An artificial intelligence-based motor operation fault detection method

Abstract

The application relates to the technical field of motor fault detection, and discloses a motor operation fault detection method based on artificial intelligence, which comprises the following steps: S1, vibration signals, current signals and temperature signals in the motor operation process are acquired, a unified clock reference is introduced at a collection end, and a time delay estimation and a dynamic time warping algorithm are used at a receiving end to perform timestamp correction on multi-source data, and calibrated multi-source alignment signals are output; S2, the multi-source alignment signals are respectively subjected to short-time Fourier transform and continuous wavelet transform processing. Through the unified clock reference and the multi-source signal timestamp, sensor sampling deviation is corrected and eliminated, so that the internal relationship between different physical quantities in the motor operation process can be more accurately captured, a reliable time sequence benchmark is provided for subsequent multi-source feature fusion and fault mode recognition, and the recognition precision of early weak faults and complex compound faults is improved.

Description

Technical Field

[0001] This invention relates to the field of motor fault detection technology, specifically to a method for detecting motor operation faults based on artificial intelligence. Background Technology

[0002] Traditional motor fault diagnosis mainly relies on expert experience, fixed threshold alarms, or signal processing techniques such as spectrum analysis. These methods often have limited ability to identify complex fault modes, are difficult to handle nonlinear and non-stationary signals, and depend on manual feature extraction, which is inefficient and prone to missed or false diagnoses. Artificial intelligence technology, especially deep learning and machine learning, can effectively capture early and weak fault signs by automatically learning deep features and hidden patterns in massive motor operation data, and achieve end-to-end fault classification and even predictive maintenance.

[0003] In actual industrial deployments, multi-sensor collaborative sensing has become the mainstream solution to more comprehensively and accurately depict the operating status of motors. By synchronously collecting multi-dimensional heterogeneous signals such as vibration, current, and temperature, a more complete equipment status profile can be constructed, enabling a comprehensive and three-dimensional depiction of the motor's operating status.

[0004] However, in existing technologies, when multiple sensors collect signals such as motor vibration, current, and temperature, they often use their own independent clock sources for sampling. The time bases of different sensors are not unified, and no special timestamp correction processing is performed on the collected multi-source data. This results in misalignment of different types of signals on the time axis, making it impossible to establish a corresponding relationship between instantaneous changes in vibration, fluctuations in current, and rises and falls in temperature at accurate time points. This causes distortion in the correlation between features, making it difficult to accurately capture the intrinsic relationship between different physical quantities during motor operation. Consequently, it affects the effect of subsequent multi-source feature fusion and the accuracy of fault mode recognition, especially the identification of early weak faults and complex composite faults, which are prone to deviation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an artificial intelligence-based method for detecting motor operation faults. This method solves the problem that existing technologies suffer from misalignment of different types of signals on the time axis, making it difficult to accurately capture the intrinsic relationships between different physical quantities during motor operation, which in turn affects the effectiveness of subsequent multi-source feature fusion and the accuracy of fault mode recognition.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting motor operation faults based on artificial intelligence, comprising the following steps: S1. Acquire vibration, current and temperature signals during motor operation. Introduce a unified clock reference at the acquisition end. At the receiving end, use time delay estimation and dynamic time warping algorithms to perform timestamp correction on the multi-source data and output a calibrated multi-source aligned signal. S2. Perform short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal to obtain time-frequency feature image. Extract robust single-mode features based on feature space enhancement strategy, and perform feature fusion using cross-modal time attention mechanism to output fused feature vector. S3. During model deployment, the statistics of the input data distribution are calculated based on the sliding window, and the distribution change of the fused feature vector is monitored using the drift detection algorithm. When a significant shift is detected, the model adaptation process is triggered. S4. In the adaptive process, the fused feature vector is updated online using the incremental learning method, and the continuity of the original feature distribution is maintained by knowledge distillation. Only the network modules related to drift are locally fine-tuned, and the updated classification model is output. S5. Use the updated classification model to identify faults in the motor's operating status, calculate the health index based on historical operating data, generate early warning information based on the health index, and achieve cloud-edge collaborative updates, outputting fault category and trend assessment results.

[0007] Preferably, in step S1, the time stamp correction of multi-source data using time delay estimation and dynamic time warping algorithms specifically includes: Vibration and current signals are acquired simultaneously using a multi-channel synchronous acquisition card, while temperature signals are acquired through a low-frequency sampling module. At the acquisition end, a unified clock source is used for timing, and at the receiving end, the signal delay is calculated based on cross-correlation, and dynamic time warping is used to achieve sequence alignment of cross-modal signals.

[0008] Preferably, in step S2, performing short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal and then fusing its features specifically includes: Short-time Fourier transforms were performed on the vibration signal and the current signal to generate time-frequency graphs, and sliding window statistical processing was performed on the temperature signal to generate trend features. The time-frequency graph is input into a convolutional neural network to extract local feature vectors, and the trend features are input into a long short-term memory network to extract time-series feature vectors. Feature fusion is performed on local feature vectors and temporal feature vectors based on cross-modal attention mechanism.

[0009] Preferably, the feature fusion using a cross-modal temporal attention mechanism specifically includes: The vibration signal feature vector is used as the query vector, the current signal feature vector is used as the key vector, and the temperature signal feature vector is used as the value vector. The cosine similarity between the query vector and the key vector is calculated by using a sliding time window, and a time alignment similarity matrix is ​​generated to generate time attention weights. The value vector is weighted and aggregated based on the time attention weights to output a cross-modal fusion feature vector.

[0010] Preferably, the feature space enhancement strategy in S2 specifically includes: Random band masking and phase perturbation are performed on the time-frequency plots of vibration and current signals; Gaussian noise perturbation is injected into the temperature trend characteristics; Regularized samples based on adversarial training to generate multimodal features; The vibration and current signals are fused using cross-modal feature combination technology, and the fused features are weighted and fused using temperature trend features.

[0011] Preferably, in step S3, monitoring the distribution changes of the fused feature vector using a drift detection algorithm specifically includes: A sliding time window is constructed to update the probability distribution of the fused feature vector, and the Kullback-Leibler divergence between the distributions is calculated. When the divergence exceeds a preset threshold, it is determined to be a concept drift and an adaptive process is triggered.

[0012] Preferably, the incremental learning online update in S4 specifically includes: The new data cached after triggering drift detection is input into the model for micro-batch training; During training, the convolutional feature extraction layer is frozen, and only the parameters of the cross-modal attention fusion layer and the classifier layer are updated; The output distribution of the historical model is introduced as a soft label constraint through knowledge distillation.

[0013] Preferably, updating only the parameters of the cross-modal attention fusion layer and the classifier layer specifically includes: Based on feature drift contribution analysis, the priority of the network modules to be updated is determined; A hierarchical learning rate adjustment is performed on the cross-modal attention fusion layer, wherein the learning rate of the temporal attention weights is set to 1.5-2 times that of the spatial attention weights; A sparse reparameterization strategy is adopted for the classifier layer. Based on the gradient contribution value, highly relevant weight parameters are selected for selective updates, while the rest are frozen.

[0014] Preferably, the generation of early warning information based on the health index and the implementation of cloud-edge collaborative updates in S5 specifically include: When the health index is continuously below the set threshold, a lifespan trend prediction curve is output and an early warning message is generated. The warning information is pushed to the operation and maintenance terminal in real time via edge devices; The system stores fault identification results, health indices, and trend prediction curves in the cloud and periodically optimizes model parameters before sending them to edge devices.

[0015] Preferably, an artificial intelligence-based motor operation fault detection system includes: The multi-source signal acquisition and alignment module is used to acquire motor vibration, current and temperature signals, and realizes multi-source data timestamp correction based on a unified clock reference; The time-frequency feature extraction and fusion module is connected to the multi-source signal acquisition and alignment module and is used to generate time-frequency feature images and perform cross-modal feature fusion and combination. The concept drift detection module, connected to the time-frequency feature extraction and fusion module, is used to monitor changes in the distribution of the fused feature vector and trigger model updates. The incremental learning and update module, connected to the concept drift detection module, is used for parameter optimization and outputting the updated classification model; The fault diagnosis and cloud-edge collaboration module is connected to the incremental learning and update module, and is used to generate health index warnings and realize cloud-edge parameter collaborative optimization.

[0016] This invention provides an artificial intelligence-based method for detecting motor operation faults. It has the following beneficial effects: 1. This invention eliminates sensor sampling deviation by using a unified clock reference and multi-source signal timestamp correction. The unified clock reference and multi-source signal timestamp correction effectively eliminate sensor sampling deviation, solve the problem of timing misalignment of heterogeneous signals such as vibration, current, and temperature, and avoid feature correlation distortion caused by timing misalignment. This allows for more accurate capture of the intrinsic relationship between different physical quantities during motor operation, providing a reliable timing reference for subsequent multi-source feature fusion and fault mode identification, and improving the identification accuracy of early weak faults and complex composite faults.

[0017] 2. This invention integrates the time-frequency and trend features of vibration current and temperature through a cross-modal time attention mechanism to construct a comprehensive feature vector. It combines a sliding window and KL divergence to monitor concept drift in real time, triggering incremental learning of the local fine-tuning model to ensure that high-precision fault identification capability is maintained continuously under scenarios such as changes in motor operating conditions and equipment aging.

[0018] 3. This invention uses short-time Fourier transform to extract the time-frequency graph of vibration current signal, generates temperature trend features through sliding window statistics, uses convolutional neural network to capture local fault texture, uses long short-term memory network to model long-term temperature dependence, and then uses attention weights to aggregate with vibration as query vector, current as key vector and temperature as value vector, to achieve deep correlation of mechanical, electrical and thermal features in the time dimension.

[0019] 4. This invention pushes health index-triggered early warning information to the operation and maintenance terminal in real time based on edge devices, and simultaneously stores fault records and trend curves in the cloud. It periodically aggregates multi-motor data to optimize model parameters and sends them down to the edge, forming a cloud-edge collaborative closed loop and reducing data transmission bandwidth requirements.

[0020] 5. This invention simulates signal distortion by performing random frequency band masking phase perturbation on the time-frequency map, injects adaptive Gaussian noise into the temperature features, generates regularized samples to expand feature diversity through adversarial training, and then constrains the incremental learning process through knowledge distillation, freezes the basic feature extractor of the convolutional layer, and only updates the attention fusion layer and the sparse reparameterized classifier, thereby suppressing noise interference and retaining the ability to recognize historical fault patterns. Attached Figure Description

[0021] Figure 1 This is a flowchart of an artificial intelligence-based motor operation fault detection method according to the present invention; Figure 2 This is an architecture diagram of an artificial intelligence-based motor operation fault detection system according to the present invention; Figure 3 This is a flowchart of the multi-source signal acquisition and timestamp correction process of the present invention; Figure 4 This is a flowchart of the time-frequency feature extraction and cross-modal attention fusion process of the present invention; Figure 5 This is a flowchart of the fusion feature vector distribution drift detection process of the present invention; Figure 6 This is a flowchart of the incremental learning and knowledge distillation model update process of the present invention; Figure 7 This is a flowchart of the fault diagnosis and cloud-edge collaborative update process of the present invention; Figure 8 This is a health index lifespan trend prediction curve for the present invention. Detailed Implementation

[0022] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see the appendix Figure 1 - Appendix Figure 8 This invention provides an artificial intelligence-based method for detecting motor operation faults, including: S1. Acquire vibration, current and temperature signals during motor operation. Introduce a unified clock reference at the acquisition end. At the receiving end, use time delay estimation and dynamic time warping algorithms to perform timestamp correction on the multi-source data and output a calibrated multi-source aligned signal. Furthermore, in S1, the time stamp correction of multi-source data using time delay estimation and dynamic time warping algorithms specifically includes: Vibration and current signals are acquired simultaneously using a multi-channel synchronous acquisition card, while temperature signals are acquired through a low-frequency sampling module. At the acquisition end, a unified clock source is used for timing, and at the receiving end, the signal delay is calculated based on cross-correlation, and dynamic time warping is used to achieve sequence alignment of cross-modal signals.

[0024] Specifically, when acquiring vibration, current, and temperature signals during motor operation, vibration signals are collected using an accelerometer installed on the motor housing, and current signals are collected using a Hall current sensor connected in series in the motor power supply circuit. Both signals are connected to a multi-channel synchronous acquisition card, and simultaneous acquisition is achieved through the synchronous sampling function of the acquisition card. Temperature signals are collected using a thermocouple sensor attached to the motor windings, and data acquisition is completed through a low-frequency sampling module. At the acquisition end, both the multi-channel synchronous acquisition card and the low-frequency sampling module are connected to the same high-precision clock source for timing. A unified clock reference ensures that the time base of the start time of each signal acquisition is consistent, thus laying the foundation for subsequent timestamp correction. After acquiring the aforementioned multi-source signals, the receiving end uses the vibration signal as a reference signal for both the vibration and current signals. The current signal is the signal to be aligned. The time delay is determined by calculating the cross-correlation function, which is defined as: ; in For time variables, This is the time offset; when the function reaches its maximum value, the corresponding... The value represents the time delay between two signals. For temperature and vibration signals, the time delay of the temperature signal relative to the vibration signal is calculated using the cross-correlation method, with the vibration signal as a reference. Then, the dynamic time warping algorithm is used to align the time sequences of the cross-modal signals with time delays. Specifically, the time sequences of the temperature signal and the vibration signal are taken as two time sequences to be matched. A distance matrix is ​​constructed, where the matrix elements are the Euclidean distances between the signal values ​​at corresponding times in the two sequences. The path with the smallest cumulative distance in the distance matrix is ​​then found. The time sequence of the temperature signal is appropriately stretched or compressed through this path, so that the vibration signal, current signal, and temperature signal are accurately aligned on the same time axis. A calibrated multi-source aligned signal is output, in which each time point contains the corresponding vibration acceleration value, current value, and temperature value. This effectively eliminates time deviations caused by differences in sensor sampling frequencies and transmission paths, enabling different types of signals to form accurate correspondences in the time dimension. This provides a time consistency guarantee for the effective fusion of multi-source features, ensuring that transient changes in vibration signals, fluctuations in current signals, and trends in temperature signals can be analyzed collaboratively at the same time point, effectively avoiding feature correlation distortion caused by time misalignment.

[0025] Please see the appendix Figure 1 and attached Figure 4 S2. Perform short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal to obtain time-frequency feature images. Extract robust single-mode features based on feature space enhancement strategy, and use cross-modal time attention mechanism to perform feature fusion and output fused feature vector. Furthermore, in S2, the process of performing short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal and then fusing its features specifically includes: Short-time Fourier transforms were performed on the vibration signal and the current signal to generate time-frequency graphs, and sliding window statistical processing was performed on the temperature signal to generate trend features. The time-frequency graph is input into a convolutional neural network to extract local feature vectors, and the trend features are input into a long short-term memory network to extract time-series feature vectors. Feature fusion is performed on local feature vectors and temporal feature vectors based on cross-modal attention mechanism.

[0026] Furthermore, feature fusion using cross-modal temporal attention mechanisms specifically includes: The vibration signal feature vector is used as the query vector, the current signal feature vector is used as the key vector, and the temperature signal feature vector is used as the value vector. The cosine similarity between the query vector and the key vector is calculated by using a sliding time window, and a time alignment similarity matrix is ​​generated to generate time attention weights. The value vector is weighted and aggregated based on temporal attention weights to output a cross-modal fusion feature vector. Furthermore, the feature space enhancement strategies in S2 specifically include: Random band masking and phase perturbation are performed on the time-frequency plots of vibration and current signals; Gaussian noise perturbation is injected into the temperature trend characteristics; Regularized samples based on adversarial training to generate multimodal features; The fusion features of vibration and current signals are generated by cross-modal feature combination technology, and the fusion features are weighted and fused using temperature trend features.

[0027] Specifically, when processing multi-source aligned signals, the vibration signal is a time-domain sequence of acceleration calibrated by time, and the current signal is a time-domain sequence of current calibrated by time. Both are processed by short-time Fourier transform to generate a time-frequency diagram. The calculation formula for short-time Fourier transform is as follows: ; in These are time-domain signal samples of vibration or current. Hanning windows are used, with a length of... Based on the motor's rated speed, ensure coverage of at least one rotational cycle to capture the frequency components associated with rotor motion. For frequency index, As a time frame index, the time-domain signal is transformed into an amplitude distribution on the time and frequency plane to obtain a time-frequency graph. The temperature signal is a time-domain temperature sequence after time calibration. It is statistically processed by a sliding time window. The window length is set according to the typical time scale of temperature change. The mean, variance and linear fitting slope are calculated in each window to generate a feature sequence containing temperature trend information.

[0028] For the generated vibration and time-frequency graphs, random frequency bands are randomly selected on the frequency axis and their amplitudes are set to zero by random frequency band masking. A random offset is applied to the phase spectrum of the time-frequency graph by phase perturbation to simulate phase distortion in signal transmission. For temperature trend features, Gaussian noise with a mean of 0 and a variance determined based on the historical fluctuation range of the temperature signal is injected to enhance robustness to temperature measurement noise. A generative adversarial network is constructed through adversarial training. The generator generates samples consistent with the distribution of real multimodal features by using random noise as input. The discriminator optimizes parameters by distinguishing between real features and generated samples, and then generates regularized samples to enrich feature diversity. The local features of vibration and current are multiplied element-wise by cross-modal feature combination technology to form a preliminary fusion feature. The time series value of temperature trend features is used as a weight to adjust the preliminary fusion feature, thereby realizing the preliminary integration of cross-modal information. The enhanced vibration and time-frequency maps are input into a convolutional neural network. The convolutional layers extract local texture features from the time-frequency maps, such as amplitude variation regions of specific fault frequencies. These features can reflect the local manifestation of motor faults in the time-frequency domain. After dimensionality reduction by pooling layers, local feature vectors of vibration and current are obtained, effectively compressing redundant information while retaining key features. The enhanced temperature trend features are input into a long short-term memory network. The gating mechanism captures the long-term dependence of temperature over time, such as the slow rise or fall trend of temperature caused by changes in motor load. The output is a time-series feature vector of temperature, which enables the effective extraction of long-term trend information in the temperature signal and provides temporal dimension support for multimodal fusion. In cross-modal temporal attention mechanisms, the local feature vectors of vibration are used as query vectors. The local eigenvectors of the current serve as the key vectors. The time-series eigenvectors of temperature are used as value vectors. By using a sliding time window to cover multiple consecutive feature moments, the calculation within the window... and cosine similarity Generate a time-aligned similarity matrix, where For the query vector at the th Characteristics of time, For the key vector at the th The matrix, which quantifies the correlation strength between vibration and current characteristics at different times, is used to normalize each row of the matrix using the softmax function to generate time attention weights. These weights highlight time points where vibration and current characteristics are highly correlated. These weights are then applied to the value vector. Weighted aggregation is performed in the time dimension to obtain a cross-modal fusion feature vector. This vector simultaneously includes the local time-frequency features of vibration, the associated frequency features of current, and the temporal trend features of temperature. This allows the mechanical, electrical, and thermal aspects of motor faults to form an organic synergy at the feature level, providing a comprehensive and closely related feature representation for subsequent fault identification.

[0029] Please see the appendix Figure 1 and attached Figure 5 S3. During model deployment, the statistics of the input data distribution are calculated based on the sliding window, and the drift detection algorithm is used to monitor the distribution changes of the fused feature vector. When a significant shift is detected, the model adaptation process is triggered. Furthermore, in S3, the specific steps for monitoring the distribution changes of the fused feature vector using the drift detection algorithm include: A sliding time window is constructed to update the probability distribution of the fused feature vectors, and the Kullback-Leibler divergence between the distributions is calculated. When the divergence exceeds a preset threshold, it is judged as concept drift and an adaptive process is triggered.

[0030] Specifically, after obtaining the cross-modal fusion feature vector, the model deployment phase begins. At this stage, a sliding time window is constructed for the fusion feature vector to dynamically track its data distribution. The window length is determined based on the number of feature samples generated within a typical motor operating cycle, ensuring that each window covers enough samples to stably estimate the distribution characteristics. The window employs a rolling update mechanism; each time a new real-time fusion feature vector is added, the earliest sample entering the window is removed. This ensures that the window always reflects the recent feature distribution state. Based on the fusion feature vector within the sliding window, two distributions are calculated using the kernel density estimation method: one is the reference distribution under the verified drift-free state at the initial model deployment, denoted as... Its probability density function is ,in To fuse feature vectors, another is the feature distribution updated in real time within the current sliding window, denoted as... Its probability density function is The difference between the two distributions is quantified by calculating the Kullback-Leibler divergence, using the following formula: The magnitude of the divergence value directly reflects the degree of deviation between the current feature distribution and the reference distribution. The preset threshold is determined based on the statistical range of KL divergence of the model on historical drift-free running data. When the divergence value calculated in real time exceeds the threshold, it is determined that concept drift has occurred, and the model adaptive process is triggered at this time. This allows for timely capture of feature distribution shifts caused by changes in motor operating conditions, equipment aging, or external environmental disturbances, ensuring that the model can initiate an update mechanism when significant changes occur in the feature distribution. This forms a coherent closed loop with the previous feature extraction and fusion processes, maintaining the continuous effectiveness of motor fault detection.

[0031] Please see the appendix Figure 1 and attached Figure 6 S4. In the adaptive process, the incremental learning method is used to update the fused feature vector online, and the continuity of the original feature distribution is maintained by knowledge distillation. Only the network modules related to drift are locally fine-tuned, and the updated classification model is output. Furthermore, the incremental learning online updates in S4 specifically include: The new data cached after triggering drift detection is input into the model for micro-batch training; During training, the convolutional feature extraction layer is frozen, and only the parameters of the cross-modal attention fusion layer and the classifier layer are updated; The output distribution of the historical model is introduced as a soft label constraint through knowledge distillation.

[0032] Furthermore, updating only the parameters of the cross-modal attention fusion layer and the classifier layer specifically includes: Based on feature drift contribution analysis, the priority of the network modules to be updated is determined; A hierarchical learning rate adjustment is performed on the cross-modal attention fusion layer, wherein the learning rate of the temporal attention weights is set to 1.5-2 times that of the spatial attention weights; A sparse reparameterization strategy is adopted for the classifier layer. Based on the gradient contribution value, highly relevant weight parameters are selected for selective updates, while the rest are frozen.

[0033] Specifically, after triggering the model adaptation process, the newly cached data during drift detection—namely, the recently generated fused feature vectors and their corresponding fault labels—is input into the model for micro-batch training. Parameters are updated iteratively through small batches to adapt to the new data distribution. Simultaneously, the parameters of the convolutional feature extraction layer are frozen during training because the basic time-frequency features of vibration and current signals extracted by this layer are relatively stable and less affected by drift. Only the parameters of the cross-modal attention fusion layer and the classifier layer are updated to specifically adjust the feature fusion method and fault classification boundary. Knowledge distillation technology is used to introduce the output distribution of the historical model as a soft label constraint. Specifically, the output probability distribution of the historical model to the input feature vector is used as a soft label, which, together with the true label of the new data (i.e., the hard label), constructs the loss function. The loss function is defined as follows: ,in For cross-entropy loss, Labels for actual motor fault categories. This represents the predicted probability distribution of the current model. For KL divergence loss, This represents the output probability distribution of the historical model for the same input. To balance the weighting coefficients of the two, this method allows the new model to retain the fault identification knowledge accumulated by the historical model while learning new feature association patterns, maintain the continuity of the original feature distribution, and avoid the decline in the ability to identify historical fault patterns due to updates. When updating the parameters of the cross-modal attention fusion layer and the classifier layer, the priority of the modules to be updated is first determined based on feature drift contribution analysis. By calculating the correlation between the output features of each module and the drift features, it is found that the cross-modal attention fusion layer is more sensitive to drift because it directly processes the dynamic correlation of multi-source features. Therefore, it is listed as the priority update target. Layered learning rate adjustment is performed on the cross-modal attention fusion layer, and the learning rate of the temporal attention weight is set higher than that of the spatial attention weight. This is because drift is often more reflected in the change of feature correlation in the time dimension, such as the change in the temporal synchronization of vibration peak and current fluctuation. A higher learning rate can make it more flexible in capturing new temporal correlation patterns. For the classifier layer, a different learning rate is used. By employing a sparse reparameterization strategy, the gradient contribution value of each parameter in the current training batch is calculated (i.e., the product of the parameter gradient and the parameter value). Parameters with larger gradient contribution values ​​are selected and selectively updated, as these parameters are highly correlated with the changes in the fault classification boundary caused by the current drift. The remaining parameters are kept frozen. This approach adapts to drift in a targeted manner while reducing unnecessary parameter fluctuations, ensuring the overall stability of the model. The final updated classification model can accurately identify fault features under the new distribution and retain the ability to identify historical fault patterns. This forms a coherent adaptive closed loop with the previous feature extraction, fusion, and drift detection processes, enabling the model to maintain effective identification capabilities for various faults during long-term motor operation.

[0034] Please see the appendix Figure 1 Appendix Figure 7 and attached Figure 8 S5. Use the updated classification model to identify faults in the motor's operating status, calculate the health index based on historical operating data, generate early warning information based on the health index and achieve cloud-edge collaborative updates, and output fault category and trend assessment results.

[0035] Furthermore, the generation of early warning information based on health indices and the implementation of cloud-edge collaborative updates in S5 specifically include: When the health index is continuously below the set threshold, a lifespan trend prediction curve is output and an early warning message is generated. Real-time early warning information is pushed to the operation and maintenance terminal via edge devices; The system stores fault identification results, health indices, and trend prediction curves in the cloud and periodically optimizes model parameters before sending them to edge devices. Specifically, when using the updated classification model to identify faults in the motor's operating status, the fused feature vector, which is collected in real time and calibrated by time, extracted and fused, is input into the model. Through the model's mapping and classification of the feature vector, the current fault category of the motor is output, such as specific fault types like bearing wear or winding short circuit. The health index is calculated by combining historical operating data, which includes characteristic parameters such as vibration signal amplitude, current harmonic content, and temperature change rate over a period of time, as well as corresponding fault identification results. The health index is calculated based on the degree of deviation between the current characteristics and those under normal operating conditions, as well as the severity of the fault identification results. The formula is as follows: ; in This is the Euclidean distance between the current feature vector and the mean vector of normal features. The failure probability is the output of the model. For balance coefficient, The distance scaling factor is used to integrate the degree of feature deviation and the failure probability into a health index between 0 and 1. The closer the index is to 1, the healthier the motor is. When the health index is continuously below a set threshold, a life trend prediction curve is generated based on the historical trend of the health index over time using a linear fitting method. This curve reflects the downward trend of the health index with operating time, and at the same time generates early warning information including the fault category, the current health index, and key nodes of the prediction curve. In the specific implementation of the motor fault detection system, the motor's health status is quantified in real time using a health index formula, whereby... Take 0.6, Using 1.2, the motor's 80-day operation monitoring data: the initial health index was 0.94, and the measured value decreased linearly over time, reaching 0.89 on day 20, 0.85 on day 40, 0.80 on day 60, and 0.76 on day 80. A fitting decay model was established. goodness of fit The health index reached 0.98, indicating that the daily average decline was 0.002 units. When the health index was continuously below the warning threshold of 0.3, the system triggered the cloud-edge collaborative warning mechanism and predicted that the motor would reach a critical state on the 320th day. By pushing early warning information to the operation and maintenance terminal in real time through edge devices, operation and maintenance personnel can keep abreast of the motor status. The results of each fault identification, real-time updated health index and life trend prediction curve are stored in the cloud. The cloud uses this accumulated data to periodically optimize the model parameters. The optimization process combines the operating data of multiple motors of the same type to improve the generalization ability of the model. The optimized parameters are then sent to the edge devices to ensure that the classification model at the edge always maintains good recognition performance. The final output of fault category and trend assessment results provides a comprehensive and dynamically updated basis for motor operation and maintenance decisions. It forms a complete closed loop with the previous feature processing, model update and drift detection process, realizing full-cycle monitoring and management of motor operating status.

[0036] Please see the appendix Figure 2 The present invention also provides an artificial intelligence-based motor operation fault detection system, comprising: The multi-source signal acquisition and alignment module is used to acquire motor vibration, current and temperature signals, and realizes multi-source data timestamp correction based on a unified clock reference; The time-frequency feature extraction and fusion module is connected to the multi-source signal acquisition and alignment module, and is used to generate time-frequency feature images and perform cross-modal feature fusion and combination; The concept drift detection module, connected to the time-frequency feature extraction and fusion module, is used to monitor changes in the distribution of the fused feature vector and trigger model updates. The incremental learning and update module, connected to the concept drift detection module, is used for parameter optimization and outputting the updated classification model; The fault diagnosis and cloud-edge collaboration module connects to the incremental learning and update module to generate health index warnings and achieve collaborative optimization of cloud-edge parameters.

[0037] Specifically, the multi-source signal acquisition and alignment module acquires vibration, current and temperature signals of the motor during operation through sensors, and performs timestamp correction on the multi-source data based on a unified clock reference to ensure the consistency of the signals in the time dimension. The time-frequency feature extraction and fusion module connects to this module. It receives the aligned signal, processes it to generate a time-frequency feature image, and then combines and integrates different types of features through cross-modal fusion technology to form a comprehensive feature vector. The concept drift detection module is connected to the feature fusion module and continuously monitors the distribution changes of the fused feature vectors. When a significant shift is detected, the model update mechanism is triggered. The incremental learning and update module responds to the trigger signal, performs local adjustments and updates to the classification model according to the parameter optimization strategy, and outputs a model adapted to the new working conditions. The fault diagnosis and cloud-edge collaboration module takes over the updated model, completes fault diagnosis and calculates health index, generates early warning information, and stores operating data, periodically optimizes model parameters and sends them to the edge through the cloud-edge collaboration mechanism.

[0038] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting motor operation faults based on artificial intelligence, characterized in that, Includes the following steps: S1. Acquire vibration, current and temperature signals during motor operation. Introduce a unified clock reference at the acquisition end. At the receiving end, use time delay estimation and dynamic time warping algorithms to perform timestamp correction on the multi-source data and output a calibrated multi-source aligned signal. S2. Perform short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal to obtain time-frequency feature image. Extract robust single-mode features based on feature space enhancement strategy, and perform feature fusion using cross-modal time attention mechanism to output fused feature vector. S3. During model deployment, the statistics of the input data distribution are calculated based on the sliding window, and the distribution change of the fused feature vector is monitored using the drift detection algorithm. When a significant shift is detected, the model adaptation process is triggered. S4. In the adaptive process, the fused feature vector is updated online using the incremental learning method, and the continuity of the original feature distribution is maintained by knowledge distillation. Only the network modules related to drift are locally fine-tuned, and the updated classification model is output. S5. Use the updated classification model to identify faults in the motor's operating status, calculate the health index based on historical operating data, generate early warning information based on the health index, and achieve cloud-edge collaborative updates, outputting fault category and trend assessment results.

2. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, In step S1, the time stamp correction of multi-source data using time delay estimation and dynamic time warping algorithms specifically includes: Vibration and current signals are acquired simultaneously using a multi-channel synchronous acquisition card, while temperature signals are acquired through a low-frequency sampling module. At the acquisition end, a unified clock source is used for timing, and at the receiving end, the signal delay is calculated based on cross-correlation, and dynamic time warping is used to achieve sequence alignment of cross-modal signals.

3. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, In step S2, performing short-time Fourier transform and continuous wavelet transform on the multi-source aligned signal and then fusing its features specifically includes: Short-time Fourier transforms were performed on the vibration signal and the current signal to generate time-frequency graphs, and sliding window statistical processing was performed on the temperature signal to generate trend features. The time-frequency graph is input into a convolutional neural network to extract local feature vectors, and the trend features are input into a long short-term memory network to extract time-series feature vectors. Feature fusion is performed on local feature vectors and temporal feature vectors based on cross-modal attention mechanism.

4. The method for detecting motor operation faults based on artificial intelligence according to claim 3, characterized in that, The feature fusion using a cross-modal temporal attention mechanism specifically includes: The vibration signal feature vector is used as the query vector, the current signal feature vector is used as the key vector, and the temperature signal feature vector is used as the value vector. The cosine similarity between the query vector and the key vector is calculated by using a sliding time window, and a time alignment similarity matrix is ​​generated to generate time attention weights. The value vector is weighted and aggregated based on the time attention weights to output a cross-modal fusion feature vector.

5. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, The feature space enhancement strategy in S2 specifically includes: Random band masking and phase perturbation are performed on the time-frequency plots of vibration and current signals; Gaussian noise perturbation is injected into the temperature trend characteristics; Regularized samples based on adversarial training to generate multimodal features; The vibration and current signals are fused using cross-modal feature combination technology, and the fused features are weighted and fused using temperature trend features.

6. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, In step S3, monitoring the distribution changes of the fused feature vector using a drift detection algorithm specifically includes: A sliding time window is constructed to update the probability distribution of the fused feature vector, and the Kullback-Leibler divergence between the distributions is calculated. When the divergence exceeds a preset threshold, it is determined to be a concept drift and an adaptive process is triggered.

7. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, The incremental learning online update in S4 specifically includes: The new data cached after triggering drift detection is input into the model for micro-batch training; During training, the convolutional feature extraction layer is frozen, and only the parameters of the cross-modal attention fusion layer and the classifier layer are updated; The output distribution of the historical model is introduced as a soft label constraint through knowledge distillation.

8. The method for detecting motor operation faults based on artificial intelligence according to claim 7, characterized in that, The specific steps of updating only the parameters of the cross-modal attention fusion layer and the classifier layer include: Based on feature drift contribution analysis, the priority of the network modules to be updated is determined; A hierarchical learning rate adjustment is performed on the cross-modal attention fusion layer, wherein the learning rate of the temporal attention weights is set to 1.5-2 times that of the spatial attention weights; A sparse reparameterization strategy is adopted for the classifier layer. Based on the gradient contribution value, highly relevant weight parameters are selected for selective updates, while the rest are frozen.

9. The method for detecting motor operation faults based on artificial intelligence according to claim 1, characterized in that, The specific steps in S5 for generating early warning information based on health indices and implementing cloud-edge collaborative updates include: When the health index is continuously below the set threshold, a lifespan trend prediction curve is output and an early warning message is generated. The warning information is pushed to the operation and maintenance terminal in real time via edge devices; The system stores fault identification results, health indices, and trend prediction curves in the cloud and periodically optimizes model parameters before sending them to edge devices.

10. An artificial intelligence-based motor operation fault detection system, used in the fault detection method according to any one of claims 1-9, characterized in that, include: The multi-source signal acquisition and alignment module is used to acquire motor vibration, current and temperature signals, and realizes multi-source data timestamp correction based on a unified clock reference; The time-frequency feature extraction and fusion module is connected to the multi-source signal acquisition and alignment module and is used to generate time-frequency feature images and perform cross-modal feature fusion and combination. The concept drift detection module, connected to the time-frequency feature extraction and fusion module, is used to monitor changes in the distribution of the fused feature vector and trigger model updates. The incremental learning and update module, connected to the concept drift detection module, is used for parameter optimization and outputting the updated classification model; The fault diagnosis and cloud-edge collaboration module is connected to the incremental learning and update module, and is used to generate health index warnings and realize cloud-edge parameter collaborative optimization.

Images