Motor fault real-time diagnosis method and system based on LSTM and random forest

Abstract

The invention belongs to the technical field of intelligent equipment fault diagnosis, and particularly relates to a motor fault real-time diagnosis method and system based on LSTM and random forest. The method comprises the steps that a vibration signal, a current signal and a temperature signal of a motor are collected and preprocessed; performing feature engineering, extracting time domain, frequency domain, time-frequency domain and cross-modal correlation features, and determining an optimal static feature subset through a hybrid screening strategy; constructing a hybrid fault prediction model comprising a random forest model and an LSTM sequential network model; fusing prediction results of the two models by adopting a dynamic credibility weighted fusion mechanism; and performing real-time decision and hierarchical feedback control based on a fusion result. According to the method, advantages of multi-source information and the model are fused, the classification precision of the lag type is remarkably improved, real-time fault diagnosis and active protection are realized, the equipment maintenance cost is reduced, and the method is suitable for motor health management of intelligent agricultural equipment such as mowers.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent equipment fault diagnosis, and specifically relates to a real-time motor fault diagnosis method and system based on LSTM and random forest. Background Art

[0002] Motors are the core power components of various types of mechanical equipment, and their operating status directly affects the performance and reliability of the equipment. Especially in smart agricultural equipment (such as lawn mowers), motors often operate in complex and changing environments, making them prone to failures such as freezing. Traditional motor fault diagnosis methods have many limitations:

[0003] Reliance on a single signal source: Many existing methods rely primarily on monitoring motor current signals and setting current thresholds to determine whether a motor stall has occurred. However, current signals alone cannot fully reflect the motor's mechanical state (such as abnormal vibration) and thermal state (such as local overheating). They are also easily affected by factors such as transient load fluctuations and grid noise, resulting in a high false alarm rate.

[0004] Inability to distinguish fault types: Threshold-based methods often fail to distinguish between temporary, recoverable jams caused by foreign matter entanglement and persistent jams requiring downtime and maintenance, such as bearing damage or winding short circuits. This distinction is crucial for taking appropriate action; misjudgment can lead to unnecessary downtime or delayed resolution, worsening the fault.

[0005] Model limitations: Some methods using machine learning models also have shortcomings. For example, traditional time series models (such as HMM) or single LSTM networks do not effectively utilize static non-time series features. Non-time series models (such as SVM and traditional random forests) struggle to effectively capture the dynamic patterns and dependencies of signal evolution over time, especially information reflecting the dynamic changes in mechanical vibration and current waveforms.

[0006] Passive response: Existing methods mostly focus on alarms after a fault occurs, lacking active protection and intervention mechanisms based on diagnostic results. This makes it difficult to take targeted measures in the early stages of a fault to mitigate damage or attempt automatic recovery.

[0007] Therefore, there is an urgent need for a motor fault diagnosis technology that can integrate multi-source information, accurately distinguish fault types, and achieve real-time diagnosis and active response, so as to improve the reliability of intelligent equipment and reduce maintenance costs. Summary of the Invention

[0008] Purpose of the invention: The purpose of the present invention is to address the deficiencies of the existing technology and provide a real-time diagnosis method for motor faults based on LSTM and random forest, aiming to achieve automatic classification and dynamic response of temporary and continuous stalls through data-driven technology.

[0009] Technical solution: The real-time motor fault diagnosis method based on LSTM and random forest described in the present invention includes the following steps:

[0010] S1: Synchronously collect the vibration signal, current signal and temperature signal of the motor through multi-modal sensors;

[0011] S2: Preprocessing the vibration signal, current signal and temperature signal to generate a preprocessing signal;

[0012] S3: performing feature engineering on the preprocessed signal to obtain an initial feature set, and processing the initial feature set using a hybrid screening strategy to generate an optimal feature subset;

[0013] S4: Build a hybrid fault prediction model that includes a random forest model and an LSTM time series network model, where:

[0014] The random forest model inputs the optimal feature subset and obtains a first prediction result based on the highly discriminative features therein;

[0015] The LSTM time series network model inputs the time series features in the initial feature set to capture the dynamic evolution law of the vibration signal and the current signal and obtain a second prediction result;

[0016] fusing the first prediction result and the second prediction result through a dynamic credibility weighted fusion mechanism to obtain a fused diagnosis result;

[0017] S5: Perform real-time decision-making and hierarchical feedback control based on the fusion diagnosis results, determine the fault type as continuous jamming, temporary jamming or normal state according to the preset threshold, and trigger the preset control strategy corresponding to the determined fault type.

[0018] To further improve the above technical solution, the multimodal sensor includes: a three-axis vibration sensor for collecting vibration signals of motor bearings or rotor shaft ends; a Hall current sensor for monitoring the three-phase current waveform of the motor; and a temperature sensor for detecting the surface temperature of the outer ring of the motor bearing or the stator winding.

[0019] Furthermore, the preprocessing includes: using discrete wavelet transform to reduce noise on the current signal, and using a bandpass filter to retain a preset key frequency band; using adaptive wavelet noise reduction, zero-phase bandpass filtering and dynamic normalization processing on the vibration signal; and using an exponential moving average algorithm to smooth the temperature signal to extract the temperature rise trend.

[0020] The system for implementing the above-mentioned real-time motor fault diagnosis method based on LSTM and random forest includes:

[0021] A multimodal sensor, including a triaxial vibration sensor, a Hall current sensor, and a temperature sensor, configured to synchronously collect vibration signals, current signals, and temperature signals of the motor;

[0022] a signal preprocessing module configured to perform noise reduction, filtering, and smoothing processing on the vibration signal, current signal, and temperature signal to generate a preprocessed signal;

[0023] a feature engineering module configured to extract time domain features, frequency domain features, time-frequency domain features, and cross-modal correlation features from the preprocessed signal to obtain an initial feature set, and process the initial feature set through a hybrid screening strategy to generate an optimal feature subset;

[0024] A hybrid fault prediction model includes: a random forest model configured to input the optimal feature subset, screen highly discriminative features, and output a first prediction result; a long short-term memory time series network model, including a vibration branch and a current branch, wherein: the vibration branch is configured to input the FFT spectrum frame features of the vibration signal in the initial feature set, using time series slices with a window length of 30; the current branch is configured to input the time domain features and frequency domain features of the current signal in the initial feature set; the vibration branch and the current branch are synchronized through a hardware trigger mechanism, respectively used to capture the dynamic evolution laws of the vibration signal and the current signal and output a second prediction result;

[0025] a dynamic weighted fusion submodule, configured to fuse the first prediction result and the second prediction result using a dynamic credibility weighted fusion mechanism to obtain a fused diagnosis result;

[0026] The decision-making and hierarchical feedback control module is configured to perform hierarchical feedback control based on the fused diagnostic result, including: if the score of the fused diagnostic result is greater than or equal to the first threshold, it is determined to be a continuous jam and power reduction protection is triggered; if the score of the fused diagnostic result is between the second threshold and the first threshold, it is determined to be a temporary jam and the reverse torque clearing program is triggered; if the temporary jam is not resolved within a preset time after the reverse torque clearing program is executed, the fault state is upgraded to a continuous fault warning.

[0027] Beneficial effects: Compared with the prior art, the advantages of the present invention are:

[0028] High diagnostic accuracy: By fusing multimodal information of vibration, current, and temperature, and combining the advantages of random forest in processing static features and LSTM in processing time series dynamic features, as well as a dynamic weighted fusion strategy, the accuracy of fault diagnosis is significantly improved, especially the ability to distinguish between temporary and persistent jams.

[0029] Comprehensive information: It covers the motor's multi-dimensional state information, including mechanical, electrical, and thermodynamics, and innovatively utilizes cross-modal correlation features to provide a more comprehensive and in-depth understanding of fault mechanisms.

[0030] Strong real-time performance: The entire diagnostic process is designed to run in real time, which can promptly capture the moment a fault occurs and respond quickly.

[0031] Active protection and intelligent response: It implements hierarchical feedback control based on diagnostic results, and can adopt the optimal response strategy for different types of jams (such as reverse torque clearance, power reduction protection, and upgraded warning). It realizes an intelligent upgrade from passive alarm to active intervention, effectively reducing the risk of equipment damage and maintenance costs.

[0032] Good robustness: Multi-source information fusion and hybrid model structure improve the robustness of the diagnostic system to single sensor noise or interference, making it more suitable for complex and changeable agricultural operation environments.

[0033] Wide applicability: Although the method is based on a lawn mower motor, the core idea and technical framework of this method can be extended to the fault diagnosis and health management of other types of motors or rotating machinery. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of the hybrid model architecture of the present invention.

[0035] Figure 2 This is a logic flow chart of the hierarchical warning system of the present invention. DETAILED DESCRIPTION

[0036] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.

[0037] Example 1: The real-time diagnosis method for motor fault types based on LSTM and random forest provided in this embodiment mainly includes the following steps:

[0038] S1. Collect vibration signals, current signals, and temperature signals of the mowing motor through a multimodal sensor to obtain real-time working parameters;

[0039] S2. Preprocessing the vibration signal, current signal and temperature signal, including noise reduction, filtering and smoothing;

[0040] S3. Perform feature engineering on the preprocessed signal to extract time domain, frequency domain, time-frequency domain and cross-modal correlation features, and determine the optimal feature subset through a hybrid screening strategy;

[0041] S4. Build a hybrid fault prediction model, including a random forest model and an LSTM time series network. Through a dynamic credibility weighted fusion mechanism, the output results of the random forest and LSTM are collaboratively diagnosed.

[0042] S5. Make real-time decisions and implement hierarchical feedback control based on the fusion results: If the diagnostic score is ≥0.8, it is judged as a continuous jam and the power reduction protection strategy is triggered; if the diagnostic score is between 0.6 and 0.8, it is judged as a temporary jam and the reverse torque clearing procedure is executed; if the temporary jam is not resolved, it is upgraded to a continuous fault warning.

[0043] Step S1: Install triaxial high-frequency vibration sensors on the mowing motor bearing housing and rotor shaft end, with a sampling frequency of 10kHz, to capture X / Y / Z triaxial vibration signals, focusing on mechanical anomalies in the 5-2000Hz frequency band. Connect a Hall effect current sensor in series to the motor's three-phase power lines, sampling at 1kHz, to record current waveforms and transient overload characteristics (such as peak current and harmonic distortion). Mount temperature sensors on the motor bearing outer ring and stator winding surface, sampling at 10Hz, to monitor the rate of temperature rise. A hardware trigger mechanism is employed, with the primary clock source using GPS PPS pulses per second (PPS), and the backup clock source using the IEEE1588 PTP protocol as a backup in the event of GPS failure. This creates a dual-redundant clock source. A hardware phase-locked loop (PLL) synchronizes the clocks of each node, eliminating clock drift differences between devices and ensuring timing alignment of vibration, current, and temperature signals, with a timestamp error of ≤1ms. The primary clock source uses GPS PPS pulses per second (PPS), leveraging its high accuracy and global coverage. The backup clock source uses the IEEE 1588PTP protocol to synchronize time over the local area network as a backup in the event of GPS failure. It uses a multi-source clock synchronization module (such as OroliaSecureSync, Microchip SyncServer), supports GPS and PTP dual inputs, and configures a high-stability crystal oscillator (such as OCXO or rubidium clock) as the local clock to smooth time jitter during the switching process.

[0044] Step S2: When the lawn mower is operating, the current signal may be affected by electromagnetic interference, sensor drift, etc. Gaussian noise can simulate such random disturbances. At the same time, in order to improve the robustness of the model to sensor noise and environmental interference, Gaussian noise with a mean of 0 and a standard deviation of 0.05A is added to the current signal during the training phase. The original current signal sequence I raw =[I 1 ,I 2 …I n ], add independent Gaussian noise ∈ i :

[0045]

[0046] Among them, ∈ i is the independent Gaussian noise added to the i-th sampling point.

[0047] The current signal is decomposed into five layers using the db8 wavelet base, and soft threshold noise reduction is performed based on the robust median absolute deviation threshold estimation. A fourth-order Butterworth filter (10-500Hz) is designed to retain the core frequency band of load fluctuations. The wavelet decomposition formula is:

[0048] {cA5,cD5,cD4,cD3,cD2,cD1}=DWT(s(t),db8,5)

[0049] Where s(t) is the original time domain signal; cA5 is the approximation coefficient of the 5th layer; cD1~cD5 are the detail coefficients of the 1st to 5th layers.

[0050] The global threshold λ formula is:

[0051] Where N is the signal length.

[0052] Soft threshold function for:

[0053] Among them, sign() is the sign function.

[0054] The vibration signal's wavelet decomposition layer number is dynamically selected based on the noise level. The noise level is quantified using the energy of the high-frequency detail coefficients. After wavelet decomposition of the vibration signal, the energy of the high-frequency detail coefficients at each layer is calculated. Noise energy is primarily concentrated in the high-frequency subband, so the noise level can be quantified by analyzing the energy distribution of the detail coefficients.

[0055] The noise level is quantified by the energy of the high-frequency detail coefficient, and the number of decomposition layers is determined according to the energy attenuation rate, ensuring that the denoising effect is maximized while retaining effective mechanical features.

[0056] The noise energy of the kth layer can be expressed as:

[0057]

[0058] like If it is significantly higher than the preset threshold of 0.2, it is considered that the current noise level is high.

[0059] Perform 8-layer wavelet decomposition on the vibration signal to generate the approximate coefficients of each layer (cA k ) and detail coefficient (cD k ).

[0060] From the 1st layer to the 8th layer, the energy attenuation rate ΔE of the detail coefficient is calculated layer by layer k :

[0061] If ΔE k<γ=0.3, it is considered that the decomposition below this layer has limited contribution to denoising, and k-1 layers are selected as the optimal decomposition layer. After selecting the number of decomposition layers, the retained detail coefficients are thresholded (such as soft threshold) and the signal is reconstructed. At the same time, a zero-phase FIR filter with a 100-2000Hz bandpass is combined to achieve adaptive noise reduction, calculate the peak-to-peak value of the vibration signal according to the sliding window, and implement amplitude adaptive calibration. Sliding peak-to-peak value Calculation formula:

[0062]

[0063] in, is the peak-to-peak value (the difference between the maximum and minimum values) at time point t; w is the sliding window length (time interval length); x(n) is the amplitude of the signal at time point n; [tw,t] is the time window interval from tw to t.

[0064] Dynamic normalization processing:

[0065]

[0066] Where x(n) is the original signal; is the dynamic peak-to-peak value at the current moment t; is the peak-to-peak value at the previous moment t-1; μ is the weight coefficient.

[0067] The temperature signal is smoothed and the EMA algorithm is applied to the original temperature data to eliminate the instantaneous disturbance of the ambient temperature and extract the temperature rise trend. The exponential moving average algorithm processing formula is:

[0068] T smooth (t) = α·T raw (t)+(1-α)·T smooth (t-1)

[0069] Among them, T raw (t) is the original temperature sampling value at the current moment, T smooth (t)) is the smoothed temperature value at the current moment, and α is the smoothing factor, which is 0.2.

[0070] Step S3: Perform time-frequency analysis on the vibration signal. Time-domain features such as the root mean square (RMS) of the vibration signal, peak factor, and pulse index reflect the amplitude characteristics of the impact event. Frequency-domain features such as FFT extract the resonance frequency and spectral entropy, and combine with wavelet packet decomposition to analyze the energy proportion of the high-frequency band. Perform time-frequency analysis on the current signal, including time-domain features such as current distortion rate and dynamic current slope, and frequency-domain features such as characteristic harmonic amplitude and instantaneous frequency tracking. Perform feature engineering on the temperature signal, such as sliding mean, temperature variance, and temperature rise rate. Cross-modal correlation features such as the correlation coefficient between the current harmonic distortion rate and the vibration RMS, and the time-delay correlation between the bearing temperature gradient and the current peak, are analyzed.

[0071] Low-variance features were eliminated based on variance analysis. A relative threshold method was used to dynamically identify low-variance features based on their variance distribution. The top 80% of features were retained, and the top 20 key features were retained using random forest importance ranking. Recursive feature elimination combined with LSTM time-series sensitivity analysis determined the final 10-dimensional optimal feature subset: current rise time, vibration peak-to-peak value, temperature rise rate, current pulse energy, third harmonic distortion value, bearing fault frequency-energy ratio, low-frequency temperature fluctuation energy, vibration-current mutual information, temperature-vibration trend correlation coefficient, and lag cointegration between current peak and temperature rise. The vibration branch LSTM input dimension is FFT spectrum frame features, using time series slices with a window length of 30. The current branch LSTM input dimension includes both time and frequency domain features. A sliding window mechanism was used to unify the time step, and a hardware trigger mechanism ensured data synchronization error of the two branches was ≤1ms.

[0072] Step S4: Configure a random forest model with 150 decision trees, the Gini coefficient as the splitting criterion, a weight of 2.0 for temporary lag samples to mitigate class imbalance, and a bootstrap sampling ratio of 70%. The LSTM network structure consists of two layers of LSTM units with 64 hidden nodes, followed by a dropout layer and a fully connected layer. The Adam optimizer is used, with a weighted cross entropy loss function, a weight of 1.5 for temporary lag, and a weight of 2.0 for persistent lag.

[0073] The random forest model uses static features as input to screen for highly discriminative features and perform preliminary classification. The random forest model measures feature importance using the Gini coefficient. When splitting, the decision tree selects the feature that reduces the Gini impurity the most. Feature importance is determined by calculating the average contribution of that feature across all trees. The formula is:

[0074]

[0075] Among them, Importance i is the importance score of the i-th feature. The higher the score, the greater the contribution of the feature to the classification decision of the random forest model and the stronger the discriminative ability. trees is the total number of decision trees in the random forest; s represents a split node in the decision tree t; S i (t) represents the set of all nodes in decision tree t that use feature i as the splitting criterion; ΔGini(s) represents the reduction in Gini impurity at node s due to the use of feature i for splitting.

[0076] The LSTM timing network model inputs time series data to capture the dynamic evolution of the vibration spectrum and current waveform; the forget gate and input gate of the LSTM dynamically adjust the memory unit state, effectively modeling the long-term trends and short-term fluctuations of the vibration spectrum and current waveform, and realizing cross-modal dynamic association learning through hidden state fusion.

[0077] The two-layer LSTM unit represents the depth of the entire LSTM timing network model. The vibration branch and the current branch are two independent processing paths within the same model, each containing two layers of LSTM units. The LSTM layers of the two branches are trained independently, but share the model framework. The hidden states (both 64-dimensional) output by the vibration and current branches at each time step are concatenated along the feature dimension to form a 128-dimensional fused hidden state.

[0078] Weighted cross entropy loss function:

[0079]

[0080] Where C is the number of categories; w c is the weight of the cth class, used to balance the class distribution; y i,c is the true label of the i-th sample; z i,c is the output logits of the model for the i-th sample in the c-th category.

[0081] Category weight:

[0082]

[0083] Calculate the weight of dynamic confidence, including the fusion weight W of the random forest model RF And the fusion weight W of the LSTM model LSTM :

[0084]

[0085] Among them, P RF is the probability output by the random forest model, P LSTM is the probability output by the LSTM model, and the maximum confidence C of the random forest model RF =max(P RF ), the maximum confidence C of the LSTM model LSTM =max(P LSTM ).

[0086] Perform dynamic confidence weighted fusion, and the final diagnosis result S is determined by the weighted score, and the formula is:

[0087] S=W RF ×P RF +W LSTM ×P LSTM.

[0088] Step S5: Perform real-time decision-making and hierarchical feedback control based on the fusion diagnosis results, determine the fault type as continuous jamming, temporary jamming or normal state according to the preset threshold, and trigger the preset and corresponding control strategy for the determined fault type.

[0089] If S≥0.8, the device will immediately enter a state of continuous freezing. In this case, the system will quickly start the power reduction process according to the preset emergency strategy. With the help of the power regulation module, the device power will be adjusted to 30% of the rated value as quickly as possible. At the same time, the system will send an encrypted specific fault code to the associated monitoring terminal and operation and maintenance platform through the 802.11ax (WiFi 6) communication protocol, providing operation and maintenance personnel with detailed equipment abnormality information to help them quickly locate the problem.

[0090] If S is in the range of 0.6≤S<0.8, the system will determine it as a temporary jam. In this case, the system will automatically trigger the reverse torque clearing program. This program relies on the motor drive control unit to drive the equipment at 120% of the rated torque for 50 milliseconds. It aims to use the effect of reverse torque to break the jamming deadlock caused by mechanical friction, uneven load and other factors. During the subsequent 5-second observation period, the system will continue to use status monitoring sensors and data analysis models to conduct real-time assessment of the equipment's operating status. Once it is found that the temporary jamming phenomenon has not been effectively resolved, the system will upgrade the fault status from a temporary jam to a continuous fault warning based on the fault escalation mechanism. This warning information will be promptly conveyed to relevant technicians and managers through various channels such as sound and light alarms, SMS push, and email notifications, so that maintenance plans can be formulated in advance to prevent further deterioration of motor equipment failures.

[0091] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A real-time motor fault diagnosis method based on LSTM and random forest, characterized by: The following steps are involved: S1: Synchronously collect the vibration signal, current signal and temperature signal of the motor through multi-modal sensors; S2: Preprocessing the vibration signal, current signal and temperature signal to generate a preprocessing signal; S3: performing feature engineering on the preprocessed signal to obtain an initial feature set, and processing the initial feature set using a hybrid screening strategy to generate an optimal feature subset; S4: Build a hybrid fault prediction model that includes a random forest model and an LSTM time series network model, where: The random forest model inputs the optimal feature subset and obtains a first prediction result based on the highly discriminative features therein; The LSTM time series network model inputs the time series features in the initial feature set to capture the dynamic evolution law of the vibration signal and the current signal and obtain a second prediction result; fusing the first prediction result and the second prediction result through a dynamic credibility weighted fusion mechanism to obtain a fused diagnosis result; S5: Perform real-time decision-making and hierarchical feedback control based on the fusion diagnosis result, determine the fault type according to a preset threshold, and trigger a preset control strategy for the determined fault type.

2. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that, The multimodal sensor comprises: Three-axis vibration sensor for collecting vibration signals of motor bearings or rotor shaft ends; Hall current sensors for monitoring the three-phase current waveforms of the motor; and Temperature sensor used to detect the surface temperature of the motor bearing outer ring or stator winding.

3. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that, The pretreatment includes: The current signal is subjected to noise reduction using discrete wavelet transform, and a bandpass filter is used to retain a preset key frequency band; Adaptive wavelet noise reduction, zero-phase bandpass filtering and dynamic normalization are applied to the vibration signal; The temperature signal is smoothed using an exponential moving average algorithm to extract the temperature rise trend.

4. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that, The feature engineering includes: extracting time domain features, frequency domain features, and time-frequency domain features from the vibration signal, extracting time domain features and frequency domain features from the current signal, extracting time domain features from the temperature signal, and extracting cross-modal correlation features between the vibration signal, the current signal, and the temperature signal; The hybrid screening strategy includes: eliminating low-variance features from the initial feature set through variance analysis; using a random forest model to rank the remaining features by importance and retaining a preset number of top-ranked key features; combining recursive feature elimination with LSTM timing sensitivity analysis to determine the final optimal static feature subset from the key features.

5. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that, The cross-modal association feature includes at least one of the following: Correlation coefficient between current harmonic distortion rate and vibration root mean square value; Time-delay correlation between bearing temperature gradient and current peak; Mutual information between vibration signal and current signal; Trend correlation coefficient between temperature signal and vibration signal; Cointegration of current peak and temperature rise lag.

6. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 4 is characterized in that, Based on the random forest feature importance ranking, the top 20 key features are retained; Generate a 10-dimensional optimal feature subset through recursive feature elimination combined with LSTM timing sensitivity analysis; The optimal feature subset includes current rise time, vibration peak-to-peak value, temperature rise rate, current pulse energy third harmonic distortion value, bearing fault frequency energy ratio, low-frequency temperature fluctuation energy, vibration current mutual information, temperature-vibration trend correlation coefficient and current peak and temperature rise lag cointegration.

7. The real-time motor fault diagnosis method based on LSTM and random forest according to claim 4 is characterized in that: The LSTM timing network model includes a vibration branch and a current branch; The vibration branch inputs the FFT spectrum frame features of the vibration signal, using time series slicing with a window length of 30; The current branch is input into a time series slice of the time domain feature and / or frequency domain feature extracted from the current signal; The vibration branch and the current branch splice the processed hidden states, and calculate the second prediction result based on the spliced ​​hidden states.

8. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that: The random forest model is configured with 150 decision trees, uses the Gini coefficient as the splitting criterion, and sets the weight of temporary stuck samples to 2.0; The LSTM temporal network model includes two layers of LSTM units, the number of hidden layer nodes is 64, and a weighted cross entropy loss function is adopted; In the dynamic credibility weighted fusion mechanism, the fusion calculation formula is: S=W RF ×P RF +W LSTM ×P LSTM Among them, C RF and C LSTM They are the maximum confidence levels of the random forest model and the LSTM temporal network model, respectively.

9. The motor fault real-time diagnosis method based on LSTM and random forest according to claim 1 is characterized in that: The hierarchical feedback control includes: if the fusion diagnosis result is greater than or equal to the first threshold, it is determined to be a continuous jam and the power reduction protection is triggered; if the fusion diagnosis result is between the second threshold and the first threshold, it is determined to be a temporary jam and the reverse torque clearing program is triggered; if the temporary jam is not relieved within a preset time after executing the reverse torque clearing program, the fault state is upgraded to a continuous fault warning.

10. A system for implementing the real-time motor fault diagnosis method based on LSTM and random forest according to claim 1, characterized in that: include: A multimodal sensor, including a triaxial vibration sensor, a Hall current sensor, and a temperature sensor, configured to synchronously collect vibration signals, current signals, and temperature signals of the motor; a signal preprocessing module configured to perform noise reduction, filtering, and smoothing processing on the vibration signal, current signal, and temperature signal to generate a preprocessed signal; a feature engineering module configured to extract time domain features, frequency domain features, time-frequency domain features, and cross-modal correlation features from the preprocessed signal to obtain an initial feature set, and process the initial feature set through a hybrid screening strategy to generate an optimal feature subset; Hybrid fault prediction model, including: A random forest model is configured to input the optimal feature subset, filter highly discriminative features and output a first prediction result; A long short-term memory time series network model includes a vibration branch and a current branch, wherein: the vibration branch is configured to input the FFT spectrum frame features of the vibration signal in the initial feature set, using time series slices with a window length of 30; the current branch is configured to input the time domain features and frequency domain features of the current signal in the initial feature set; the vibration branch and the current branch are synchronized through a hardware trigger mechanism, respectively used to capture the dynamic evolution laws of the vibration signal and the current signal and output a second prediction result; a dynamic weighted fusion submodule, configured to fuse the first prediction result and the second prediction result using a dynamic credibility weighted fusion mechanism to obtain a fused diagnosis result; The decision-making and hierarchical feedback control module is configured to perform hierarchical feedback control based on the fused diagnostic result, including: if the score of the fused diagnostic result is greater than or equal to the first threshold, it is determined to be a continuous jam and power reduction protection is triggered; if the score of the fused diagnostic result is between the second threshold and the first threshold, it is determined to be a temporary jam and the reverse torque clearing program is triggered; if the temporary jam is not resolved within a preset time after the reverse torque clearing program is executed, the fault state is upgraded to a continuous fault warning.

Images