Intelligent mechanical equipment fault diagnosis method and system fused with machine learning algorithm

Abstract

The invention provides an intelligent mechanical equipment fault diagnosis method fused with a machine learning algorithm, and the method comprises the steps: obtaining a multi-source signal, and generating a complete machine-level anomaly detection result through dynamic threshold comparison and machine learning anomaly detection model processing; the multi-source signal comprises a vibration signal, a temperature signal and a pressure signal; performing signal decomposition and multi-source parameter fusion analysis processing based on a complete machine level anomaly detection result, and determining a physical subsystem of an anomaly source; time-frequency features are extracted from the operation data of the physical subsystem, and a part-level anomaly positioning result and a microdefect classification result are generated through hierarchical machine learning model processing; and performing multi-granularity data fusion processing based on a complete machine level anomaly detection result, a physical subsystem and a microdefect classification result, and generating a hierarchical diagnosis report. By adopting the method, progressive diagnosis can be carried out on the mechanical equipment hierarchically, and the accuracy of fault diagnosis is improved.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical fault diagnosis, and in particular relates to an intelligent mechanical equipment fault diagnosis method and system integrating a machine learning algorithm. Background Art

[0002] Fault diagnosis of intelligent mechanical equipment is an indispensable research direction in the modern industrial field. Its core lies in ensuring the reliability and safety of equipment operation, which is directly related to the improvement of production efficiency and economic benefits. With the increasing complexity of industrial systems, the importance of fault diagnosis has become increasingly prominent, and it has become a key technical pillar to promote the development of intelligent manufacturing. However, traditional fault diagnosis methods often expose significant limitations when facing complex and changeable mechanical systems. Existing solutions mostly rely on a single level of analysis, which is either limited to rough detection at the system level and lacks in-depth analysis of specific components; or they focus too much on details and ignore the overall operating status, resulting in one-sided and inefficient diagnostic results. These methods are difficult to find a balance between comprehensiveness and accuracy, especially in resource-constrained scenarios, and are difficult to meet actual needs.

[0003] In this context, the core challenges facing this field have gradually surfaced. The most critical technical factors include the hierarchical depth of diagnostic analysis, the synergy between algorithms, and the optimal allocation of computing resources. Due to the lack of systematic analysis capabilities from macro to micro, the current methods are difficult to effectively respond to the detection needs of multi-level faults such as the entire system, subsystem components, and even the microstructure of materials. In addition, the lack of an efficient linkage mechanism between coarse-grained and fine-grained algorithms has led to a disconnect between abnormality detection and precise positioning. The uneven distribution of computing resources has further exacerbated the contradiction between diagnostic efficiency and accuracy. These unresolved technical problems make it difficult to achieve comprehensive and accurate application of intelligent fault diagnosis in complex equipment.

[0004] Therefore, how to build a multi-granularity collaborative intelligent diagnosis system to achieve a hierarchical diagnosis process from whole machine abnormality detection to component-specific analysis and then to micro-defect identification has become a key issue that needs to be broken through. This problem not only requires the algorithm to be seamlessly connected between different granularities, but also requires high-precision analysis of key parts while ensuring comprehensive coverage, and achieving the goal of efficient operation under resource constraints.

[0005] Existing methods are difficult to simultaneously cover the multi-level fault detection requirements of the entire system, subsystem components and material microstructures. Summary of the invention

[0006] Based on this, it is necessary to provide an intelligent mechanical equipment fault diagnosis method and system that integrates machine learning algorithms to address the above-mentioned technical problems, which can perform progressive diagnosis of mechanical equipment in a hierarchical manner and improve the accuracy of fault diagnosis.

[0007] In a first aspect, the present application provides an intelligent mechanical equipment fault diagnosis method integrating machine learning algorithms, including:

[0008] Obtain multi-source signals and process them through dynamic threshold comparison and a machine learning anomaly detection model to generate an overall machine-level anomaly detection result; the multi-source signals include vibration signals, temperature signals, and pressure signals;

[0009] Based on the overall machine-level anomaly detection result, perform signal decomposition and multi-source parameter fusion analysis to determine the physical subsystem where the anomaly source is located;

[0010] Extract time-frequency features from the operation data of the physical subsystem and process them through a hierarchical machine learning model to generate a component-level anomaly localization result and a microscopic defect classification result;

[0011] Based on the overall machine-level anomaly detection result, the physical subsystem, and the microscopic defect classification result, perform multi-granularity data fusion processing to generate a hierarchical diagnosis report.

[0012] In a possible embodiment, obtaining multi-source signals and processing them through dynamic threshold comparison and a machine learning anomaly detection model to generate an overall machine-level anomaly detection result includes:

[0013] Based on the peak-to-peak value and kurtosis of the vibration signal, generate a dynamic threshold through a time series prediction model;

[0014] Perform sliding window statistical processing on the temperature signal and the pressure signal to generate auxiliary anomaly determination parameters;

[0015] Input the dynamic threshold and the auxiliary anomaly determination parameters into a random forest classifier for processing to generate an overall machine-level anomaly detection result.

[0016] In a possible embodiment, based on the overall machine-level anomaly detection result, performing signal decomposition and multi-source parameter fusion analysis to determine the physical subsystem where the anomaly source is located includes:

[0017] Based on the variational mode decomposition algorithm, separate the vibration signal of the whole machine into multiple sub-band components to generate a mode decomposition result;

[0018] Perform Pearson correlation analysis on the energy ratio of the mode decomposition result, temperature, and pressure parameters to generate a subsystem correlation score;

[0019] Input the subsystem correlation score into a logistic regression classifier for processing to determine the physical subsystem where the anomaly source is located.

[0020] In a possible embodiment, extracting time-frequency features from the operation data of the physical subsystem and processing them through a hierarchical machine learning model to generate a component-level anomaly localization result and a microscopic defect classification result includes:

[0021] Perform time-frequency analysis on the operation data of the physical subsystem based on the short-time Fourier transform to generate the frequency and amplitude feature vectors of key components;

[0022] Input the frequency and amplitude feature vectors into the pre-trained ResNet-34 model for processing to generate component-level anomaly localization results;

[0023] Calculate the multi-scale wavelet entropy of the high-frequency signal corresponding to the component-level anomaly localization result and process it through a cascade classifier to obtain the microscopic defect classification result.

[0024] In one possible embodiment, the generation and update of the dynamic threshold include:

[0025] Generate the dynamic threshold according to the following formula:

[0026]

[0027] where T(t) is the dynamic threshold, T0 is the initial threshold, α is the attenuation factor, μ is the mean of the vibration signal, σ is the standard deviation of the vibration signal, k is the statistical confidence coefficient, and t is the time step;

[0028] Use the following formula to perform a comparison process based on the dynamic threshold, peak-to-peak value, and kurtosis to generate a vibration determination parameter, and the vibration judgment parameter is used to characterize the vibration abnormal state:

[0029]

[0030] where P vib is the comparison determination parameter, X pp is the peak-to-peak value, K is the kurtosis, K th is the preset kurtosis threshold, and else is used to represent other situations;

[0031] When the value of the vibration judgment parameter is 1, use the following formula to update the attenuation factor according to the cumulative operation time of the device to obtain the updated attenuation value:

[0032] α(t) = α0 - Δ α ·sigmoid(t / τ)

[0033] where α(t) is the updated attenuation value, Δ α is the attenuation amplitude coefficient, and τ is the time attenuation coefficient;

[0034] Recalculate the dynamic threshold based on the updated attenuation value.

[0035] In one possible embodiment, calculating the multi-scale wavelet entropy of the high-frequency signal corresponding to the component-level anomaly localization result and processing it through a cascade classifier to obtain the microscopic defect classification result includes:

[0036] Perform multi-scale decomposition processing on high-frequency signals based on the Morlet wavelet transform to generate a wavelet coefficient matrix corresponding to the defect scales of different materials;

[0037] Use the following formula to perform energy distribution statistical processing on the wavelet coefficient matrix to obtain the energy entropy of each scale:

[0038]

[0039] where E(s) is the energy entropy, s is the scale, N is the number of signal sampling points, and t i is the time point index;

[0040] Input the energy entropy into a pre-trained cascade classifier for multi-scale feature fusion to generate a microscopic defect classification result.

[0041] In a possible embodiment, perform multi-granularity data fusion processing based on the whole-machine-level anomaly detection result, physical subsystems, and microscopic defect classification result to generate a hierarchical diagnostic report, including:

[0042] Extract the attention heat map of the component anomaly region from the ResNet-34 model through the gradient class activation map;

[0043] Intercept high-frequency signal segments based on the significant region of the attention heat map, perform wavelet entropy calculation processing, and generate a microscopic defect classification result;

[0044] Perform weighted fusion processing on the confidence levels of the whole-machine-level anomaly detection result, the positioning result of the physical subsystem, the component-level anomaly positioning result, and the microscopic defect classification result to generate a comprehensive failure probability;

[0045] Calculate the difference between the comprehensive failure probability and the historical failure distribution based on the KL divergence. If the difference exceeds the threshold, trigger model recalibration.

[0046] In a second aspect, the present application also provides an intelligent mechanical equipment fault diagnosis system integrating machine learning algorithms, including:

[0047] A whole-machine anomaly detection module, configured to obtain multi-source signals and process them through dynamic threshold comparison and a machine learning anomaly detection model to generate a whole-machine-level anomaly detection result; the multi-source signals include vibration signals, temperature signals, and pressure signals;

[0048] A subsystem anomaly determination module, configured to perform signal decomposition and multi-source parameter fusion analysis processing based on the whole-machine-level anomaly detection result to determine the physical subsystem where the anomaly source is located;

[0049] The micro - anomaly processing module is used to extract time - frequency features from the operation data of the physical subsystem, and through a hierarchical machine - learning model for processing, generate component - level anomaly location results and micro - defect classification results;

[0050] The multi - granularity hierarchical diagnosis module is used to perform multi - granularity data fusion processing based on the whole - machine - level anomaly detection results, physical subsystems, and micro - defect classification results, and generate a hierarchical diagnosis report.

[0051] In a third aspect, the present application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the intelligent mechanical equipment fault diagnosis method of the above - mentioned fusion machine - learning algorithm.

[0052] In a fourth aspect, the present application also provides a computer - readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the intelligent mechanical equipment fault diagnosis method of the above - mentioned fusion machine - learning algorithm.

[0053] The intelligent mechanical equipment fault diagnosis method and system of the above - mentioned fusion machine - learning algorithm are based on the collaborative processing of dynamic threshold comparison of vibration, temperature, and pressure multi - source signals and a machine - learning anomaly detection model to achieve a preliminary identification of the whole - machine - level abnormal state, providing a global anomaly location benchmark for subsequent analysis; through signal decomposition and multi - source parameter fusion analysis of the whole - machine abnormal data, a subsystem relevance evaluation mechanism is established to accurately lock the abnormal physical subsystem, completing the hierarchical progression from the whole machine to the subsystem; for the time - frequency characteristics of the abnormal subsystem operation, a hierarchical machine - learning model is used to simultaneously perform component - level anomaly location and micro - defect classification, achieving refined diagnosis from the subsystem to components and micro - structures; through multi - granularity data fusion technology, the hierarchical diagnosis results of the whole machine, subsystems, components, and micro - defects are integrated to form a comprehensive analysis conclusion covering the full - level state of the mechanical equipment. The above - mentioned method improves the comprehensiveness and accuracy of intelligent mechanical equipment fault diagnosis through a progressive diagnosis process, gradually narrowing the fault range and increasing the analysis granularity layer by layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following will briefly introduce the drawings required for use in the description of the embodiments or related technologies. Obviously, the following - described drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0055] Figure 1 It is a flowchart of an intelligent mechanical equipment fault diagnosis method of a fusion machine - learning algorithm provided by an embodiment of the present invention;

[0056] Figure 2 This is a schematic structural diagram of an intelligent mechanical equipment fault diagnosis system integrating machine learning algorithms provided by an embodiment of the present invention. Detailed implementation manners

[0057] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the following further details the present application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0058] First, a brief introduction is made to the nouns involved in the embodiments of the present application.

[0059] Vibration signal is an important physical quantity generated during the operation of a mechanical system, which can reflect the mechanical state. It is presented in the form of a time series, generated and transmitted by the vibration of mechanical components. Among them, the peak-to-peak value, as a key parameter of the vibration signal, refers to the difference between the maximum value and the minimum value of the signal within a complete cycle, which can intuitively and quantitatively show the amplitude range of the vibration, helping technicians judge the vibration intensity during the operation of the machine. Kurtosis is an important index used to characterize the impact characteristics of the vibration signal. It is extremely sensitive to the impact components in the signal. When mechanical components have faults such as wear and cracks, abnormal impacts will be caused, and at this time, the kurtosis value will change significantly. Through in-depth analysis of the vibration signal and its characteristics such as peak-to-peak value and kurtosis, it can provide key basis for the fault diagnosis and performance evaluation of mechanical equipment.

[0060] The random forest classifier is a machine learning algorithm based on ensemble learning, which consists of multiple decision trees. These decision trees are constructed by randomly sampling the training data with replacement, and each decision tree is independently trained and classified. When classifying, the random forest combines the classification results of each decision tree, usually by voting, and takes the category with the most votes as the final classification result. The random forest classifier has high accuracy and stability, can effectively process high-dimensional data and data with noise, and can also well avoid the overfitting phenomenon, and has a wide range of applications in the field of mechanical equipment fault diagnosis and so on.

[0061] The Variational Mode Decomposition (VMD) algorithm is an adaptive signal processing method. It decomposes complex signals into multiple Intrinsic Mode Function (IMF) components with different center frequencies and bandwidths. By constructing a variational model and solving it using the Alternating Direction Method of Multipliers (ADMM), the bandwidth of each IMF component is minimized under the constraint conditions, thus achieving effective decomposition of the signal. This algorithm has good anti-noise performance and decomposition accuracy, can adaptively extract the characteristic information of the signal, and performs excellently in processing non-stationary and non-linear signals. It is widely used in many fields such as mechanical fault diagnosis, speech processing, and image processing.

[0062] The ResNet-34 model is a deep residual neural network model belonging to the ResNet series. It consists of 34 stacked convolutional layers, pooling layers, fully connected layers, etc. By introducing residual connections, it solves the problems of gradient vanishing and degradation that occur as the network depth increases, enabling the model to be trained deeper and thus learn more complex feature representations. This model performs excellently in fields such as image recognition, can automatically extract high-level features of images, accurately classify and locate objects in images, and has high accuracy and generalization ability.

[0063] The KL divergence, namely the Kullback-Leibler divergence, is an index used to measure the degree of difference between two probability distributions. When applied in fields such as mechanical equipment fault diagnosis, it quantifies the deviation between the currently generated comprehensive fault probability distribution and the historical fault distribution by comparing them. The larger the KL divergence value, the more significant the difference between the two distributions, indicating a larger deviation of the current fault mode from the historical situation.

[0064] According to the above noun explanations, the implementation environment of an intelligent mechanical equipment fault diagnosis method integrating machine learning algorithms provided by the embodiments of this application is described. Exemplarily, this implementation environment includes: a sensor group and a processor. Among them, the processor group is signal-connected to the sensor group and the storage device through a network; the sensor group can be an acceleration sensor, a thermocouple sensor, a pressure sensor, etc.; the processor includes but is not limited to a central processing unit, a multi-core processor, or an artificial intelligence chip, etc., which is not limited here.

[0065] Combined with the above noun explanations and implementation environment, the application scenarios of the embodiments of this application are described. An intelligent mechanical equipment fault diagnosis method integrating machine learning algorithms provided by the embodiments of this application can be applied to the following scenarios including but not limited to:

[0066] In the precision machining scenario, this method can monitor the vibration of the machine tool spindle, cutting force, and temperature rise data in real time, and identify the overall operation deviation through a dynamic threshold model; when abnormal spindle vibration energy is detected, the characteristic frequency bands of the spindle bearing and tool clamping system are separated, and the abnormal subsystem is located by combining the change of cutting load; for tool wear or spindle eccentricity problems, the energy distribution characteristics of high-frequency acoustic emission signals are extracted, and the historical defect maps are matched through a transfer learning model to realize early warning of microscopic damages such as tool chipping and spindle microcracks, ensuring machining accuracy and equipment life.

[0067] In energy power equipment, such as steam turbines and generators in power stations, as well as compressors and pumps in the petrochemical industry, the stability of their operation is crucial to the entire energy supply and production process. Using this method can analyze the operation data of the equipment and timely diagnose potential faults such as rotor imbalance, blade damage, and seal leakage. Taking a wind turbine generator as an example, by collecting vibration, temperature, and rotational speed data of components such as the wind turbine, gearbox, and generator, combined with machine learning algorithms, gear wear of the gearbox, winding faults of the generator, etc. can be predicted, realizing preventive maintenance, reducing maintenance costs and downtime, and ensuring the stable supply of energy.

[0068] Exemplarily, the intelligent mechanical equipment fault diagnosis method integrating machine learning algorithms provided in the embodiments of the present application can also be applied to other application scenarios. Only examples are given here, and the specific application scenarios are not limited.

[0069] In an exemplary embodiment, as Figure 1 shown, an intelligent mechanical equipment fault diagnosis method integrating machine learning algorithms is provided. Taking the application of this method to the aforementioned processor as an example for illustration. In this embodiment, this method includes the following steps 101 to step 104:

[0070] Step 101, obtain multi-source signals and process them through dynamic threshold comparison and machine learning anomaly detection model to generate an overall machine-level anomaly detection result; the multi-source signals include vibration signals, temperature signals, and pressure signals.

[0071] Specifically, install corresponding sensors at key parts of the mechanical equipment, such as vibration sensors, temperature sensors, and pressure sensors, which are respectively used to collect vibration signals, temperature signals, and pressure signals. Dynamically set the normal range thresholds of each signal according to the historical operation data and current working conditions of the equipment, train the machine learning anomaly detection model with historical data, input the real-time collected multi-source signals into the trained model, and the model judges whether the signals are abnormal according to the characteristic distribution of the data; comprehensively combine the results of dynamic threshold comparison and model judgment to generate an overall machine-level anomaly detection result, and mark whether there is an anomaly in the overall machine.

[0072] Step 102: Based on the whole-machine-level anomaly detection results, perform signal decomposition and multi-source parameter fusion analysis to determine the physical subsystem where the anomaly source lies.

[0073] Specifically, when the whole-machine-level anomaly detection results indicate the existence of an anomaly, decompose the collected multi-source signals for a clearer analysis of the frequency components and characteristics of the signals. Conduct a fusion analysis on the characteristic parameters of multi-source signals such as vibration, temperature, and pressure to find the correlation relationships between different signal characteristics. By analyzing the changes in the signal characteristics corresponding to each physical subsystem and combining the structure and working principle of the equipment, determine the physical subsystem where the anomaly source lies.

[0074] Step 103: Extract time-frequency characteristics from the operation data of the physical subsystem and process them through a hierarchical machine learning model to generate component-level anomaly localization results and microscopic defect classification results.

[0075] Specifically, extract time-frequency characteristics from the operation data of the physical subsystem, such as the peak value, mean value, variance, and spectral characteristics of the signal, to reflect the operation status and fault characteristics of the component.

[0076] Step 104: Based on the whole-machine-level anomaly detection results, physical subsystem, and microscopic defect classification results, perform multi-granularity data fusion processing to generate a hierarchical diagnostic report.

[0077] Specifically, fuse the whole-machine-level anomaly detection results, the anomaly information of the physical subsystem, and the microscopic defect classification results at the component level. Comprehensively consider the data information at different levels and adopt data fusion algorithms such as the weighted average method and D-S evidence theory to obtain more comprehensive and accurate diagnostic information. Further, the whole-machine health status, anomaly subsystem identifier, component localization results, and defect types can be integrated into a visualization interface to automatically generate a hierarchical diagnostic report including fault severity grading and maintenance suggestions, providing a complete decision-making chain from macroscopic anomaly warning to microscopic defect determination and improving the operation and maintenance personnel's understanding and handling efficiency of complex faults.

[0078] The intelligent mechanical equipment fault diagnosis method and system integrating the machine learning algorithm, through the collaborative processing of the dynamic threshold comparison of multi-source signals of vibration, temperature, and pressure and the machine learning anomaly detection model, realizes the preliminary identification of the overall machine-level abnormal state, providing a global abnormal location benchmark for subsequent analysis; through the signal decomposition and multi-source parameter fusion analysis of the overall machine abnormal data, a subsystem correlation evaluation mechanism is established to accurately lock the abnormal physical subsystem and complete the hierarchical progression from the overall machine to the subsystem; for the running time-frequency characteristics of the abnormal subsystem, a hierarchical machine learning model is used to simultaneously perform component-level abnormal location and microscopic defect classification, realizing the refined diagnosis from the subsystem to the component and the microscopic structure; through the multi-granularity data fusion technology, the hierarchical diagnosis results of the overall machine, subsystem, component, and microscopic defect are integrated to form a comprehensive analysis conclusion covering the full hierarchical state of the mechanical equipment. The above method improves the comprehensiveness and accuracy of the intelligent mechanical equipment fault diagnosis through a progressive diagnosis process, gradually narrowing the fault range and increasing the analysis granularity layer by layer.

[0079] In a possible embodiment, obtaining multi-source signals and processing them through dynamic threshold comparison and the machine learning anomaly detection model to generate the overall machine-level anomaly detection result may include:

[0080] Step 201, generating a dynamic threshold through a time series prediction model based on the peak-to-peak value and kurtosis of the vibration signal.

[0081] Specifically, obtain the vibration signal of the mechanical equipment and calculate its peak-to-peak value and kurtosis in real time as time domain features. Among them, the peak-to-peak value is used to characterize the extreme fluctuation of the vibration amplitude, and the kurtosis is used to quantify the sharpness of the signal distribution. Input the historical vibration data into the time series prediction model (such as the ARIMA model) to predict the dynamic threshold under the current working condition; exemplarily, the dynamic threshold can be adaptively adjusted according to the equipment running time, load status, and historical fault data. The dynamic threshold generated through this step can track the equipment aging or working condition fluctuation in real time, avoid false alarms or missed alarms caused by a fixed threshold, fuse the equipment degradation trend through the time series model, and improve the flexibility and environmental adaptability of the threshold setting to enhance the sensitivity of the overall machine anomaly detection.

[0082] Step 202, performing sliding window statistical processing on the temperature signal and the pressure signal to generate auxiliary anomaly determination parameters.

[0083] Specifically, temperature signals and pressure signals can be acquired, and statistical features such as mean, variance, and rate of change can be calculated based on a sliding time window. For temperature signals, the average temperature rise rate and local extreme values within the window are extracted; for pressure signals, the fluctuation amplitude and periodic change trend are calculated. By comparing the above statistical features with the preset normal operating condition range, temperature anomaly scores and pressure anomaly scores are generated as auxiliary determination parameters. The statistical processing in this step can effectively smooth the instantaneous noise interference, capture the slow-varying anomaly patterns of temperature and pressure, construct an auxiliary determination system through multi-dimensional statistical parameters, make up for the detection blind spots of single vibration signals, improve the comprehensiveness of anomaly detection under complex working conditions, and reduce the risk of misjudgment.

[0084] Step 203: Input the dynamic threshold and the auxiliary anomaly determination parameters into a random forest classifier for processing to generate an overall machine-level anomaly detection result.

[0085] Specifically, the dynamic threshold (the basis for vibration signal determination) and the temperature and pressure auxiliary anomaly determination parameters are concatenated to form a multi-dimensional input vector, which is then input into a pre-trained random forest classifier. This classifier can be trained based on historical normal and fault data, and comprehensively determines whether the current state is abnormal through the voting mechanism of multiple decision trees. When the vibration signal exceeds the dynamic threshold, or the cumulative anomaly scores of the temperature and pressure auxiliary parameters exceed the classification threshold, an overall machine-level anomaly detection result is output. It can utilize the high inclusiveness and anti-overfitting ability of the random forest for multi-source heterogeneous data to achieve collaborative decision-making of vibration, temperature, and pressure signals, improve the accuracy and reliability of overall machine anomaly determination, and maintain stable output especially in multi-parameter conflict scenarios.

[0086] In a possible embodiment, based on the overall machine-level anomaly detection result, signal decomposition and multi-source parameter fusion analysis processing are performed to determine the physical subsystem of the anomaly source, which may include:

[0087] Step 301: Separate the vibration signal of the overall machine into multiple sub-band components based on the variational mode decomposition algorithm to generate a modal decomposition result.

[0088] Specifically, through variational mode decomposition processing of the overall machine vibration signal, the vibration signal is adaptively separated into multiple sub-band components according to the signal frequency domain characteristics. Each sub-band component corresponds to different physical subsystems in the mechanical equipment (such as bearings, gearboxes, motor rotors, etc.). The decomposition process ensures that each sub-band component is independent in the time-frequency domain and matches the characteristic frequencies of the physical subsystems through preset modal number and bandwidth constraint parameters. It can effectively avoid the modal aliasing problem of traditional methods (such as wavelet decomposition) through adaptive frequency band separation technology, accurately extract the vibration characteristics of each subsystem, and provide a high-resolution signal basis for subsequent fault location.

[0089] Step 302: Perform Pearson correlation analysis on the energy proportion of the modal decomposition result and the temperature and pressure parameters to generate a subsystem correlation score.

[0090] Specifically, the system calculates the energy proportion of each sub-band component (i.e., the ratio of the component energy to the total vibration energy), and simultaneously obtains the real-time measurement values of the temperature and pressure parameters. Through Pearson correlation coefficient analysis, the statistical correlation degree between the energy proportion of each component and the temperature and pressure parameters is analyzed to generate a subsystem correlation score. For example, if the energy proportion of a certain component shows a strong positive correlation with the change in bearing temperature, it is determined that the component corresponds to an abnormal bearing subsystem. This method can eliminate the limitations of single-signal analysis through the joint analysis of multi-source parameters (vibration, temperature, pressure), enhance the reliability of subsystem positioning, and is particularly suitable for abnormal source tracing in multi-subsystem coupling fault scenarios.

[0091] Step 303: Input the subsystem correlation score into a logistic regression classifier for processing to determine the physical subsystem where the abnormality source is located.

[0092] Specifically, the subsystem correlation score can be used as a feature vector and input into a pre-trained logistic regression classifier for classification decision-making. This classifier establishes a mapping relationship between the score and the physical subsystem based on historical fault data, and determines the subsystem where the abnormality source is located through probability output. For example, when the probability of the bearing subsystem exceeds 80%, it is determined as abnormal. The logistic regression model can efficiently process structured data, quickly identify strongly correlated subsystems, and at the same time provide an interpretable decision basis through probabilistic output, improving the real-time performance and accuracy of fault location and enhancing the fault diagnosis efficiency.

[0093] In a possible embodiment, time-frequency features are extracted from the operation data of the physical subsystem and processed through a hierarchical machine learning model to generate component-level abnormal location results and microscopic defect classification results, including:

[0094] Step 401: Perform time-frequency analysis on the operation data of the physical subsystem based on the short-time Fourier transform to generate a frequency and amplitude feature vector of the key components.

[0095] Specifically, perform short-time Fourier transform (STFT) processing on the vibration signal of the physical subsystem (such as a bearing or a gearbox). By setting the Hanning window function and a fixed time window length (such as 20 ms), the time-domain signal is converted into a time-frequency matrix, and the amplitude change characteristics of the key frequency components are extracted. The frequency and amplitude feature vector includes the main frequency band energy distribution, the harmonic component amplitude ratio, and the instantaneous frequency volatility, which are used to characterize the operating state of the component. Capturing the local characteristics of non-stationary signals through time-frequency analysis can improve the sensitivity to transient faults (such as impact damage), and at the same time maintain the balance between frequency resolution and time resolution, providing high-information-density input data for component-level abnormal location.

[0096] Step 402: Input the frequency and amplitude feature vectors into a pre-trained ResNet-34 model for processing to generate component-level anomaly localization results.

[0097] Specifically, the ResNet-34 model can be trained by transfer learning based on historical fault data (such as inner ring cracks of bearings, broken teeth of gears, etc.), extract deep spatial features through the residual connection structure, and output component-level anomaly localization results (such as "inner ring fault of bearing" or "tooth surface wear of gear"). The deep network structure of ResNet-34 can effectively capture the subtle differences in anomaly patterns in the time-frequency diagram, combine the transfer learning strategy to solve the small sample training problem, and achieve high-precision and low-latency component-level fault localization, especially suitable for multi-component concurrent fault scenarios under complex working conditions.

[0098] Step 403: Calculate the multi-scale wavelet entropy of the high-frequency signal corresponding to the component-level anomaly localization result and process it through a cascade classifier to obtain the microscopic defect classification result.

[0099] Specifically, the randomness and non-linear characteristics of microscopic defects are quantified by wavelet entropy, combined with the two-stage decision-making mechanism of the cascade classifier, which can reduce the computational complexity while ensuring the classification accuracy, effectively distinguish microscopic defects with similar morphologies (such as fatigue cracks and mechanical scratches), and improve the credibility of defect classification.

[0100] In a possible embodiment, the generation and update of the dynamic threshold may include:

[0101] Step 501: Generate a dynamic threshold according to the following formula:

[0102]

[0103] where \(T(t)\) is the dynamic threshold, \(T_0\) is the initial threshold, \(\alpha\) is the attenuation factor, \(\mu\) is the mean of the vibration signal, \(\sigma\) is the standard deviation of the vibration signal, \(k\) is the statistical confidence coefficient, and \(t\) is the time step.

[0104] Specifically, by combining the time decay factor with the statistical extreme value compensation, the threshold is adaptively adjusted with the aging of the equipment and the fluctuation of the working conditions, avoiding the limitations of the fixed threshold.

[0105] Step 502: Use the following formula to perform a comparison process based on the dynamic threshold, peak-to-peak value, and kurtosis to generate a vibration determination parameter, and the vibration judgment parameter is used to characterize the vibration abnormal state.

[0106]

[0107] where \(P\) vib is the comparison determination parameter, \(X\) ppis the peak-to-peak value, K is the kurtosis, and K th is the preset kurtosis threshold, and else is used to represent other cases.

[0108] Specifically, by combining the time-domain feature (peak-to-peak value) and the statistical feature (kurtosis) to construct a multi-dimensional determination condition, the synchronous capture ability of transient shock and steady-state anomaly is enhanced, and the misjudgment risk of a single index is reduced.

[0109] Step 503, when the value of the vibration judgment parameter is 1, use the following formula to update the attenuation factor according to the cumulative operation time of the device to obtain the updated attenuation value:

[0110] α(t) = α0 - Δ α ·sigmoid(t / τ)

[0111] where α(t) is the updated attenuation value, and Δ α is the attenuation amplitude coefficient, and τ is the time attenuation coefficient.

[0112] Specifically, the nonlinear change of the attenuation rate is controlled by the Sigmoid function. It responds quickly to device mutations in the initial stage and slowly tracks long-term degradation in the later stage, balancing sensitivity and stability.

[0113] Step 504, recalculate the dynamic threshold based on the updated attenuation value.

[0114] Specifically, through the above method, the attenuation factor is dynamically updated based on anomalies, realizing the continuous self-optimization of the threshold parameters. It can adapt to complex working conditions and the requirements of the entire life cycle management of the device, and continuously maintain the effectiveness and accuracy of device anomaly detection.

[0115] In a possible embodiment, performing multi-scale wavelet entropy calculation on the high-frequency signal corresponding to the component-level anomaly localization result and processing it through a cascade classifier to generate a microscopic defect classification result may include:

[0116] Step 601, perform multi-scale decomposition processing on the high-frequency signal based on the Morlet wavelet transform to generate a wavelet coefficient matrix corresponding to different material defect scales.

[0117] Specifically, perform Morlet wavelet transform processing on the high-frequency vibration signal corresponding to the component-level anomaly localization result. By presetting multiple scale parameters, the high-frequency signal is decomposed into sub-signal components of different scales to generate a wavelet coefficient matrix. Among them, the scale parameter is inversely proportional to the physical size of the material defect (for example, a large scale corresponds to a micro crack, and a small scale corresponds to a macroscopic spall), so as to realize the targeted feature extraction of defects of different sizes. The above method adapts to the size diversity of microscopic defects through multi-scale decomposition, overcoming the problem of insufficient characterization ability of traditional single-scale analysis for complex defects.

[0118] Step 602: Use the following formula to perform energy distribution statistical processing on the wavelet coefficient matrix to obtain the energy entropy of each scale:

[0119]

[0120] where E(s) is the energy entropy, s is the scale, N is the number of signal sampling points, and t i is the time point index.

[0121] Specifically, the energy entropy is used to quantify the randomness of the energy distribution of the signal at a specific scale: the higher the entropy value, the more complex the defect morphology (such as irregular cracks); the lower the entropy value, the corresponding uniform damage (such as uniform wear). By quantifying the physical characteristics of the defect through the entropy value, the traditional qualitative analysis is transformed into a computable quantitative index, improving the objectivity of the classification basis.

[0122] Step 603: Input the energy entropy into a pre-trained cascade classifier for multi-scale feature fusion to generate a microscopic defect classification result.

[0123] Exemplarily, the cascade classifier can be composed of a support vector machine (SVM) preliminary screening module and a convolutional neural network (CNN) fine classification module. Among them, the SVM preliminary screening module quickly screens the major defect categories based on the statistical features of the energy entropy; the CNN fine classification module is used to perform convolution operations on the multi-scale entropy value sequence corresponding to the preliminary screening result, extract cross-scale spatial features, and output a fine-grained classification result. The above method can balance efficiency and accuracy through a cascade architecture. The SVM quickly excludes irrelevant categories, and the CNN focuses on high-value features, solving the problem of insufficient fitting of a single model to multi-scale data.

[0124] In a possible embodiment, multi-granularity data fusion processing is performed based on the whole-machine-level anomaly detection result, the physical subsystem, and the microscopic defect classification result to generate a hierarchical diagnostic report, which may include:

[0125] Step 701: Extract the attention heat map of the component anomaly area from the ResNet-34 model through the gradient class activation map.

[0126] Specifically, using the gradient class activation map technique, the gradient information of the feature map is extracted from the last convolutional layer of the pre-trained ResNet-34 model to generate the attention heat map of the component abnormal area. By superimposing this heat map on the original time-frequency map, it visually identifies the key areas (such as the high-frequency vibration area of the bearing inner ring or the damaged area of the gear tooth surface) that lead to the abnormal determination of the component. This method transforms the "black box" decision of the deep learning model into an interpretable heat distribution map through a visualization attention mechanism, which helps engineers verify the rationality of abnormal localization, provides a spatial localization basis for subsequent high-frequency signal interception, and can combine the model feature response with the physical signal feature to improve the credibility and traceability of component abnormal localization.

[0127] Step 702: Based on the significant area of the attention heat map, intercept the high-frequency signal segment, perform wavelet entropy calculation processing, and generate the microscopic defect classification result.

[0128] Specifically, according to the time range of the high-weight area in the attention heat map, intercept the corresponding segment from the original high-frequency vibration signal. Perform Morlet wavelet transform on the intercepted signal segment, calculate the multi-scale energy entropy value, and input it into the pre-trained cascade classifier for fine-grained classification to achieve the precise alignment of "model decision - physical signal", avoid the waste of resources in the full-time signal analysis. This method can focus on the key signal segments, improve the microscopic defect classification efficiency, and reduce the misclassification risk caused by noise interference.

[0129] Step 703: Perform weighted fusion processing on the confidence levels of the whole-machine-level abnormal detection result, the localization result of the physical subsystem, the component-level abnormal localization result, and the microscopic defect classification result to generate the comprehensive fault probability.

[0130] Specifically, the confidence levels of the diagnostic results at different levels are weighted and fused to generate the comprehensive fault probability, which fully considers the diagnostic information at multiple levels such as the whole-machine level, the physical subsystem level, the component level, and the microscopic defect level. The diagnostic results at different levels reflect the operating state of the mechanical equipment from different angles. Through weighted fusion, various aspects of information can be comprehensively weighed, and the possibility of the overall equipment failure can be evaluated more comprehensively and accurately, avoiding the one-sidedness that may be brought by relying only on the diagnostic result of a single level and improving the reliability of fault diagnosis.

[0131] Step 704: Calculate the difference between the comprehensive fault probability and the historical fault distribution based on the KL divergence. If the difference exceeds the threshold, trigger model recalibration.

[0132] Specifically, the divergence between the comprehensive fault probability and the historical fault distribution is calculated based on the KL divergence, and the model recalibration is triggered according to the divergence, enabling the diagnostic model to adaptively accommodate changes in the operating state of mechanical equipment. When new fault modes occur in the equipment or the operating environment changes significantly, the model can promptly detect the differences from the historical situation and optimize itself through the recalibration mechanism, maintaining the accuracy and effectiveness of fault diagnosis, improving the adaptability and robustness of the diagnostic model, and ensuring long-term stable and reliable service for mechanical equipment fault diagnosis.

[0133] In summary, the intelligent mechanical equipment fault diagnosis method incorporating machine learning algorithms provided by the embodiments of this application generates the overall machine-level anomaly detection results based on multi-source signal dynamic threshold comparison and random forest classifiers, initially locking in the global anomalies; separates the subsystem characteristic frequency bands through variational mode decomposition, and locates the abnormal physical subsystems by combining the correlation analysis of temperature and pressure parameters; further extracts the time-frequency characteristics of subsystem operation, uses the ResNet-34 model to achieve component-level anomaly location, and analyzes the microscopic defect types based on multi-scale wavelet entropy and cascade classifiers; guides the focusing of high-frequency signals through attention heat maps, combines multi-level confidence weighted fusion and the model calibration mechanism driven by KL divergence, generates a hierarchical diagnostic report covering the whole machine, subsystems, components, and microscopic defects, and realizes the dynamic optimization of algorithm parameters. The above technical solutions achieve all-round accurate diagnosis from the whole machine to the micro level through a hierarchical and progressive technical architecture, can gradually focus on the fault scope, integrate multi-source signal complementary verification and physical feature quantitative modeling, decouple cross-level fault characteristics and achieve high-precision matching under complex working conditions, and improve the comprehensiveness and discriminant accuracy of fault diagnosis.

[0134] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are sequentially shown according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments, and the execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0135] Based on the same inventive concept, an embodiment of the present application further provides an intelligent mechanical equipment fault diagnosis system for implementing the fusion machine learning algorithm for the intelligent mechanical equipment fault diagnosis method involved above. The implementation solution provided by this system for solving problems is similar to the implementation solution recorded in the above method. Therefore, the specific limitations in one or more of the following intelligent mechanical equipment fault diagnosis method system embodiments for the fusion machine learning algorithm can refer to the limitations for the intelligent mechanical equipment fault diagnosis method for the fusion machine learning algorithm in the above text, and will not be elaborated here.

[0136] In an exemplary embodiment, as Figure 2 shown, an intelligent mechanical equipment fault diagnosis system 20 for the fusion machine learning algorithm is provided, including the following modules:

[0137] The whole machine abnormal detection module 21 is used to obtain multi-source signals and process them through dynamic threshold comparison and machine learning abnormal detection model to generate the whole machine level abnormal detection result; the multi-source signals include vibration signals, temperature signals and pressure signals.

[0138] The subsystem abnormal determination module 22 is used to perform signal decomposition and multi-source parameter fusion analysis processing based on the whole machine level abnormal detection result to determine the physical subsystem where the abnormal source is located.

[0139] The microscopic abnormal processing module 23 is used to extract time-frequency features from the operation data of the physical subsystem and process them through a hierarchical machine learning model to generate the component level abnormal location result and the microscopic defect classification result.

[0140] The multi-granularity hierarchical diagnosis module 24 is used to perform multi-granularity data fusion processing based on the whole machine level abnormal detection result, the physical subsystem and the microscopic defect classification result to generate a hierarchical diagnosis report.

[0141] In a possible embodiment, the whole machine abnormal detection module 21 may include:

[0142] The dynamic threshold generation unit 211 is used to generate a dynamic threshold through a time series prediction model based on the peak-to-peak value and kurtosis of the vibration signal.

[0143] The auxiliary abnormal judgment parameter generation unit 212 is used to perform sliding window statistical processing on the temperature signal and the pressure signal to generate auxiliary abnormal judgment parameters.

[0144] The whole machine abnormal detection unit 213 is used to input the dynamic threshold and the auxiliary abnormal judgment parameter into a random forest classifier for processing to generate the whole machine level abnormal detection result.

[0145] In a possible embodiment, the subsystem abnormal determination module 22 may include:

[0146] A modal decomposition unit 221, configured to separate the vibration signal of the whole machine into multiple sub-band components based on the variational mode decomposition algorithm, and generate a modal decomposition result.

[0147] A subsystem correlation analysis unit 222, configured to perform Pearson correlation analysis on the energy ratio of the modal decomposition result and the temperature and pressure parameters, and generate a subsystem correlation score.

[0148] A logistic regression classification unit 223, configured to input the subsystem correlation score into a logistic regression classifier for processing to determine the physical subsystem of the abnormal source.

[0149] In a possible embodiment, the microscopic anomaly processing module 23 may include:

[0150] A time-frequency analysis unit 231, configured to perform time-frequency analysis on the operation data of the physical subsystem based on the short-time Fourier transform, and generate a frequency and amplitude feature vector of the key components.

[0151] A component anomaly location unit 232, configured to input the frequency and amplitude feature vector into a pre-trained ResNet-34 model for processing to generate a component-level anomaly location result.

[0152] A microscopic defect classification unit 233, configured to calculate the multi-scale wavelet entropy of the high-frequency signal corresponding to the component-level anomaly location result and process it through a cascade classifier to obtain a microscopic defect classification result.

[0153] In a possible embodiment, the dynamic threshold generation unit 211 may include:

[0154] A threshold generation subunit 2111, configured to generate a dynamic threshold according to the following formula:

[0155]

[0156] where T(t) is the dynamic threshold, T0 is the initial threshold, α is the attenuation factor, μ is the mean of the vibration signal, σ is the standard deviation of the vibration signal, k is the statistical confidence coefficient, and t is the time step.

[0157] A vibration anomaly judgment subunit 2112, configured to use the following formula to perform a comparison process based on the dynamic threshold, the peak-to-peak value, and the kurtosis, and generate a vibration judgment parameter, where the vibration judgment parameter is used to characterize the vibration anomaly state:

[0158]

[0159] where P vib is the comparison judgment parameter, X pp is the peak-to-peak value, K is the kurtosis, K this a preset kurtosis threshold, and else is used to represent other situations.

[0160] The attenuation factor update subunit 2113 is used to update the attenuation factor according to the cumulative operation time of the device using the following formula to obtain an updated attenuation value when the value of the vibration judgment parameter is 1:

[0161] α(t) = α0 - Δ α ·sigmoid(t / τ)

[0162] where α(t) is the updated attenuation value, Δ α is the attenuation amplitude coefficient, and τ is the time attenuation coefficient.

[0163] The update threshold subunit 2114 is used to recalculate the dynamic threshold based on the updated attenuation value.

[0164] In a possible embodiment, the micro-defect classification unit 233 may include:

[0165] The wavelet transform subunit 2331 is used to perform multi-scale decomposition processing on the high-frequency signal based on the Morlet wavelet transform to generate a wavelet coefficient matrix corresponding to different material defect scales.

[0166] The energy entropy calculation subunit 2332 is used to perform energy distribution statistical processing on the wavelet coefficient matrix using the following formula to obtain the energy entropy of each scale:

[0167]

[0168] where E(s) is the energy entropy, s is the scale, N is the number of signal sampling points, and t i is the time point index.

[0169] The multi-scale feature fusion subunit 2333 is used to input the energy entropy into a pre-trained cascade classifier for multi-scale feature fusion to generate a micro-defect classification result.

[0170] In a possible embodiment, the multi-granularity hierarchical diagnosis module 24 may include:

[0171] The heatmap extraction unit 241 is used to extract the attention heatmap of the component abnormal area from the ResNet-34 model through the gradient class activation map.

[0172] The wavelet entropy calculation unit 242 is used to intercept the high-frequency signal segment based on the significant area of the attention heatmap and perform wavelet entropy calculation processing to generate a micro-defect classification result.

[0173] The comprehensive failure probability generation unit 243 is configured to perform weighted fusion processing on the confidence levels of the overall machine-level anomaly detection results, the localization results of physical subsystems, the component-level anomaly localization results, and the microscopic defect classification results to generate a comprehensive failure probability.

[0174] The difference degree detection unit 244 is configured to calculate the difference degree between the comprehensive failure probability and the historical failure distribution based on the KL divergence, and trigger model recalibration if the difference degree exceeds a threshold.

[0175] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of a method for fault diagnosis of an intelligent mechanical equipment integrating a machine learning algorithm as described above are implemented.

[0176] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0177] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can refer to the partial descriptions of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present disclosure. A person of ordinary skill in the art can understand and implement it without creative efforts.

[0178] The above embodiments only represent several implementation manners of the embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the embodiments of the application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the embodiments of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the embodiments of the present application.

Claims

1. A fault diagnosis method for intelligent mechanical equipment integrating machine learning algorithm, characterized in that: The method comprises: Acquire multi-source signals and generate whole-machine-level anomaly detection results by comparing dynamic thresholds with machine learning anomaly detection models; the multi-source signals include vibration signals, temperature signals, and pressure signals; Based on the whole machine level abnormality detection result, signal decomposition and multi-source parameter fusion analysis and processing are performed to determine the physical subsystem where the abnormality comes from; Extracting time-frequency features from the operation data of the physical subsystem, and processing them through a hierarchical machine learning model to generate component-level anomaly location results and micro-defect classification results; Multi-granularity data fusion processing is performed based on the whole-machine level abnormality detection results, the physical subsystem and the microscopic defect classification results to generate a hierarchical diagnosis report.

2. The method according to claim 1, characterized in that The method acquires multi-source signals and processes them through dynamic threshold comparison and machine learning anomaly detection model to generate whole-machine level anomaly detection results, including: Based on the peak-to-peak value and kurtosis of the vibration signal, a dynamic threshold is generated by a time series prediction model; Performing sliding window statistical processing on the temperature signal and the pressure signal to generate auxiliary abnormality determination parameters; The dynamic threshold and the auxiliary abnormality determination parameter are input into a random forest classifier for processing to generate the whole machine level abnormality detection result.

3. The method according to claim 2, characterized in that The signal decomposition and multi-source parameter fusion analysis processing based on the whole machine level abnormality detection result to determine the physical subsystem of the abnormality source includes: Separating the vibration signal of the whole machine into multiple sub-band components based on a variational modal decomposition algorithm to generate a modal decomposition result; Performing Pearson correlation analysis on the energy proportion, temperature and pressure parameters of the modal decomposition results to generate a subsystem correlation score; The subsystem correlation scores are input into a logistic regression classifier to determine the physical subsystem of the anomaly source.

4. The method according to claim 1, characterized in that: The time-frequency features are extracted from the operation data of the physical subsystem, and processed by a hierarchical machine learning model to generate component-level anomaly location results and micro-defect classification results, including: Performing time-frequency analysis on the operation data of the physical subsystem based on short-time Fourier transform to generate frequency and amplitude feature vectors of key components; Inputting the frequency and amplitude feature vectors into a pre-trained ResNet-34 model for processing to generate the component-level anomaly positioning result; The high-frequency signal corresponding to the component-level abnormality positioning result is subjected to multi-scale wavelet entropy calculation and processed through a cascade classifier to obtain the micro defect classification result.

5. The method according to claim 2, characterized in that: The generation and updating of the dynamic threshold includes: The dynamic threshold is generated according to the following formula: Wherein, T(t) is the dynamic threshold, T0 is the initial threshold, α is the attenuation factor, μ is the mean of the vibration signal, σ is the standard deviation of the vibration signal, k is the statistical confidence coefficient, and t is the time step; The following formula is used to generate a vibration determination parameter based on a comparison between the dynamic threshold, the peak-to-peak value, and the kurtosis. The vibration determination parameter is used to characterize an abnormal vibration state: Among them, P vib is the comparison judgment parameter, X pp is the peak-to-peak value, K is the kurtosis, K th is the preset kurtosis threshold, and else is used to represent other situations; When the value of the vibration judgment parameter is 1, the attenuation factor is updated according to the accumulated running time of the device using the following formula to obtain an updated attenuation value: α(t)=α0-Δ α ·sigmoid(t / τ) Wherein, α(t) is the updated attenuation value, Δ α is the attenuation amplitude coefficient, τ is the time attenuation coefficient; The dynamic threshold is recalculated based on the updated attenuation value.

6. The method according to claim 4, characterized in that The multi-scale wavelet entropy calculation of the high-frequency signal corresponding to the component-level abnormality positioning result and processing through a cascade classifier to obtain the micro defect classification result includes: Performing multi-scale decomposition processing on the high-frequency signal based on Morlet wavelet transform to generate wavelet coefficient matrices corresponding to different material defect scales; The energy entropy of each scale is obtained by performing energy distribution statistics on the wavelet coefficient matrix using the following formula: Where E(s) is the energy entropy, s is the scale, N is the number of signal sampling points, t i Index for time points; The energy entropy is input into the pre-trained cascade classifier for multi-scale feature fusion to generate the micro defect classification result.

7. The method according to claim 6, characterized in that The multi-granularity data fusion processing is performed based on the whole machine level abnormality detection result, the physical subsystem and the micro defect classification result to generate a hierarchical diagnosis report, including: Extracting an attention heat map of the abnormal area of ​​the component from the ResNet-34 model through a gradient class activation map; Based on the significant area of ​​the attention heat map, high-frequency signal segments are intercepted and processed by wavelet entropy calculation to generate the micro defect classification result; Performing weighted fusion processing on the confidence of the whole-machine-level anomaly detection result, the physical subsystem positioning result, the component-level anomaly positioning result, and the micro-defect classification result to generate a comprehensive fault probability; The difference between the comprehensive fault probability and the historical fault distribution is calculated based on the KL divergence, and if the difference exceeds a threshold, a model recalibration is triggered.

8. An intelligent mechanical equipment fault diagnosis system integrating machine learning algorithms, characterized in that: The system comprises: The whole machine anomaly detection module is used to obtain multi-source signals and generate whole machine level anomaly detection results by comparing dynamic thresholds with machine learning anomaly detection models; the multi-source signals include vibration signals, temperature signals and pressure signals; A subsystem anomaly determination module is used to perform signal decomposition and multi-source parameter fusion analysis based on the whole machine level anomaly detection result to determine the physical subsystem where the anomaly comes from; A micro-anomaly processing module is used to extract time-frequency features from the operation data of the physical subsystem, and generate component-level anomaly location results and micro-defect classification results through hierarchical machine learning model processing; The multi-granularity hierarchical diagnosis module is used to perform multi-granularity data fusion processing based on the whole machine level abnormality detection results, the physical subsystem and the micro defect classification results to generate a hierarchical diagnosis report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images