A Cloud-Based Method and System for Detecting the Inertial Response of Asynchronous Motors

Abstract

This invention relates to the fields of motor control and cloud computing technology, specifically to a cloud-based method and system for detecting the inertia response of an asynchronous motor. The method includes: initializing a prior probability distribution based on preset parameters of the motor; generating optimal excitation signal parameters; synthesizing an excitation signal and injecting it into the motor based on the received optimal excitation signal parameters, synchronously collecting raw dynamic response data, generating a feature dataset; uploading the dataset to a cloud-based inertia inference center; calculating the posterior probability distribution; comparing the variance of the posterior probability distribution with a preset convergence accuracy threshold; if the variance is less than the convergence accuracy threshold, updating the controller parameters of the edge-side control execution unit based on the final inertia identification result; if the variance is not less than the convergence accuracy threshold, using the currently calculated posterior probability distribution as the prior probability distribution for the next iteration, and returning to the step of generating the optimal excitation signal parameters. This invention achieves high-precision parameter identification under extremely low disturbance conditions.

Description

Technical Field

[0001] This invention relates to the fields of motor control and cloud computing technology, specifically to a cloud-based method and system for detecting the inertia response of asynchronous motors. Background Technology

[0002] In modern industrial automation, asynchronous motors are widely used as core power components, and their control performance directly affects the accuracy and efficiency of the entire system. The moment of inertia of a motor is a key parameter determining its dynamic response characteristics; however, in practical applications, changes in load often cause dynamic changes in the total moment of inertia of the motor. To achieve high-performance control of the motor, it is essential to be able to acquire its current moment of inertia online, quickly, and accurately.

[0003] Traditional methods for identifying motor inertia typically rely on injecting specific disturbance signals and analyzing the system's response. These methods generally suffer from inherent technical contradictions: on the one hand, to obtain high identification accuracy, a disturbance signal with sufficiently strong energy needs to be injected, but this inevitably interferes with the smoothness of the motor's normal operation; on the other hand, if the disturbance signal is weakened to ensure operating performance, the signal-to-noise ratio of the response signal will decrease, thereby affecting the identification accuracy and convergence speed. At the same time, existing identification strategies mostly use fixed or preset detection signals and lack the ability to adaptively adjust according to the degree of uncertainty in the identification process, resulting in low information acquisition efficiency.

[0004] With the development of cloud computing technology, although complex optimization and identification algorithms can be deployed in the cloud to utilize its powerful computing capabilities, this application model faces severe challenges in industrial real-time control scenarios. Communication latency and bandwidth limitations between the cloud and field motor controllers have become the main technical bottlenecks restricting the implementation of intelligent algorithms. Uploading raw data containing high-frequency dynamic responses from the edge to the cloud for analysis not only places extremely high demands on the network, but the communication latency also makes it difficult to achieve real-time closed-loop control based on cloud computing results.

[0005] Therefore, how to provide an online inertia detection method for asynchronous motors that can coordinate the detection of disturbances and operational performance, balance identification speed and accuracy, and effectively solve the problem of implementing remote intelligent algorithms in industrial real-time control scenarios is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention discloses a cloud computing-based method and system for detecting the inertia response of asynchronous motors. Specifically, the technical solution of this invention is as follows:

[0007] A cloud-based method for detecting the inertia response of asynchronous motors includes:

[0008] In the cloud-based inertia inference center, the prior probability distribution is initialized based on the motor's preset parameters;

[0009] At the cloud-based inertia inference center, the optimal excitation signal parameters are generated based on the prior probability distribution of the current round.

[0010] The optimal excitation signal parameters are sent to the edge-side control execution unit;

[0011] The edge-side control execution unit synthesizes an excitation signal and injects it into the motor based on the received optimal excitation signal parameters, and simultaneously collects the original dynamic response data;

[0012] The edge-side control execution unit preprocesses the raw dynamic response data to generate a feature dataset.

[0013] Upload the feature dataset to the cloud-based inertia inference center;

[0014] In the cloud-based inertia inference center, the posterior probability distribution is calculated by Bayesian update by combining the feature dataset with the prior probability distribution of the current round.

[0015] Extract the variance of the posterior probability distribution and compare it with a preset convergence accuracy threshold;

[0016] If the variance is less than the convergence accuracy threshold, the mean of the posterior probability distribution is extracted as the final inertia identification result, and the controller parameters of the edge-side control execution unit are updated based on the final inertia identification result.

[0017] If the variance is not less than the convergence accuracy threshold, then the currently calculated posterior probability distribution is used as the prior probability distribution for the next iteration, and the process returns to the step of generating the optimal excitation signal parameters.

[0018] Preferably, initializing the prior probability distribution includes:

[0019] Experienced knowledge of the motor's nameplate parameters and mechanical structure;

[0020] Based on the acquired nameplate parameters and empirical knowledge, a Gaussian distribution is constructed as the prior probability distribution;

[0021] The mean of the Gaussian distribution is calculated based on the parameters of the motor and the no-load system, and its initial variance is used to characterize the uncertainty of the inertia value in the initial state.

[0022] Preferably, generating the optimal excitation signal parameters includes:

[0023] Construct a constrained optimization problem with the objective of maximizing the expected information gain;

[0024] Under the constraints of satisfying the preset maximum allowable torque ripple and maximum detection time, the constraint optimization problem is solved to generate the optimal excitation signal parameters.

[0025] Preferably, the expected information gain is quantified by the determinant of the Fisher information matrix, and the torque ripple is evaluated by a pre-constructed mapping function.

[0026] Preferably, the raw dynamic response data is preprocessed, including:

[0027] The original dynamic response data is digitally filtered;

[0028] The feature dataset is generated by extracting the velocity response amplitude and phase information at the excitation signal frequency from the filtered data.

[0029] Preferably, the posterior probability distribution is calculated using Bayesian update, including:

[0030] Construct a likelihood function to quantify the probability of observing the feature dataset given the inertia assumption;

[0031] The posterior probability distribution is calculated by combining the likelihood function and the prior probability distribution.

[0032] Preferably, constructing the likelihood function includes:

[0033] Using a built-in digital twin model of the motor in the cloud, the output response is predicted under a given assumption of inertia.

[0034] Calculate the residual between the predicted output response and the feature dataset;

[0035] The likelihood function is determined based on the residual.

[0036] Preferably, the method performs cloud-edge communication through parameterized signaling and characteristic data, including:

[0037] The cloud-based inertia inference center only sends out the parameters of the optimal excitation signal;

[0038] The edge-side control execution unit only uploads the feature dataset.

[0039] Preferably, the convergence accuracy threshold is set according to the requirements of the application scenario for the accuracy of the controller parameters.

[0040] A cloud-based asynchronous motor inertia response detection system includes:

[0041] The model initialization module is used to initialize the prior probability distribution based on the preset parameters of the motor.

[0042] The incentive design module is used to generate optimal incentive signal parameters based on the prior probability distribution of the current round;

[0043] The excitation injection and acquisition module is used to synthesize an excitation signal based on the optimal excitation signal parameters and inject it into the motor, while simultaneously acquiring the original dynamic response data.

[0044] The data preprocessing module is used to preprocess the original dynamic response data to generate a feature dataset;

[0045] The Bayesian inference module is used to combine the feature dataset with the prior probability distribution of the current round to calculate the posterior probability distribution.

[0046] The convergence decision and control module is used to determine whether the variance of the posterior probability distribution is less than a preset convergence accuracy threshold. If it is less, the final inertia identification result is generated to update the controller parameters. If it is not less, the excitation design module is triggered to enter the next iteration.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. This invention constructs a constrained optimization problem with the goal of maximizing expected information, and can proactively design the optimal excitation and detection signal while satisfying the constraints of stable system operation. This solves the inherent contradiction between detection disturbance and motor operating performance in traditional methods, and realizes high-precision parameter identification under extremely low disturbance.

[0049] 2. This invention utilizes a cloud-edge collaborative architecture to construct a self-optimizing closed-loop identification process from uncertainty assessment, optimal detection, data fusion to convergence decision; by incorporating prior engineering knowledge into the identification starting point and combining it with Bayesian iterative updates, the convergence speed of the identification process is significantly accelerated, enabling rapid online identification of motor inertia under dynamically changing loads.

[0050] 3. By adopting an asymmetric communication strategy of parameterized signaling downlink and characteristic data uplink, this invention minimizes the amount of data interaction between the cloud and the edge. This design effectively overcomes the challenges of limited network bandwidth and unstable communication latency in industrial sites, and solves the application problem of remote intelligent algorithms in industrial real-time control scenarios.

[0051] 4. The convergence accuracy of the present invention can be flexibly set according to the specific requirements of the controller parameter accuracy in different application scenarios; this mechanism of setting accuracy on demand enables the identification resources to be used optimally, avoids unnecessary detection and time waste, and achieves a precise match between identification efficiency and application requirements while ensuring that the final control performance is met. Attached Figure Description

[0052] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0053] Figure 1 This is a flowchart of the method of the present invention;

[0054] Figure 2 This is a flowchart of the system of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. 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.

[0056] Example 1:

[0057] Please see Figure 1 A cloud computing-based method for detecting the inertia response of an asynchronous motor, comprising:

[0058] In the cloud-based inertia inference center, the prior probability distribution is initialized based on the motor's preset parameters;

[0059] At the cloud-based inertia inference center, the optimal excitation signal parameters are generated based on the prior probability distribution of the current round.

[0060] The optimal excitation signal parameters are sent to the edge-side control execution unit;

[0061] The edge-side control execution unit synthesizes an excitation signal and injects it into the motor based on the received optimal excitation signal parameters, while simultaneously acquiring the original dynamic response data;

[0062] At the edge-side control execution unit, the raw dynamic response data is preprocessed to generate a feature dataset;

[0063] Upload the feature dataset to the cloud-based inertia inference center;

[0064] In the cloud-based inertia inference center, the posterior probability distribution is calculated by combining the feature dataset with the prior probability distribution of the current round through Bayesian updates.

[0065] Extract the variance of the posterior probability distribution and compare it with a preset convergence accuracy threshold;

[0066] If the variance is less than the convergence accuracy threshold, the mean of the posterior probability distribution is extracted as the final inertia identification result, and the controller parameters of the edge-side control execution unit are updated based on the final inertia identification result.

[0067] If the variance is not less than the convergence accuracy threshold, the currently calculated posterior probability distribution is used as the prior probability distribution for the next iteration, and the process returns to the step of generating the optimal excitation signal parameters.

[0068] A cloud-based asynchronous motor inertia response detection method aims to utilize a cloud-edge collaborative architecture to construct a self-optimizing closed-loop identification process, thereby achieving online, adaptive, fast, and accurate motor rotational inertia identification. In this embodiment, the method is deployed in a system consisting of a cloud-based inertia inference center and an edge-side control execution unit. The cloud-based inertia inference center refers to a computing center with powerful computing capabilities deployed on a cloud server, which performs complex optimization calculations and Bayesian inference tasks. The edge-side control execution unit refers to a local controller or industrial computer located close to the physical entity of the asynchronous motor, which performs real-time control command injection, high-frequency data acquisition, and preprocessing.

[0069] The complete closed-loop iterative process of this method is implemented in this embodiment through the following steps:

[0070] At the cloud-based inertia inference center, the prior probability distribution is initialized based on the motor's preset parameters. This step provides a mathematical starting point for the entire Bayesian iterative process. The prior probability distribution quantifies the system's initial understanding or degree of uncertainty regarding the true value of the motor's inertia when there is no dynamic test data.

[0071] Entering the iterative loop, at the cloud-based inertia inference center, the optimal excitation signal parameters are generated based on the prior probability distribution of the current round. This step is the core embodiment of the active information exploration idea of ​​this invention. Its purpose is not to use fixed or random perturbation signals, but to adaptively design a detection signal that is expected to bring the maximum information based on the current uncertainty of inertia and described by the prior probability distribution. This signal is defined by a set of key parameters, rather than a complete signal waveform.

[0072] The optimal excitation signal parameters are sent to the edge control execution unit. The communication here is parameterized signaling with a very small amount of data. For example, it only contains a few values ​​such as the amplitude, frequency, and duration of the signal, which effectively reduces the requirements for cloud-edge communication bandwidth and avoids the impact of latency on real-time control.

[0073] The edge-side control execution unit synthesizes an excitation signal and injects it into the motor based on the received optimal excitation signal parameters, and synchronously collects the original dynamic response data; the edge-side controller is responsible for synthesizing the received parameters into a complete, high-resolution excitation signal waveform in real time, and superimposing it into the control command for normal motor operation; at the same time, the data acquisition module synchronously records the dynamic response of the motor under this excitation at a high sampling frequency, such as the timing data of angular velocity and armature current, to form the original dynamic response data;

[0074] After the data acquisition is completed, the edge-side control execution unit preprocesses the raw dynamic response data to generate a feature dataset. The raw dynamic response data refers to high-frequency sampled, unprocessed time-series data with a large data volume. The feature dataset refers to key information extracted from the raw data that is strongly correlated with the inertia parameter and whose data volume has been significantly compressed. This step aims to further reduce uplink bandwidth pressure and remove irrelevant information such as noise.

[0075] Uploading the feature dataset to the cloud inertia inference center, since it is highly condensed feature data rather than raw time-series data, also achieves high efficiency in cloud-edge communication.

[0076] After receiving the feature data in the cloud, the cloud inertia inference center combines the feature dataset with the prior probability distribution of the current round and calculates the posterior probability distribution through Bayesian update. This step is the core of knowledge iteration. It uses Bayes' theorem to integrate newly acquired experimental evidence with previous knowledge, thereby obtaining an updated posterior probability distribution with lower uncertainty.

[0077] The convergence decision is made by extracting the variance of the posterior probability distribution and comparing it with a preset convergence accuracy threshold. The variance of the posterior probability distribution is a quantitative indicator of the uncertainty of the current inertia estimate. The smaller the variance, the higher the confidence of the estimate. The convergence accuracy threshold is a preset standard used to judge whether the identification accuracy meets the application requirements.

[0078] Based on the comparison results, different operations are performed: if the variance is less than the convergence accuracy threshold, the mean of the posterior probability distribution is extracted as the final inertia identification result, and the controller parameters of the edge-side control execution unit are updated based on the final inertia identification result; this marks the successful termination of the identification process, and the identification result is applied to the actual motor control to improve system performance; if the variance is not less than the convergence accuracy threshold, the currently calculated posterior probability distribution is used as the prior probability distribution for the next iteration, and the process returns to the step of generating the optimal excitation signal parameters; this forms a closed loop, and the system will use the updated understanding to design the next more targeted detection experiment until the accuracy requirements are met;

[0079] The method disclosed in this embodiment overcomes the contradictions between detection disturbance and operational performance, and between identification speed and accuracy in traditional methods by constructing a cloud-edge collaborative, active detection, and closed-loop iterative identification framework. It can dynamically optimize the detection strategy based on its own knowledge state, achieving rapid, accurate, and low-disturbance online identification of motor inertia under dynamically changing loads. Furthermore, it solves the problem of implementing remote intelligent algorithms in industrial real-time control scenarios through an efficient cloud-edge communication strategy. Currently, this method mainly focuses on modeling scenarios where load characteristics are mainly rotational inertia. For loads with complex characteristics such as significant nonlinear friction, this identification framework also has the potential for expansion. Synchronous identification can be achieved by adding a corresponding friction coefficient to the parameters to be identified and adjusting the digital twin model and likelihood function.

[0080] Example 2:

[0081] Initialize the prior probability distribution, including:

[0082] Experienced knowledge of the motor's nameplate parameters and mechanical structure;

[0083] Based on the obtained nameplate parameters and empirical knowledge, a Gaussian distribution is constructed as the prior probability distribution;

[0084] The mean of the Gaussian distribution is calculated based on the parameters of the motor and the no-load system, and its initial variance is used to characterize the uncertainty of the inertia value under the initial state.

[0085] In the method of Example 1, the specific way to initialize the prior probability distribution is as follows:

[0086] The purpose of this step is to provide a starting point for Bayesian iteration that is as information-rich as possible and in line with physical reality, rather than a uniform distribution with no information at all; in this embodiment, this step is achieved by integrating prior engineering knowledge.

[0087] Specifically, this involves acquiring the nameplate parameters and mechanical structure experience of the motor. Nameplate parameters refer to the rated parameters marked on the equipment when the motor leaves the factory, such as rated power, rated speed, rotor resistance, etc., which can be obtained by directly reading the physical label of the motor or consulting its specification manual. Mechanical structure experience refers to the known information of the load system connected to the motor, such as the transmission ratio, the approximate range of the basic load inertia, etc., which is usually obtained from equipment design documents or the engineering experience of senior engineers.

[0088] Based on the acquired nameplate parameters and empirical knowledge, a Gaussian distribution is constructed as the prior probability distribution. The Gaussian distribution is chosen because of its favorable mathematical properties, ease of Bayesian conjugate updates, and ability to intuitively express the expected value and uncertainty of the parameters through the mean and variance. The specific form of this Gaussian distribution is as follows:

[0089] ;

[0090] in, The initial prior probability distribution is a probability density function describing the possible values ​​of the moment of inertia J before any measurement is performed; it originates from the calculation results of this step. J is the moment of inertia, a core physical parameter to be identified, with units of 1000 ppm. ;

[0091] The mean of the Gaussian distribution is calculated based on the parameters of the motor and the no-load system, and its initial variance is used to characterize the uncertainty of the inertia value under the initial state. Specifically... The initial mean of the Gaussian distribution represents the initial best guess of the inertia value, which is derived from the theoretical inertia value calculated based on the motor nameplate parameters and known no-load or light-load system information. The initial variance is a Gaussian distribution, representing the degree of uncertainty regarding the inertia value in the initial state. It is determined empirically; for example, a larger initial variance is set if the load varies greatly, and a smaller value is set if the load varies less. This ensures that the initial model can accommodate the large deviations that may exist in the actual inertia value.

[0092] Initializing the prior probability distribution in this way, compared to using uninformative priors such as uniform distribution, allows known physical information and engineering experience to be incorporated into the starting point of identification. This provides a better starting search region for subsequent iterations, thereby reducing unnecessary exploration, significantly accelerating the convergence speed of the entire identification process, and improving the overall efficiency of the method.

[0093] Example 3:

[0094] Generate optimal excitation signal parameters, including:

[0095] Construct a constrained optimization problem with the objective of maximizing expected information;

[0096] Under the constraints of satisfying the preset maximum allowable torque ripple and maximum detection time, solve the constraint optimization problem to generate the optimal excitation signal parameters;

[0097] Expected information is quantified using the determinant of the Fisher information matrix, and torque ripple is evaluated using a pre-built mapping function.

[0098] In the method of Example 1, the step of generating the optimal excitation signal parameters is the key to maximizing active information. Its specific implementation involves constructing and solving a constrained optimization problem with the goal of maximizing expected information. This step works in conjunction with the technical features of information quantification and constraint evaluation methods.

[0099] To further clarify, the purpose of constructing a constrained optimization problem with the objective of maximizing expected information is to find a set of excitation signal parameters, such as the amplitude of a sinusoidal excitation. and frequency This makes it possible to have a current understanding of inertia. Based on this understanding, we expect to obtain information about [the next measurement] from [the previous measurement]. The amount of information reaches its maximum;

[0100] Expected information is quantified using the determinant of the Fisher information matrix. It is a measure used in statistics and systems identification to measure how much information about unknown parameters can be obtained through a set of measurements. The core tool for information content, its value is related to the system response and inertia. The sensitivity is positively correlated, and its origin is calculated based on a cloud-based dynamic mathematical model of the motor. Specifically, its elements can be obtained from the output of the motor model regarding the parameters. The integral of the sensitivity is used for calculation; its determinant is... The size of the parameter directly reflects the effect of the identification experiment on the parameter. The distinguishability; therefore, the optimization objective is constructed as maximizing the mathematical expectation of the determinant of the Fisher information matrix over the current inertia probability distribution;

[0101] Meanwhile, torque ripple is evaluated using a pre-built mapping function. In actual operation, the injected excitation signal will cause additional ripple in the motor torque, and excessive ripple will affect the smooth operation of the system. In order to efficiently evaluate this effect during the optimization process, this embodiment introduces a mapping function. This function refers to a low-computational-cost approximation model, such as a polynomial or neural network, which directly maps the parameters of the excitation signal to a scalar value that approximates the square of the torque ripple norm it causes. This value is constructed before system deployment through offline simulation or fitting of experimental data from a limited number of trials. To ensure the effectiveness of the optimization results, this mapping function needs to be fully sampled and verified throughout the entire feasible domain of the excitation parameters during its construction to ensure that its approximation error is within an acceptable range. The accuracy of this function will directly affect the effectiveness of the constraint boundary; lower accuracy may cause the actual torque ripple to slightly exceed or fall short of the preset threshold.

[0102] In summary, this constrained optimization problem can be formally expressed as:

[0103] ;

[0104] And satisfy the following constraints: and ;

[0105] in: The optimal excitation signal parameters are the solution results of this optimization problem and will be sent to the edge side.

[0106] Let be the mathematical expectation operator, representing all historical observation data up to the (k-1)th round. Under the given conditions, obtain the objective function with respect to inertia. posterior probability distribution The average value, its function is to Given that the true value is unknown, find one among all possible values. The excitation signal with the best overall performance in terms of value;

[0107] The motor model parameters, such as stator resistance and flux linkage, are obtained from offline calibration or as input as known parameters.

[0108] The square of the maximum permissible torque ripple norm is an external constraint threshold set according to the operational stability requirements of the specific application scenario.

[0109] The maximum detection duration is an external constraint threshold set according to the application's requirements for real-time identification.

[0110] Solving the constrained optimization problem to generate the optimal excitation signal parameters; the cloud-based inertia inference center utilizes its computing power and employs numerical optimization algorithms such as sequential quadratic programming to solve the above problem, thereby obtaining the optimal combination of excitation signal parameters. ;

[0111] This embodiment solves the problem of intelligent questioning by actively designing excitation signals through constraint optimization. It also provides a concrete and calculable implementation path: on the one hand, it uses the Fisher information matrix determinant to provide a solid mathematical quantitative basis for the abstract goal of information; on the other hand, it uses a pre-constructed mapping function to evaluate torque ripples, cleverly circumventing the challenge of high-cost dynamic simulation during optimization iterations. The combination of these two approaches enables the system to design the most informative detection signal with extremely high efficiency while meeting key performance constraints, thereby greatly improving the convergence speed and smoothness of identification, and achieving a harmonious unity between detection behavior and main task performance.

[0112] Example 4:

[0113] Preprocessing of the raw dynamic response data includes:

[0114] Digital filtering is applied to the raw dynamic response data;

[0115] The velocity response amplitude and phase information at the excitation signal frequency are extracted from the filtered data to generate a feature dataset.

[0116] In the method of Example 1, the step of preprocessing the raw dynamic response data is performed in the edge-side control execution unit. Its core purpose is to refine the data, that is, to extract the most effective and concise feature information for inertia identification from the massive amount of raw data before uploading it to the cloud, so as to reduce the occupation of cloud-edge communication bandwidth.

[0117] The specific implementation includes the following steps:

[0118] First, digital filtering is performed on the raw dynamic response data, which is typically the motor angular velocity containing high-frequency noise. and armature current The purpose of digital filtering for timing signals is to remove noise introduced by measurement sensors or electromagnetic interference. In this embodiment, a low-pass filter or a band-pass filter can be used, and the cutoff frequency of the filter is determined by the frequency of the injected excitation signal. Adaptive settings are implemented to suppress noise to the greatest extent possible while preserving effective signal response;

[0119] Second, the amplitude and phase information of the velocity response at the frequency of the excitation signal are extracted from the filtered data to generate a feature dataset; since a specific frequency is injected... For sinusoidal excitation, according to linear system theory, the amplitude and phase of the frequency component in the steady-state response of the system output, relative to the input signal, are directly related to the system's transfer function, while the moment of inertia... These are the key parameters of the transfer function; therefore, these two eigenvalues ​​contain information about... The rich information; in this embodiment, the frequency of the filtered velocity signal can be accurately extracted from the filtered velocity signal using Fast Fourier Transform (FFT) or lock-in amplifier algorithms. The response amplitude and phase at the point are used as the final feature dataset, with these two scalar values ​​(or their complex representations) serving as the final feature set.

[0120] This preprocessing step compresses high-frequency time-series data, which may originally contain thousands of data points, into a feature dataset containing only a few key values. This not only greatly reduces the demand for uplink communication bandwidth, making the method applicable to bandwidth-constrained industrial wireless network environments, but also improves the signal-to-noise ratio of the uploaded data through filtering and feature extraction. This allows Bayesian inference in the cloud to be based on cleaner and more effective information, thereby improving the robustness and accuracy of identification.

[0121] Example 5:

[0122] The posterior probability distribution is calculated using Bayesian updates, including:

[0123] Construct a likelihood function to quantify the probability of observing the feature dataset given the inertia assumption;

[0124] By combining the likelihood function and the prior probability distribution, the posterior probability distribution can be calculated.

[0125] Constructing the likelihood function includes:

[0126] Using a built-in digital twin model of the motor in the cloud, the output response is predicted under a given assumption of inertia.

[0127] Calculate the residual between the predicted output response and the feature dataset;

[0128] The likelihood function is determined based on the residuals.

[0129] In the method of Example 1, the mathematical core of realizing knowledge iteration and uncertainty reduction is to calculate the posterior probability distribution through Bayesian update; this step is executed in the cloud inertia inference center and is closely integrated with the likelihood function construction method.

[0130] The purpose of this process is to transfer the feature dataset uploaded from the edge side. The new information contained herein, along with the system's existing knowledge, namely the prior probability distribution, The fusion is performed; this is achieved through Bayes' theorem:

[0131] ;

[0132] in, This is the target output of this step, namely the posterior probability distribution; the key to completing this calculation lies in how to construct... , i.e., the likelihood function;

[0133] Construct a likelihood function to quantify the value of a given inertia hypothesis. The observed feature dataset The probability of the prior probability distribution is calculated by combining the likelihood function with the prior probability distribution.

[0134] The process of constructing the likelihood function is as follows:

[0135] Using a built-in digital twin model of the motor in the cloud, the output response is predicted under a given inertia assumption. The digital twin model of the motor refers to a mathematical model running in the cloud that can accurately simulate the dynamic behavior of a real asynchronous motor; this model includes a variable inertia parameter. In addition, its static parameters can be pre-input through offline calibration; for any assumed inertia value This model can simulate and predict the corresponding characteristic responses. For example, the predicted velocity response amplitude and phase;

[0136] It should be noted that the model in this embodiment assumes that the motor parameters other than the moment of inertia are precisely known constants. In practical applications, if these parameters change significantly in real time, such as due to temperature, it may affect the identification accuracy. In this case, an online parameter correction algorithm can be used to improve the model's fidelity.

[0137] Calculate the residual between the predicted output response and the feature dataset; upload the true measurement features from the edge side. With the assumption of inertia The model predicts features By comparison, the residuals are obtained. ;

[0138] The likelihood function is determined based on the residuals; the residuals Primarily caused by measurement noise and model uncertainty, it can usually be reasonably assumed that it follows a zero-mean Gaussian distribution with a covariance of . , Representing the noise level, therefore, the likelihood function, i.e., at inertia of Observed at time The probability of can be expressed as:

[0139] ;

[0140] If a hypothetical value of inertia This enables the model's predicted response to closely approximate the actual measured value, i.e., the residual. It's very small, so this The probability of it being a positive value is high; conversely, the probability of it being a negative value is low.

[0141] After obtaining the specific expression for the likelihood function, the cloud-based inference module can multiply it by the prior distribution and calculate the final posterior probability distribution using numerical methods such as Markov chain Monte Carlo or grid integration. ;

[0142] This embodiment provides a clear path for deeply integrating physical models with probabilistic inference. By constructing a likelihood function based on a digital twin model, the abstract Bayesian formula can be closely associated with the specific dynamic behavior of the motor. This method not only provides a solid physical foundation for the inertia identification process, but also, through a probabilistic framework, can naturally handle measurement noise and model uncertainty, making the identification results more robust and reliable. It combines data-driven updates with model-driven predictions to achieve accurate inference of the motor state.

[0143] Example 6:

[0144] The method enables cloud-edge communication through parameterized signaling and characteristic data, including:

[0145] The cloud-based inertia inference center only sends out the parameters of the optimal excitation signal;

[0146] The edge-side control execution unit only uploads the feature dataset.

[0147] In the method of Embodiment 1, in order to optimize the performance of the cloud-edge collaborative architecture, the present invention has made special design at the communication layer, that is, the method performs cloud-edge communication through parameterized signaling and characteristic data;

[0148] This design is reflected in two aspects:

[0149] In terms of downlink communication, the cloud-based inertia inference center only sends the parameters of the optimal excitation signal; the cloud-based system, through complex optimization calculations, does not obtain a long sequence of signal waveform data, but only the key parameters defining this waveform, such as... These parameters have a very small data size, usually only a few floating-point numbers;

[0150] In terms of uplink communication, the edge-side control execution unit only uploads the feature dataset. As in the implementation of Example 4, the edge side preprocesses the massive amount of raw dynamic response data locally and extracts core features such as response amplitude and phase at a specific frequency; only these compressed and refined feature values ​​are uploaded.

[0151] This asymmetric communication strategy, characterized by downlink parameterized signaling and uplink characteristic data, is one of the keys to the successful application of this invention in real-world industrial environments. It precisely matches the capabilities of the cloud and edge: complex optimization and inference tasks are placed in the cloud, while real-time control and data acquisition tasks are placed at the edge, and the information interaction between the two is designed to be as simple and efficient as possible. This strategy minimizes the amount of data in cloud-edge communication, effectively overcoming the challenges of limited network bandwidth and unstable communication latency that may exist in industrial settings, and ensuring the real-time performance and feasibility of the entire closed-loop identification system.

[0152] Example 7:

[0153] The convergence accuracy threshold is set according to the requirements of the application scenario for the accuracy of the controller parameters.

[0154] In the method of Example 1, the termination condition of the iterative process is determined by the convergence accuracy threshold, which directly affects the accuracy of the final identification result and the time consumed by the entire identification process.

[0155] In this embodiment, the convergence accuracy threshold is set according to the application scenario's requirements for controller parameter accuracy. This is a function-driven rather than fixed setting logic; the specific setting process is as follows:

[0156] Analyze the specific application scenarios served by the motor; for example, a servo motor used for the feed axis of a CNC machine tool has extremely high requirements for the accuracy of its inertia parameters, because inertia mismatch will directly lead to machining trajectory errors and vibrations; while the control requirements for a motor used to drive a fan are relatively relaxed.

[0157] Through controller theory analysis or system simulation, the sensitivity of controller parameters to inertia identification error is determined. For example, it might be concluded that to ensure the dynamic response performance of the system, such as overshoot less than 5% and settling time less than 50ms, the identification error of the inertia parameters must be controlled within a certain range. within;

[0158] This performance requirement is translated into a variance threshold for the posterior probability distribution. The method for determining this is, for example, if the ratio of the standard deviation to the mean of the identification results is required to be less than 2%, then a setting can be made. When the calculated posterior variance Smaller than this When the recognition accuracy meets the application requirements, the iteration ends.

[0159] By linking the convergence accuracy threshold to specific application requirements, this method becomes highly adaptable. It avoids unnecessary detection and time waste caused by setting the threshold too high, and also avoids insufficient identification accuracy and impact on final control performance caused by setting the threshold too low. This mechanism of setting accuracy on demand allows identification resources to be used optimally, completing the identification task as quickly as possible while meeting performance requirements.

[0160] Example 8:

[0161] Please see Figure 2 A cloud-based asynchronous motor inertia response detection system includes:

[0162] The model initialization module is used to initialize the prior probability distribution based on the preset parameters of the motor.

[0163] The incentive design module is used to generate optimal incentive signal parameters based on the prior probability distribution of the current round.

[0164] The excitation injection and acquisition module is used to synthesize an excitation signal based on the optimal excitation signal parameters and inject it into the motor, while simultaneously acquiring the original dynamic response data.

[0165] The data preprocessing module is used to preprocess the raw dynamic response data to generate a feature dataset;

[0166] The Bayesian inference module is used to combine the feature dataset with the prior probability distribution of the current round to calculate the posterior probability distribution.

[0167] The convergence decision and control module is used to determine whether the variance of the posterior probability distribution is less than the preset convergence accuracy threshold. If it is less, the final inertia identification result is generated to update the controller parameters. If it is not less, the excitation design module is triggered to enter the next iteration.

[0168] A cloud-based asynchronous motor inertia response detection system is a combination of hardware and software designed to perform the above-described method. It logically or physically comprises the following cooperative modules:

[0169] The model initialization module is deployed in the cloud inertia inference center. Its function is to initialize the prior probability distribution based on the preset parameters of the motor. This module receives the input motor nameplate parameters and mechanical structure experience knowledge, and builds an initial Gaussian probability distribution model accordingly, providing a starting point for the entire identification process.

[0170] The excitation design module is deployed in the cloud-based inertia inference center. As the core decision-making unit of the system, this module is used to generate the optimal excitation signal parameters based on the prior probability distribution of the current cycle. This module has a built-in digital twin model of the motor and a constraint optimization solver, which performs calculations with the goal of maximizing expected information and outputs the optimal excitation signal parameters.

[0171] The excitation injection and acquisition module is deployed on the edge control execution unit. As the physical interface and sensing unit of the system, this module is used to synthesize excitation signals based on the optimal excitation signal parameters and inject them into the motor, and synchronously acquire raw dynamic response data. This module is responsible for translating cloud-based decisions into actions in the physical world and sensing the results of those actions.

[0172] The data preprocessing module is deployed on the edge control execution unit. Its function is to preprocess the raw dynamic response data to generate refined feature data for uploading. This module refines the massive amount of raw data collected locally through digital filtering and feature extraction algorithms.

[0173] The Bayesian inference module is deployed in the cloud-based inertia inference center. As the core inference engine of the system, this module is used to combine the feature dataset with the prior probability distribution of the current round to calculate the posterior probability distribution. Through Bayesian update rules, this module integrates new data evidence into the existing knowledge system to achieve iterative deepening of the understanding of inertia.

[0174] The convergence decision and control module is deployed in the cloud-based inertia inference center. As the main control unit of the entire closed-loop process, this module is used to determine whether the variance of the posterior probability distribution is less than the preset convergence accuracy threshold. If it is less, the final inertia identification result is generated to update the controller parameters, and an update command is sent to the edge side. If it is not less, the stimulus design module is triggered to enter the next iteration, and the posterior probability distribution calculated in this round is passed to the stimulus design module as the prior for the next round. This module is responsible for terminating or continuing the decision process.

[0175] This system clearly implements the method of Example 1 in a modular manner. Each module has a clear function and performs its own duties. Through cloud-edge collaborative deployment, computing tasks and real-time tasks are reasonably allocated, giving full play to the advantages of the powerful computing power of the cloud and the low-latency control of the edge. This system architecture ensures that the entire inertial response detection method can operate efficiently, stably and reliably, providing a complete system-level solution for realizing online adaptive motor control in complex industrial environments.

[0176] This system exhibits a certain degree of robustness: Firstly, based on the iterative characteristics of the Bayesian framework, even with significant deviations in the initial prior distribution, the identification results can gradually converge to near the true value through multiple rounds of effective data updates; secondly, the digital filtering and feature extraction steps on the edge side can effectively suppress measurement noise at a normal level. However, under extreme conditions, such as extremely low signal-to-noise ratios or severe mismatches in motor model parameters, the convergence speed and final accuracy of the identification may be affected. Furthermore, for cases where the inertia value to be identified is at the maximum or minimum boundary, special configuration of the optimization solver is required to prevent numerical computation problems. Before actual deployment, the applicable boundaries of the system should be calibrated through sufficient simulations and experiments.

[0177] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0178] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A cloud computing-based method for detecting the inertia response of an asynchronous motor, characterized in that, include: In the cloud-based inertia inference center, the prior probability distribution is initialized based on the motor's preset parameters; At the cloud-based inertia inference center, the optimal excitation signal parameters are generated based on the prior probability distribution of the current round. The optimal excitation signal parameters are sent to the edge-side control execution unit; The edge-side control execution unit synthesizes an excitation signal and injects it into the motor based on the received optimal excitation signal parameters, and simultaneously collects the original dynamic response data; The edge-side control execution unit preprocesses the raw dynamic response data to generate a feature dataset. Upload the feature dataset to the cloud-based inertia inference center; In the cloud-based inertia inference center, the posterior probability distribution is calculated by Bayesian update by combining the feature dataset with the prior probability distribution of the current round. Extract the variance of the posterior probability distribution and compare it with a preset convergence accuracy threshold; If the variance is less than the convergence accuracy threshold, the mean of the posterior probability distribution is extracted as the final inertia identification result, and the controller parameters of the edge-side control execution unit are updated based on the final inertia identification result. If the variance is not less than the convergence accuracy threshold, then the currently calculated posterior probability distribution is used as the prior probability distribution for the next iteration, and the process returns to the step of generating the optimal excitation signal parameters.

2. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, Initializing the prior probability distribution includes: Experienced knowledge of the motor's nameplate parameters and mechanical structure; Based on the acquired nameplate parameters and empirical knowledge, a Gaussian distribution is constructed as the prior probability distribution; The mean of the Gaussian distribution is calculated based on the parameters of the motor and the no-load system, and its initial variance is used to characterize the uncertainty of the inertia value in the initial state.

3. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, Generating the optimal excitation signal parameters includes: Construct a constrained optimization problem with the objective of maximizing the expected information gain; Under the constraints of satisfying the preset maximum allowable torque ripple and maximum detection time, the constraint optimization problem is solved to generate the optimal excitation signal parameters.

4. The asynchronous motor inertia response detection method based on cloud computing according to claim 3, characterized in that, The expected information gain is quantified by the determinant of the Fisher information matrix, and the torque ripple is evaluated by a pre-constructed mapping function.

5. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, Preprocessing the raw dynamic response data includes: The original dynamic response data is digitally filtered; The feature dataset is generated by extracting the velocity response amplitude and phase information at the excitation signal frequency from the filtered data.

6. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, The posterior probability distribution is calculated using Bayesian updates, including: Construct a likelihood function to quantify the probability of observing the feature dataset given the inertia assumption; The posterior probability distribution is calculated by combining the likelihood function and the prior probability distribution.

7. The asynchronous motor inertia response detection method based on cloud computing according to claim 6, characterized in that, Constructing the likelihood function includes: Using a built-in digital twin model of the motor in the cloud, the output response is predicted under a given assumption of inertia. Calculate the residual between the predicted output response and the feature dataset; The likelihood function is determined based on the residual.

8. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, The method enables cloud-edge communication through parameterized signaling and characteristic data, including: The cloud-based inertia inference center only sends out the parameters of the optimal excitation signal; The edge-side control execution unit only uploads the feature dataset.

9. The asynchronous motor inertia response detection method based on cloud computing according to claim 1, characterized in that, The convergence accuracy threshold is set according to the requirements of the application scenario for the accuracy of the controller parameters.

10. A cloud-based asynchronous motor inertia response detection system, applied to the cloud-based asynchronous motor inertia response detection method described in claim 1, characterized in that, include: The model initialization module is used to initialize the prior probability distribution based on the preset parameters of the motor. The incentive design module is used to generate optimal incentive signal parameters based on the prior probability distribution of the current round; The excitation injection and acquisition module is used to synthesize an excitation signal based on the optimal excitation signal parameters and inject it into the motor, while simultaneously acquiring the original dynamic response data. The data preprocessing module is used to preprocess the original dynamic response data to generate a feature dataset; The Bayesian inference module is used to combine the feature dataset with the prior probability distribution of the current round to calculate the posterior probability distribution. The convergence decision and control module is used to determine whether the variance of the posterior probability distribution is less than a preset convergence accuracy threshold. If it is less, the final inertia identification result is generated to update the controller parameters. If it is not less, the excitation design module is triggered to enter the next iteration.

Images