An intelligent agent-based constant current detection method and system for a frequency converter

Abstract

The application relates to the technical field of automation control, and discloses a frequency converter constant current detection method and system based on an intelligent agent. The method obtains multi-channel electrical signal data, and obtains a load dynamic electrical characteristic data set after filtering and denoising; time sequence decomposition is performed on the data set to determine a dynamic coupling strength index sequence. When the sequence exceeds a preset threshold, an abnormal data segment is intercepted, and a current fluctuation precursor characteristic is identified; after training and prediction, a current trend category is determined, a pre-adjustment instruction sequence is generated accordingly, and frequency converter parameters are dynamically adjusted to form an optimized regulation and control scheme. When a control deviation occurs, the deviation data is extracted and classified, a correction pre-adjustment parameter set is determined, and the internal control logic of the frequency converter is updated. The system determines the mode transition result in combination with real-time feedback data, finally integrates the electrical characteristic data and the correction parameter set, and realizes continuous and stable constant current control. The method can intelligently predict and actively suppress current fluctuation, and improves system control precision and operation stability.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to a method and system for constant current detection of frequency converters based on intelligent agents. Background Technology

[0002] Currently, in the field of modern industrial automation, Supervisory Control and Data Acquisition (SCADA) systems, as upper-level monitoring and management platforms, together with lower-level control devices such as frequency converters, form a complete automation architecture. Among these, the frequency converter, as the core component of motor control, directly determines the efficiency of the production line and the quality of products. Especially in scenarios with extremely stringent requirements for current stability, such as electroplating, special welding, and precision material processing, achieving high-precision constant current control is the cornerstone of ensuring process consistency and safe equipment operation.

[0003] In existing technologies, mainstream constant current control methods rely on real-time feedback adjustment of the output current. Control parameters are dynamically adjusted by detecting the deviation between the current value and the setpoint to maintain current stability. However, in specific high-dynamic scenarios such as high-speed, posture-changing grinding by robotic arms and periodic impact loads on stamping presses, sudden changes in mechanical load can trigger rapid reconstruction of the electromagnetic relationships within the motor. Feedback control, essentially an error-based post-compensation mechanism, is limited by the current loop's response bandwidth and sampling delay, making it difficult to completely suppress current spikes at the moment of impact. Although the current loop bandwidth of modern frequency converters has significantly improved, they still suffer from limitations in passive response when dealing with periodic impacts with obvious precursory characteristics.

[0004] Therefore, in certain scenarios with high inertia, large hysteresis, or periodic impact loads, existing technologies suffer from response delays when relying solely on feedback control, making it difficult to completely eliminate current spikes or meet ultra-high dynamic accuracy requirements. Summary of the Invention

[0005] This invention provides a constant current detection method and system for frequency converters based on intelligent agents, in order to solve the problem that in the prior art, under certain scenarios with high inertia, large hysteresis or periodic impact loads, simply relying on feedback control results in response delay, making it difficult to completely eliminate current spikes or meet ultra-high dynamic accuracy requirements.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a constant current detection method for frequency converters based on intelligent agents, comprising: Acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; The electrical characteristic data is decomposed in time series to extract the main features, the dynamic coupling strength between the feature components is calculated, and the index sequence of the dynamic coupling strength between electrical characteristics is determined. Candidate precursor features are obtained from the index sequence and compared with a pre-established library of typical fluctuation precursor features to determine the precursor features of the upcoming current fluctuation. The precursor features are subjected to inference, prediction, and classification processing to determine the current trend category; Based on the current trend category, a pre-adjustment command sequence is generated, and the inverter control parameters are dynamically adjusted to obtain a current regulation scheme. Obtain deviation data from the current regulation scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification results; Based on the pre-adjusted parameter set, real-time feedback data is obtained, and the internal control logic of the frequency converter is updated based on the real-time feedback data. If the updated logic control parameters meet the preset operating state threshold, the real-time feedback data is classified and analyzed to determine the mode transition result. Based on the mode transition results, the electrical characteristic data and the preset parameter set are integrated and classified for mode analysis to determine a continuous and stable constant current control result.

[0007] Secondly, the present invention provides an agent-based inverter constant current detection system, comprising: The data acquisition and filtering module is used to acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; The timing analysis and coupling evaluation module is used to perform timing decomposition on the electrical characteristic data, extract the main features, calculate the dynamic coupling strength between feature components, and determine the index sequence of the dynamic coupling strength between electrical characteristics. An anomaly detection and extraction module is used to obtain candidate precursor features from the indicator sequence and compare them with a pre-established typical fluctuation precursor feature library to determine the precursor features of the upcoming current fluctuation. The trend prediction and classification module is used to perform inference prediction and classification processing on the precursor features to determine the current trend category. The control command generation module is used to generate a pre-adjustment command sequence based on the current trend category, dynamically adjust the inverter control parameters, and obtain a current regulation scheme. The deviation correction and parameter optimization module is used to acquire deviation data in the current control scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification results. The control logic update and feedback module is used to obtain real-time feedback data based on the pre-adjusted parameter set, update the internal control logic of the frequency converter based on the real-time feedback data, and if the updated logic control parameters meet the preset operating state threshold, classify and analyze the real-time feedback data to determine the mode transition result. The system integration and constant current control module is used to integrate the electrical characteristic data and the preset parameter set according to the mode transition result, and perform classification processing and mode analysis to determine a continuous and stable constant current control result.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention achieves multi-dimensional coverage of load characteristics through multi-channel acquisition, and improves data purity by filtering to adapt to dynamic noise; then, it decomposes the components at different time scales through time-series decomposition, calculates the coupling strength and forms a time-series sequence, and fully presents the dynamic correlation evolution. It solves the problems of one-sided data acquisition, limited filtering effect and static feature analysis in the prior art, greatly improves the integrity and purity of data, realizes the accurate characterization of the dynamic coupling relationship of electrical characteristics, and provides a high-quality data foundation and core basis for subsequent anomaly analysis and control.

[0009] (2) By introducing dynamic coupling strength analysis and trend prediction, this invention upgrades current fluctuation control from a single ex-post feedback regulation to a collaborative control mode that combines feedforward prediction and feedback correction. In scenarios with strong nonlinearity and large inertia loads that traditional PID control struggles to handle, this invention can identify early signs of fluctuations and generate feedforward compensation, assisting the main controller to respond more quickly and effectively reducing overshoot, thereby further improving the dynamic control accuracy and operational stability of the system based on existing hardware.

[0010] (3) This invention updates the internal control logic of the frequency converter according to the modified pre-adjusted parameter set, determines the mode transition result through real-time feedback loop data, integrates the load dynamic electrical characteristic dataset and the modified parameter set to obtain a continuous and stable constant current control state, integrates the modified parameters into the control logic, monitors the operating effect through real-time feedback and confirms the effectiveness of the mode transition, and then deeply integrates the load data and modified parameters to construct a dynamically balanced control system. This realizes the dynamic iteration of the frequency converter control logic, makes the control system adaptive, breaks through the bottleneck of insufficient stability of constant current control in the prior art, ensures the continuity and stability of the constant current control state, and significantly improves the reliability of load operation. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of a constant current detection method for frequency converters based on intelligent agents provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a variable frequency drive constant current detection system based on an intelligent agent, provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a constant current detection method for frequency converters based on intelligent agents, including the following steps: S11: Acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; S12, perform time-series decomposition on the electrical characteristic data, extract the main features, calculate the dynamic coupling strength between feature components, and determine the index sequence of the dynamic coupling strength between electrical characteristics; S13, Obtain candidate precursor features from the index sequence and compare them with a pre-established library of typical fluctuation precursor features to determine the precursor features of the upcoming current fluctuation. S14, perform reasoning, prediction and classification processing on the precursor features to determine the current trend category; S15, Based on the current trend category, generate a pre-adjustment command sequence, dynamically adjust the inverter control parameters, and obtain a current regulation scheme; S16, Obtain deviation data in the current regulation scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification result; S17. Based on the pre-adjusted parameter set, obtain real-time feedback data, update the internal control logic of the frequency converter based on the real-time feedback data, and if the updated logic control parameters meet the preset operating state threshold, classify and analyze the real-time feedback data to determine the mode transition result. S18. Based on the mode transition result, integrate the electrical characteristic data and the preset parameter set, and perform classification processing and mode analysis to determine a continuous and stable constant current control result.

[0014] In step S11, acquiring multi-channel electrical signal data and performing filtering and sequence correlation processing to obtain electrical characteristic data includes: Multi-channel electrical signal data and load status tags are acquired from the sensor array at the inverter output and inside the load. The electrical signal data is filtered to obtain a purified electrical sequence; Based on the voltage and current amplitude ratio and phase difference in the purification electrical sequence, the impedance value and inductive reactance and capacitive reactance are calculated point by point to determine the impedance value sequence and inductive reactance and capacitive reactance record. The impedance value sequence and inductive and capacitive reactance records are sequentially associated with the load state labels to generate electrical characteristic data.

[0015] It should be noted that, firstly, the multi-channel electrical signal data is multi-channel voltage and current timing data collected from the inverter output and the sensor array inside the load. The load status label is time-synchronized with this data and identifies the specific operating conditions of the load, such as no-load operation, load capture, fault abnormality, etc.

[0016] Secondly, the electrical signal data is processed using an adaptive Kalman filter. Specifically, the process noise covariance matrix Q of the Kalman filter is preset, which includes the process noise of the voltage and current signals. The observed noise covariance matrix R is set according to the sensor manual. Then, the voltage U is estimated based on the state value from the previous moment. k ₋1. Current I k ₋1 Predict the current state value and predict the voltage Ũ k Predicting current Ĩ k Combined with the observed voltage U at the current moment k Observation current I k The predicted values ​​are corrected to obtain the filtered voltage and current sequences. After filtering, the signal-to-noise ratio (SNR) of the purified signal is calculated. Based on the general standard for industrial electrical signal processing, the SNR threshold is set to 40dB. If the SNR after the first filtering is 35dB (not meeting the standard), the Q matrix parameters are automatically adjusted, and filtering is re-executed until the SNR ≥ 40dB, finally obtaining the purified electrical sequence.

[0017] Next, the impedance, inductive reactance, and capacitive reactance are calculated. Specifically, first, according to Ohm's law, the impedance amplitude Z = U1 / I1, where U1 and I1 are the purified electrical sequences from the previous steps. Then, the reactance X = Z * sinφ, the inductive reactance XL = ωL, and the capacitive reactance XC = 1 / ωC, where ω is the angular frequency (rad / s), derived from the power supply frequency f (ω = 2πf), where f is a known power supply parameter, L is the inductance (H), an inherent parameter of the inductor in the circuit, provided by the component datasheet. φ is the impedance angle (rad), i.e., the phase difference between voltage and current in the circuit, directly measured by a phase measuring instrument. C is the capacitance (F), an inherent parameter of the capacitor in the circuit, provided by the component datasheet. When φ > 0, X > 0, XC = 0; when φ < 0, X < 0, XL = 0; when φ = 0, X = 0, XL = XC = 0. Finally, the impedance value Z(t) of each sampling point is recorded in chronological order to form an impedance value sequence; at the same time, the values ​​and sign changes of XL and XC at each moment are recorded to generate inductive reactance and capacitive reactance records.

[0018] Finally, based on the timestamp of the load status label, the data time window is divided, with each label corresponding to a continuous time segment. For example, the time range corresponding to the label "grabbing a 5kg workpiece and accelerating" is 10.5s-11.2s. The impedance value change sequence and inductive reactance-capacitive reactance interaction record within this time window are extracted. Then, the field structure of the dataset is constructed. Each record includes: operating condition identifier, such as "grabbing a 5kg workpiece and accelerating"; time window information, including start time, end time, and number of sampling points; impedance value change sequence, storing the Z value of each sampling point in array form; inductive reactance-capacitive reactance interaction record, including key time points, X_L / X_C values, characteristic switching type, such as inductive → capacitive; and original feature summary, such as the average and peak values ​​of voltage and current within this time window. Finally, electrical characteristic data is generated.

[0019] In step S12, the step of performing time-series decomposition on the electrical characteristic data, extracting key features, calculating the dynamic coupling strength between feature components, and determining the index sequence of the dynamic coupling strength between electrical characteristics includes: The electrical characteristic data is decomposed into time series to extract the main feature components and obtain the feature sequence; Based on the feature sequence, calculate the Pearson correlation coefficient between each feature component. If the Pearson correlation coefficient is greater than a preset coefficient threshold, it is determined that there is a coupling relationship, and the feature coupling matrix is ​​determined. Based on the feature coupling matrix, the dynamic coupling strength between each feature is calculated, and an index sequence representing the dynamic coupling strength is generated.

[0020] It should be noted that, firstly, the electrical characteristic dataset is decomposed using wavelet transform. Specifically, the appropriate wavelet basis function db4 is selected first. Then, the number of decomposition layers is determined to be 4: the high-frequency detail components cD1 correspond to 12.5-25kHz, cD2 to 6.25-12.5kHz, cD3 to 3.125-6.25kHz, and cD4 to 1.5625-3.125kHz, while the low-frequency approximation component cA4 corresponds to 0-1.5625kHz. Next, the time-series sequence of each electrical parameter is processed layer by layer. The first layer uses low-pass and high-pass filtering to obtain cA1 and cD1. Then, using cA1 as input, it is recursively decomposed to obtain cA2 and cD2, cA3 and cD3, and cA4 and cD4, respectively. Simultaneously, impedance, inductive reactance, and capacitive reactance are processed in parallel. Finally, all decomposition results are integrated to form the first decomposition sequence.

[0021] Secondly, PCA is used to extract the main feature components of the first decomposition sequence. Specifically, an original data matrix X of dimension N×15 is first constructed, where... This represents the value of the j-th component at the i-th time point. Then, through... Standardization processing is performed, among which It is the mean of the j-th column. It is the standard deviation of the j-th column. Let be the element in the i-th row and j-th column of the standardized data matrix. Then, calculate the covariance matrix of the standardized data matrix using the formula: ,in, Let N be the covariance matrix, and N be the number of rows of observations. For the standardized data matrix, For the standardized data matrix The transpose of the matrix is ​​then used. The eigenvalues ​​and eigenvectors are then calculated from the covariance matrix Cov. Finally, the variance explained by each eigenvalue is calculated using the following formula: ,in The above steps calculate the eigenvalues, where k represents the first k eigenvalues. Then, a scree plot is drawn, and the inflection point is identified. Typically, a threshold of 95% for the cumulative variance explained is set, and the minimum k value that first maximizes the cumulative explained variance is selected. Finally, the standardized raw data is projected onto the selected directions of the first k principal components, using the formula: ,in These are the eigenvectors corresponding to the aforementioned eigenvalues. We generate k new features (principal components). We then combine the calculated score sequences of these k principal components to form a feature sequence.

[0022] Subsequently, the Pearson correlation coefficients between each characteristic component are calculated. Under normal operating conditions, the correlation coefficients of physically related characteristic pairs, such as PC1 and PC2 corresponding to load inertia and motion acceleration, are generally stable above 0.8, reaching 0.95 under heavy load and rapid motion, and approximately 0.85 under light load and stable conditions. Therefore, a preset coefficient threshold of 0.8 is set, and a k×k symmetric matrix is ​​created. This indicates the coupling state between the i-th and j-th feature components: if the correlation coefficient is greater than a threshold, then... , indicates coupling; otherwise This is the characteristic coupling matrix.

[0023] Finally, the dynamic coupling strength is calculated. Specifically, all feature pairs marked as coupled are first extracted, and then the strength index of each coupling pair is dynamically calculated using a sliding time window method. The window size is set to 1000 data points, corresponding to a 20ms time span, based on a 50kHz sampling rate, and progressed along the time axis with a 50% overlap rate. Within each window, corresponding data subsets of the two feature sequences are extracted, and their Pearson correlation coefficient is calculated as the coupling strength value for that time period. By continuously sliding the window, a series of coupling strength values ​​arranged in chronological order are generated, ultimately forming a sequence of dynamic coupling strength indices.

[0024] In step S13, the step of obtaining candidate precursor features from the index sequence and comparing them with a pre-established library of typical fluctuation precursor features to determine the precursor features of an impending current fluctuation includes: If the value of any data point in the index sequence exceeds the preset intensity threshold, then an abnormal data segment is extracted from the index sequence to obtain an abnormal index segment. The abnormal index segments are subjected to multidimensional scaling to obtain a discrete point set representing nonlinear correlations; Clustering is performed on the discrete point set. If multiple density-separable clusters are formed in the clustering results, the geometric center and shape parameters of each cluster are combined to obtain the candidate precursor features. The candidate precursor features are compared with a pre-established database of typical fluctuation precursor features. Features with high matching degree are selected and identified as precursor features of the upcoming current fluctuation.

[0025] It should be noted that, firstly, based on a large amount of clean historical data from normal robot operations, statistical analysis revealed that under 99% of normal operating conditions, the dynamic coupling strength index is below 0.9. Therefore, a preset strength threshold of 0.9 was set. When any data point exceeds the limit, all consecutive data points within that time period are extracted, starting from the first exceeding point and ending at the last exceeding point, forming an abnormal index segment.

[0026] Secondly, a sliding window is used to process outlier segments. For example, with a window size of 5 and a step size of 1, a segment of 100 data points is slid across, generating 96 5-dimensional vectors. The Euclidean distances between each pair of these 96 high-dimensional vectors are then calculated, forming a 96×96 distance matrix. The value in the i-th row and j-th column of the matrix represents the distance between the i-th and j-th vectors. Finally, the distance matrix is ​​mapped onto a two-dimensional plane to reconstruct the positions of each point, ensuring that the distances between the reconstructed points reflect the distance relationships between the original high-dimensional vectors as closely as possible. The final output is the coordinates of the 96 data points on the two-dimensional plane, forming a discrete point set representing the nonlinear correlation.

[0027] Subsequently, the two-dimensional discrete point set is clustered using density-based clustering (such as DBSCAN). First, a neighborhood radius and a minimum number of points are set, automatically identifying high-density regions and sparse noise points. For each valid cluster, the mean coordinates of all points within the cluster are calculated as the geometric center and shape parameters. Principal component analysis (PCA) is used to calculate the major axis, minor axis, and eccentricity of the cluster to determine its shape characteristics. The major and minor axes and eccentricity of each cluster constitute potential precursor features. For example, cluster A has similar major and minor axis lengths, resembling a circle. Cluster B is highly stretched in one direction, resembling an ellipsoid.

[0028] Then, the similarity between the candidate precursor features and each typical feature in the pre-established typical wave precursor feature library is calculated, which can be done using Euclidean distance. The degree of proximity of the center coordinates is measured and then converted into a similarity score. ,in To calculate the similarity, The Euclidean distance between precursor features and typical features. The maximum possible distance is determined through statistical analysis of historical data. Then, the similarity of the proportions of the principal axes of the shapes is calculated. ,in The major and minor axes of the candidate precursor features, , These are the major and minor axes, which are typical features. Then, the similarity of eccentricities is calculated. ,in It is the eccentricity of the candidate precursor features. The center coordinates are used for typical features. Finally, the similarity is calculated. If the similarity between a candidate feature and a typical feature in the library exceeds a preset matching threshold of 0.9, where the preset matching threshold of 0.9 is an empirical balance point determined by statistical analysis of the similarity distribution of known matching pairs in the historical validation set, ensuring a sufficiently high recall rate while keeping the false alarm rate within an acceptable range, then the candidate feature is a precursor feature of an impending current fluctuation.

[0029] It is worth noting that the pre-established typical fluctuation precursor feature library is built by systematically simulating various known minor fault modes in a controlled experimental environment, such as current loop gain drift, encoder signal interference, and power supply voltage ripple. Abnormal data fragments of dynamic coupling strength indicators under each fault are collected, and after undergoing multidimensional scaling and cluster analysis processes that are completely consistent with online detection, representative cluster geometric centers and shape parameter features are extracted. After multiple experimental verifications to ensure the stability and repeatability of the features, a benchmark database containing the mapping relationship between fault modes and features is finally established.

[0030] In step S14, the precursor features are subjected to inference prediction and classification processing to determine the current trend category, including: The precursor features are normalized to obtain standardized features; Calculate the similarity between the standardized feature and the pre-established first feature. If the similarity is less than a preset similarity threshold, then infer the standardized feature to obtain a short-term trend vector. Obtain the historical trend vector, and perform a weighted fusion of the short-term trend vector and the historical trend vector to obtain the fused trend vector; The fused trend vector is input into a pre-established trend classifier, and the current trend category is determined based on the output fluctuation type.

[0031] It should be noted that, firstly, the normalization of precursor features specifically involves extracting the mean and standard deviation of each dimension of this type of feature from the historical database. For example, the mean of the three dimensions is [0.5, 0.5, 0.6], and the standard deviation is [0.1, 0.1, 0.15]. Then, Z-score standardization is applied to each dimension of the current precursor feature vector, and the result is (current value minus the historical mean of that dimension) divided by the historical standard deviation of that dimension to obtain the standardized feature.

[0032] Secondly, it should be noted that the first feature is a benchmark library composed of massive feature data collected and standardized through thousands of normal operation cycles under the robot's healthy state. Similarity is calculated by first calculating the KL divergence, using the following formula: ; Where k is the feature dimension (e.g., k=3 for three-dimensional features); The mean of each dimension of the standardized feature. The mean of each dimension of the first feature, The covariance matrix of the standardized feature set. The covariance matrix of the first feature. for The inverse matrix is ​​given by tr(⋅), which is the trace of the matrix (i.e., the sum of the elements on the main diagonal), and det(⋅) is the determinant of the matrix.

[0033] If the calculated KL divergence value is less than the preset similarity threshold of 0.05, the current abnormal pattern is determined to be homologous to historically known patterns. The standardized feature set sequence is then input into a pre-trained GRU network. The GRU network outputs a short-term trend vector, which contains quantified trend information for the future short term (e.g., within 150ms). For example, [+0.8A, -5rpm, 150ms] physically means that the average current increases by 0.8A and the rotational speed decreases by 5rpm within the next 150ms.

[0034] It is worth noting that the similarity threshold of 0.05 is a balanced value determined based on multi-dimensional quantitative analysis of the robot current fluctuation monitoring scenario. Specifically, the analysis of 120,000 historical anomalies and interference data showed that the natural boundary of the divergence distribution between abnormal and normal samples is concentrated around 0.05, with 92% of the samples exceeding this value being abnormal. Secondly, in terms of business risk, this threshold is a key node for balancing production efficiency and equipment safety, ensuring that the false alarm rate and false negative rate are controlled within acceptable ranges of 3.2% and 1.8%, respectively. Through cross-validation with the GRU model, the 0.05 threshold improves the F1-score to 95.7%, achieving a better balance between recognition performance and false alarm control compared to other candidate thresholds (such as 0.03 and 0.07). Finally, this threshold is highly compatible with the multi-dimensional normal distribution characteristics of the three-dimensional feature vector, corresponding to the 97.5 quantile of the distribution curve, which can cover 98.3% of typical anomaly patterns.

[0035] In one embodiment, a gated recurrent unit (GRU) network is used for inference on standardized features. Its specific structure and parameters are as follows: the input dimension is 6-dimensional, corresponding to the 6 standardized features; there are 2 hidden layers, each containing 128 hidden units; the output dimension is 3-dimensional, corresponding to the change in current, the change in rotational speed, and the predicted duration; the activation function is tanh, and the dropout probability is 0.2. This network receives standardized features that have passed similarity checks as input and outputs a short-term trend vector by capturing the temporal dependencies of the sequence data.

[0036] For example, the model output vector is [+0.8A, -5rpm, 150ms], which means that in the next 150 milliseconds, the average current of the robot joint motor will increase by 0.8 amperes and the motor speed will decrease by 5 revolutions per minute, providing the system with a quantitative basis for the magnitude and direction of change.

[0037] For example, to smooth out potential noise and spikes from a single prediction, a weighted average method is used for fusion. Assume the short-term trend vector generated at the current moment is Vt = [+0.8A, -5rpm], while the historical trend vector, i.e., the weighted average of the past 10 predictions, is Vh = [+0.7A, -4.5rpm]. The new prediction is assigned a higher weight, for example, 0.6, while the historical trend has a weight of 0.4. The fused trend vector is then obtained. This fusion process makes the final trend judgment more robust and less susceptible to transient interference. Finally, this fused trend vector, which includes the latest prediction information and short-term historical trends, is input into a pre-trained trend classifier, such as a multi-class support vector machine (SVM). The trained SVM classifier outputs the class of the input vector based on its position in the feature space.

[0038] It is worth noting that the GRU model is trained based on historical operating data of the frequency converter system, and a supervised learning framework is used to build its predictive capabilities. The training data covers electrical signal sequences under multiple operating conditions. After Z-score standardization and sequence segmentation preprocessing, the training, validation, and test sets are divided in a 7:1.5:1.5 ratio. The training uses the mean squared error loss function and the Adam optimizer, along with an initial learning rate of 0.001 and an early stopping strategy to avoid overfitting. After training, the model is validated using metrics such as MAE and R². The model is considered complete when the current prediction error is ≤0.1A. The SVM classifier is trained based on historical fused trend vectors and corresponding current fluctuation categories labeled by domain experts. The training process first prepares a labeled dataset, where the input is the fused trend vector and the output is the predefined category. The data is calibrated using the Z-score standard to ensure consistent feature dimensions. A one-to-many strategy is used to handle multi-classification problems. A radial basis function (RBF) kernel function is selected to map the feature space, and the hyperparameter penalty coefficient C and kernel parameter γ are optimized through grid search. The sequence minimum optimization algorithm (SMO) is used to minimize the loss function, and cross-validation is employed to prevent overfitting.

[0039] For example, for the fused trend vector of [+0.76A, -4.8rpm, 148ms], the final output category of SVM is control parameter misalignment oscillation, thus completing the accurate judgment of the type of current fluctuation that is about to occur.

[0040] In step S15, generating a pre-adjustment command sequence based on the current trend category and dynamically adjusting the inverter control parameters to obtain a current regulation scheme includes: Based on the current trend category, a fuzzy instruction set is obtained from a preset parameter mapping library; The fuzzy instruction set is subjected to timing constraint detection. If the parameter value in the fuzzy instruction set exceeds a preset safety threshold, the parameter value is adjusted to obtain a preliminary instruction sequence. The initial instruction sequence is converted into a binary data frame to generate a pre-tuned instruction sequence; The pre-adjustment instruction sequence and the current trend category are associated and encapsulated, and a timestamp for scheduling execution is added to obtain the current regulation scheme.

[0041] In this embodiment, the current trend categories include stable, rising, falling, and oscillating. Stable indicates that the current value fluctuates within a preset normal operating range without significant trend changes; rising indicates that the current value shows a continuous or step-like increasing trend; falling indicates that the current value shows a continuous or step-like decreasing trend; and oscillating indicates that the current value fluctuates frequently and irregularly within a short period of time.

[0042] It should be noted that the preset parameter mapping library is a lookup table built based on historical operating data. This mapping library establishes a correspondence between current trend categories and a set of initial inverter control parameters. The construction process of the mapping library is as follows: First, a large amount of motor operating data covering different operating conditions, loads, and equipment health states is collected. Then, for each data record, simulation is performed based on the physical model of the motor and load to determine a set of control parameter adjustment actions that have been verified to effectively stabilize current, improve energy efficiency, or protect equipment as the reference adjustment strategy for that record. Adjustable control parameters include target frequency, output voltage, current limit, speed loop proportional gain, and speed loop integral time. Finally, the data is grouped according to current trend categories, and the reference adjustment parameters within each group are statistically analyzed to form a typical parameter set corresponding to that category, i.e., a fuzzy instruction set. For example, for the rising category, its fuzzy instruction set includes instructions such as "reduce target frequency" and "maintain current limit." The parameter values ​​in the fuzzy instruction set are given in the form of offsets or percentages relative to the current operating values; for example, target frequency reduction... .

[0043] It should be noted that timing constraint detection is to ensure that the fuzzy instruction set obtained from the mapping library is feasible and safe in terms of parameter values. This step should more accurately be called parameter safety threshold detection. The preset safety threshold is set based on the rated parameters, thermal characteristics, and industry safety standards of the frequency converter and motor. For example, the safety threshold range for the output frequency is typically [range missing]. ,in and The upper limit of the safety threshold for current limits is determined based on the motor nameplate and manufacturing process requirements. The dynamic setting needs to comprehensively consider the motor's rated current, the short-time overload multiple allowed by the insulation class, the inverter's overload capacity, and the real-time thermal capacity margin calculated based on the thermal model. During detection, the suggested parameter values ​​in the fuzzy instruction set are compared with the corresponding safety thresholds. If any parameter value exceeds the safety threshold, the parameter value is adjusted. This embodiment uses amplitude limiting processing for adjustment: the parameter value that exceeds the limit is directly set to the closest safety threshold boundary value. For example, if the calculated target frequency... It is 65Hz, and If the frequency is 60Hz, then the amplitude is limited to 60Hz. After completing this type of detection and necessary adjustments for all parameters, a set of determined control parameter values ​​that meet safety constraints is obtained, forming a preliminary command sequence.

[0044] Subsequently, the initial instruction sequence is converted into binary data frames conforming to the communication protocol. The conversion process includes parsing the parameters of each instruction (such as the target frequency adjustment value and speed command) and encoding them according to a specific communication protocol, such as Modbus RTU, CANopen, or EtherCAT. For example, the speed command is converted into 16-bit binary data, and a protocol header, checksum, and trailer frame are added to generate the pre-tuned instruction sequence.

[0045] Finally, the pre-tuning instruction sequence is associated and encapsulated with the final current trend category to form a structured data packet. This association encapsulation includes adding metadata such as trend category identifier, instruction sequence version number, and control target description. Simultaneously, a scheduling execution timestamp is appended, calculated based on the system real-time clock and the expected execution delay. For example, based on historical data or current load conditions, the execution time might be set to 150 milliseconds, but considering response latency, a buffer time is reserved, adjusting it to 170 milliseconds. Ultimately, an optimized output current control scheme is obtained.

[0046] In step S16, acquiring deviation data in the current regulation scheme, classifying the deviation data, and determining the pre-adjustment parameter set based on the classification results includes: Obtain the deviation data in the current regulation scheme; The deviation data is classified to obtain a set of classified deviation patterns; Potential risk factors are separated from the set of deviation patterns. If the potential risk factors exceed a preset activation threshold, a set of risk factors is obtained. Based on the set of risk factors, parameter adjustment rules are obtained from the parameter mapping library and modified to obtain the modified pre-adjusted parameter set.

[0047] It should be noted that when a deviation is detected in the optimized output current control scheme, relevant deviation data is collected. For example, during the robot's joint motor's task execution, the deviation information between the predicted current and the actual current is acquired in real time. This deviation data constitutes a time series vector containing multi-dimensional information such as current deviation value, deviation change rate, and current motor load. Taking the rapid grasping action of the robotic arm as an example, within 100 milliseconds of the action's duration, the continuously collected deviation vector sequence shows a deviation vector sequence with a small initial deviation, followed by a rapid increase and periodic decay. Secondly, the multi-dimensional time series data of current deviation value, deviation change rate, and motor load are classified using a Support Dimension Vector Machine (SVM). First, outliers are removed and missing values ​​are filled. Then, the Z-score is used to standardize the units and unify the dimensions, and the sequences are aligned according to a fixed time window. Subsequently, key features are extracted, such as the mean, variance, extreme values, peak frequencies of each dimension sequence, as well as the maximum rate of change, duration of continuous deviation, and main oscillation frequency, which are combined into a fixed-dimensional feature vector. Next, a pre-trained SVM is invoked to map the input features to a high-dimensional space for matching. If the features exhibit high fluctuation frequency and current peaks close to the mean, they are classified as high-frequency oscillation type. If they show a slow, continuously accumulating bias, they are classified as integral saturation type. Finally, the classification results are output, forming a set of bias patterns.

[0048] It is worth noting that the pre-trained Support Vector Machine (SVM) structure includes an input layer that receives a fixed-dimensional feature vector as input. The dimension is determined by the number of features. For example, a 12-dimensional feature may contain 3 means, 3 variances, 3 extreme values, and one each of peak frequency, maximum rate of change, duration of persistent bias, and main oscillation frequency. Subsequently, the kernel mapping layer maps the low-dimensional input to a high-dimensional feature space using the radial basis function (RBF). The decision boundary layer constructs the optimal classification hyperplane based on the support vector set S, using the Lagrange multiplier α and the bias term b. The output layer generates binary classification results, with labels corresponding to high-frequency oscillation type (0) and integral saturation type (1), respectively. Specifically, the penalty coefficient C and kernel parameter γ are first set and optimized through grid search and 5-fold cross-validation. The optimal values ​​are C=10 and γ=0.1. Then, the low-dimensional feature vector is mapped to a high-dimensional feature space using the radial basis function (RBF). Subsequently, the goal is to minimize the SVM loss function. Where ω is the normal vector of the decision hyperplane in the high-dimensional space, and C is the optimal penalty coefficient. This is used to allow for a small number of samples with classification bias in order to improve model robustness. The soft-margin slack variables are used to balance the two, achieving the training objective of high classification accuracy and low model complexity. Then, by solving the Lagrangian dual problem of this loss function, the original convex quadratic programming problem is transformed into a more easily solvable dual problem. Finally, support vectors that play a crucial supporting role in constructing the decision boundary are selected from the training set, and the corresponding Lagrange multipliers α and the bias term b of the decision hyperplane are determined. The output SVM model after training is shown.

[0049] If the actual value of a potential risk factor exceeds a preset activation threshold, which includes an oscillation peak value threshold of 0.5A and an oscillation frequency threshold of 100Hz, then the risk factor is included in the potential risk factor set. For example, if the current oscillation peak value is 0.6A, exceeding the 0.5A activation threshold, then the high oscillation peak value is included in the risk factor set. It should be noted that the activation threshold is a safety threshold set based on the physical characteristics of the motor and driver, historical operating data statistics, and the stability boundary of control performance. Statistical analysis shows that the 0.5A oscillation peak value threshold can minimize overheating and mechanical wear while ensuring system response speed, and the 100Hz oscillation frequency threshold is to avoid the inherent resonant frequency of the mechanical structure and prevent oscillation from being amplified.

[0050] Finally, based on the set of potential risk factors, the parameter mapping knowledge base is queried, and parameter adjustment rules corresponding to the potential risk factors are extracted from the knowledge base. The relevant parameters are then adjusted according to these rules. For example, the rule for high oscillation peaks is to appropriately reduce the proportional gain Kp of the PID controller while increasing the derivative gain Kd. Kp is reduced from the current 5.2 to 4.8, and Kd is increased from 1.5 to 1.8, ultimately yielding the corrected set of preset parameters.

[0051] In step S17, the real-time feedback data is obtained according to the pre-adjusted parameter set, and the internal control logic of the frequency converter is updated according to the real-time feedback data. If the updated logic control parameters meet the preset operating state threshold, the real-time feedback data is classified and analyzed to determine the mode transition result, including: Based on the pre-tuned parameter set, real-time feedback data is obtained and pre-processed to obtain a feedback data set; Based on the feedback data set, the internal control logic of the frequency converter is updated to obtain the updated control logic parameters; If the control logic parameters meet the preset operating state threshold, the feedback data set is classified to obtain the classified operating state data. The operational status data is analyzed to determine the mode transition result.

[0052] It should be noted that, firstly, multi-dimensional feedback data is continuously collected in real time from the robot's joint motors driven by the frequency converter through sensors and ADC converters. This includes real-time motor torque, rotor angular velocity, and DC bus voltage fluctuations. For torque signals that may contain high-frequency noise, a moving average is used for smoothing. Specifically, a sliding window is first set (e.g., a window size of 10 data points), the arithmetic mean of the data within the window is calculated as the smoothed value for the current point, and then Z-score standardization is used to form a standardized real-time feedback data set. Secondly, the internal control logic of the frequency converter is updated. Several control logic templates are preset. For example, Template 1 is a high-rigidity response template, suitable for scenarios requiring rapid start-up and shutdown and suppression of external disturbances. Its characteristics are high controller gain and short response time. Template 2 is a flexible and smooth template, suitable for scenarios involving precise fitting or handling of fragile items. Its characteristics are lower acceleration and a smoother torque output curve. When the real-time feedback data set shows that the motor is starting from a standstill with a large load, its rapid torque increase and drastic speed changes will highly match the high-rigidity response template, and this template will then be invoked to update the control logic parameters inside the frequency converter.

[0053] Subsequently, the real-time feedback data set is categorized. Specifically, the current margin index A is calculated first, followed by the calculation of the peak current. ,in The torque constant is obtained from the motor's operation manual. This is the maximum torque that the motor can output under the expected load. To estimate the peak current, the peak current is subtracted from the real-time feedback current and then divided by the peak current. Subsequently, the voltage margin index B is calculated by subtracting the minimum required voltage from the actual bus voltage and then dividing by the rated voltage, where the minimum required voltage... ,in The back electromotive force constant is obtained from the motor's operation manual. This refers to the real-time speed of the motor. This is the peak current. Let \(R\) be the winding resistance, which is obtained from the motor operation manual. Calculate the heat dissipation margin \(C\) by subtracting the predicted peak temperature from the maximum allowable temperature and then dividing by the maximum allowable temperature. For the three indicators of current margin \(A\), voltage margin \(B\), and heat dissipation margin \(C\), first set their respective critical values \(x_0\). For example, \(x_0\) of \(A\) means the predicted current reaches the allowable maximum value, and \(x_1 = 0.5\) means the predicted current is only 50% of the allowable maximum value. Then map the actual calculated value \(x\) of each indicator to the interval \([0,1]\) according to the following rules: if \(x\leq x_0\), the normalized value is 0; if \(x_0 < x < x_1\), the normalized value is \((x - x_0) / (x_1 - x_0)\); if \(x\geq x_1\), the normalized value is 1. Finally, the system dynamic margin is synthesized by weighting the above sub-indicators as \(w_1\times A + w_2\times B+w_3\times C\), where the weight coefficients \(w_1 = 0.5\), \(w_2 = 0.3\), \(w_3 = 0.2\). This weight setting can be determined according to the priority of the system's requirements for current capacity, voltage stability, and heat dissipation performance. If the calculated dynamic margin is higher than the preset operating state threshold of 0.75, which is a verified value based on the engineering safety margin principle to cope with system model uncertainty, sensor errors, and sudden load disturbances and achieve an optimal balance between performance and long-term reliability and has been verified through simulation experiments, then the new parameters are judged to be safe and available.

[0054] Finally, the SVM classifier is used to judge the control mode. The characteristics of the lag compensation mode are that after an external disturbance occurs, key indicators (such as angular velocity) first deviate significantly from the set value (such as a decrease of more than 5%), and then the control quantity (such as current) is adjusted significantly. The characteristics of the pre-intervention mode are that the system monitors precursor signals such as small torque fluctuations and specific frequency resonances, and fine-tunes the control quantity such as the voltage vector angle before the key indicators deviate. The input of the SVM mode judgment is the mode-related features extracted from the classified operating state data, such as deviation duration, intervention timing, etc., and the output is a binary classification result of lag compensation or pre-intervention. The judgment rule is that if the SVM outputs pre-intervention and the confidence level is higher than the threshold of 0.9, and the confidence level of 0.9 is a common high certainty standard, then the mode transition is confirmed and a mode transition result report is generated.

[0055] It should be noted that the training process of this SVM classifier is as follows: Based on historical operating status data, a training set is formed by collecting labeled operating condition samples through the system. The training data comes from the operating records of the frequency converter under different control modes. Each sample contains a mode feature vector extracted from the operating status data, such as deviation duration, intervention timing, oscillation frequency, etc., and is labeled with its actual mode category: hysteresis compensation or pre-intervention. Before training, the feature vectors need to be standardized to eliminate the influence of dimensions. A one-to-many strategy is adopted to solve the binary classification problem, and a radial basis function (RBF) is selected to map the features to a high-dimensional space. The hyperparameter penalty coefficient C and kernel parameter γ are optimized through grid search and cross-validation. The optimal classification hyperplane is solved using a sequential minimum optimization algorithm to complete the model training.

[0056] In step S18, the process of integrating the electrical characteristic data and the preset parameter set based on the mode transition result, performing classification processing and mode analysis, and determining a continuously stable constant current control result includes: Based on the mode transition results, the electrical characteristic data and the preset parameter set are integrated to obtain an optimized dataset; The optimized dataset is classified into states to obtain classified constant current control data; Pattern analysis is performed on the constant current control data to determine a continuously stable constant current control result.

[0057] It should be noted that, firstly, the high-dimensional load dynamic electrical characteristic data and the corrected preset parameter set are integrated using a weighted average fusion method with confidence weighting. Four types of load dynamic electrical characteristic data are extracted first: basic current data, dq-axis current components, current harmonic spectrum, and power factor (100Hz sampling). Then, the corrected preset parameter set is extracted, consisting of at least three Kp / Ki combinations for different operating conditions and matched current loop bandwidths. Each parameter set is associated with a material label and speed range. Next, the weights are determined: load data weight 0.7, preset parameter weight 0.3. Then, weighted fusion values ​​are calculated for each indicator. For current-related indicators, such as the q-axis current fusion value = load data q-axis current mean × load weight + preset theoretical q-axis current × preset weight. Control parameters, such as the Kp fusion value = load data derived Kp × load weight + pre-tuning Kp × pre-tuning weight, where the load derived Kp is calculated based on the current error and rate of change using the Ziegler-Nichols method. Finally, the initial closed-loop optimization dataset is constructed according to current-type indicators, control parameter-type indicators, and mode label structure.

[0058] Secondly, density-based DBSCAN clustering is used to classify the dataset. Specifically, feature vectors including current values, harmonic features, power factor, and PI parameters as auxiliary features are extracted from the initial closed-loop optimization dataset. Then, density-based DBSCAN clustering is used for state classification, with parameters including neighborhood radius (eps, e.g., 0.5 after standardization) and minimum number of samples (min_samples, e.g., 5). During clustering, all data points are scanned, and points with density connections are grouped into the same cluster. Typical output clusters include: Cluster 1 (entry stage), characterized by rapid current increase and high-frequency harmonic components (e.g., a large amplitude of the 5th harmonic); Cluster 2 (steady-state polishing stage), characterized by stable current values, I_d close to zero, and low harmonic components; and Cluster 3 (corner processing stage), characterized by periodic fluctuations in I_q and high amplitudes of specific frequency harmonics (e.g., the second harmonic). Next, clusters are matched with the preset parameter set. The average electrical characteristics of each cluster are calculated, and similarity matching is performed with parameter groups in the preset parameter set using Euclidean distance. For example, cluster 2 (steady-state grinding stage) may have the highest matching degree with a set of parameters with high Ki (e.g., 0.1) and low Kp (e.g., 0.01) to minimize steady-state error. Finally, the classified constant current control data is generated.

[0059] Finally, pattern analysis is performed on the classified constant current control data to determine the continuously stable constant current control result. Specifically, sample entropy is extracted from the classified constant current control data, and the Lyapunov exponent is calculated. Stability is then evaluated using preset stability criteria. The stability evaluation results can be divided into two categories: sustainable stability (low sample entropy, negative Lyapunov exponent) and transient pseudo-stability (low sample entropy, zero Lyapunov exponent or high sample entropy, negative Lyapunov exponent). The remaining combinations are then classified as unstable. If a combination is classified as sustainable stability, the continuously stable constant current control result is determined.

[0060] In summary, this invention discloses a constant current detection method for frequency converters based on intelligent agents. The method includes acquiring multi-channel electrical signal data, filtering and denoising it to obtain a dynamic electrical characteristic dataset of the load; performing time-series decomposition to determine a dynamic coupling strength index sequence. When the sequence exceeds a preset threshold, abnormal data segments are extracted and current fluctuation precursors are identified; after training and prediction, the current trend category is determined, and a pre-adjustment command sequence is generated accordingly to dynamically adjust the frequency converter parameters to form an optimized control scheme. When control deviations occur, deviation data is extracted and classified, a correction pre-adjustment parameter set is determined, and the internal control logic of the frequency converter is updated. The system combines real-time feedback data to determine the mode transition result, and finally integrates the electrical characteristic data and the correction parameter set to achieve constant current control of the frequency converter.

[0061] Reference Figure 2 The second embodiment of the present invention provides a variable frequency drive constant current detection system based on an intelligent agent, comprising: The data acquisition and filtering module is used to acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; The timing analysis and coupling evaluation module is used to perform timing decomposition on the electrical characteristic data, extract the main features, calculate the dynamic coupling strength between feature components, and determine the index sequence of the dynamic coupling strength between electrical characteristics. An anomaly detection and extraction module is used to obtain candidate precursor features from the indicator sequence and compare them with a pre-established typical fluctuation precursor feature library to determine the precursor features of the upcoming current fluctuation. The trend prediction and classification module is used to perform inference prediction and classification processing on the precursor features to determine the current trend category. The control command generation module is used to generate a pre-adjustment command sequence based on the current trend category, dynamically adjust the inverter control parameters, and obtain a current regulation scheme. The deviation correction and parameter optimization module is used to acquire deviation data in the current control scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification results. The control logic update and feedback module is used to obtain real-time feedback data based on the pre-adjusted parameter set, update the internal control logic of the frequency converter based on the real-time feedback data, and if the updated logic control parameters meet the preset operating state threshold, classify and analyze the real-time feedback data to determine the mode transition result. The system integration and constant current control module is used to integrate the electrical characteristic data and the preset parameter set according to the mode transition result, and perform classification processing and mode analysis to determine a continuous and stable constant current control result.

[0062] This invention also provides an electronic device. The electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps described in the embodiments of the agent-based inverter constant current detection method, for example... Figure 1 The step S11 shown. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the above-described device embodiments.

[0063] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0064] The electronic device may be a desktop computer, laptop, handheld computer, or smart tablet, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above components are merely examples of electronic devices and do not constitute a limitation on the electronic device. It may include more or fewer components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0065] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the electronic device, connecting various parts of the electronic device through various interfaces and lines.

[0066] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0067] Wherein, if the modules / units integrated in the electronic device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0068] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0069] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that 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 for those skilled in the art.

Claims

1. A constant current detection method for frequency converters based on intelligent agents, characterized in that, include: Acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; The electrical characteristic data is decomposed into time series, the main features are extracted, the dynamic coupling strength between the feature components is calculated, and the index sequence of the dynamic coupling strength between electrical characteristics is determined. Candidate precursor features are obtained from the index sequence and compared with a pre-established library of typical fluctuation precursor features to determine the precursor features of the upcoming current fluctuation. The precursor features are subjected to inference, prediction, and classification processing to determine the current trend category; Based on the current trend category, a pre-adjustment command sequence is generated, and the inverter control parameters are dynamically adjusted to obtain a current regulation scheme. Obtain deviation data from the current regulation scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification results; Based on the pre-adjusted parameter set, real-time feedback data is obtained, and the internal control logic of the frequency converter is updated based on the real-time feedback data. If the updated logic control parameters meet the preset operating state threshold, the real-time feedback data is classified and analyzed to determine the mode transition result. Based on the mode transition results, the electrical characteristic data and the preset parameter set are integrated and classified for mode analysis to determine a continuous and stable constant current control result.

2. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The process of acquiring multi-channel electrical signal data, performing filtering and sequence correlation processing to obtain electrical characteristic data includes: Multi-channel electrical signal data and load status tags are acquired from the sensor array at the inverter output and inside the load. The electrical signal data is filtered to obtain a purified electrical sequence; Based on the voltage and current amplitude ratio and phase difference in the purification electrical sequence, the impedance value and inductive reactance and capacitive reactance are calculated point by point to determine the impedance value sequence and inductive reactance and capacitive reactance record. The impedance value sequence and inductive and capacitive reactance records are sequentially associated with the load state labels to generate electrical characteristic data.

3. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The step of performing time-series decomposition on the electrical characteristic data, extracting key features, calculating the dynamic coupling strength between feature components, and determining the index sequence of the dynamic coupling strength between electrical characteristics includes: The electrical characteristic data is decomposed into time series to extract the main feature components and obtain the feature sequence; Based on the feature sequence, calculate the Pearson correlation coefficient between each feature component. If the Pearson correlation coefficient is greater than a preset coefficient threshold, it is determined that there is a coupling relationship, and the feature coupling matrix is ​​determined. Based on the feature coupling matrix, the dynamic coupling strength between each feature is calculated, and an index sequence representing the dynamic coupling strength is generated.

4. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The step of obtaining candidate precursor features from the index sequence and comparing them with a pre-established library of typical fluctuation precursor features to determine the precursor features of an impending current fluctuation includes: If the value of any data point in the index sequence exceeds the preset intensity threshold, then an abnormal data segment is extracted from the index sequence to obtain an abnormal index segment. The abnormal index segments are subjected to multidimensional scaling to obtain a discrete point set representing nonlinear correlations; Clustering is performed on the discrete point set. If multiple density-separable clusters are formed in the clustering results, the geometric center and shape parameters of each cluster are combined to obtain the candidate precursor features. The candidate precursor features are compared with a pre-established database of typical fluctuation precursor features. Features with high matching degree are selected and identified as precursor features of the upcoming current fluctuation.

5. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The process of reasoning, predicting, and classifying the precursor features to determine the current trend category includes: The precursor features are normalized to obtain standardized features; Calculate the similarity between the standardized feature and the pre-established first feature. If the similarity is less than a preset similarity threshold, then infer the standardized feature to obtain a short-term trend vector. Obtain the historical trend vector, and perform a weighted fusion of the short-term trend vector and the historical trend vector to obtain the fused trend vector; The fused trend vector is input into a pre-established trend classifier, and the current trend category is determined based on the output fluctuation type.

6. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The step of generating a pre-adjustment command sequence based on the current trend category and dynamically adjusting the inverter control parameters to obtain a current regulation scheme includes: Based on the current trend category, a fuzzy instruction set is obtained from a preset parameter mapping library; The fuzzy instruction set is subjected to timing constraint detection. If the parameter value in the fuzzy instruction set exceeds a preset safety threshold, the parameter value is adjusted to obtain a preliminary instruction sequence. The initial instruction sequence is converted into a binary data frame to generate a pre-tuned instruction sequence; The pre-adjustment instruction sequence and the current trend category are associated and encapsulated, and a timestamp for scheduling execution is added to obtain the current regulation scheme.

7. The variable frequency drive constant current detection method based on intelligent agents according to claim 6, characterized in that, The process of acquiring deviation data in the current regulation scheme, classifying the deviation data, and determining the pre-adjustment parameter set based on the classification results includes: Obtain the deviation data in the current regulation scheme; The deviation data is classified to obtain a set of classified deviation patterns; Potential risk factors are separated from the set of deviation patterns. If the potential risk factors exceed a preset activation threshold, a set of risk factors is obtained. Based on the set of risk factors, parameter adjustment rules are obtained from the parameter mapping library and modified to obtain the modified pre-adjusted parameter set.

8. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The process involves obtaining real-time feedback data based on the pre-adjusted parameter set, updating the inverter's internal control logic based on the real-time feedback data, and if the updated logic control parameters meet a preset operating state threshold, then classifying and analyzing the real-time feedback data to determine the mode transition result, including: Based on the pre-tuned parameter set, real-time feedback data is obtained and pre-processed to obtain a feedback data set; Based on the feedback data set, the internal control logic of the frequency converter is updated to obtain the updated control logic parameters; If the control logic parameters meet the preset operating state threshold, the feedback data set is classified to obtain the classified operating state data. The operational status data is analyzed to determine the mode transition result.

9. The variable frequency drive constant current detection method based on intelligent agents according to claim 1, characterized in that, The process of integrating the electrical characteristic data and the preset parameter set based on the mode transition results, performing classification processing and mode analysis, and determining a continuously stable constant current control result includes: Based on the mode transition results, the electrical characteristic data and the preset parameter set are integrated to obtain an optimized dataset; The optimized dataset is classified into states to obtain classified constant current control data; Pattern analysis is performed on the constant current control data to determine a continuously stable constant current control result.

10. A variable frequency drive constant current detection system based on an intelligent agent, characterized in that, include: The data acquisition and filtering module is used to acquire multi-channel electrical signal data, and perform filtering and sequence correlation processing to obtain electrical characteristic data; The timing analysis and coupling evaluation module is used to perform timing decomposition on the electrical characteristic data, extract the main features, calculate the dynamic coupling strength between feature components, and determine the index sequence of the dynamic coupling strength between electrical characteristics. An anomaly detection and extraction module is used to obtain candidate precursor features from the indicator sequence and compare them with a pre-established typical fluctuation precursor feature library to determine the precursor features of the upcoming current fluctuation. The trend prediction and classification module is used to perform inference prediction and classification processing on the precursor features to determine the current trend category. The control command generation module is used to generate a pre-adjustment command sequence based on the current trend category, dynamically adjust the inverter control parameters, and obtain a current regulation scheme. The deviation correction and parameter optimization module is used to acquire deviation data in the current control scheme, classify the deviation data, and determine the pre-adjustment parameter set based on the classification results. The control logic update and feedback module is used to obtain real-time feedback data based on the pre-adjusted parameter set, update the internal control logic of the frequency converter based on the real-time feedback data, and if the updated logic control parameters meet the preset operating state threshold, classify and analyze the real-time feedback data to determine the mode transition result. The system integration and constant current control module is used to integrate the electrical characteristic data and the preset parameter set according to the mode transition result, and perform classification processing and mode analysis to determine a continuous and stable constant current control result.

Images