor array deployed at key parts of the equipment. To ensure compatibility between low-frequency thermal inertial signals and extremely high-frequency acoustic / impact signals within the same system, this embodiment aligns the multi-source signals using a preset sampling clock to obtain a synchronization signal segment. It should be noted that the preset sampling clock is set based on the requirements for capturing high-frequency transient impacts. This embodiment extracts the highest rotating fundamental frequency specified on the industrial equipment nameplate and multiplies this highest rotating fundamental frequency by a preset high-frequency oversampling coefficient (e.g., a constant between 200 and 512) to obtain a high-resolution reference sampling frequency for capturing microsecond-level physical delays. This embodiment generates a discrete timestamp sequence based on this high-resolution reference sampling frequency and determines this timestamp sequence as the preset sampling clock. For low-frequency sensors such as temperature sensors with sampling rates much lower than this clock, this embodiment uses zero-order hold interpolation to map their values ​​onto the high-frequency time axis, obtaining the synchronization signal segment aligned within a unified time window.

[0017] It is worth noting that due to the response delay and data transmission jitter of different types of sensors, the synchronization signal segment may have missing data at certain timestamp nodes. This embodiment uses a linear interpolation algorithm to interpolate the synchronization signal segment to obtain a complete signal sequence. The specific calculation process is as follows: This embodiment traverses the timestamp sequence in the synchronization signal segment, locates the timestamp node with missing data, obtains the first value of the previous valid sampling time and the second value of the next valid sampling time for the missing node; calculates the difference between the second value and the first value, and divides the difference by the time difference between the next valid sampling time and the previous valid sampling time to obtain the slope of the value change in that interval; multiplies the slope by the time difference between the missing node and the previous valid sampling time, and adds the first value to calculate the simulated value of the missing node; fills all the calculated simulated values ​​into the corresponding missing nodes of the synchronization signal segment to obtain the complete signal sequence that is continuous and uninterrupted in the time domain.

[0018] In one implementation, this embodiment utilizes an adaptive filtering algorithm to denoise the complete signal sequence to obtain an intermediate signal set. Specifically, the denoising logic is as follows: This embodiment acquires a background noise sequence pre-collected while the industrial equipment is stopped, and uses this background noise sequence as a reference input; it then dynamically adjusts the filter weight vector using a normalized minimum mean square error (MMS) algorithm. This embodiment calculates the error value between the current filter output and the corresponding sampling point in the complete signal sequence, and updates the current weight vector using the product of this error value and the reference input, along with a preset step size factor. It should be noted that the preset step size factor is determined through offline simulation testing. In this embodiment, after normalizing the original signal amplitude to the -1 to 1 range offline, Gaussian white noise is injected into the standard sinusoidal test signal. By traversing the step size values ​​within the 0.01 to 0.1 range, the MMS error of the output signal under different step sizes is calculated, and the step size value that minimizes the MMS error is selected as the preset step size factor. In this embodiment, the converged weight vector is used to perform convolution and subtraction operations on the complete signal sequence to remove the broadband steady-state noise component and obtain the intermediate signal set.

[0019] For example, in this embodiment, the intermediate signal set is adjusted using a preset gain rule to obtain a purified signal set. It should be noted that this embodiment uses an adaptive adjustment strategy based on envelope kurtosis to construct the preset gain rule. Specifically, this embodiment calculates the envelope kurtosis index of the intermediate signal set within a set frequency band. This envelope kurtosis index physically corresponds to the prominence of periodic impulse components in the signal. This embodiment sets a base gain value of 1. If the calculated envelope kurtosis index is greater than a preset kurtosis threshold, the base gain value is multiplied and amplified according to the ratio of the envelope kurtosis index to the kurtosis threshold to obtain a target gain coefficient. If the envelope kurtosis index is not greater than the kurtosis threshold, the target gain coefficient is maintained at the base gain value. This embodiment uses the above segmented mapping logic as the preset gain rule. The preset kurtosis threshold is determined by statistically analyzing the maximum kurtosis value recorded during historical fault-free operation cycles. In this embodiment, the value of each sampling point in the intermediate signal set is multiplied by the corresponding target gain coefficient, and a secondary amplification operation is performed to obtain the purified signal set.

[0020] In step S12, the purified signal set is filtered by propagation characteristics according to preset propagation constraints to obtain a filtered signal set, including: The purified signal set is time-domain mapped using a preset acquisition interval to obtain a discrete multidimensional column; The abnormal fluctuation segment is obtained by extracting the data segment whose amplitude exceeds the preset amplitude threshold in the discrete multidimensional column using the time-domain segmentation algorithm; The time difference and amplitude attenuation value are calculated by analyzing the arrival time and amplitude data of the abnormal fluctuation segment in the preset sensor array. The propagation characteristic value is obtained by combining the time difference value and the amplitude attenuation value. If the propagation characteristic value satisfies the preset propagation constraint, then the abnormal fluctuation segment is determined as the filter signal set.

[0021] In one implementation, this embodiment maps the purification signal set to a discrete multidimensional column using a preset acquisition interval in the time domain. It should be noted that the preset acquisition interval is determined based on the physical quantity with the lowest rate of change in the purification signal set. This embodiment extracts the temperature signal acquisition period from the purification signal set and determines the temperature signal acquisition period as the preset acquisition interval. Using the preset acquisition interval as a time window reference, this embodiment maps the high-frequency vibration data in the purification signal set to the corresponding low-frequency temperature data time window, constructing the discrete multidimensional column reflecting the multidimensional physical state of the equipment.

[0022] It is worth noting that this embodiment utilizes a time-domain segmentation algorithm to extract abnormal fluctuation segments from data segments in the discrete multidimensional column whose amplitudes exceed a preset amplitude threshold. Regarding the determination of the preset amplitude threshold, this embodiment acquires historical amplitude data of the device under calibrated normal operating conditions, calculates the arithmetic mean and standard deviation of the historical amplitude data, and adds three times the standard deviation to the arithmetic mean to obtain the preset amplitude threshold. This embodiment uses a sliding time window to traverse along the time axis of the discrete multidimensional column, truncating and separating waveform sequences that continuously exceed the preset amplitude threshold to obtain the abnormal fluctuation segments.

[0023] For example, this embodiment calculates the time difference and amplitude attenuation value by analyzing the arrival time and amplitude data of the abnormal fluctuation segment in a preset sensor array. This embodiment extracts the absolute timestamp of the signal rise triggered by the abnormal fluctuation segment at the first sensor node and the absolute timestamp of the signal rise triggered by the abnormal fluctuation segment at the second sensor node in the preset sensor array. The value of the latter is subtracted from the value of the former, and the absolute value is taken to obtain the time difference, which is always a non-negative number. This embodiment extracts the peak amplitude of the abnormal fluctuation segment at the first sensor node and the peak amplitude at the second sensor node. The absolute value of the difference between the peak amplitude at the first sensor node and the peak amplitude at the second sensor node is calculated, and this absolute value is divided by the peak amplitude at the first sensor node to obtain the amplitude attenuation value. This embodiment encapsulates and combines the time difference and the amplitude attenuation value into an array to obtain propagation characteristic values.

[0024] In one implementation, if the propagation characteristic value satisfies a preset propagation constraint, this embodiment determines the abnormal fluctuation segment as the filtered signal set. It should be noted that the preset propagation constraint is constructed based on the device's physical sound velocity model and error limits. This embodiment obtains the theoretical sound velocity of the device's casing material and the physical straight-line distance between the first and second sensing nodes. The theoretical propagation delay is obtained by dividing the physical straight-line distance by the theoretical sound velocity; the theoretical amplitude attenuation is obtained by multiplying the physical straight-line distance by a preset material damping coefficient.

[0025] It should be noted that regarding the acquisition methods of the above-mentioned constraint parameters, the preset material damping coefficient is obtained by consulting the standard mechanical engineering material handbook and extracting the corresponding standard acoustic attenuation rate based on the actual physical material of the equipment casing (such as cast steel or ductile iron). In this embodiment, the hardware datasheet of the data acquisition card is obtained, the maximum clock jitter time of multi-channel synchronous sampling is extracted, and it is determined as the preset synchronization tolerance. If the hardware datasheet does not provide the maximum clock jitter time, the standard deviation of the phase difference between each channel is calculated by synchronously acquiring standard sine signals through multiple channels, and three times the standard deviation is determined as the preset synchronization tolerance. In this embodiment, the factory calibration certificate of the sensor is obtained, its percentage of the absolute error limit across the full range is extracted, and it is determined as the preset measurement tolerance. If the sensor does not provide a calibration certificate, the preset measurement tolerance is determined by on-site comparison. By simultaneously acquiring the same stable signal from the sensor under test and the high-precision reference sensor, the percentage of the absolute value of the amplitude difference between the two to the amplitude of the reference sensor is calculated, and the maximum value of multiple measurement results is taken as the preset measurement tolerance. In this embodiment, the theoretical propagation delay is superimposed with the preset synchronization tolerance to form a first constraint interval, and the theoretical amplitude attenuation is superimposed with the preset measurement tolerance to form a second constraint interval. If the calculated time difference value falls within the first constraint interval and the amplitude attenuation value falls within the second constraint interval, this embodiment determines that the propagation characteristic value satisfies the preset propagation constraints, retains the signal that conforms to the physical space attenuation law, and identifies the abnormal fluctuation segment as the filtered signal set.

[0026] It is worth noting that when obtaining the preset material damping coefficient, if the manual provides a clear attenuation rate value, it is used directly; if an attenuation rate range is provided, the median value is taken; if there is no corresponding material data, it is obtained through on-site impact test, that is, two sensing nodes are arranged on the equipment casing, the amplitude attenuation of the impact signal between the two nodes is measured, and the damping coefficient is deduced.

[0027] In step S13, the filtered signal set is processed using short-time Fourier transform to obtain a time-frequency distribution matrix. The time-frequency distribution matrix is ​​then differentially calculated with a preset background noise to obtain a correction spectrum column. The correction spectrum column is then subjected to drift analysis to obtain a frequency drift mode.

[0028] In one implementation, this embodiment uses a short-time Fourier transform to process the filtered signal set to obtain a time-frequency distribution matrix. This embodiment extracts the discrete-time series from the filtered signal set, uses a Hamming window as a sliding analysis window, sets the window length to the number of sampling points containing at least two complete device rotation cycles, and uses half of this window length as the sliding step size. This embodiment slides the analysis window along the discrete-time series, performs a fast Fourier transform calculation on the truncated signal within the window, and concatenates the calculated one-dimensional spectrum in chronological order to construct a three-dimensional data structure containing time, frequency, and amplitude dimensions. This three-dimensional data structure is then defined as the time-frequency distribution matrix.

[0029] It should be noted that in this embodiment, the time-frequency distribution array and the preset background noise are differentially calculated to obtain the corrected spectrum. Regarding the determination of the preset background noise, in this embodiment, when the equipment is in an unloaded or stopped state, the same sensor nodes are used to collect steady-state environmental signals. The same short-time Fourier transform operation is performed on this steady-state environmental signal, and the arithmetic mean of the energy of each frequency band is extracted to form a basic noise spectrum. This basic noise spectrum is determined as the preset background noise. Specifically, environmental signals are collected under different load conditions such as no-load, half-load, and full-load, as well as at different speeds. For each condition, a short-time Fourier transform is performed, and the arithmetic mean of the energy of each frequency band is calculated to form a two-dimensional noise spectrum of condition-frequency band. During differential calculation, the corresponding background noise spectrum is matched according to the current condition. In this embodiment, the amplitude of the corresponding frequency band in the preset background noise is subtracted from the amplitude at each time frequency node in the time-frequency distribution array to remove the broadband steady-state components present in the equipment operating environment, resulting in a pure time-frequency sequence. This pure time-frequency sequence is determined as the corrected spectrum.

[0030] Among them, the frequency drift mode is obtained by performing a shift analysis on the corrected spectrum, including: The harmonic peaks in the corrected spectrum are extracted to obtain a dynamic trajectory, and the frequency drift characteristics are obtained by tracking and analyzing the dynamic trajectory. The frequency change rate is obtained by calculating the rate of change of the frequency drift feature; If the rate of change of frequency exhibits nonlinear fluctuations, the corresponding state characteristic is determined to be a frequency drift mode.

[0031] In one implementation, this embodiment extracts harmonic peak values ​​from the calibration spectrum to obtain a dynamic trajectory, and tracks and analyzes the dynamic trajectory to obtain frequency drift characteristics. This embodiment obtains the design reference frequency of the equipment spindle and locks the target frequency band corresponding to the integer multiples of this design reference frequency in the calibration spectrum. This embodiment traverses the target frequency band along the time axis, extracts the frequency coordinate point with the largest amplitude at each time segment, and connects the frequency coordinate points under consecutive timestamps in chronological order to generate the dynamic trajectory reflecting the change of the local dominant frequency over time. This embodiment extracts the upper frequency limit, lower frequency limit, and instantaneous center frequency point of the dynamic trajectory within the target frequency band, and combines the upper frequency limit, lower frequency limit, and instantaneous center frequency point to obtain the frequency drift characteristics; the upper and lower frequency limits in the frequency drift characteristics are used to evaluate the amplitude range of the frequency drift, and the instantaneous center frequency point is used to calculate the frequency change rate; these three are used for different diagnostic analysis dimensions in subsequent steps.

[0032] For example, in this embodiment, the frequency change rate is obtained by calculating the rate of change of the frequency drift feature. In this embodiment, the instantaneous center frequency points of adjacent time windows in the frequency drift feature are extracted, the difference between the instantaneous center frequency point value of the later time window and the instantaneous center frequency point value of the previous time window is calculated, and the difference is divided by the physical time interval between adjacent time windows to calculate the frequency change rate representing the degree of instantaneous frequency drift.

[0033] It is worth noting that in this embodiment, if the frequency change rate exhibits nonlinear fluctuations, the corresponding state characteristic is determined to be a frequency drift mode. Regarding the determination criteria for the nonlinear fluctuations, this embodiment extracts multiple consecutive frequency change rate values ​​within a preset monitoring period; the preset monitoring period includes at least twenty consecutive sampling points. This embodiment calculates the difference between adjacent frequency change rate values ​​as the frequency acceleration (dimensions: 1 / 2). If more than half of the calculated absolute values ​​of frequency acceleration are greater than a preset acceleration threshold, and their positive and negative signs alternate more than twice within the preset monitoring period, this embodiment determines that it is in a nonlinear fluctuation state.

[0034] It should be noted that the preset acceleration threshold is determined by obtaining the upper limit of the normal mechanical angular acceleration allowed by the equipment drive system under full load conditions (in units of 1 / 2 Ω·cm). ), divide it by After converting to the fundamental frequency change rate, multiply it by the harmonic multiple (i.e., the harmonic order) of the currently tracked harmonic peak to calculate the equivalent preset acceleration threshold.

[0035] In step S14, feature extraction is performed on the filtered signal set to obtain modulation phase features, and the modulation phase features are sequenced to obtain an early feature group, including: The filtered signal set is decomposed using a wavelet packet decomposition algorithm to obtain a sub-band spectrum set; Frequency abrupt changes are obtained by extracting abrupt change points from the sub-band spectrum set. The amplitude modulation rate and phase offset are calculated from the amplitude envelope change value and phase lag value at the frequency abrupt change point. The amplitude modulation rate and the phase offset are combined to obtain the modulation phase characteristics; The modulation phase features are serialized and mapped to obtain an early feature set.

[0036] In one implementation, this embodiment extracts the feature values ​​contained in the frequency drift pattern and calculates the absolute value of the difference between the feature value and a preset benchmark value as the deviation value. It should be noted that the preset benchmark value is a standard state scalar parameter obtained by performing a statistical arithmetic mean operation on multiple batches of feature data of the equipment under historical fault-free operating conditions. This embodiment compares the calculated deviation value with a preset error limit. The preset error limit is mathematically set based on the maximum confidence interval (three times the standard deviation) of normal operating condition fluctuations in historical data; specifically, feature data of the equipment are collected over at least ten consecutive fault-free production cycles, and their arithmetic mean and standard deviation are calculated. The product of three times the standard deviation and the mean is used as the error limit. If the deviation value is greater than the preset error limit, this embodiment determines that the frequency drift pattern deviates from the preset benchmark value and uses a wavelet packet decomposition algorithm to decompose the filtered signal set to obtain a sub-band spectrum set. The specific decomposition operation is as follows: In this embodiment, the Daubechies wavelet is selected as the mother wavelet basis function, and the filtered signal set is subjected to discrete wavelet packet transform calculation according to the preset number of three decomposition layers to obtain multiple orthogonal and independent discrete sequences from low frequency to high frequency. The multiple orthogonal and independent discrete sequences are determined as the sub-band spectrum set.

[0037] In one implementation, this embodiment extracts frequency abrupt change points from the sub-band spectrum set. Specifically, this embodiment calculates the local energy variance of each sequence in the sub-band spectrum set within a sliding time window, and extracts and determines the time coordinates of the local maxima where the local energy variance exhibits a discontinuous jump and its first derivative changes from positive to negative as the frequency abrupt change point.

[0038] For example, this embodiment calculates the amplitude modulation rate and phase offset by the amplitude envelope change value and phase lag value at the frequency abrupt change point. Specifically, the calculation logic is as follows: This embodiment extracts the instantaneous peak amplitude of the envelope at the frequency abrupt change point and divides this instantaneous peak amplitude by the average envelope amplitude within the time window of the sub-band spectrum set; the calculated ratio is the amplitude modulation rate. Simultaneously, this embodiment uses Hilbert transform to extract the instantaneous phase angle at the frequency abrupt change point and calculates the algebraic difference between this instantaneous phase angle and the reference carrier phase angle; this algebraic difference is determined as the phase offset. It should be noted that the reference carrier phase angle is not set empirically. This embodiment obtains the absolute rotational speed pulse signal collected by a key phase sensor deployed on the device spindle, performs zero-crossing detection on the absolute rotational speed pulse signal to extract a zero-phase reference frame, and determines this zero-phase reference frame as the reference carrier phase angle. This embodiment combines the calculated amplitude modulation rate and phase offset into an array to obtain the two-dimensional modulation phase feature.

[0039] In one implementation, this embodiment performs serialization mapping on the modulation phase features to obtain an early feature set. Since frequency abrupt changes are random discrete events on the time axis, this embodiment abandons fixed-length matrix mapping and instead constructs an initially empty multidimensional discrete feature point set. This embodiment uses the extracted amplitude modulation rate and phase offset as independent two-dimensional spatial coordinate points, and adds the absolute timestamp of the abrupt change as a third-dimensional label, adding them one by one to the multidimensional discrete feature point set. After completing the traversal of the current analysis cycle, the resulting spatial point cloud set is determined as the early feature set capable of structurally characterizing the dynamic evolution of device anomalies.

[0040] It is worth noting that if the frequency drift mode does not deviate from the preset benchmark value, the device is determined to be in a stable operating state. The subsequent early feature group construction, clustering matching and model update operations are not performed, and the normal operating status identifier is directly output.

[0041] In step S15, the early feature groups are clustered using a preset clustering model to obtain an initial feature cluster. Boundary shrinkage is performed on the initial feature cluster to obtain an optimized cluster. The optimized cluster is then matched with a preset mechanism feature library to obtain a fault category set.

[0042] The process includes clustering the early feature groups using a preset clustering model to obtain initial feature clusters, and then performing boundary shrinking on the initial feature clusters to obtain optimized clusters, including: The spatial density value is calculated for each data point in the early feature group using a preset clustering model. Data points with spatial density values ​​greater than a preset clustering threshold are identified as target data points. The target data points are then aggregated to obtain an initial feature cluster. The intersection ratio value is obtained by comparing the initial feature cluster with the preset historical noise model; If the intersection ratio is greater than a preset tolerance threshold, then the edge data points of the initial feature cluster are shrunk towards the preset geometric center using a preset shrinkage ratio to obtain an optimized cluster.

[0043] In one implementation, the clustering model preset in this embodiment specifically uses the DBSCAN algorithm to calculate the spatial density value of each data point in the early feature group. Specifically, this embodiment constructs a spatial hypersphere with any data point in the early feature group as the center and a preset neighborhood radius as the radius. The absolute number of adjacent data points falling within the range of this spatial hypersphere is counted and calculated, and this absolute number is determined as the spatial density value of the corresponding data point.

[0044] It should be noted that the preset neighborhood radius is determined by plotting a K-distance map on the sample dataset and extracting the coordinates of the inflection point where the curve curvature is maximum; the preset clustering threshold is obtained by extracting the feature space dimension value of the early feature group and multiplying this feature space dimension value by a constant 2; specifically, the clustering threshold is set to twice the feature space dimension, and this multiple is determined by evaluating the clustering effect by iterating through 50 sets of simulation data between integers 2 and 6, and selecting the multiple corresponding to the largest silhouette coefficient as the constant 2; the neighborhood radius is determined by the K-distance map inflection point method. In this embodiment, data points with spatial density values ​​greater than the preset clustering threshold are identified as target data points, and all density-connected target data points are aggregated and connected to obtain the initial feature cluster.

[0045] It is worth noting that in this embodiment, the initial feature cluster is compared with a preset historical noise model to obtain the intersection ratio. Regarding the preset historical noise model, this embodiment acquires random discrete noise data collected by the equipment under calibration shutdown or no-load conditions. Kernel density estimation is performed on this random discrete noise data to generate a spatial probability distribution function, which is then used as the preset historical noise model. The bandwidth of the kernel density estimation is determined using the Silverman rule, with the bandwidth value being the sample standard deviation multiplied by the sample size to the power of -1 / 5, and then multiplied by 0.9. This embodiment calculates the overlapping spatial volume between the hypervolume bounding box occupied by the initial feature cluster and the high probability density interval of the preset historical noise model, and divides this overlapping spatial volume by the total volume of the initial feature cluster; the quotient obtained is the intersection ratio.

[0046] For example, if the intersection ratio is greater than a preset tolerance threshold, this embodiment uses a preset shrinkage ratio to shrink the distance of the edge data points of the initial feature cluster towards a preset geometric center. It should be noted that the preset tolerance threshold is set based on the upper limit of the false positive rate in Receiver Operational Characteristic (ROC) analysis. Specifically, at least one hundred known fault samples and one hundred known normal samples are obtained, and an ROC curve is plotted on the intersection ratio of the initial feature cluster and the historical noise model. The intersection ratio corresponding to the maximum true positive rate when the false positive rate does not exceed five percent is selected as the preset tolerance threshold. The preset shrinkage ratio is calculated by normalizing the product of the average distance from all edge data points in the initial feature cluster to the geometric center and the average local density gradient descent rate at the edge data points. The normalization is achieved by sorting the products of the average density gradient descent rate and the average distance of all edge data points, taking the maximum value as the denominator, and dividing the product of all edge data points by this maximum value to obtain the normalized shrinkage ratio. In this embodiment, edge data points located on the outermost convex hull of the initial feature cluster are identified, the geometric center coordinates of the cluster are obtained, the distance vector from the edge data points to the geometric center is calculated, and the magnitude of the distance vector is multiplied by a coefficient equal to 1 minus the preset shrinkage ratio to obtain the updated spatial coordinate position. This completes the edge shrinkage and yields the optimized cluster with a compact spatial shape. If the intersection ratio is not greater than the preset tolerance threshold, the initial feature cluster is directly used as the optimized cluster without performing the boundary shrinkage operation.

[0047] The optimized clusters are matched with a preset mechanistic feature library to obtain a fault category set, including: The centroid features are obtained by extracting the geometric center coordinates of the optimized clusters; The variance feature is obtained by calculating the dispersion value of each data point in the optimized cluster relative to the centroid feature; The centroid features and variance features are input into the preset mechanism feature library for matching and retrieval; If the centroid feature and the variance feature fall within a preset matching interval, the corresponding fault category is determined, and the fault categories are combined to obtain a fault category set.

[0048] In one implementation, this embodiment extracts the centroid features from the geometric center coordinates of the optimized cluster. Specifically, this extraction process involves calculating the numerical arithmetic mean of all data points within the optimized cluster on each independent feature dimension, combining the arithmetic mean values ​​of each dimension in dimensional order to form a multidimensional coordinate vector, and determining this multidimensional coordinate vector as the centroid feature.

[0049] For example, this embodiment calculates the variance feature by measuring the dispersion of each data point in the optimized cluster relative to the centroid feature. Specifically, the calculation logic is as follows: this embodiment calculates the square of the Euclidean distance between each data point in the optimized cluster and the centroid feature; sums the squares of the Euclidean distances of all data points; and divides this sum by the total number of data points in the cluster to obtain the variance feature that quantifies the dispersion of the distribution.

[0050] In one implementation, this embodiment inputs the centroid features and variance features into the preset mechanistic feature library for matching and retrieval. It should be noted that the preset mechanistic feature library is a pre-constructed relational database. Its construction method is as follows: This embodiment acquires a test device after it has been physically injected with a specific fault (such as rotor thermal bending or gear micro-pitting) on ​​a laboratory bench. During a full life-cycle aging test, it collects multi-sensor signal sequences from the test device, performs clustering calibration on the extracted cluster centers and distribution variance, records the standard centroid coordinates and allowable variance intervals corresponding to various fault types, and constructs the preset mechanistic feature library from the above data. When constructing the preset mechanistic feature library, each fault type must have at least... Twenty samples from different aging stages were collected, ensuring that the samples covered the entire process from fault initiation to failure. For faults that could not be simulated on the bench, real fault data confirmed in historical maintenance records were used to supplement the data. After the system was built, 10 test samples that were not included in the system were used for verification. If the matching accuracy was less than 95%, the sample size was expanded until the standard was met. The allowable variance interval was calculated by taking the historical samples of each fault category and calculating the variance value of each sample cluster. The fifth percentile and the ninety-fifth percentile of the variance value were used as the lower limit and upper limit of the allowable variance interval, respectively.

[0051] This embodiment performs spatial distance comparison operations in the preset mechanism feature library. The Mahalanobis distance between the currently input centroid feature and the standard centroid coordinates of a specific fault in the library is calculated, along with the absolute difference between the current variance feature and the standard variance scalar. If the calculated Mahalanobis distance is less than a preset spatial bounding radius (e.g., set to 3.0), and the absolute difference is less than a preset variance tolerance percentage, the feature is determined to fall within a preset matching interval for that fault category, thus identifying the corresponding physical fault category. The preset spatial bounding radius is calculated by obtaining the standard centroid coordinates of various faults in the mechanism feature library, calculating the minimum Euclidean distance between the standard centroids of various faults, and using half of this minimum distance as the spatial bounding radius. The preset variance tolerance percentage is calculated by taking historical samples of each fault type in the mechanism feature library, calculating the coefficient of variation of their variance values, and multiplying the arithmetic mean of all category coefficients of variation by 20% to obtain the variance tolerance percentage.

[0052] In step S16, a feature drift value is obtained by comparing the fault category set and the historical time series, including: Frequency band locking is performed on the fault category set to obtain the target frequency band interval, and peak extraction is performed on the filtered signal set within the target frequency band interval to obtain the dominant spectral features; The historical time series is used to extract data within the target frequency band to obtain a dynamic energy series; The dominant spectral features are numerically analyzed to obtain the feature amplitude, and the dynamic energy column is boundary extracted to obtain the upper boundary value; The feature drift is obtained by calculating the difference between the current feature amplitude and the upper boundary value. The deviation correction value is obtained by weighting the feature drift using preset correction weight values; The stability index value is obtained by using the deviation correction value to perform a weighted product calculation on the preset initial confidence level.

[0053] In one implementation, this embodiment performs frequency band locking on the fault category set to obtain a target frequency band interval. It should be noted that this embodiment extracts specific fault tags from the fault category set, performs a search and matching in a pre-constructed fault frequency band mapping table, obtains the lower frequency limit value and upper frequency limit value bound to the specific fault tag, and determines the interval formed by the lower frequency limit value and the upper frequency limit value as the target frequency band interval. It is worth noting that the pre-constructed fault frequency band mapping table is a relational database constructed by performing spectral analysis on real fault vibration signals from equipment maintenance records over the years, extracting the dominant frequency band range with an energy proportion exceeding 80% of the total energy, and associating it with the corresponding fault tags. This embodiment extracts peak values ​​from the filtered signal set within the target frequency band interval to obtain dominant spectral features. Specifically, this embodiment traverses all discrete amplitude data points of the filtered signal set within the target frequency band interval, locates the discrete data point with the largest amplitude using local extremum search logic, and combines the frequency value and amplitude value corresponding to this discrete data point to determine the dominant spectral feature.

[0054] In one implementation, this embodiment acquires historical time series segments and extracts data from these segments within the target frequency band to obtain a dynamic energy sequence. It should be noted that this embodiment acquires the load parameters of the current processing condition of the equipment, extracts fault-free operating data with the same load parameters from the database of the manufacturing execution system within a consecutive preset period, and determines this as the historical time series segment. The preset period is determined based on the standard time taken for the equipment to complete a full production batch. This embodiment uses a bandpass filter to filter the historical time series segment, extracts the time-series signal within the target frequency band, and performs a square calculation on each instantaneous amplitude of the time-series signal to generate an energy data set arranged in chronological order. This energy data set is then determined as the dynamic energy sequence.

[0055] For example, this embodiment performs numerical analysis on the dominant spectral features to obtain the feature amplitude, and performs boundary extraction on the dynamic energy column to obtain the upper boundary value. This embodiment extracts the scalar value corresponding to the amplitude dimension in the dominant spectral features and determines it as the feature amplitude. For the extraction of the upper boundary value, this embodiment calculates the arithmetic mean and standard deviation of all energy data in the dynamic energy column, multiplies the standard deviation by a constant 3, and sums the resulting product with the arithmetic mean. The sum is determined as the upper boundary value. The constant 3 is obtained by performing a normality test on the dynamic energy column data collected during historical fault-free operation. If the data distribution is approximately normal, three times the standard deviation is used to construct the upper boundary; if the data distribution is skewed, the 95th percentile is used as the upper boundary value. This embodiment calculates the feature drift by difference between the feature amplitude at the current moment and the upper boundary value. The specific calculation logic is as follows: in this embodiment, the upper boundary value is subtracted from the feature amplitude to obtain the algebraic difference value representing the deviation of the current feature from the historical normal boundary, and the algebraic difference value is determined as the feature drift amount.

[0056] It is worth noting that this embodiment uses a preset correction weight value to calculate the deviation correction value by weighting the feature drift amount. To prevent extreme feature drift amounts from causing a dimensional explosion in the correction value, this embodiment does not use a simple linear multiplication. The preset correction weight value is a scaling sensitivity constant obtained before the system leaves the factory by performing Logistic regression fitting on a large number of fault drift samples; by obtaining the maximum absolute value of the feature drift amount in historical fault samples, the correction weight value is set to the reciprocal of this maximum value, so that the absolute value of the product of the feature drift amount and the correction weight value does not exceed 6, ensuring that the Sigmoid function output is in the effective range of 0.002 to 0.998. The embodiment uses the Logistic Sigmoid nonlinear activation function to calculate the deviation correction value, multiplying the feature drift amount by the preset correction weight value and taking the opposite number, adding the constant 1 to the opposite power of the natural constant e, and finally taking its reciprocal. This calculation process strictly compresses the infinite drift amount in physical dimensions to the (0,1) interval, which is then determined as the deviation correction value. In this embodiment, the deviation correction value is used to perform a weighted product calculation on the preset initial confidence level to obtain a stable index value.

[0057] It should be noted that the preset initial confidence level is the basic confidence score statically configured by the system. Its specific value (e.g., 0.85) is obtained by statistically analyzing the clustering accuracy of normal samples in the historical test set. The clustering accuracy is obtained by comparing the clustering results with the pre-labeled real fault labels and calculating the proportion of samples correctly clustered into the corresponding fault clusters to the total number of samples. Since the bias correction value is mathematically constrained to the range (0,1), this embodiment multiplies it by the preset initial confidence level to obtain the stable index value.

[0058] In step S17, the clustering threshold in the clustering model is updated according to the stability index value to obtain an optimized clustering model, including: The target deviation value is obtained by calculating the deviation between the stability index value and the preset ideal convergence value; The parameter compensation amount is obtained by performing proportional-integral-derivative calculations on the target deviation value using a PID control algorithm. The updated threshold is obtained by superimposing the parameter compensation amount onto the clustering threshold in the clustering model; The updated threshold is used to replace the clustering threshold in the clustering model to obtain an optimized clustering model.

[0059] In one implementation, this embodiment calculates the target deviation value by measuring the deviation between the stability index value and the preset ideal convergence value. It should be noted that the preset ideal convergence value represents an ideal baseline state where the diagnostic results are absolutely stable and without systematic decay. To prevent the PID controller from getting stuck in the mathematical divergence deadlock of integral saturation, this embodiment numerically forces the preset ideal convergence value to be equal to the preset initial confidence level configured in the preceding steps, for example, both are set to 0.85. This embodiment subtracts the current stability index value from the preset ideal convergence value to obtain the target deviation value used to characterize the current system's adaptive tracking error. Under this logic, when a real serious fault occurs in the device, the stability index value reaches the upper limit of 0.85, the target deviation is zero, and the model parameters are frozen to lock the fault characteristics; while when the device is in a state of normal aging causing slight drift, the target deviation is not zero, driving the system to perform adaptive parameter compensation; it should be noted that when the integral term value exceeds the preset integral upper limit or falls below the preset integral lower limit, the integral accumulation stops, and the integral term is clamped to the corresponding limit value.

[0060] For example, this embodiment uses a PID control algorithm to perform proportional-integral-derivative (PI) calculations on the target deviation value to obtain the parameter compensation amount. It should be noted that in this embodiment, the proportional coefficient, integral coefficient, and derivative coefficient of the PID controller are obtained in advance. The proportional coefficient, integral coefficient, and derivative coefficient are obtained through tuning using the Ziegler-Nichols critical proportionality rule in offline mode. Specifically, the PI calculation process is as follows: In this embodiment, the target deviation value is multiplied by the proportional coefficient to obtain the proportional term value; the target deviation values ​​of the current period and all historical sampling periods are summed, and the sum is multiplied by the integral coefficient to obtain the integral term value; the current target deviation value is subtracted from the historical target deviation value of the previous sampling period, and the difference is multiplied by the derivative coefficient to obtain the derivative term value; in this embodiment, the proportional term value, the integral term value, and the derivative term value are algebraically added to calculate the parameter compensation amount, which is expressed as a continuous floating-point number.

[0061] It is worth noting that in this embodiment, the parameter compensation amount is superimposed on the clustering threshold in the clustering model to obtain the updated threshold. In this step, since the clustering threshold in the clustering model corresponds to the number of data points required in the DBSCAN algorithm, its data type is discrete positive integers, while the parameter compensation amount is a continuous floating-point number. Specifically, the superposition operation is as follows: In this embodiment, the currently effective integer form of the clustering threshold in the clustering model is extracted, and the clustering threshold and the floating-point form of the parameter compensation amount are algebraically added to obtain a floating-point sum; In this embodiment, the floating-point sum is rounded to the nearest integer, forcibly mapped to a discrete positive integer, and this discrete positive integer is determined as the updated threshold; It should be noted that when the absolute value of the parameter compensation amount is less than 0.5, the parameter compensation amount is accumulated in a preset integral residual register. When the absolute value of the accumulated value in the integral residual register reaches 0.5 or more, a threshold update is performed, and the integral residual register is cleared to zero. In this embodiment, the updated threshold is used to replace the clustering threshold in the preset clustering model to obtain an optimized clustering model. In this embodiment, the updated threshold is directly overwritten into the underlying parameter configuration file of the clustering model. In the next execution cycle, the clustering algorithm is guided to divide the input data based on the new discrete point boundary, and output an optimized clustering model with adaptive characteristics.

[0062] In summary, this invention discloses an intelligent monitoring and fault diagnosis method and system for industrial equipment production processes. This invention acquires multi-source signals from equipment and innovatively introduces spatial propagation constraint verification based on a physical sound velocity model and error limits. This enables precise removal of strong background noise and random common-mode interference in complex workshop environments, ensuring the physical fidelity and high signal-to-noise ratio of abnormal fluctuation signals from the data source. By fusing short-time Fourier background difference and wavelet packet multi-scale decomposition techniques, it deeply deconstructs the weak frequency drift trajectory and modulation phase characteristics that are easily masked in the early stages of faults, achieving high-sensitivity capture of early nonlinear minor instability states. By constructing a joint matching architecture based on DBSCAN density clustering, edge shrinkage denoising, and spatial mapping with a priori mechanism feature library, it achieves precise analysis of high-dimensional abstract fluctuation features into specific physical fault types, effectively overcoming the problem of misjudgment caused by the mutual adhesion of multiple concurrent abnormal features under complex operating conditions. Furthermore, this invention constructs an adaptive feedback closed loop based on the dominant spectrum historical energy deviation and PID control algorithm, transforming the real-time diagnostic stability residual into a dynamic compensation parameter for the underlying clustering threshold, realizing the autonomous evolution of the intelligent diagnostic decision boundary as the equipment ages or the operating condition changes. Ultimately, this solution effectively addresses the pain points of early fault underreporting and generalized failure caused by traditional monitoring technologies that rely on a single fixed threshold or macroscopic spectrum, giving the industrial equipment health management system a dynamic adaptability that becomes more stable and accurate with use.

[0063] Reference Figure 2 The second embodiment of the present invention provides an intelligent monitoring and fault diagnosis system for industrial equipment production processes, comprising: The preprocessing module is used to acquire multi-source signals from the device and preprocess the multi-source signals to obtain a purified signal set. The signal filtering module is used to filter the purified signal set according to the propagation characteristics based on preset propagation constraints to obtain a filtered signal set. The mode determination module is used to process the filtered signal set using short-time Fourier transform to obtain a time-frequency distribution matrix, perform differential calculation between the time-frequency distribution matrix and a preset background noise to obtain a correction spectrum column, and perform shift analysis on the correction spectrum column to obtain a frequency drift mode. The feature construction module is used to extract modulation phase features from the filtered signal set if the frequency drift mode deviates from a preset reference value, and to construct an early feature group from the modulation phase features. The clustering matching module is used to cluster the early feature groups using a preset clustering model to obtain an initial feature cluster, perform boundary shrinking on the initial feature cluster to obtain an optimized cluster, and match the optimized cluster with a preset mechanism feature library to obtain a fault category set. The indicator calculation module is used to obtain historical time series segments, compare the feature drift amount with the fault category set and the historical time series segments, and obtain stable indicator values. The model optimization module is used to update the clustering threshold in the clustering model based on the stability index value, so as to obtain an optimized clustering model.

[0064] It should be noted that the intelligent monitoring and fault diagnosis system for industrial equipment production process provided in this embodiment of the invention is used to execute all process steps of the intelligent monitoring and fault diagnosis method for industrial equipment production process in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0065] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0066] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for intelligent monitoring and fault diagnosis of industrial equipment production processes, characterized in that, include: Acquire multi-source signals from the device, and preprocess the multi-source signals to obtain a purification signal set; The purified signal set is filtered by propagation feature screening according to preset propagation constraints to obtain a filtered signal set; The filtered signal set is processed using short-time Fourier transform to obtain a time-frequency distribution matrix. The time-frequency distribution matrix is ​​then differentially calculated with a preset background noise to obtain a correction spectrum column. The correction spectrum column is then subjected to shift analysis to obtain a frequency drift mode. If the frequency drift mode deviates from the preset reference value, then the filtered signal set is used to extract the modulation phase features, and the modulation phase features are used to construct an early feature group. The early feature groups are clustered using a preset clustering model to obtain an initial feature cluster. Boundary shrinkage is performed on the initial feature cluster to obtain an optimized cluster. The optimized cluster is matched with a preset mechanism feature library to obtain a fault category set. The process involves acquiring historical time series segments, comparing feature drift amounts with the fault category set, and obtaining a stability index value. This includes: frequency band locking of the fault category set to obtain a target frequency band interval; peak extraction of the filtered signal set within the target frequency band interval to obtain dominant spectral features; data extraction of the historical time series within the target frequency band interval to obtain a dynamic energy column; numerical analysis of the dominant spectral features to obtain feature amplitudes; boundary extraction of the dynamic energy column to obtain upper boundary values; difference calculation of the current feature amplitude and the upper boundary value to obtain feature drift amounts; weighted calculation of the feature drift amounts using preset correction weight values ​​to obtain deviation correction values; and weighted product calculation of the deviation correction values ​​with preset initial confidence levels to obtain a stability index value. The clustering threshold in the clustering model is updated based on the stability index value to obtain an optimized clustering model.

2. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The preprocessing of the multi-source signals from the device to obtain the purified signal set includes: The multi-source signals of the device are aligned using a preset sampling clock to obtain a synchronization signal segment; The complete signal sequence is obtained by interpolating the synchronization signal segment using a linear interpolation algorithm; An intermediate signal set is obtained by denoising the complete signal series using an adaptive filtering algorithm; The intermediate signal set is adjusted using a preset gain rule to obtain a purified signal set.

3. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The step of filtering the purified signal set according to preset propagation constraints to obtain a filtered signal set includes: The purified signal set is time-domain mapped using a preset acquisition interval to obtain a discrete multidimensional column; The abnormal fluctuation segment is obtained by extracting the data segment whose amplitude exceeds the preset amplitude threshold in the discrete multidimensional column using the time-domain segmentation algorithm; The time difference and amplitude attenuation value are calculated by analyzing the arrival time and amplitude data of the abnormal fluctuation segment in the preset sensor array. The propagation characteristic value is obtained by combining the time difference value and the amplitude attenuation value. If the propagation characteristic value satisfies the preset propagation constraint, then the abnormal fluctuation segment is determined as the filter signal set.

4. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The step of performing a shift analysis on the corrected spectrum to obtain the frequency drift mode includes: The harmonic peaks in the corrected spectrum are extracted to obtain a dynamic trajectory, and the frequency drift characteristics are obtained by tracking and analyzing the dynamic trajectory. The frequency change rate is obtained by calculating the rate of change of the frequency drift feature; If the rate of change of frequency exhibits nonlinear fluctuations, the corresponding state characteristic is determined to be a frequency drift mode.

5. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The process of extracting modulation phase features from the filtered signal set and constructing an early feature set from the modulation phase features includes: The filtered signal set is decomposed using a wavelet packet decomposition algorithm to obtain a sub-band spectrum set; Frequency abrupt changes are obtained by extracting abrupt change points from the sub-band spectrum set. The amplitude modulation rate and phase offset are calculated from the amplitude envelope change value and phase lag value at the frequency abrupt change point. The amplitude modulation rate and the phase offset are combined to obtain the modulation phase characteristics; The modulation phase features are serialized and mapped to obtain an early feature set.

6. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The process of clustering the early feature groups using a preset clustering model to obtain initial feature clusters, and then performing boundary shrinking on the initial feature clusters to obtain optimized clusters, includes: The spatial density value is calculated for each data point in the early feature group using a preset clustering model. Data points with spatial density values ​​greater than a preset clustering threshold are identified as target data points. The target data points are then aggregated to obtain an initial feature cluster. The intersection ratio value is obtained by comparing the initial feature cluster with the preset historical noise model; If the intersection ratio is greater than a preset tolerance threshold, then the edge data points of the initial feature cluster are shrunk towards the preset geometric center using a preset shrinkage ratio to obtain an optimized cluster.

7. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The step of matching the optimized clusters with a preset mechanistic feature library to obtain a fault category set includes: The centroid features are obtained by extracting the geometric center coordinates of the optimized clusters; The variance feature is obtained by calculating the dispersion value of each data point in the optimized cluster relative to the centroid feature; The centroid features and variance features are input into the preset mechanism feature library for matching and retrieval; If the centroid feature and the variance feature fall within a preset matching interval, the corresponding fault category is determined, and the fault categories are combined to obtain a fault category set.

8. The intelligent monitoring and fault diagnosis method for industrial equipment production process according to claim 1, characterized in that, The step of updating the clustering threshold in the clustering model based on the stability index value to obtain an optimized clustering model includes: The target deviation value is obtained by calculating the deviation between the stability index value and the preset ideal convergence value; The parameter compensation amount is obtained by performing proportional-integral-derivative calculations on the target deviation value using a PID control algorithm. The updated threshold is obtained by superimposing the parameter compensation amount onto the clustering threshold in the clustering model; The updated threshold is used to replace the clustering threshold in the clustering model to obtain an optimized clustering model.

9. An intelligent monitoring and fault diagnosis system for industrial equipment production processes, characterized in that, include: The preprocessing module is used to acquire multi-source signals from the device and preprocess the multi-source signals to obtain a purified signal set. The signal filtering module is used to filter the purified signal set according to the propagation characteristics based on preset propagation constraints to obtain a filtered signal set. The mode determination module is used to process the filtered signal set using short-time Fourier transform to obtain a time-frequency distribution matrix, perform differential calculation between the time-frequency distribution matrix and a preset background noise to obtain a correction spectrum column, and perform shift analysis on the correction spectrum column to obtain a frequency drift mode. The feature construction module is used to extract modulation phase features from the filtered signal set if the frequency drift mode deviates from a preset reference value, and to construct an early feature group from the modulation phase features. The clustering matching module is used to cluster the early feature groups using a preset clustering model to obtain an initial feature cluster, perform boundary shrinking on the initial feature cluster to obtain an optimized cluster, and match the optimized cluster with a preset mechanism feature library to obtain a fault category set. The index calculation module is used to acquire historical time series segments, compare the feature drift amount with the fault category set and the historical time series segments to obtain a stable index value, including: frequency band locking of the fault category set to obtain a target frequency band interval; peak extraction of the filtered signal set within the target frequency band interval to obtain the dominant spectral feature; data extraction of the historical time series within the target frequency band interval to obtain a dynamic energy column; numerical analysis of the dominant spectral feature to obtain the feature amplitude; boundary extraction of the dynamic energy column to obtain the upper boundary value; difference calculation of the feature amplitude at the current time and the upper boundary value to obtain the feature drift amount; weighted calculation of the feature drift amount using a preset correction weight value to obtain a deviation correction value; and weighted product calculation of the deviation correction value and a preset initial confidence level to obtain a stable index value. The model optimization module is used to update the clustering threshold in the clustering model based on the stability index value, so as to obtain an optimized clustering model.

Images