AI-driven industrial equipment fault analysis method and system

Abstract

The invention relates to the technical field of industrial equipment diagnosis, in particular to an AI-driven industrial equipment fault analysis method and system, and solves the problems of insufficient multi-source data fusion, weak working condition adaptability and inaccurate fault positioning in the prior art. According to the method, high-frequency time sequence signals are synchronously collected through a vibration sensor, a temperature sensor and the like, working condition parameters are marked, a dynamic confidence coefficient weight is generated based on a working condition matching historical fault library, and cross-modal features are extracted through a space-time convolutional network after multi-source data are fused. A fault mode separation module is used for decoupling the fault contribution degree of mechanical transmission, electric power and heat dissipation systems, precise diagnosis is achieved in combination with frequency domain fingerprints, time domain waveform analysis and cross decision verification, finally, the remaining life is predicted through a long-short-term memory network and a Weibull model, and the system integrates multi-source sensing, dynamic fusion and intelligent diagnosis units. The fault early warning accuracy and the equipment maintenance scientificity under complex working conditions are improved, and a solution is provided for intelligent diagnosis of industrial equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial equipment diagnosis, and more specifically, to an AI-driven industrial equipment fault analysis method and system. Background Art

[0002] Industrial equipment diagnostic technology is an important technology. In modern industrial production, industrial equipment fault analysis technology is a core means to ensure the continuous operation of production lines and reduce maintenance costs by monitoring the operating status of equipment in real time and identifying potential fault hazards. As the intelligence and complexity of equipment increase, multi-sensor data fusion analysis and early fault warning become key needs.

[0003] However, traditional fault analysis technologies suffer from the core problems of insufficient multi-source data fusion and weak adaptability to operating conditions. Existing solutions independently analyze single sensor data such as vibration and temperature, and do not establish a correlation model between operating condition parameters and sensor signals. When the load intensity or ambient temperature and humidity change, the signal-to-noise ratio of the sensor data fluctuates, leading to feature extraction bias and an increase in the misdiagnosis rate of mechanical faults. When fixed-weight multi-source data is fused, the contribution of fault features from different sensors changes when the equipment speed changes, increasing the missed detection rate of power system faults. In addition, traditional methods lack the ability to decouple the equipment level and are unable to distinguish the fault contributions of mechanical transmission, power, and cooling systems. When multiple systems fail concurrently, the positioning accuracy is low, resulting in increased blind maintenance and extended unplanned downtime. This rough data fusion and lack of adaptability to operating conditions ultimately result in low fault warning accuracy for traditional solutions in complex industrial scenarios, high maintenance costs, and difficulty meeting the reliability requirements of intelligent manufacturing. To address this technical problem, we provide an AI-driven industrial equipment fault analysis method and system. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI-driven industrial equipment fault analysis method and system to solve the problems raised in the above background technology.

[0005] 1. Due to the insufficient multi-source data fusion of traditional solutions and the deviation in feature extraction when the working conditions change, this case fuses multi-source data through dynamic confidence weighting and combines the working condition parameters with the historical fault library to improve the accuracy of data fusion under complex working conditions and reduce the misdiagnosis rate of mechanical faults.

[0006] 2. Because traditional methods lack device-level decoupling capabilities and cannot accurately locate concurrent faults in multiple systems, this case uses a decoupling neural network to distinguish the fault contribution of each subsystem, which can achieve precise fault location, reduce maintenance blindness, and improve fault location accuracy.

[0007] To achieve the above objectives, one of the objectives of the present invention is to provide an AI-driven industrial equipment fault analysis method, comprising the following steps:

[0008] S1. Vibration sensors, temperature sensors, acoustic sensors, and current sensors embedded in industrial equipment are used to synchronously collect high-frequency timing signals and mark equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed.

[0009] S2. Based on the operating condition parameters, the historical fault database is matched and the characteristic correlation coefficient between each sensor data and the typical fault mode under the current operating condition is calculated. The real-time signal-to-noise ratio of each sensor signal is dynamically evaluated using a transfer learning model. The dynamic confidence weight of each data source is generated based on the characteristic correlation coefficient and the real-time signal-to-noise ratio.

[0010] S3. Input the weighted multi-source data into the spatiotemporal convolutional network to extract cross-modal fusion features. The fusion features are decoupled at the device level through the fault mode separation module to distinguish the fault contributions of the mechanical transmission system, power system, and cooling system. The fault location results and remaining service life prediction with mechanism explainability are output.

[0011] As a further improvement of this technical solution, the acquisition of the dynamic confidence weight in S2 includes:

[0012] The load intensity, ambient temperature and humidity, and operating speed are input into a fully connected neural network, and a 128-dimensional operating condition feature vector is output. The cosine similarity between the current operating condition feature vector and the fault mode labels in the historical fault database is calculated to generate a feature correlation coefficient. The spectral kurtosis and envelope entropy of the sensor signal are extracted through a transfer learning model, and a real-time signal-to-noise ratio score is generated based on the stability of the spectral energy distribution. The feature correlation coefficient and the real-time signal-to-noise ratio score are normalized and weighted and summed to output the dynamic confidence weight of each data source.

[0013] As a further improvement of this technical solution, the cross-modal fusion feature extraction in S3 includes parallel extraction of spatiotemporal features and dynamic weight injection:

[0014] The spatiotemporal features include spatial and temporal dimensions. The spatial dimension is obtained by using a dilated convolutional layer to capture the spatial topological association of multi-sensor data, with the convolution kernel span being the number of sensors. The temporal dimension is obtained by using a gated recurrent unit to capture the long-term dependency of high-frequency time series signals.

[0015] The dynamic confidence weight of each data source is converted into a channel attention coefficient, and the feature map channel response of the convolutional layer is weighted and adjusted.

[0016] As a further improvement of this technical solution, it further includes frequency domain fingerprint extraction, time domain waveform analysis and cross-decision verification:

[0017] Frequency domain fingerprint extraction: performing fast Fourier transform on the weighted vibration sensor signal to extract the energy proportion at the harmonic multiple of the fundamental frequency as the mechanical fault fingerprint;

[0018] The time domain waveform analysis: calculates the waveform distortion rate of the current sensor signal, and when the distortion rate exceeds a threshold, triggers the secondary verification mechanism of the frequency domain fingerprint;

[0019] The cross-decision verification automatically increases the confidence weight of the acoustic sensor for arbitration when the time domain distortion rate conflicts with the frequency domain energy distribution.

[0020] As a further improvement of this technical solution, the device-level decoupling in S3 includes:

[0021] A three-layer decoupled neural network is constructed, consisting of an input layer, a hidden layer, and an output layer. The input layer receives the cross-modal fusion feature vector, the hidden layer separates the feature subspaces of the mechanical transmission system, the power system, and the heat dissipation system through a sparse autoencoder, and the output layer calculates the fault activation intensity of each subsystem and outputs the fault contribution percentage.

[0022] As a further improvement of the present technical solution, the mechanical transmission system fault diagnosis further includes:

[0023] If the harmonic energy of the vibration signal at 1-3 times the rotation frequency increases by more than 80% of the normal rate, and the signal-to-noise ratio of the acoustic sensor in this frequency band is greater than or equal to 20dB, it is determined to be a bearing inner ring fault;

[0024] When the acoustic signal shows a modulation side frequency at the meshing frequency sideband, the gear wear level can be confirmed in combination with the temperature gradient change rate.

[0025] As a further improvement of this technical solution, the power system fault diagnosis further includes:

[0026] When the third harmonic component of the current sensor increases to more than 15% of the fundamental wave and the temperature sensor detects a temperature rise rate greater than 5°C / min at the motor winding, an insulation aging warning is triggered;

[0027] If the current waveform shows a millisecond-level pulse drop and the vibration sensor detects impact energy in the motor axis, it is located as a short circuit between winding turns.

[0028] As a further improvement of the present technical solution, the heat dissipation system fault diagnosis further includes:

[0029] The turbulent sound pressure level in the 200-500 Hz frequency band is separated from the acoustic signal to construct a cooling air volume attenuation curve. When the air volume attenuation rate is greater than 30% and the temperature difference between the inlet and outlet of the temperature sensor is reduced by 40%, the radiator blockage location is determined.

[0030] As a further improvement of this technical solution, the remaining useful life prediction includes:

[0031] The fault contribution of the mechanical transmission system, power system, and heat dissipation system is input into the long short-term memory network and Weibull hybrid model to calculate the remaining life of each subsystem. The minimum value of the remaining life of the three subsystems is taken as the overall remaining life of the equipment, and the dominant fault subsystem is marked.

[0032] A second object of the present invention is to provide a system for implementing an AI-driven industrial equipment fault analysis method, comprising:

[0033] The multi-source sensing unit integrates four types of industrial-grade sensors: vibration, temperature, acoustics, and current. It is equipped with an operating parameter acquisition module to collect equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed.

[0034] The dynamic fusion unit deploys a working condition feature mapping neural network and a transfer learning signal-to-noise ratio assessment module, outputs a dynamic confidence weight matrix, and injects it into the spatiotemporal convolutional network in real time;

[0035] The intelligent diagnosis unit includes a cross-modal feature fusion module, a fault decoupling engine and a life predictor, and outputs a fault location map with subsystem contribution and a remaining life confidence interval.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention uses multi-source sensors to synchronously collect high-frequency time series signals such as vibration and temperature and mark operating parameters to provide multi-dimensional data support for fault analysis, ensure the comprehensiveness of data and its relevance to operating conditions, match historical fault libraries based on operating condition parameters, calculate feature correlation coefficients and generate dynamic confidence weights in combination with transfer learning models to achieve adaptive fusion of multi-source data, improve the accuracy of data fusion under complex operating conditions, and reduce feature extraction deviations caused by changes in operating conditions. The weighted data is input into the spatiotemporal convolutional network to extract cross-modal fusion features, and the equipment level is decoupled through the fault mode separation module. It can accurately distinguish the fault contribution of mechanical transmission, power, heat dissipation and other systems, and achieve accurate fault positioning and mechanism explainability analysis. Combined with frequency domain fingerprint extraction, time domain waveform analysis and cross-decision verification mechanism, the reliability of fault diagnosis is improved, and the remaining service life is predicted through long-short-term memory network and Weibull hybrid model, providing a scientific basis for equipment maintenance and ensuring safe operation of equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall workflow diagram of the present invention;

[0039] Figure 2It is a schematic diagram of the overall structure of the present invention;

[0040] The meaning of each number in the figure is:

[0041] 1. Multi-source perception unit; 2. Dynamic fusion unit; 3. Intelligent diagnosis unit. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] See also Figure 1-Figure 2 As shown, one of the purposes of this embodiment is to provide an AI-driven industrial equipment fault analysis method, including the following steps:

[0044] S1. Vibration sensors, temperature sensors, acoustic sensors, and current sensors embedded in industrial equipment are used to synchronously collect high-frequency timing signals and mark equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed.

[0045] To accurately analyze industrial equipment failures, it is first necessary to synchronously collect high-frequency time series signals and mark operating parameters through multiple types of sensors. The specific implementation is as follows:

[0046] In order to obtain comprehensive equipment operating status data, vibration sensors, temperature sensors, acoustic sensors and current sensors are embedded in key parts of industrial equipment. The four types of sensors use a unified clock source to synchronously collect vibration acceleration, temperature, sound pressure level and current signals, providing time- and space-consistent raw data for subsequent cross-modal analysis. While collecting sensor signals, the equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed, are obtained in real time through the industrial bus. The system adds operating parameter labels to each set of sensor time series data and uses hardware timestamp synchronization technology for marking. The collection time of the operating parameter is accurately matched with the sensor signal sampling point, ensuring that subsequent analysis can evaluate the signal characteristics based on specific operating conditions, providing accurate operating condition input for dynamic confidence weight calculation, and effectively solving the problems of incomplete data collection and lack of operating condition association in traditional solutions.

[0047] S2. Based on the operating condition parameters, the historical fault database is matched and the characteristic correlation coefficient between each sensor data and the typical fault mode under the current operating condition is calculated. The real-time signal-to-noise ratio of each sensor signal is dynamically evaluated using a transfer learning model. The dynamic confidence weight of each data source is generated based on the characteristic correlation coefficient and the real-time signal-to-noise ratio.

[0048] To solve the problem of fixed multi-source data fusion weights and poor adaptability to working conditions in traditional solutions, it is necessary to generate dynamic confidence weights based on working condition parameters and sensor signal characteristics. The acquisition of dynamic confidence weights in S2 includes:

[0049] In order to map the multi-dimensional working condition parameters into a computable feature vector, the load intensity, ambient temperature, ambient humidity and operating speed are input into a three-layer fully connected neural network, where the input layer has 4 neurons, the hidden layer has 64 neurons, and the output layer has 128 neurons. The hidden layer uses the ReLU activation function, and the output layer is not activated. For example, at a certain moment, the working condition parameters are load 80%, temperature 35°C, humidity 50% RH and speed 1800rpm. After network calculation, a 128-dimensional working condition feature vector Vc = [v1, v2, ..., v128] is output. This vector compresses the spatiotemporal correlation characteristics of the working condition parameters. ,To measure the correlation degree between the current working condition and ,historical faults, the cosine similarity between the working condition ,feature vector and the label vector of each fault mode in the ,historical fault database is calculated.,To adapt to the signal quality fluctuation caused by ,working condition changes, the pre-trained transfer learning model ,is used to extract the spectral kurtosis and envelope ,entropy of the sensor signal, and calculate the standard of ,spectral energy distribution as the stability index.,A real-time signal-to-noise ratio score of 0-1 is generated by ,linear mapping.,The feature correlation coefficient matrix C and the real-time signal-to-noise ,ratio score S are normalized and then weighted and summed,with the weight coefficient α=0.6, the working condition correlation weight,β=0.4, and the signal-to-noise ,ratio weight, formula is W. i =α·(c i -min(C) / max(C)-min(C))+β·s i , where W i is the dynamic confidence weight of the i-th sensor, c i is the characteristic correlation coefficient matrix of the i-th sensor, s i The real-time signal-to-noise ratio score of the i-th type of sensor is calculated. This weight increases the signal weight of high signal-to-noise ratio sensors under highly correlated working conditions, effectively reducing the feature extraction deviation caused by working condition changes and solving the problem of misdiagnosis and missed diagnosis caused by fixed weights in traditional solutions.

[0050] S3. Input the weighted multi-source data into the spatiotemporal convolutional network to extract cross-modal fusion features. The fusion features are decoupled at the device level through the fault mode separation module to distinguish the fault contributions of the mechanical transmission system, power system, and cooling system. The fault location results and remaining service life prediction with mechanism explainability are output.

[0051] In order to extract deep cross-modal features from multi-source sensor data and improve the accuracy of fault diagnosis, it is necessary to extract spatiotemporal features in parallel from the weighted multi-source data and inject dynamic weights to adjust the feature response. The cross-modal fusion features extracted in S3 include spatiotemporal feature parallel extraction and dynamic weight injection. The specific implementation is as follows:

[0052] In order to obtain the spatial correlation features between multi-sensor data, the dilated convolution layer is used to process the weighted sensor signal matrix. The dimension of the sensor signal matrix is ​​the number of sampling points × the number of sensors. For example, for four types of sensors, vibration, temperature, acoustics, and current, the number of sampling points is 1024, then the matrix dimension is 1024×4. The convolution kernel span of the dilated convolution layer is set to the number of sensors, that is, 4, and the dilation rate is set to 1. The convolution kernel weight is obtained through training. For example, the dimension of the convolution kernel is [4×1×64], where 64 is the number of output channels. The convolution kernel is convolved with the input matrix. The product operation captures the spatial correlation features of different sensor data at the same time, such as the coordinated change pattern of vibration and current signals. Compared with the traditional convolution layer, the dilated convolution layer expands the capture range of spatial features by 2 times without increasing the amount of computation, which is used to improve the accuracy of extracting multi-sensor spatial topological associations. To process the time series features of high-frequency time series signals, a gated recurrent unit network is used to process the output features of the dilated convolution layer. The gated recurrent unit contains an update gate and a reset gate, and captures the long-term dependencies of the signal through forgetting and memory mechanisms. The input is the feature sequence output by the dilated convolution layer, that is, feature dimension × number of sampling points. For example, if the feature dimension is 64 and the number of sampling points is 1024, the hidden layer dimension of the gated recurrent unit is set to 128. The feature vector containing the time dependency is obtained through forward propagation calculation. For example, for the vibration signal sequence during the development of a bearing fault, the gated recurrent unit can effectively memorize the temporal correlation between the early weak impact features and the later violent vibration features. To strengthen the feature channels corresponding to high-confidence sensors, the dynamic confidence weights of each data source are converted into channel attention coefficients. The dynamic confidence weights are first normalized. For example, the weights of vibration, temperature, acoustics, and current are [0.9, 0.7, 0.8, 0.6], which are normalized to [0.32, 0.25, 0.29, 0.14]. The weights are then expanded to the same dimension as the number of channels in the output feature map of the convolutional layer. The feature map response is adjusted through weighted operations on the channel dimension to obtain the weighted feature map. Through the above-mentioned parallel extraction of spatiotemporal features and dynamic weight injection operations, the system realizes cross-modal feature fusion of multi-source sensor data, providing highly discriminative feature representation for subsequent fault location and life prediction.

[0053] To further improve the reliability of fault diagnosis, after extracting cross-modal fusion features, it is necessary to combine frequency domain fingerprints, time domain waveforms, and cross-decision verification to perform multi-dimensional fault feature verification. The specific implementation method is as follows:

[0054] Further includes frequency domain fingerprint extraction, time domain waveform analysis and cross-decision verification:

[0055] To capture the characteristic frequency components of mechanical faults, a fast Fourier transform is performed on the weighted vibration sensor signal to convert the time domain signal into a frequency domain spectrum. First, the fundamental frequency f0 is calculated based on the equipment operating speed (speed / 60). Then, the energy proportions of the fundamental frequency and its 2nd to 5th harmonics, namely 50Hz, 75Hz, 100Hz, and 125Hz, are extracted as the mechanical fault fingerprint. The energy proportion calculation formula is: Where X(f i ) is the spectrum amplitude of the i-th harmonic, n is the number of spectrum points, in order to monitor the abnormality of the power system, the waveform distortion rate of the current sensor signal is calculated, first the fundamental component is extracted through the sliding window, and then the root mean square error between the actual waveform and the fundamental component is calculated. The distortion rate formula is: distortion rate = root mean square error / fundamental amplitude × 100%, the preset distortion rate threshold is 10%, when the calculated value exceeds the threshold, such as the distortion rate is 15% at a certain moment, the secondary verification mechanism of the frequency domain fingerprint is automatically triggered to ensure the accurate distinction of the fault type, when the current signal distortion rate conflicts with the frequency domain energy distribution of the vibration signal, such as the distortion rate is greater than 10% but there is no obvious abnormality in the frequency domain energy, the system automatically enhances the acoustic sensor The arbitration mechanism is based on the following rules: if the time domain distortion rate exceeds the threshold and the frequency domain energy change is less than 20%, the acoustic sensor's confidence weight is increased from the initial value of 0.8 to 0.95, and the sound pressure level characteristics in the 200-500Hz frequency band are extracted for auxiliary judgment. For example, under certain operating conditions, the current distortion rate is 18% but the vibration frequency domain energy is normal. After increasing the acoustic weight, it is found that the gear meshing frequency sideband energy has increased by 15dB. The final judgment is that the gear is slightly worn rather than a power failure. This effectively solves the limitations of single-feature diagnosis, provides an accurate diagnostic basis for the rapid maintenance of industrial equipment, and effectively avoids unplanned downtime caused by misjudgment of a single feature.

[0056] To accurately locate industrial equipment faults and address the inaccurate location of concurrent multi-system faults in traditional methods, a three-layer decoupling neural network is required to perform device-level decoupling of cross-modal fusion features. The specific implementation is as follows:

[0057] Device-level decoupling in S3 includes:

[0058] In order to distinguish the fault contribution of mechanical transmission, power and cooling systems, a three-layer decoupling neural network is constructed, which includes input layer, hidden layer and output layer. The input layer receives 128-dimensional cross-modal fusion feature vector, which contains the features of multi-sensor spatiotemporal correlation extracted by spatiotemporal convolutional network. The number of neurons in the input layer is consistent with the feature dimension. The feature vector is transferred to the hidden layer through full connection. The time delay of input processing is controlled within 10m to ensure real-time performance. The hidden layer separates the feature subspaces of mechanical transmission system, power system and cooling system through sparse autoencoder, which contains 256 neurons. Decoupling is achieved through the following steps: First, the encoder is used to map the input features to the latent space, and the encoder weight matrix Bias The encoding formula is h=σ(W e ·F+b e ), where σ is the ReLU activation function. Then, the input features are reconstructed by the decoder, and the decoder weight matrix Bias The reconstruction formula is During training, a sparse constraint term is introduced to make only 10% of the hidden layer neurons active, thus forcing the network to learn the independent feature representation of each subsystem. The output layer contains 3 neurons, corresponding to the fault activation intensity of the mechanical transmission system, power system, and cooling system respectively. The hidden layer feature h is mapped to the output layer through the fully connected layer. The weight matrix Bias The activation function is Softmax, and the calculation formula is a=Softmax(W o h+b o ), where a=[a1,a2,a3], a i Expressed as the fault activation intensity of the i-th subsystem, the final fault contribution percentage p i =a i ×100%, the calculation of the output layer makes the quantization error of each subsystem fault contribution less than 5%, providing an explainable quantitative basis for fault location, and effectively solving the problems of fuzzy fault location and high maintenance blindness in traditional methods.

[0059] After completing device-level decoupling, fault diagnosis of mechanical transmission systems requires combining vibration and acoustic signal characteristics with the temperature gradient change rate to accurately identify bearing and gear faults. The specific implementation methods are as follows:

[0060] Mechanical transmission system fault diagnosis further includes:

[0061] To avoid misjudgment based on a single signal, the frequency domain characteristics of the vibration signal and the signal-to-noise ratio of the acoustic signal are combined to determine the bearing inner ring fault. First, the 1-3 times speed frequency range is calculated based on the equipment operating speed. For example, at a speed of 1800 rpm, the fundamental frequency is 30 Hz, and the 1-3 times speed frequency is 30-90 Hz. Then, the harmonic energy growth rate within this frequency band is calculated using the following formula: Among them E 当前 The current moment is 1-3 times the frequency energy, E 基准 is the baseline energy of the equipment in normal state, △t is the time interval. When the growth rate exceeds 80% of the normal rate and the signal-to-noise ratio of the acoustic sensor in this frequency band is greater than or equal to 20dB, signal-to-noise ratio = signal energy / noise energy, it is determined to be a bearing inner ring fault. To quantify the degree of gear wear, the wear level is confirmed by coordinating the acoustic signal modulation sideband and the temperature gradient change rate. First, the gear meshing frequency in the acoustic signal is extracted. For example, when the speed is 1500rpm and the number of teeth is 20, the meshing frequency is 500Hz. Then, the sideband is detected to see if there is a modulation side frequency, i.e., f m ±n×rotation frequency, n=1, 2, 3, and calculate the gearbox temperature gradient change rate at the same time. The formula is Among them, T 当前 is the current temperature, T 1小时前 The temperature is 1 hour ago. The wear level is determined according to the number of modulation side frequencies and the temperature gradient change rate: 1 modulation side frequency and a change rate of less than 1°C / hour is light wear, 2 side frequencies and a change rate of 1-3°C / hour is moderate wear, and 3 or more side frequencies and a change rate of >3°C / hour is heavy wear, which provides an accurate grading basis for gear maintenance.

[0062] After completing device-level decoupling, power system fault diagnosis requires combining current harmonic components, motor winding temperature, and vibration impact characteristics to accurately identify insulation aging and inter-turn short-circuit faults. The specific implementation method is as follows:

[0063] Power system fault diagnosis further includes:

[0064] In order to detect the hidden dangers of motor insulation aging at an early stage, a joint warning method of the current third harmonic component ratio and the winding temperature rise rate is used. First, a fast Fourier transform is performed on the current sensor signal to calculate the energy ratio of the third harmonic component to the fundamental wave. The formula is: Among them, f0 is the fundamental frequency of the current, which is determined by the grid frequency, and X(f) is the spectrum amplitude at frequency f. When the proportion of the third harmonic increases to more than 15% of the fundamental wave, and the temperature sensor detects a temperature rise rate greater than 5°C / min at the motor winding, the insulation aging warning is triggered, avoiding motor burning accidents caused by insulation failure. In order to accurately locate the short-circuit fault between the winding turns, the current waveform pulse drop and the motor axial vibration impact are jointly judged. First, a sliding window analysis is performed on the current signal with a window length of 10ms and a step size of 1ms to detect whether there is a millisecond-level pulse drop, that is, the drop amplitude is greater than 20% of the fundamental wave amplitude and the duration is 1-5ms. For example, the normal current amplitude is 10A, and a 2A pulse drop occurs and The impact energy is expressed in g·ms by the product of the peak value of the time-domain signal and the duration. When the impact energy is greater than 0.5 g·ms, the fault is located as a short circuit between winding turns. In this fault, the short-circuit current causes a sudden change in the electromagnetic force, triggering an axial vibration impact. The time difference between the two is less than 10 ms. For example, at a certain moment, a 2 ms pulse drop occurs in the current waveform, and the amplitude drops from 8 A to 6 A. At the same time, the axial vibration impact energy reaches 0.8 g·ms, locating the short circuit in the third winding turn. This provides accurate fault location information for rapid maintenance of the motor, effectively reducing the loss of unplanned downtime caused by power failures.

[0065] After completing device-level decoupling, cooling system fault diagnosis requires combining the turbulent sound pressure level of the acoustic signal, the cooling air volume attenuation curve, and the temperature difference change of the temperature sensor to accurately determine the radiator blockage location. The specific implementation method is as follows:

[0066] Cooling system fault diagnosis further includes:

[0067] To monitor the air volume changes in the cooling system, the turbulent sound pressure level in the 200-500 Hz frequency band is first separated from the acoustic sensor signal. This frequency band is positively correlated with the turbulence intensity of the airflow in the radiator. The signal is filtered using a bandpass filter, and the sound pressure level amplitude in this frequency band is calculated by the root mean square. For example, under normal working conditions, the sound pressure level is 70 dB. Then, based on historical data, a cooling air volume attenuation curve is constructed, with the sound pressure level as the horizontal axis and the actual air volume as the vertical axis. Through the calibration of the air volume sensor, a mapping relationship Q = f(L p ), where Q is the air volume, L p The sound pressure level is 200-500Hz. When the radiator is gradually blocked, the air turbulence increases, the sound pressure level rises, the air volume decreases, and the slope of the curve The result is obtained by fitting 100 sets of calibration data, which can accurately reflect the dynamic relationship between air volume and sound pressure level. In order to accurately determine the blockage position of the radiator, the combined threshold of the air volume attenuation rate and the inlet / outlet temperature difference change rate of the temperature sensor is calculated. The formula for the air volume attenuation rate is: Among them, Q 基准 Q is the reference air volume when the radiator is clean. 当前 is the real-time air volume, and the temperature sensor inlet / outlet temperature difference reduction rate formula is: where △T 基准 The reference temperature difference is 20°C for inlet and outlet temperature in clean state, △T 当前 For real-time temperature difference, when the air volume attenuation rate is greater than 30% and the temperature difference reduction rate is greater than 40%, the radiator is determined to be blocked, and the blockage position is located through the spatial distribution of sound pressure levels: if the sound pressure level on the left side of the radiator is 5dB higher than that on the right side, the blockage position is located in the left 1 / 3 area, providing accurate blockage location information for maintenance of the cooling system, effectively improving the cooling efficiency and operational reliability of the equipment.

[0068] After completing the fault contribution analysis of each subsystem, in order to scientifically predict the remaining service life of the equipment and guide maintenance decisions, the fault contribution needs to be input into the long-short-term memory network and the Weibull hybrid model for remaining service life prediction. The specific implementation method is as follows:

[0069] Remaining useful life forecasts include:

[0070] In order to capture the time series characteristics of fault development, the historical data of fault contribution of mechanical transmission, power and heat dissipation systems are input into the long short-term memory network. The long short-term memory network contains 2 hidden layers, 128 neurons in each layer, and the input dimension is 3, corresponding to the contribution of 3 subsystems. The output dimension is 1, that is, the remaining life prediction value. The mean square error loss function is used during training, the optimizer is adaptive moment estimation, and the learning rate is 0.001. For example, the fault contribution of a mechanical transmission system gradually increases from 10% to 30%. The long short-term memory network predicts its remaining life to be 150 hours by memorizing the historical growth trend. The prediction error of the long short-term memory network for nonlinear degradation trends is less than 10%. Compared with the traditional linear extrapolation method, the prediction accuracy of accelerated degradation scenarios is improved. In order to quantify the uncertainty of the remaining life prediction, the Weibull mixture model is used to calculate the probability distribution of the remaining life of each subsystem. The Weibull distribution probability density function is Among them, λ is the characteristic life, k is the shape parameter, that is, dimensionless. Through maximum likelihood estimation, λ and k are obtained based on the subsystem failure history data. For the current fault contribution, the characteristic life is adjusted by the mapping function λ'=λ×(1-contribution). The hybrid model generates a probability density curve of the remaining life by weighting multiple Weibull distributions, with the weights determined by the proportion of failure modes, providing a quantitative risk assessment basis for maintenance decisions. The overall remaining life of the equipment is the minimum value of the remaining life of the three subsystems, and the formula is L 整体 =min(L 机械 , L 电力 , L 散热 ); where L机械 , L 电力 , L 散热 The remaining life prediction values ​​of the three subsystems are given respectively, and the dominant fault subsystem, that is, the subsystem with the largest contribution, is marked. This judgment method conforms to the physical law that the shortest life subsystem of the equipment determines the overall life, effectively avoiding the problems of excessive maintenance or delayed maintenance. Through the above remaining life prediction method, the system realizes the full chain intelligence from fault diagnosis to life prediction, provides scientific quantitative support for predictive maintenance of industrial equipment, and effectively reduces maintenance costs and unplanned downtime losses.

[0071] A second object of the present invention is to provide a system for implementing an AI-driven industrial equipment fault analysis method as described above, comprising:

[0072] The multi-source sensing unit 1 integrates four types of industrial-grade sensors: vibration, temperature, acoustics, and current. It is equipped with an operating parameter acquisition module to collect equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed.

[0073] Dynamic fusion unit 2 deploys the working condition feature mapping neural network and transfer learning signal-to-noise ratio evaluation module, outputs the dynamic confidence weight matrix, and injects it into the spatiotemporal convolutional network in real time;

[0074] The intelligent diagnosis unit 3 includes a cross-modal feature fusion module, a fault decoupling engine and a life predictor, and outputs a fault location map with subsystem contribution and a remaining life confidence interval.

[0075] The present invention uses multiple sensors such as vibration and temperature to synchronously collect high-frequency time series signals and mark operating parameters. It generates dynamic confidence weights based on the operating condition matching historical fault library, extracts cross-modal features through a spatiotemporal convolutional network after fusing multi-source data, and uses a fault mode separation module to decouple the fault contribution of mechanical transmission, power, and heat dissipation systems. It combines frequency domain fingerprinting, time domain waveform analysis, and cross-decision verification to achieve accurate diagnosis. Finally, it predicts the remaining service life through a long short-term memory network and a Weibull model, thereby improving the accuracy of fault warning and the scientific nature of equipment maintenance under complex working conditions, and providing a solution for intelligent diagnosis of industrial equipment.

[0076] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven industrial equipment fault analysis method, characterized in that: The following steps are involved: S1. Vibration sensors, temperature sensors, acoustic sensors, and current sensors embedded in industrial equipment are used to synchronously collect high-frequency timing signals and mark equipment operating parameters, including load intensity, ambient temperature and humidity, and operating speed. S2. Based on the operating condition parameters, the historical fault database is matched and the characteristic correlation coefficient between each sensor data and the typical fault mode under the current operating condition is calculated. The real-time signal-to-noise ratio of each sensor signal is dynamically evaluated using a transfer learning model. The dynamic confidence weight of each data source is generated based on the characteristic correlation coefficient and the real-time signal-to-noise ratio. S3. Input the weighted multi-source data into the spatiotemporal convolutional network to extract cross-modal fusion features. The fusion features are decoupled at the device level through the fault mode separation module to distinguish the fault contributions of the mechanical transmission system, power system, and cooling system. The fault location results and remaining service life prediction with mechanism explainability are output.

2. The AI-driven industrial equipment fault analysis method according to claim 1, characterized in that: The acquisition of the dynamic confidence weight in S2 includes: The load intensity, ambient temperature and humidity, and operating speed are input into a fully connected neural network, and a 128-dimensional operating condition feature vector is output. The cosine similarity between the current operating condition feature vector and the fault mode labels in the historical fault database is calculated to generate a feature correlation coefficient. The spectral kurtosis and envelope entropy of the sensor signal are extracted through a transfer learning model, and a real-time signal-to-noise ratio score is generated based on the stability of the spectral energy distribution. The feature correlation coefficient and the real-time signal-to-noise ratio score are normalized and weighted and summed to output the dynamic confidence weight of each data source.

3. The AI-driven industrial equipment fault analysis method according to claim 1, characterized in that: Extracting cross-modal fusion features in S3 includes parallel extraction of spatiotemporal features and dynamic weight injection: The spatiotemporal features include spatial and temporal dimensions. The spatial dimension is obtained by using a dilated convolutional layer to capture the spatial topological association of multi-sensor data, with the convolution kernel span being the number of sensors. The temporal dimension is obtained by using a gated recurrent unit to capture the long-term dependency of high-frequency time series signals. The dynamic confidence weight of each data source is converted into a channel attention coefficient, and the feature map channel response of the convolutional layer is weighted and adjusted.

4. The AI-driven industrial equipment fault analysis method according to claim 3, characterized in that: Further includes frequency domain fingerprint extraction, time domain waveform analysis and cross-decision verification: Frequency domain fingerprint extraction: performing fast Fourier transform on the weighted vibration sensor signal to extract the energy proportion at the harmonic multiples of the fundamental frequency as the mechanical fault fingerprint; The time domain waveform analysis: calculates the waveform distortion rate of the current sensor signal, and when the distortion rate exceeds a threshold, triggers the secondary verification mechanism of the frequency domain fingerprint; The cross-decision verification automatically increases the confidence weight of the acoustic sensor for arbitration when the time domain distortion rate conflicts with the frequency domain energy distribution.

5. The AI-driven industrial equipment fault analysis method according to claim 1, characterized in that: The device-level decoupling in S3 includes: A three-layer decoupled neural network is constructed, consisting of an input layer, a hidden layer, and an output layer. The input layer receives the cross-modal fusion feature vector, the hidden layer separates the feature subspaces of the mechanical transmission system, the power system, and the heat dissipation system through a sparse autoencoder, and the output layer calculates the fault activation intensity of each subsystem and outputs the fault contribution percentage.

6. The AI-driven industrial equipment fault analysis method according to claim 5, characterized in that: The mechanical transmission system fault diagnosis further includes: If the harmonic energy of the vibration signal at 1-3 times the rotation frequency increases by more than 80% of the normal rate, and the signal-to-noise ratio of the acoustic sensor in this frequency band is greater than or equal to 20dB, it is determined to be a bearing inner ring fault; When the acoustic signal shows a modulation side frequency at the meshing frequency sideband, the gear wear level can be confirmed in combination with the temperature gradient change rate.

7. The AI-driven industrial equipment fault analysis method according to claim 5, characterized in that: The power system fault diagnosis further includes: When the third harmonic component of the current sensor increases to more than 15% of the fundamental wave and the temperature sensor detects a temperature rise rate greater than 5°C / min at the motor winding, an insulation aging warning is triggered; If the current waveform shows a millisecond-level pulse drop and the vibration sensor detects impact energy in the motor axis, it is located as a short circuit between winding turns.

8. The AI-driven industrial equipment fault analysis method according to claim 5, characterized in that: The heat dissipation system fault diagnosis further includes: The turbulent sound pressure level in the 200-500 Hz frequency band is separated from the acoustic signal to construct a cooling air volume attenuation curve. When the air volume attenuation rate is greater than 30% and the temperature difference between the inlet and outlet of the temperature sensor is reduced by 40%, the radiator blockage location is determined.

9. The AI-driven industrial equipment fault analysis method according to claim 1, characterized in that: The remaining useful life forecast includes: The fault contribution of the mechanical transmission system, power system, and heat dissipation system is input into the long short-term memory network and Weibull hybrid model to calculate the remaining life of each subsystem. The minimum value of the remaining life of the three subsystems is taken as the overall remaining life of the equipment, and the dominant fault subsystem is marked.

10. A system for implementing an AI-driven industrial equipment fault analysis method according to any one of claims 1 to 9, characterized in that: include: The multi-source sensing unit (1) integrates four types of industrial-grade sensors, namely vibration, temperature, acoustics and current, and is equipped with a working condition parameter acquisition module to collect equipment working condition parameters, wherein the equipment working condition parameters include load intensity, ambient temperature and humidity, and operating speed; The dynamic fusion unit (2) deploys the working condition feature mapping neural network and the transfer learning signal-to-noise ratio evaluation module, outputs the dynamic confidence weight matrix, and injects it into the spatiotemporal convolutional network in real time; The intelligent diagnosis unit (3) includes a cross-modal feature fusion module, a fault decoupling engine and a life predictor, and outputs a fault location map with subsystem contribution and a remaining life confidence interval.

Images