A multi-scenario-based full lifecycle management method for frameless torque motors

Abstract

This invention relates to a multi-scenario-based full lifecycle management method for frameless torque motors, specifically in the field of motor full lifecycle management. This solution achieves refined perception of complex operating environments by fusing multi-source operating data and generating quantitative scenario codes. Rapid model adaptation guided by scenario codes can extract health degradation features strongly correlated with specific scenarios from sensor data. A lifecycle prediction network combined with scenario conditions significantly improves the accuracy and rationality of remaining lifecycle predictions under different operating environments. The resulting closed-loop learning and scenario knowledge base enable the prediction model to continuously evolve and utilize historical experience to provide accurate initialization and prior knowledge for new scenarios, thereby achieving autonomous, precise, and evolvable intelligent management of the health status of frameless torque motors throughout their entire lifecycle.

Description

Technical Field

[0001] This invention relates to the field of motor lifecycle management, and more specifically, to a method for managing the entire lifecycle of a frameless torque motor based on multiple scenarios. Background Technology

[0002] As the high-end equipment manufacturing industry develops towards intelligence and flexibility, frameless torque motors, with their advantages of high torque density, low cogging torque, and ease of integration, have become core drive components for robot joints, direct-drive rotary tables of high-end CNC machine tools, precision automated equipment, and special mobile platforms. Their application scenarios are highly complex and diverse. For example, in laboratory collaborative robots, motors typically operate in clean, temperature-controlled environments, performing long-term, medium-to-low load precision movements. In the slewing mechanisms of large port lifting equipment, they must withstand severe impact loads, continuous dust pollution, and high humidity corrosion. Furthermore, when applied to field inspection robots or special vehicles, motors must maintain reliable operation under extreme conditions of wide temperature range, strong vibration, and large day-night temperature differences. This cross-scenario application means that the stress types, load spectra, and environmental disturbances experienced by motors during actual service are completely different, directly leading to significant differences in their performance degradation paths and dominant failure modes.

[0003] However, current health management of in-service frameless torque motors mainly relies on alarm mechanisms based on fixed thresholds and periodic preventive maintenance strategies. Although predictive maintenance technologies based on vibration signal analysis have been attempted in some high-end equipment, these methods generally have limitations. Specifically, most existing predictive models rely on in-depth monitoring data such as vibration and noise collected in specific scenarios, or are trained based on laboratory-standard accelerated life test data. When faced with new application environments that differ significantly from the training scenarios, these models fail to effectively incorporate the contextual information of the actual operating scenario, such as dynamic load spectra, variable temperature and humidity environments, and specific start-stop cycles. Its generalization ability is seriously insufficient. A life prediction model trained on a constant temperature and humidity test bench with insulation aging as the main failure mode cannot accurately capture the degradation characteristics dominated by mechanical wear when deployed on port machinery with high vibration and high dust. This leads to a sharp increase in the error of remaining service life prediction, making it difficult to support accurate predictive maintenance decisions. Therefore, how to build an intelligent framework that can deeply integrate multi-dimensional operating scenario information and adapt to different external environments and internal load conditions to achieve accurate health status assessment and remaining service life prediction has become a key technical bottleneck to improve the reliability management and service efficiency of frameless torque motors throughout their entire life cycle. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a multi-scenario-based full lifecycle management method for frameless torque motors, thereby resolving the issues raised in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps: Step S1: Obtain multi-source heterogeneous data of the target frameless torque motor during operation. The multi-source heterogeneous data includes time-series sensing data collected by the motor body sensors, structured operation log data recorded by the device controller, and external environment data obtained through the Internet of Things gateway. Use a multi-modal scene perception and quantization coding model to fuse the multi-source heterogeneous data and generate a first scene coding vector representing the current operating scene. Step S2: Obtain the pre-trained meta-feature extraction model under multiple known scenarios, and use the time-series sensing data and the first scene encoding vector to perform fast fine-tuning of the meta-feature extraction model based on the scenario to obtain the target feature extraction model adapted to the target scenario; use the target feature extraction model to process the time-series sensing data and output the first health index sequence corresponding to the target frameless torque motor. Step S3: Construct a conditional deep degradation network model, using the first health indicator sequence as the main input sequence and the first scenario encoding vector as the conditional input; use the conditional deep degradation network model to jointly analyze the main input sequence and the conditional input to predict the conditional remaining service life of the target frameless torque motor under the current operating scenario. Step S4: When the target frameless torque motor reaches the end of its lifespan or a maintenance node, collect its complete lifecycle data. The lifecycle data includes at least multi-source heterogeneous data, a first health indicator sequence, a conditional remaining service life prediction value, and a recorded actual service life value, which together constitute a closed-loop verification sample. Using the closed-loop verification sample, update the parameters of the conditional deep degradation network model based on an online learning strategy to obtain the updated conditional deep degradation network model. Extract the degradation pattern contained in the first health indicator sequence in the closed-loop verification sample as the degradation mode feature. Store the parameters of the updated conditional deep degradation network model, the first scene encoding vector, and the degradation mode feature as an associated record in a pre-built scene knowledge base. The scene knowledge base is used for model initialization and prediction prior guidance for new operating scenarios. In a preferred embodiment, in step S1, the timing sensing data includes at least the three-phase current data and winding temperature data of the motor, the structured operation log data includes at least the real-time torque command value sequence, speed value sequence and position value sequence, and the external environment data includes at least the ambient temperature data, relative humidity data and dust concentration data.

[0006] In a preferred embodiment, the process of fusing multi-source heterogeneous data using a multimodal scene perception and quantization coding model specifically includes the following operations performed sequentially: First, perform multimodal feature extraction: process the time-series sensor data, structured operation log data and external environment data to generate corresponding initial sensor feature vectors, initial log feature vectors and initial environment feature vectors respectively; Specifically, a one-dimensional convolutional neural network unit in the multimodal scene perception and quantization coding model is used to process the time-series sensing data to capture its local fluctuations and short-term dependency patterns, forming an initial sensing feature vector. A statistical feature aggregation unit in the same model calculates and combines the mean, peak value, and ripple amplitude of the torque command value sequence, speed value sequence, and position value sequence within a set time window to form an initial log feature vector. Finally, a standardization unit in the same model performs standardization transformation on the external environment data to form an initial environmental feature vector. Next, a feature concatenation operation is performed: the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector are concatenated to obtain an original mixed feature vector.

[0007] In a preferred embodiment, after obtaining the original mixed feature vector, the dynamic fusion and encoding operation based on the attention mechanism is further performed to generate the first scene encoding vector. This operation specifically includes: An attention fusion unit in a multimodal scene perception and quantization coding model is used to generate a query vector based on the original mixed feature vector; An attention fusion unit is used to apply different learnable linear transformations to the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector to generate their respective key vectors and value vectors. Calculate the dot product between the query vector and each key vector to obtain multiple dot product results; Each dot product result is scaled by dividing by the square root of a predetermined dimension of the key vector; all scaled dot product results are then processed by a normalized exponential function to assign dynamic adaptive attention weight coefficients to the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector. By using the adaptive attention weight coefficients assigned to each type of feature vector, the corresponding value vectors are weighted respectively, and all weighted value vectors are summed to obtain a primary fusion scene vector; Finally, the encoding generation unit in the multimodal scene perception and quantization coding model is used to input the primary fused scene vector into a multilayer perceptron. The multilayer perceptron performs nonlinear transformation and dimension normalization operations to generate a first scene encoding vector with a predetermined fixed dimension.

[0008] In a preferred embodiment, step S2 involves obtaining a pre-trained meta-feature extraction model for multiple known scenarios. The pre-training process specifically includes: We collected time-series sensing data samples and corresponding health status labels of frameless torque motors under multiple different operating scenarios. The data samples of each operating scenario constituted an independent meta-learning task. We used a model-independent meta-learning algorithm framework to train the meta-feature extraction model. In each training iteration, for each sampled meta-learning task, its data is divided into a support set and a query set; the temporal sensing data in the support set and the corresponding first scene encoding vector are input into the meta-feature extraction model, the support set loss is calculated based on the model output, and the model parameters are updated by gradient descent to obtain temporary adapted model parameters, wherein the first scene encoding vector is used to dynamically adjust the internal operation of the model. The temporal sensing data in the query set is forward-propagated using temporary adaptive model parameters, and the query set loss is calculated as the meta-loss. By adjusting the model parameters by minimizing the expected value of the meta-loss on all sampling tasks, a pre-trained meta-feature extraction model is obtained.

[0009] In a preferred embodiment, the process of rapidly fine-tuning the meta-feature extraction model based on the scene using time-series sensing data and a first scene encoding vector specifically involves: A short-term time-series sensing data is collected from the target frameless torque motor as target scene support data, and the corresponding first scene encoding vector is obtained. The target scene support data and the first scene encoding vector are input into the meta-feature extraction model. The feature extraction focus of the meta-feature extraction model is adjusted by the first scene encoding vector. An adaptive loss is calculated, and the model parameters are updated by gradient descent a very small number of times using this adaptive loss to obtain the target feature extraction model. The very small number of times is one to five. Then, the complete time-series sensing data generated by the continuous operation of the target frameless torque motor is input into the target feature extraction model; The target feature extraction model outputs a single-value scalar for each time point or a data segment of a continuous time window in the time-series sensing data. This single-value scalar represents a summary of deep features related to the motor health status extracted from the data of the corresponding time period. In chronological order, all the output single-value scalars are concatenated to form the first health indicator sequence.

[0010] In a preferred embodiment, the specific operation of constructing the conditional deep degradation network model and using the first scene encoding vector as the conditional input in step S3 is as follows: The backbone of the conditional deep degradation network model is composed of multiple stacked temporal feature extraction units, which are used to process the first health indicator sequence. The first scene encoding vector is injected through a conditional normalization layer set in multiple network layers within the conditional deep degenerate network model; In each conditional normalization layer, the first scene encoding vector is input into two independent small parameter generation networks to generate scaling parameter vectors and translation parameter vectors respectively. Both scaling parameter vectors and translation parameter vectors have the same number of channels as the intermediate features to be processed in the current network layer. For the intermediate features input to this layer, they are first normalized, then multiplied element-wise with the scaling parameter vector, and then added element-wise with the translation parameter vector to output the features modulated by the scene conditions.

[0011] In a preferred embodiment, the specific process of jointly analyzing the main input sequence and the conditional input using a conditional deep degradation network model to predict the conditional remaining useful life is as follows: During the training phase of the conditional deep degradation network model, optimization is performed by minimizing a multi-objective loss function; the multi-objective loss function consists of a weighted sum of a prediction loss term and a physical guidance loss term. The difference between the predicted conditional remaining useful life and the actual remaining useful life is calculated using the smoothed average absolute error loss function. The physics-guided loss term constrains the latent degradation trajectories learned internally by the model and related to the scene, ensuring they conform to physical priors. Its calculation process is as follows: From the dataset used to train the conditional deep degradation network model, which includes temporal sensor data samples, health status labels, and corresponding first-scene encoding vectors, multiple first-scene encoding vectors representing severe operating conditions are selected to form a severe scene sample set, and multiple first-scene encoding vectors representing mild operating conditions are selected to form a mild scene sample set. The penalty term corresponding to each first-scene encoding vector in the severe scene sample set is calculated. The calculation process is as follows: first calculate the penalty term corresponding to each first-scene encoding vector in the severe scene sample set... The second derivative of the latent degradation trajectory of the conditional deep degradation network model with respect to time under the first scene encoding vector condition is calculated, and then a linear rectified function is applied to the negative of the second derivative. The penalty term corresponding to each first scene encoding vector in the mild scene sample set is calculated. The calculation process is as follows: first, calculate the second derivative of the latent degradation trajectory of the conditional deep degradation network model with respect to time under the first scene encoding vector condition, and then apply a linear rectified function to the second derivative itself. The physical guidance loss term is equal to the sum of the average of all sample penalty terms in the severe scene sample set and the average of all sample penalty terms in the mild scene sample set.

[0012] In a preferred embodiment, step S4, which involves updating the parameters of the conditional deep degradation network model using closed-loop validation samples based on an online learning strategy, specifically comprises: An online learning strategy is implemented using an elastic weight consolidation algorithm, which assigns an importance weight to each model parameter in the conditional deep degenerate network model. Construct an online learning loss function, which consists of a weighted sum of a prediction loss term and a regularization penalty term; The predicted loss term is calculated based on the first health indicator sequence, the first scenario encoding vector, and the actual lifespan value in the closed-loop validation samples. The regularization penalty term is as follows: for each model parameter in the conditional deep degenerate network model, calculate the square of the difference between its updated value and its original value, multiply it by the importance weight of that parameter, and finally sum it over all parameters; The updated conditional deep degradation network model is obtained by updating the model parameters by minimizing the online learning loss function.

[0013] In a preferred embodiment, the specific process of storing the updated parameters of the conditional deep degradation network model, the first scene encoding vector, and the degradation mode features as an associated record in the scene knowledge base is as follows: Construct a temporal autoencoder model, including an encoder network and a decoder network; The first health indicator sequence in the closed-loop verification sample is input into the encoder network, and the low-dimensional degradation pattern features are output. The degradation mode features, the corresponding first scene encoding vector, the actual lifetime value, and the parameters of the updated conditional deep degradation network model are stored as a knowledge record in the scene knowledge base. When a new running scenario appears, the cosine similarity between its first scenario encoding vector and all historical vectors is calculated in the scenario knowledge base, and the K historical knowledge records with the highest similarity are retrieved. Support is provided by using search results, including: using model parameters associated with the most similar historical records as initial parameters for model training in new scenarios; and referring to actual lifetime values ​​and degradation pattern features in search records to provide prior information for new scenario prediction.

[0014] The beneficial effects of this invention are as follows: By fusing multi-source operating data and generating quantitative scene codes, this solution achieves refined perception of complex operating environments; rapid model adaptation guided by scene codes can extract health degradation features strongly correlated with the scene from sensor data; the lifespan prediction network combined with scene conditions significantly improves the accuracy and rationality of remaining lifespan prediction under different operating environments; the formed closed-loop learning and scene knowledge base enables the prediction model to continuously evolve and utilize historical experience to provide accurate initialization and prior knowledge for new scenes, thereby achieving autonomous, accurate, and evolvable intelligent control of the health status of the frameless torque motor throughout its entire life cycle. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1 The method for full lifecycle management of frameless torque motors based on multiple scenarios is shown, and specifically includes the following steps: Step S1: Obtain multi-source heterogeneous data of the target frameless torque motor during operation. The multi-source heterogeneous data includes time-series sensing data collected by the motor body sensors, structured operation log data recorded by the device controller, and external environment data obtained through the Internet of Things gateway. Use a multi-modal scene perception and quantization coding model to fuse the multi-source heterogeneous data and generate a first scene coding vector representing the current operating scene. Step S2: Obtain the pre-trained meta-feature extraction model under multiple known scenarios, and use the time-series sensing data and the first scene encoding vector to perform fast fine-tuning of the meta-feature extraction model based on the scenario to obtain the target feature extraction model adapted to the target scenario; use the target feature extraction model to process the time-series sensing data and output the first health index sequence corresponding to the target frameless torque motor. Step S3: Construct a conditional deep degradation network model, using the first health indicator sequence as the main input sequence and the first scenario encoding vector as the conditional input; use the conditional deep degradation network model to jointly analyze the main input sequence and the conditional input to predict the conditional remaining service life of the target frameless torque motor under the current operating scenario. Step S4: When the target frameless torque motor reaches the end of its lifespan or a maintenance node, collect its complete lifecycle data. The lifecycle data includes at least multi-source heterogeneous data, a first health indicator sequence, a conditional remaining service life prediction value, and a recorded actual service life value, which together constitute a closed-loop verification sample. Using the closed-loop verification sample, update the parameters of the conditional deep degradation network model based on an online learning strategy to obtain the updated conditional deep degradation network model. Extract the degradation pattern contained in the first health indicator sequence in the closed-loop verification sample as the degradation mode feature. Store the parameters of the updated conditional deep degradation network model, the first scene encoding vector, and the degradation mode feature as an associated record in a pre-built scene knowledge base. The scene knowledge base is used for model initialization and prediction prior guidance for new operating scenarios.

[0020] In this embodiment, it should be specifically noted that in step S1, the timing sensing data includes at least the three-phase current data and winding temperature data of the motor, the structured operation log data includes at least the real-time torque command value sequence, speed value sequence and position value sequence, and the external environment data includes at least the ambient temperature data, relative humidity data and dust concentration data. The three-phase current data and winding temperature data in the timing sensing data are directly acquired by the motor body sensors. The current sensor and temperature sensor included in the motor body sensors are embedded in the motor stator. The sampling frequency of the three-phase current data and winding temperature data is set according to the rated speed of the target frameless torque motor and the bandwidth of the servo control system. The sampling frequency setting range is 1kHz to 10kHz to ensure that the acquired timing sensing data can completely capture the current harmonic components and the thermal dynamic changes of the winding during motor operation. The real-time torque command value sequence, speed value sequence, and position value sequence in the structured operation log data are obtained in real time from the equipment controller. The equipment controller is the control loop inside the motor driver or the upper-level motion controller connected to it. The update frequency of the real-time torque command value sequence, speed value sequence, and position value sequence is consistent with the control frequency of the current control loop or speed control loop of the equipment controller. The ambient temperature, relative humidity, and dust concentration data in the external environment data are obtained through the Internet of Things (IoT) gateway. Specifically, the data can be obtained by directly reading the data through IoT environmental sensor nodes connected to the IoT gateway and deployed near the motor, or by the IoT gateway calling the corresponding meteorological service application programming interface and industrial environmental monitoring service application programming interface online based on the device's geographical location information. The update cycle of the external environment data is set to 1 to 10 minutes to balance the timeliness of environmental status monitoring with the resource consumption of network communication. Multimodal scene perception and quantization coding models are used to fuse heterogeneous data from multiple sources. Specifically, the following operations are performed sequentially: First, perform multimodal feature extraction: process the time-series sensor data, structured operation log data and external environment data to generate corresponding initial sensor feature vectors, initial log feature vectors and initial environment feature vectors respectively; The process employs a one-dimensional convolutional neural network unit in a multimodal scene perception and quantization coding model to process time-series sensing data, capturing its local fluctuations and short-term dependency patterns to form an initial sensing feature vector. The one-dimensional convolutional neural network unit can contain 2 to 4 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The kernel size can be set to 3, 5, or 7. By stacking convolutional layers, the receptive field is gradually expanded to extract multi-scale frequency and time domain features related to potential fault modes such as motor electrical imbalance and local overheating from high-frequency sampled current and temperature signals. Finally, a global average pooling layer aggregates the time-varying feature map into a fixed-length initial sensing feature vector. The statistical feature aggregation unit in the multimodal scene perception and quantization coding model is used to calculate and combine the statistical features of the mean, peak value, and ripple amplitude of the torque command value sequence, speed value sequence, and position value sequence within a set time window to form the initial log feature vector. The length of the set time window can be determined according to the main working cycle of the motor, such as 5 seconds, 30 seconds, or 1 minute. The mean value reflects the average load level, the peak value reflects the maximum stress impact, and the ripple amplitude (the difference between the peak value and the valley value) reflects the intensity of the dynamic fluctuation of the load. The calculated 9 statistical features (3 sequences × 3 statistics) are combined into a 9-dimensional initial log feature vector, which quantifies the mechanical load condition of the motor within the window period from the perspective of energy and dynamics. The external environmental data is standardized using the standardized unit in the multimodal scene perception and quantization coding model to form an initial environmental feature vector. The standardization transformation usually adopts Z-score standardization, which is to subtract the mean of the historical data of each environmental parameter (temperature, humidity, dust concentration) and divide it by its standard deviation, so that the mean of each dimension of the processed data is 0 and the variance is 1. This helps to eliminate the influence of the differences in the dimensions and numerical ranges of different environmental parameters on the stability of subsequent fusion model training. The generated initial environmental feature vector can more effectively represent the environmental stress level. Next, a feature concatenation operation is performed: the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector are concatenated to obtain an original mixed feature vector. The concatenation operation is a vector concatenation. For example, if the initial sensor feature vector has a dimension of 64, the initial log feature vector has a dimension of 9, and the initial environment feature vector has a dimension of 3, then the original mixed feature vector obtained after concatenation has a dimension of 76. This operation initially integrates three types of features from different physical meanings and different processing methods, providing a foundation for subsequent attention-based fine fusion. After obtaining the original mixed feature vector, a dynamic fusion and encoding operation based on an attention mechanism is performed to generate the first scene encoding vector. This operation is specifically as follows: An attention fusion unit in a multimodal scene perception and quantization coding model is used to generate a query vector based on the original mixed feature vector. The generation of the query vector is achieved by applying a trainable fully connected layer (linear transformation) to the original mixed feature vector. This query vector serves as an "intent" representation, aiming to retrieve the part most relevant to the current integrated scene from multi-source information. An attention fusion unit is used to apply different learnable linear transformations to the initial feature vectors of the sensors, logs, and environment, respectively, to generate corresponding key vectors and value vectors. The key vectors and value vectors have a predetermined dimension. The predetermined dimension is usually set to the same value, such as 64 dimensions, to ensure consistency in subsequent calculations. The key and value transformation matrices are set independently for each type of initial feature vector, allowing the model to learn the different representations that are most suitable for matching (key) and contributing information (value) for each type of data. Calculate the dot product between the query vector and each key vector to obtain multiple dot product results; Each dot product result is scaled by dividing by the square root of a predetermined dimension of the key vector. This scaling is a standard step in the attention mechanism, preventing the dot product results from becoming too large and causing the gradient of the Softmax function to vanish, thus stabilizing the model training process. A normalized exponential function is applied to all scaled dot product results to assign dynamic adaptive attention weight coefficients to the initial sensor feature vector, initial log feature vector, and initial environmental feature vector. The normalized exponential function ensures that the sum of the weight coefficients assigned to all feature vectors is one. The adaptive attention weight coefficients are scalars, and their magnitudes dynamically reflect the relative importance of the information carried by the corresponding category feature vectors when generating the current scene encoding. For example, in high-temperature, heavy-load scenarios, the winding temperature feature representing thermal load and the log feature representing mechanical load may receive higher attention weights; while in high-dust scenarios, the environmental dust concentration feature may receive higher attention weights. This dynamic weight allocation mechanism is key to achieving "scene-aware" fusion in this method. By using adaptive attention weight coefficients assigned to each type of feature vector, the corresponding value vectors are weighted respectively, and all weighted value vectors are summed to obtain a primary fusion scene vector. The weighted summation operation realizes the selective fusion of information. The value vector with high weight dominates in the final fusion result, while the value vector with low weight contributes less. The primary fusion scene vector generated in this way is not a simple average of all features, but a weighted combination that focuses on the most critical information in the current scene, and its information density and scene representation ability are significantly improved. Finally, the encoding generation unit in the multimodal scene perception and quantization coding model is used to input the primary fused scene vector into a multilayer perceptron. The multilayer perceptron performs nonlinear transformation and dimension normalization operations to generate a first scene encoding vector with a predetermined fixed dimension. The multilayer perceptron is usually composed of 1 to 3 fully connected layers. Nonlinearity is introduced between the layers using activation functions such as ReLU. Its functions are twofold: first, to perform deeper nonlinear transformation and feature abstraction on the primary fused scene vector; and second, to normalize the vector dimension to a unified, pre-set fixed dimension (e.g., 128 dimensions) so that it can be stably received and processed by the model in subsequent steps S2 and S3. The output of this fixed dimension is the "first scene encoding vector" that finally quantitatively represents the complex running scene.

[0021] In this embodiment, it is specifically necessary to explain step S2, which involves obtaining a pre-trained meta-feature extraction model under multiple known scenarios (such as a laboratory constant temperature scenario, a factory vibration scenario, etc.). The pre-training process is as follows: We collected time-series sensing data samples and corresponding health status labels of frameless torque motors under multiple different operating scenarios. The data samples of each operating scenario constituted an independent meta-learning task. We trained the meta-feature extraction model using a model-independent meta-learning algorithm framework. In each training iteration, we randomly sampled a batch of meta-learning tasks from all meta-learning tasks. For each sampled meta-learning task, we divided the data of the meta-learning task into a support set and a query set. The ratio of the support set to the query set is usually 1:1 or 1:2. For example, 70% of the data of each meta-learning task is used as the support set and 30% as the query set, in order to simulate a meta-learning paradigm in which fast learning is performed with only a small amount of labeled data in new scenarios and the adaptation effect is evaluated on more data. First, the temporal sensing data from the support set and the first scene encoding vector generated in step S1, which matches the running scenario corresponding to the meta-learning task, are input into the meta-feature extraction model. Based on the model output and the health status labels corresponding to the support set, the support set loss is calculated, and the current parameters of the meta-feature extraction model are updated once or several times using gradient descent according to this support set loss, resulting in temporary adaptive model parameters suitable for the meta-learning task. The first scene encoding vector is used to dynamically adjust the computation of the neural network layers within the meta-feature extraction model, so that the model's feature extraction behavior is associated with the running scenario of the meta-learning task. One specific implementation of this dynamic adjustment is "conditional batch normalization," which is applied after the convolutional or fully connected layers of the meta-feature extraction model. Before the activation function, a conditional batch normalization layer is added. The first scene encoding vector is passed through a small fully connected network to generate a scaling factor vector and a bias factor vector specific to this conditional batch normalization layer. These are used to perform affine transformations on the normalized features, thereby injecting scene information into the feature stream of the network and guiding the direction of feature extraction. Gradient descent update calculates the gradient of the support set loss with respect to the current parameters of the meta-feature extraction model, and subtracts the product of a preset inner loop step size and the gradient from the current parameters to calculate the temporary adaptation model parameters. The preset inner loop step size is a small positive number, ranging from 0.001 to 0.1, with a typical value of 0.01, to ensure that the rapid adaptation of the inner loop is a stable and small-amplitude adjustment, rather than a complete retraining. Subsequently, the temporary adapted model parameters are used to perform forward propagation on the temporal sensing data in the query set, and the query set loss of the model on the query set is calculated. This query set loss is the meta-loss. The training objective of the meta-feature extraction model is to minimize the expected value of the meta-loss calculated on all sampled meta-learning tasks by adjusting its parameters in multiple iterations. "Minimizing the expected value of the meta-loss" is achieved through stochastic gradient descent or its variant. In each training iteration, the average value of the meta-loss of all meta-learning tasks in the current batch is calculated, and the gradient of this average value with respect to the initial parameters of the meta-feature extraction model is calculated. Then, gradient descent is performed in the outer loop to update the model. This process forces the initial parameters of the model to be optimized so that when facing any new task, only a few gradient updates are needed on the support set of that task to obtain temporary adapted model parameters that perform well on the query set of that task. Through the training process, a pre-trained meta-feature extraction model is obtained, whose parameters constitute a set of initial parameters that can quickly adapt to new operating scenarios. The process of rapidly fine-tuning the meta-feature extraction model based on the scene using time-series sensing data and the first scene encoding vector is as follows: After obtaining the pre-trained meta-feature extraction model and its parameters, a short-term time-series sensing data is collected from the target frameless torque motor as target scene support data, and the first scene encoding vector corresponding to the current target scene generated in step S1 is obtained. The target scene support data and the first scene encoding vector are input into the meta-feature extraction model. The first scene encoding vector is applied to the internal feature representation of the model through the conditional adjustment mechanism contained in the meta-feature extraction model, so that the model, guided by the first scene encoding vector, adjusts the focus of feature extraction to match the operating environment and load characteristics represented by the first scene encoding vector. An initial input is obtained through forward propagation. "Short time range" typically refers to a duration that covers several typical working cycles of the target motor, such as collecting time-series sensor data for 10 seconds to 5 minutes to ensure that the data reflects typical operating conditions in the scenario. Based on this initial output, an adaptive loss is calculated in conjunction with a preset adaptive target. The adaptive target is set in at least one of the following ways: based on a small number of health status reference values ​​attached to the target scenario support data, based on the reconstruction error generated by encoding and reconstructing the target scenario support data, or based on a contrastive learning target constructed on unlabeled target scenario support data. If the method based on a small number of health status reference values ​​is adopted, the reference values ​​can be manually labeled or based on simple rules. The initial health score is estimated. If a reconstruction error-based approach is used, a decoder network needs to be connected after the meta-feature extraction model to form an autoencoder structure. The adaptive loss is the mean squared error between the target scene support data and the decoder output of the reconstructed input data. If a contrastive learning objective-based approach is used, data augmentation is needed to generate positive sample pairs for the target scene support data, and data from different time segments or different augmented views are used as negative samples. The adaptive loss is the contrastive loss that encourages positive sample pairs to have similar features and negative sample pairs to have different features. Using the adaptive loss, the gradient of the adaptive loss with respect to the parameters of the meta-feature extraction model is calculated, and the meta-features are adjusted according to a fine-tuning step size. Extracting all or part of the model's parameters and performing gradient descent updates a very small number of times, typically one to five times, allows for effective scene adaptation with minimal computational overhead, avoiding overfitting risks. This update process mimics the inner loop adaptation process of the meta-feature extraction model's pre-training phase and is conditionally adjusted by the first scene encoding vector, ensuring that the updated model parameters are specialized in extracting scene-related health degradation features from the temporal sensing data of the current target scene, thus obtaining a target feature extraction model specifically designed for the target scene. Subsequently, the complete time-series sensing data generated by the continuous operation of the target frameless torque motor is input into the target feature extraction model. The last layer of the target feature extraction model is configured to output a single-valued scalar. For each time point or a data segment within a continuous time window in the time-series sensing data, the target feature extraction model outputs a corresponding single-valued scalar. Each single-valued scalar represents a summary of deep features related to the motor's health status extracted by the model from the data of the corresponding time period. The length of the "continuous time window" can differ from the short time series used during rapid fine-tuning. It is typically set to a fixed, relatively short analysis window, such as 1 second or 0.1 seconds, to achieve higher temporal resolution monitoring of the health status. The single-valued scalar represents the target feature. The output of the last fully connected layer of the model is extracted. Its value is usually monotonically related to the degree of degradation of the motor from health to failure. For example, insulation aging may cause the value to increase monotonically, while increased friction due to lubrication failure may cause it to decrease monotonically. In chronological order, the single-value scalars corresponding to all time points or time windows are sequentially connected to form a time series sequence, which is the first health index sequence. The first health index sequence is a one-dimensional signal that varies with time. It shows a monotonically increasing or decreasing trend, which is strongly correlated with the overall health degradation process of the motor. The first health index sequence filters out noise and operating condition fluctuations contained in the time series sensing data and directly reveals the inherent performance degradation trajectory of the motor.

[0022] In this embodiment, it is specifically necessary to explain the following steps in step S3: constructing the conditional deep degradation network model and using the first scene encoding vector as the conditional input. The Conditional Deep Degradation Network (CDDN) model employs an encoder structure. Its backbone network consists of stacked layers of temporal feature extraction units, which receive and process the first health indicator sequence as the main input sequence to capture its long-range dependencies. The temporal feature extraction units can be dilated causal convolutional layers, long short-term memory (LSTM) network layers, or Transformer encoder layers, with dilated causal convolutional layers being preferred. By increasing the dilation rate layer by layer, the receptive field is expanded exponentially, thereby effectively modeling the long-term degradation trend dependencies in the first health indicator sequence that may span hundreds or even thousands of time steps, while maintaining computational efficiency. The first scene encoding vector is injected into the model's computation process through conditional normalization layers set in multiple network layers within the CDDN model. In each conditional normalization layer, the following steps are performed to generate scene-conditionally modulated output features: the first scene encoding vector is input into two independent small-parameter generation networks (SPGs). The first SPG outputs a scaling parameter vector, and the second SPG outputs a translation parameter vector. Both the scaling and translation parameter vectors have the same number of channels as the intermediate features to be processed in the current network layer. The SPGs are typically single-layer or two-layer fully connected neural networks. Their input is the first scene encoding vector, and their output dimension is equal to the number of channels in the intermediate features of the current layer. This design allows high-dimensional scene information to be compressed and converted into features that can be directly modulated. The affine transformation parameters enable fine-grained, channel-level control of scene information on the network's internal feature representation. For the intermediate features of the input conditional normalization layer, their mean and standard deviation over time are calculated. The mean is subtracted from each element of the intermediate feature, and the result of the subtraction is divided by the standard deviation to obtain a normalized feature with zero mean and unit variance. Then, the scaling parameter vector is multiplied element-wise with the normalized feature to obtain the scaled feature. Finally, the translation parameter vector is added element-wise with the scaled feature to output the final scene-conditionally modulated feature. This modulation process allows the same intermediate features to undergo different scaling and translation transformations under the influence of first scene encoding vectors with different values, thereby deeply integrating scene information into the network's analytical path for health degradation trends. In this way, the analysis of the first health indicator sequence by the conditional deep degradation network model is no longer static. For example, for an intermediate feature sequence representing "slight increase in vibration energy," under the influence of the first scene encoding vector representing "port heavy crane," it may be amplified and interpreted as "strong signal of early bearing wear," resulting in a shorter predicted remaining life; while under the influence of the first scene encoding vector representing "laboratory precision platform," it may be suppressed and interpreted as "normal operating condition fluctuations," resulting in a longer predicted remaining life. This dynamic conditional analysis capability is the core of the model's ability to achieve accurate scene adaptive prediction. The specific process of using a conditional deep degradation network model to jointly analyze the main input sequence and conditional input to predict the conditional remaining useful life is as follows: During the training phase of the Conditional Deep Degradation Network (CDDN) model, the joint analysis and prediction process is optimized by minimizing a multi-objective loss function. This multi-objective loss function consists of a prediction loss term and a physical guidance loss term, where the weight of the physical guidance loss term in the total loss is controlled by a non-negative hyperparameter. This non-negative hyperparameter, i.e., the weight coefficient, typically ranges from 0.01 to 0.5; for example, it can be set to 0.1. The setting of this coefficient needs to balance direct prediction accuracy and physical plausibility. A coefficient that is too small may result in weak physical constraints, while a coefficient that is too large may excessively interfere with the model's learning of the main prediction task. The prediction loss term calculates the conditional remaining lifetime predicted by the model. The difference between the predicted value and the actual remaining lifetime is calculated using the smoothed mean absolute error loss function. This function uses a squared term when the absolute value of the error is less than 1 and a linear term when it is greater than 1. This makes it more robust to outliers in the training data than the mean squared error loss, thus improving the model's generalization performance. A physics-guided loss term constrains the scene-related latent degradation trajectories learned within the model, ensuring they conform to engineering physics priors. This is achieved through the following calculation process: From the dataset used to train the conditional deep degradation network model, which includes time-series sensor data samples, health status labels, and corresponding first-scene encoding vectors, multiple first-scene encoding vectors representing adverse operating conditions are selected to form an adverse scenario. A mild scenario sample set is formed by selecting a set of scenario samples and multiple first scenario encoding vectors representing mild operating conditions. The distinction between "severe operating conditions" and "mild operating conditions" can be based on prior knowledge by setting thresholds. For example, scenarios with an average load rate consistently above 80% of the rated value, an ambient temperature consistently above 40 degrees Celsius, or a total vibration value consistently above a certain threshold can be labeled as severe. Scenarios with an average load rate below 30% of the rated value, an ambient temperature between 20-25 degrees Celsius, and stable vibration can be labeled as mild. First scenario encoding vectors corresponding to data samples that meet these conditions are selected from the entire training set, forming two sets respectively. The first scenario encoding vectors in each severe scenario sample set are calculated separately. The penalty term corresponding to the quantity is calculated as follows: First, calculate the second derivative of the hidden degradation trajectory of the conditional deep degradation network model with respect to time under the encoding vector of the first scene. Then, apply the linear rectifier function to the negative of the second derivative. The "hidden degradation trajectory" is not the direct output of the model, but can be indirectly reflected from the activation values ​​or gradient information of the intermediate layers of the model. A feasible approximation method is to regard the change of the feature vector before the final fully connected layer of the model in the time series as a low-dimensional expression of the degradation trajectory, and calculate the second difference of the feature vector sequence with respect to the time index to approximate the second derivative. The role of the linear rectifier function is to apply a positive penalty only when the approximate second derivative is greater than or equal to zero (i.e., it does not show accelerated degradation).If the value is less than zero (indicating accelerated degradation), the penalty is zero, thus unilaterally encouraging the model to learn accelerated degradation patterns in harsh scenarios. The penalty term for each first-scene encoding vector in the mild-scene sample set is calculated as follows: first, the second derivative of the latent degradation trajectory of the conditional deep degradation network model with respect to time under the given first-scene encoding vector is calculated; then, a linear rectified function is applied to this second derivative. Similarly, this linear rectified function ensures that a penalty is applied only when the approximate second derivative is less than or equal to zero (i.e., no decelerating degradation is observed); if it is greater than zero (indicating a decelerating degradation trend), the penalty is applied. If the penalty is zero, then the model is unilaterally encouraged to learn a deceleration degradation pattern in mild scenarios. The physical guidance loss term is equal to the sum of the average penalty terms of all samples in the harsh scenario sample set and the average penalty terms of all samples in the mild scenario sample set. The second derivative of the implicit degradation trajectory with respect to time is negative if it indicates an accelerating degradation trend and positive if it indicates a decelerating degradation trend. The linear rectified function ensures that a non-zero penalty is generated only when the sign of the second derivative violates the physical prior (i.e., the second derivative is non-negative in harsh scenarios or non-positive in mild scenarios). The specific structure of the multi-objective loss function is as follows: The predicted loss term is added to the product of a weight coefficient and a physical guidance loss term, where the weight coefficient is the non-negative hyperparameter controlling the weight of the physical guidance loss term. The conditional deep degradation network model is trained end-to-end by minimizing the multi-objective loss function. The trained conditional deep degradation network model can receive the first health indicator sequence as the main input sequence and the first scene encoding vector as the conditional input, and output a conditional remaining lifetime prediction value that accurately reflects the scene's impact. End-to-end training means that the gradients of the predicted loss term and the physical guidance loss term can be used to jointly update all parameters in the conditional deep degradation network model through the backpropagation algorithm, including the parameters of the backbone temporal feature extraction unit, the parameters of the small parameter generation network in each conditional normalization layer, and the parameters of the final output layer. This joint optimization makes the mapping function learned by the model not only numerically close to the real lifetime, but its internal inference process also tends to conform to the physical law of "accelerating failure in severe scenes and delaying failure in mild scenes", enhancing the model's prediction rationality and extrapolation credibility in data-scarce or extreme scenarios.

[0023] In this embodiment, it is specifically necessary to explain that in step S4, the process of updating the parameters of the conditional deep degradation network model based on an online learning strategy using closed-loop verification samples is as follows: An online learning strategy is implemented using an elastic weight consolidation algorithm. This algorithm assigns an importance weight to each model parameter in the Conditional Deep Degradation Network (CDDN) model and constructs an online learning loss function containing a regularization penalty term. The algorithm calculates the approximate values ​​of the diagonal elements of the Fisher information matrix of each model parameter on historical data accumulated in the scene knowledge base, using these approximate values ​​as the importance weights. This reflects the curvature effect of each model parameter on the historical prediction task loss function; a larger value indicates a greater impact of slight perturbations on the parameter's historical prediction performance, meaning the parameter is more "important" and should be kept as stable as possible in subsequent online learning to prevent catastrophic forgetting. The importance weight characterizes the degree to which the model parameter is important to the predictive ability already learned by the model on the data corresponding to the historical closed-loop validation samples stored in the scene knowledge base. An online learning loss function is constructed when updating the parameters of the CDDN model using closed-loop validation samples. This online learning loss function consists of a prediction loss term and a regularization penalty term. The regularization penalty term is composed of weighted sums; the prediction loss term is calculated based on the difference between the newly predicted conditional remaining lifespan value of the first health indicator sequence and the first scene encoding vector in the closed-loop validation sample and the actual lifespan value recorded in the closed-loop validation sample; the prediction loss term here is consistent with the prediction loss term function used when training the conditional deep degradation network model in step S3, for example, both use smoothed average absolute error loss to ensure the consistency between the online learning objective and the original training objective; the regularization penalty term constrains the update magnitude of the model parameters, and the specific calculation process is as follows: for each model parameter in the model, calculate the square of the difference between its updated new parameter value and its old parameter value before the update, then multiply this squared difference by the importance weight of the parameter, and finally sum the calculation results of all model parameters; the specific composition of the online learning loss function is: the prediction loss term, a preset non-negative regularization strength coefficient, and the product of the regularization penalty term are added together; the preset non-negative regularization strength coefficient is a key hyperparameter, and its typical value range is between 0.1 and 1.0, for example, it can be set to 0.5. This coefficient determines the balance between retaining old knowledge and learning new knowledge. If the coefficient is set too small, the regularization penalty term has a weak effect, and the model is prone to over-adapting to new closed-loop validation samples and forgetting historical knowledge. If the coefficient is set too large, it will excessively restrict the model parameter updates, making it difficult for the model to learn effectively from new samples and affecting its adaptability to new scenarios. A suitable value needs to be determined through small-scale experiments or cross-validation. Among them, the prediction loss term is a measure of the difference between the model's predicted value and the actual lifetime value. The regularization penalty term is the sum of the square of the difference between the updated parameter value and the original parameter value for each model parameter in the conditional deep degradation network model, multiplied by the importance weight of that parameter, over all model parameters. The preset non-negative regularization strength coefficient is used to adjust the relative contribution of the regularization penalty to the total loss. By minimizing the online learning loss function, the parameters of the conditional deep degradation network model are updated to obtain the updated conditional deep degradation network model. The specific process of storing the updated parameters of the conditional deep degradation network model, the first scene encoding vector, and the degradation mode features as an associated record in the scene knowledge base is as follows: A temporal autoencoder (TAC) model is constructed to extract degradation patterns. The TAC model comprises an encoder network and a decoder network. The encoder network, typically composed of stacked one-dimensional convolutional layers or long short-term memory (LSTM) layers, maps the variable-length first health indicator sequence to a fixed-length low-dimensional vector. The decoder network, symmetrically structured, reconstructs the original-length sequence from this low-dimensional vector. The encoder and decoder are trained together to minimize the reconstruction error, thus forcing the intermediate low-dimensional vector, i.e., the degradation pattern feature, to capture the most representative temporal structure information of the input sequence. The first health indicator sequence from the loop closure verification sample is input into the encoder network of the TAC model. The encoder network outputs a low-dimensional vector, which represents the degradation pattern feature, characterizing the core temporal degradation pattern of the first health indicator sequence. Simultaneously, the degradation pattern feature is input into the decoder network of the TAC model, and the decoder network outputs a reconstructed sequence. By minimizing the reconstruction error between the reconstructed sequence and the original first health index sequence input to the temporal autoencoder model, the temporal autoencoder model is trained, enabling the degradation mode features to effectively compress and retain the key morphological information of the degradation trajectory. The reconstruction error is usually calculated using mean squared error. Through this unsupervised training method, the extracted degradation mode features can strip away random noise and short-term fluctuations in the health index sequence, focusing on characterizing the essential attributes of the overall degradation trend, such as the shape and curvature change rate, for example, whether it is a linear decline, exponential acceleration, or a clear inflection point. This feature is more expressive and discriminative than simple statistical values. From the closed-loop verification samples, the degradation mode features, the corresponding first scene encoding vector, and the recorded actual lifespan value are extracted. These information are interconnected and jointly characterize the correspondence between the specified degradation mode exhibited by the frameless torque motor and its final actual lifespan under the specified operating environment. The storage and guidance process of the scene knowledge base is as follows: the parameters of the updated conditional deep degradation network model, the first scene encoding vector corresponding to the loop closure verification sample, and the extracted degradation mode features are collectively treated as a knowledge record, associated with the actual lifetime value recorded in the loop closure verification sample, and stored in the pre-built scene knowledge base; the scene knowledge base uses the first scene encoding vector in each knowledge record as the index key; to achieve efficient retrieval, the scene knowledge base can be constructed using specialized vector database technology, which is optimized for similarity search of high-dimensional vectors, storing the first scene encoding vector in each knowledge record separately as an indexable vector. This involves establishing a mapping relationship between the new scenario and other data (model parameters, degradation mode features, and actual lifetime values) in the same record. When a new running scenario emerges and its corresponding first scenario encoding vector is generated, a cosine similarity-based nearest neighbor retrieval operation is performed in the scenario knowledge base. This involves calculating the cosine similarity between the first scenario encoding vector of the new scenario and all historical first scenario encoding vectors in the scenario knowledge base, sorting them from high to low similarity, and retrieving the K historical knowledge records with the highest similarity, where K is a positive integer. The value of K is a preset threshold, usually set according to the size of the knowledge base and the trade-off between guidance accuracy and efficiency. For example, it could be set to a threshold value of K. 5, 10, or 20, cosine similarity can effectively measure the directional consistency of the encoding vectors of two scenes in multidimensional space. The closer the directions, the more similar the comprehensive operating conditions of the two scenes are. Using the retrieved K historical knowledge records, new operating scenarios are supported. Support methods include model initialization support and prediction prior guidance support. Model initialization support refers to using the updated parameters of the conditional deep degradation network model associated with the historical knowledge record with the highest similarity in the search results as the initial parameters for training the conditional deep degradation network model built for the new operating scenario. This replaces random initialization, achieving a hot start for the model, thereby increasing... The fast model converges quickly in new scenarios; this hot-start mechanism utilizes the prior assumption that "in similar scenarios, the degradation patterns of motors and the parameters of the optimal prediction model should also have a certain similarity." Since the initial parameters are close to the optimal solution in the new scenario, the model only needs fewer training rounds and data to achieve good prediction performance, significantly reducing the model training cost and data requirements in new scenarios; the prediction prior guidance support refers to referring to the actual lifespan values ​​and degradation pattern characteristics associated with each of the K retrieved historical knowledge records to provide a prior trend reference for the expected lifespan range and a confidence interval reference for evaluating the credibility of the prediction results in the new scenario.For example, the median or weighted average of K historical actual lifespan values ​​(weighted by similarity) can be used as a rough reference for the expected lifespan in a new scenario. Simultaneously, analyzing the commonalities in the characteristics of the K historical degradation patterns (e.g., through clustering or calculating average features) can infer the possible degradation curve shape under the new scenario. Furthermore, the dispersion (e.g., standard deviation) of the K historical lifespan values ​​can be used to construct the confidence interval for the prediction results. Greater dispersion indicates potentially greater lifespan differences between similar scenarios, leading to higher prediction uncertainty and providing users with a more comprehensive basis for decision-making. The scenario knowledge base continuously expands with the accumulation of motor lifecycle data, and its retrieval accuracy and guidance effect on new scenarios also continuously improve.

[0024] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0025] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0026] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0027] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0028] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0029] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0030] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for full lifecycle management of frameless torque motors based on multiple scenarios, characterized in that, Specifically, the steps include the following: Step S1: Obtain multi-source heterogeneous data of the target frameless torque motor during operation. The multi-source heterogeneous data includes time-series sensing data collected by the motor body sensors, structured operation log data recorded by the device controller, and external environment data obtained through the Internet of Things gateway. Use a multi-modal scene perception and quantization coding model to fuse the multi-source heterogeneous data and generate a first scene coding vector representing the current operating scene. Step S2: Obtain the pre-trained meta-feature extraction model under multiple known scenarios, and use the time-series sensing data and the first scenario encoding vector to perform rapid fine-tuning of the meta-feature extraction model based on the scenario to obtain the target feature extraction model adapted to the target scenario. The target feature extraction model is used to process time-series sensor data and output the first health index sequence corresponding to the target frameless torque motor. Step S3: Construct a conditional deep degradation network model, using the first health indicator sequence as the main input sequence and the first scenario encoding vector as the conditional input; use the conditional deep degradation network model to jointly analyze the main input sequence and the conditional input to predict the conditional remaining service life of the target frameless torque motor under the current operating scenario. Step S4: When the target frameless torque motor reaches the end of its life or maintenance node, collect its complete life cycle data. The life cycle data includes at least multi-source heterogeneous data, the first health indicator sequence, the conditional remaining service life prediction value, and the recorded actual service life value, which together constitute a closed-loop verification sample. Using closed-loop validation samples, the parameters of the conditional deep degradation network model are updated based on an online learning strategy to obtain the updated conditional deep degradation network model; the degradation pattern contained in the first health indicator sequence in the closed-loop validation samples is extracted as the degradation mode feature. The parameters of the updated conditional deep degradation network model, the first scene encoding vector, and the degradation mode features are stored as an associated record in a pre-built scene knowledge base. The scene knowledge base is used for model initialization and prediction prior guidance for new running scenes.

2. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 1, characterized in that: In step S1, the timing sensing data includes at least the three-phase current data and winding temperature data of the motor, the structured operation log data includes at least the real-time torque command value sequence, speed value sequence and position value sequence, and the external environment data includes at least the ambient temperature data, relative humidity data and dust concentration data.

3. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 2, characterized in that: The process of fusing multi-source heterogeneous data using a multimodal scene perception and quantization coding model specifically includes the following operations performed sequentially: First, perform multimodal feature extraction: process the time-series sensor data, structured operation log data and external environment data to generate corresponding initial sensor feature vectors, initial log feature vectors and initial environment feature vectors respectively; Specifically, a one-dimensional convolutional neural network unit in the multimodal scene perception and quantization coding model is used to process the time-series sensing data to capture its local fluctuations and short-term dependency patterns, forming an initial sensing feature vector. A statistical feature aggregation unit in the same model calculates and combines the mean, peak value, and ripple amplitude of the torque command value sequence, speed value sequence, and position value sequence within a set time window to form an initial log feature vector. Finally, a standardization unit in the same model performs standardization transformation on the external environment data to form an initial environmental feature vector. Next, a feature concatenation operation is performed: the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector are concatenated to obtain an original mixed feature vector.

4. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 3, characterized in that: After obtaining the original mixed feature vector, the next step is to perform a dynamic fusion and encoding operation based on an attention mechanism to generate a first scene encoding vector. This operation specifically involves: An attention fusion unit in a multimodal scene perception and quantization coding model is used to generate a query vector based on the original mixed feature vector; An attention fusion unit is used to apply different learnable linear transformations to the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector to generate their respective key vectors and value vectors. Calculate the dot product between the query vector and each key vector to obtain multiple dot product results; Each dot product result is scaled by dividing by the square root of a predetermined dimension of the key vector; all scaled dot product results are then processed by a normalized exponential function to assign dynamic adaptive attention weight coefficients to the initial sensor feature vector, the initial log feature vector, and the initial environment feature vector. By using the adaptive attention weight coefficients assigned to each type of feature vector, the corresponding value vectors are weighted respectively, and all weighted value vectors are summed to obtain a primary fusion scene vector; Finally, the encoding generation unit in the multimodal scene perception and quantization coding model is used to input the primary fused scene vector into a multilayer perceptron. The multilayer perceptron performs nonlinear transformation and dimension normalization operations to generate a first scene encoding vector with a predetermined fixed dimension.

5. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 4, characterized in that: In step S2, a meta-feature extraction model pre-trained under multiple known scenarios is obtained. The pre-training process is as follows: We collected time-series sensing data samples and corresponding health status labels of frameless torque motors under multiple different operating scenarios. The data samples of each operating scenario constituted an independent meta-learning task. We used a model-independent meta-learning algorithm framework to train the meta-feature extraction model. In each training iteration, for each sampled meta-learning task, its data is divided into a support set and a query set; the temporal sensing data in the support set and the corresponding first scene encoding vector are input into the meta-feature extraction model, the support set loss is calculated based on the model output, and the model parameters are updated by gradient descent to obtain temporary adapted model parameters, wherein the first scene encoding vector is used to dynamically adjust the internal operation of the model. The temporal sensing data in the query set is forward-propagated using temporary adaptive model parameters, and the query set loss is calculated as the meta-loss. By adjusting the model parameters by minimizing the expected value of the meta-loss on all sampling tasks, a pre-trained meta-feature extraction model is obtained.

6. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 5, characterized in that: The process of rapidly fine-tuning the meta-feature extraction model based on the scene using time-series sensing data and the first scene encoding vector is as follows: A short-term time-series sensing data is collected from the target frameless torque motor as target scene support data, and the corresponding first scene encoding vector is obtained. The target scene support data and the first scene encoding vector are input into the meta-feature extraction model. The feature extraction focus of the meta-feature extraction model is adjusted by the first scene encoding vector. An adaptive loss is calculated, and the model parameters are updated by gradient descent a very small number of times using this adaptive loss to obtain the target feature extraction model. The very small number of times is one to five. Then, the complete time-series sensing data generated by the continuous operation of the target frameless torque motor is input into the target feature extraction model; The target feature extraction model outputs a single-value scalar for each time point or a data segment of a continuous time window in the time-series sensing data. This single-value scalar represents a summary of deep features related to the motor health status extracted from the data of the corresponding time period. In chronological order, all the output single-value scalars are concatenated to form the first health indicator sequence.

7. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 6, characterized in that: In step S3, the specific operation of constructing the conditional deep degradation network model and using the first scene encoding vector as the conditional input is as follows: The backbone of the conditional deep degradation network model is composed of multiple stacked temporal feature extraction units, which are used to process the first health indicator sequence. The first scene encoding vector is injected through a conditional normalization layer set in multiple network layers within the conditional deep degenerate network model; In each conditional normalization layer, the first scene encoding vector is input into two independent small parameter generation networks to generate scaling parameter vectors and translation parameter vectors respectively. Both scaling parameter vectors and translation parameter vectors have the same number of channels as the intermediate features to be processed in the current network layer. For the intermediate features input to this layer, they are first normalized, then multiplied element-wise with the scaling parameter vector, and then added element-wise with the translation parameter vector to output the features modulated by the scene conditions.

8. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 7, characterized in that: The specific process of using a conditional deep degradation network model to jointly analyze the main input sequence and conditional input to predict the conditional remaining useful life is as follows: During the training phase of the conditional deep degradation network model, optimization is performed by minimizing a multi-objective loss function; the multi-objective loss function consists of a weighted sum of a prediction loss term and a physical guidance loss term. The difference between the predicted conditional remaining useful life and the actual remaining useful life is calculated using the smoothed average absolute error loss function. The physical guidance loss term is used to constrain the latent degradation trajectory learned by the model that is related to the scene, so that it conforms to the physical prior. Its calculation process is as follows: From the data set used to train the conditional deep degradation network model, which includes time-series sensor data samples, health status labels and corresponding first scene encoding vectors, multiple first scene encoding vectors representing severe operating conditions are selected to form a severe scene sample set, and multiple first scene encoding vectors representing mild operating conditions are selected to form a mild scene sample set. The penalty term for each first-scene encoding vector in the harsh scene sample set is calculated as follows: first, the second derivative of the latent degradation trajectory of the conditional deep degradation network model with respect to time under the condition of the first-scene encoding vector is calculated, and then a linear rectified function is applied to the negative of the second derivative. The penalty term for each first-scene encoding vector in the mild scene sample set is calculated as follows: first, the second derivative of the latent degradation trajectory of the conditional deep degradation network model with respect to time under the condition of the first-scene encoding vector is calculated, and then a linear rectified function is applied to the second derivative itself. The physical guidance loss term is equal to the sum of the average of all penalty terms in the harsh scene sample set and the average of all penalty terms in the mild scene sample set.

9. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 8, characterized in that: In step S4, the process of updating the parameters of the conditional deep degradation network model based on an online learning strategy using closed-loop validation samples is as follows: An online learning strategy is implemented using an elastic weight consolidation algorithm, which assigns an importance weight to each model parameter in the conditional deep degenerate network model. Construct an online learning loss function, which consists of a weighted sum of a prediction loss term and a regularization penalty term; The predicted loss term is calculated based on the first health indicator sequence, the first scenario encoding vector, and the actual lifespan value in the closed-loop validation samples. The regularization penalty term is as follows: for each model parameter in the conditional deep degenerate network model, calculate the square of the difference between its updated value and its original value, multiply it by the importance weight of that parameter, and finally sum it over all parameters; The updated conditional deep degradation network model is obtained by updating the model parameters by minimizing the online learning loss function.

10. The method for full life-cycle management of frameless torque motors based on multiple scenarios as described in claim 9, characterized in that: The specific process of storing the updated parameters of the conditional deep degradation network model, the first scene encoding vector, and the degradation mode features as an associated record in the scene knowledge base is as follows: Construct a temporal autoencoder model, including an encoder network and a decoder network; The first health indicator sequence in the closed-loop verification sample is input into the encoder network, and the low-dimensional degradation pattern features are output. The degradation mode features, the corresponding first scene encoding vector, the actual lifetime value, and the parameters of the updated conditional deep degradation network model are stored as a knowledge record in the scene knowledge base. When a new running scenario appears, the cosine similarity between its first scenario encoding vector and all historical vectors is calculated in the scenario knowledge base, and the K historical knowledge records with the highest similarity are retrieved. Support is provided by using search results, including using model parameters associated with the most similar historical records as initial parameters for model training in new scenarios; In addition, the actual lifetime values ​​and degradation pattern characteristics in the reference retrieval records provide prior information for predicting new scenarios.

Images